1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
|
/*
* Copyright (C) 2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "shared/source/command_container/command_encoder.h"
#include "shared/source/command_stream/linear_stream.h"
#include "shared/source/command_stream/preemption.h"
#include "shared/source/execution_environment/execution_environment.h"
#include "shared/source/gmm_helper/gmm_helper.h"
#include "shared/source/helpers/hw_helper.h"
#include "shared/source/helpers/simd_helper.h"
#include "shared/source/helpers/state_base_address.h"
#include "shared/source/kernel/dispatch_kernel_encoder_interface.h"
#include "pipe_control_args.h"
#include <algorithm>
namespace NEO {
template <typename Family>
void EncodeDispatchKernel<Family>::encode(CommandContainer &container,
const void *pThreadGroupDimensions, bool isIndirect, bool isPredicate, DispatchKernelEncoderI *dispatchInterface,
uint64_t eventAddress, Device *device, PreemptionMode preemptionMode) {
using MEDIA_STATE_FLUSH = typename Family::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
using MI_BATCH_BUFFER_END = typename Family::MI_BATCH_BUFFER_END;
using STATE_BASE_ADDRESS = typename Family::STATE_BASE_ADDRESS;
auto &kernelDescriptor = dispatchInterface->getKernelDescriptor();
auto sizeCrossThreadData = dispatchInterface->getCrossThreadDataSize();
auto sizePerThreadData = dispatchInterface->getPerThreadDataSize();
auto sizePerThreadDataForWholeGroup = dispatchInterface->getPerThreadDataSizeForWholeThreadGroup();
const HardwareInfo &hwInfo = device->getHardwareInfo();
LinearStream *listCmdBufferStream = container.getCommandStream();
size_t sshOffset = 0;
size_t estimatedSizeRequired = estimateEncodeDispatchKernelCmdsSize(device);
if (container.getCommandStream()->getAvailableSpace() < estimatedSizeRequired) {
auto bbEnd = listCmdBufferStream->getSpaceForCmd<MI_BATCH_BUFFER_END>();
*bbEnd = Family::cmdInitBatchBufferEnd;
container.allocateNextCommandBuffer();
}
WALKER_TYPE cmd = Family::cmdInitGpgpuWalker;
auto idd = Family::cmdInitInterfaceDescriptorData;
{
auto alloc = dispatchInterface->getIsaAllocation();
UNRECOVERABLE_IF(nullptr == alloc);
auto offset = alloc->getGpuAddressToPatch();
idd.setKernelStartPointer(offset);
idd.setKernelStartPointerHigh(0u);
}
EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), true);
EncodeStates<Family>::adjustStateComputeMode(*container.getCommandStream(), container.lastSentNumGrfRequired, nullptr, false, false);
EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), false);
auto numThreadsPerThreadGroup = dispatchInterface->getNumThreadsPerThreadGroup();
idd.setNumberOfThreadsInGpgpuThreadGroup(numThreadsPerThreadGroup);
EncodeDispatchKernel<Family>::programBarrierEnable(idd,
kernelDescriptor.kernelAttributes.hasBarriers,
hwInfo);
auto slmSize = static_cast<typename INTERFACE_DESCRIPTOR_DATA::SHARED_LOCAL_MEMORY_SIZE>(
HwHelperHw<Family>::get().computeSlmValues(dispatchInterface->getSlmTotalSize()));
idd.setSharedLocalMemorySize(
dispatchInterface->getSlmTotalSize() > 0
? slmSize
: INTERFACE_DESCRIPTOR_DATA::SHARED_LOCAL_MEMORY_SIZE_ENCODES_0K);
{
uint32_t bindingTableStateCount = kernelDescriptor.payloadMappings.bindingTable.numEntries;
uint32_t bindingTablePointer = 0u;
if (bindingTableStateCount > 0u) {
auto ssh = container.getHeapWithRequiredSizeAndAlignment(HeapType::SURFACE_STATE, dispatchInterface->getSurfaceStateHeapDataSize(), BINDING_TABLE_STATE::SURFACESTATEPOINTER_ALIGN_SIZE);
sshOffset = ssh->getUsed();
bindingTablePointer = static_cast<uint32_t>(EncodeSurfaceState<Family>::pushBindingTableAndSurfaceStates(
*ssh, bindingTableStateCount,
dispatchInterface->getSurfaceStateHeapData(),
dispatchInterface->getSurfaceStateHeapDataSize(), bindingTableStateCount,
kernelDescriptor.payloadMappings.bindingTable.tableOffset));
}
idd.setBindingTablePointer(bindingTablePointer);
uint32_t bindingTableStatePrefetchCount = 0;
if (EncodeSurfaceState<Family>::doBindingTablePrefetch()) {
bindingTableStatePrefetchCount = std::min(31u, bindingTableStateCount);
}
idd.setBindingTableEntryCount(bindingTableStatePrefetchCount);
}
PreemptionHelper::programInterfaceDescriptorDataPreemption<Family>(&idd, preemptionMode);
auto heap = container.getIndirectHeap(HeapType::DYNAMIC_STATE);
UNRECOVERABLE_IF(!heap);
uint32_t samplerStateOffset = 0;
uint32_t samplerCount = 0;
if (kernelDescriptor.payloadMappings.samplerTable.numSamplers > 0) {
samplerCount = kernelDescriptor.payloadMappings.samplerTable.numSamplers;
samplerStateOffset = EncodeStates<Family>::copySamplerState(heap, kernelDescriptor.payloadMappings.samplerTable.tableOffset,
kernelDescriptor.payloadMappings.samplerTable.numSamplers,
kernelDescriptor.payloadMappings.samplerTable.borderColor,
dispatchInterface->getDynamicStateHeapData());
}
idd.setSamplerStatePointer(samplerStateOffset);
auto samplerCountState =
static_cast<typename INTERFACE_DESCRIPTOR_DATA::SAMPLER_COUNT>((samplerCount + 3) / 4);
idd.setSamplerCount(samplerCountState);
auto numGrfCrossThreadData = static_cast<uint32_t>(sizeCrossThreadData / sizeof(float[8]));
idd.setCrossThreadConstantDataReadLength(numGrfCrossThreadData);
auto numGrfPerThreadData = static_cast<uint32_t>(sizePerThreadData / sizeof(float[8]));
DEBUG_BREAK_IF(numGrfPerThreadData <= 0u);
idd.setConstantIndirectUrbEntryReadLength(numGrfPerThreadData);
uint32_t sizeThreadData = sizePerThreadDataForWholeGroup + sizeCrossThreadData;
uint64_t offsetThreadData = 0u;
{
auto heapIndirect = container.getIndirectHeap(HeapType::INDIRECT_OBJECT);
UNRECOVERABLE_IF(!(heapIndirect));
heapIndirect->align(WALKER_TYPE::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
auto ptr = container.getHeapSpaceAllowGrow(HeapType::INDIRECT_OBJECT, sizeThreadData);
UNRECOVERABLE_IF(!(ptr));
offsetThreadData = heapIndirect->getHeapGpuStartOffset() + static_cast<uint64_t>(heapIndirect->getUsed() - sizeThreadData);
memcpy_s(ptr, sizeCrossThreadData,
dispatchInterface->getCrossThreadData(), sizeCrossThreadData);
if (isIndirect) {
void *gpuPtr = reinterpret_cast<void *>(heapIndirect->getHeapGpuBase() + heapIndirect->getUsed() - sizeThreadData);
EncodeIndirectParams<Family>::setGroupCountIndirect(container, kernelDescriptor.payloadMappings.dispatchTraits.numWorkGroups, gpuPtr);
EncodeIndirectParams<Family>::setGlobalWorkSizeIndirect(container, kernelDescriptor.payloadMappings.dispatchTraits.globalWorkSize, gpuPtr, dispatchInterface->getGroupSize());
}
if (kernelDescriptor.payloadMappings.bindingTable.numEntries > 0) {
patchBindlessSurfaceStateOffsets(sshOffset, dispatchInterface->getKernelDescriptor(), reinterpret_cast<uint8_t *>(ptr));
}
ptr = ptrOffset(ptr, sizeCrossThreadData);
memcpy_s(ptr, sizePerThreadDataForWholeGroup,
dispatchInterface->getPerThreadData(), sizePerThreadDataForWholeGroup);
}
auto slmSizeNew = dispatchInterface->getSlmTotalSize();
bool dirtyHeaps = container.isAnyHeapDirty();
bool flush = container.slmSize != slmSizeNew || dirtyHeaps;
if (flush) {
PipeControlArgs args(true);
if (dirtyHeaps) {
args.hdcPipelineFlush = true;
}
MemorySynchronizationCommands<Family>::addPipeControl(*container.getCommandStream(), args);
if (dirtyHeaps) {
STATE_BASE_ADDRESS sba;
EncodeStateBaseAddress<Family>::encode(container, sba);
container.setDirtyStateForAllHeaps(false);
}
if (container.slmSize != slmSizeNew) {
EncodeL3State<Family>::encode(container, slmSizeNew != 0u);
container.slmSize = slmSizeNew;
if (container.nextIddInBlock != container.getNumIddPerBlock()) {
EncodeMediaInterfaceDescriptorLoad<Family>::encode(container);
}
}
}
uint32_t numIDD = 0u;
void *ptr = getInterfaceDescriptor(container, numIDD);
memcpy_s(ptr, sizeof(idd), &idd, sizeof(idd));
cmd.setIndirectDataStartAddress(static_cast<uint32_t>(offsetThreadData));
cmd.setIndirectDataLength(sizeThreadData);
cmd.setInterfaceDescriptorOffset(numIDD);
EncodeDispatchKernel<Family>::encodeThreadData(cmd,
nullptr,
static_cast<const uint32_t *>(pThreadGroupDimensions),
dispatchInterface->getGroupSize(),
kernelDescriptor.kernelAttributes.simdSize,
kernelDescriptor.kernelAttributes.numLocalIdChannels,
dispatchInterface->getNumThreadsPerThreadGroup(),
dispatchInterface->getThreadExecutionMask(),
true,
false,
isIndirect,
dispatchInterface->getRequiredWorkgroupOrder());
cmd.setPredicateEnable(isPredicate);
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, hwInfo);
PreemptionHelper::applyPreemptionWaCmdsBegin<Family>(listCmdBufferStream, *device);
auto buffer = listCmdBufferStream->getSpace(sizeof(cmd));
*(decltype(cmd) *)buffer = cmd;
PreemptionHelper::applyPreemptionWaCmdsEnd<Family>(listCmdBufferStream, *device);
{
auto mediaStateFlush = listCmdBufferStream->getSpace(sizeof(MEDIA_STATE_FLUSH));
*reinterpret_cast<MEDIA_STATE_FLUSH *>(mediaStateFlush) = Family::cmdInitMediaStateFlush;
}
}
template <typename Family>
void EncodeMediaInterfaceDescriptorLoad<Family>::encode(CommandContainer &container) {
using MEDIA_STATE_FLUSH = typename Family::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
auto heap = container.getIndirectHeap(HeapType::DYNAMIC_STATE);
auto mediaStateFlush = container.getCommandStream()->getSpaceForCmd<MEDIA_STATE_FLUSH>();
*mediaStateFlush = Family::cmdInitMediaStateFlush;
MEDIA_INTERFACE_DESCRIPTOR_LOAD cmd = Family::cmdInitMediaInterfaceDescriptorLoad;
cmd.setInterfaceDescriptorDataStartAddress(static_cast<uint32_t>(ptrDiff(container.getIddBlock(), heap->getCpuBase())));
cmd.setInterfaceDescriptorTotalLength(sizeof(INTERFACE_DESCRIPTOR_DATA) * container.getNumIddPerBlock());
auto buffer = container.getCommandStream()->getSpace(sizeof(cmd));
*(decltype(cmd) *)buffer = cmd;
}
template <typename Family>
bool EncodeDispatchKernel<Family>::isRuntimeLocalIdsGenerationRequired(uint32_t activeChannels,
size_t *lws,
std::array<uint8_t, 3> walkOrder,
bool requireInputWalkOrder,
uint32_t &requiredWalkOrder,
uint32_t simd) {
requiredWalkOrder = 0u;
return true;
}
template <typename Family>
void EncodeDispatchKernel<Family>::encodeThreadData(WALKER_TYPE &walkerCmd,
const uint32_t *startWorkGroup,
const uint32_t *numWorkGroups,
const uint32_t *workGroupSizes,
uint32_t simd,
uint32_t localIdDimensions,
uint32_t threadsPerThreadGroup,
uint32_t threadExecutionMask,
bool localIdsGenerationByRuntime,
bool inlineDataProgrammingRequired,
bool isIndirect,
uint32_t requiredWorkGroupOrder) {
if (isIndirect) {
walkerCmd.setIndirectParameterEnable(true);
} else {
walkerCmd.setThreadGroupIdXDimension(static_cast<uint32_t>(numWorkGroups[0]));
walkerCmd.setThreadGroupIdYDimension(static_cast<uint32_t>(numWorkGroups[1]));
walkerCmd.setThreadGroupIdZDimension(static_cast<uint32_t>(numWorkGroups[2]));
}
if (startWorkGroup) {
walkerCmd.setThreadGroupIdStartingX(static_cast<uint32_t>(startWorkGroup[0]));
walkerCmd.setThreadGroupIdStartingY(static_cast<uint32_t>(startWorkGroup[1]));
walkerCmd.setThreadGroupIdStartingResumeZ(static_cast<uint32_t>(startWorkGroup[2]));
}
walkerCmd.setSimdSize(getSimdConfig<WALKER_TYPE>(simd));
auto localWorkSize = workGroupSizes[0] * workGroupSizes[1] * workGroupSizes[2];
if (threadsPerThreadGroup == 0) {
threadsPerThreadGroup = static_cast<uint32_t>(getThreadsPerWG(simd, localWorkSize));
}
walkerCmd.setThreadWidthCounterMaximum(threadsPerThreadGroup);
uint64_t executionMask = threadExecutionMask;
if (executionMask == 0) {
auto remainderSimdLanes = localWorkSize & (simd - 1);
executionMask = maxNBitValue(remainderSimdLanes);
if (!executionMask)
executionMask = ~executionMask;
}
constexpr uint32_t maxDword = std::numeric_limits<uint32_t>::max();
walkerCmd.setRightExecutionMask(static_cast<uint32_t>(executionMask));
walkerCmd.setBottomExecutionMask(maxDword);
}
template <typename Family>
void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor,
uint32_t value,
const HardwareInfo &hwInfo) {
interfaceDescriptor.setBarrierEnable(value);
}
template <typename Family>
void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const HardwareInfo &hwInfo, WALKER_TYPE &walkerCmd) {}
template <typename Family>
void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTOR_DATA *pInterfaceDescriptor, const HardwareInfo &hwInfo, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize) {}
template <typename Family>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const HardwareInfo &hwInfo) {}
template <typename Family>
size_t EncodeDispatchKernel<Family>::estimateEncodeDispatchKernelCmdsSize(Device *device) {
using MEDIA_STATE_FLUSH = typename Family::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
using MI_BATCH_BUFFER_END = typename Family::MI_BATCH_BUFFER_END;
size_t issueMediaInterfaceDescriptorLoad = sizeof(MEDIA_STATE_FLUSH) + sizeof(MEDIA_INTERFACE_DESCRIPTOR_LOAD);
size_t totalSize = sizeof(WALKER_TYPE);
totalSize += PreemptionHelper::getPreemptionWaCsSize<Family>(*device);
totalSize += sizeof(MEDIA_STATE_FLUSH);
totalSize += issueMediaInterfaceDescriptorLoad;
totalSize += EncodeStates<Family>::getAdjustStateComputeModeSize();
totalSize += EncodeWA<Family>::getAdditionalPipelineSelectSize(*device);
totalSize += EncodeIndirectParams<Family>::getCmdsSizeForIndirectParams();
totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupCountIndirect();
totalSize += EncodeIndirectParams<Family>::getCmdsSizeForSetGroupSizeIndirect();
totalSize += sizeof(MI_BATCH_BUFFER_END);
return totalSize;
}
template <typename Family>
void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd) {
EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), true);
auto gmmHelper = container.getDevice()->getGmmHelper();
StateBaseAddressHelper<Family>::programStateBaseAddress(
&sbaCmd,
container.isHeapDirty(HeapType::DYNAMIC_STATE) ? container.getIndirectHeap(HeapType::DYNAMIC_STATE) : nullptr,
container.isHeapDirty(HeapType::INDIRECT_OBJECT) ? container.getIndirectHeap(HeapType::INDIRECT_OBJECT) : nullptr,
container.isHeapDirty(HeapType::SURFACE_STATE) ? container.getIndirectHeap(HeapType::SURFACE_STATE) : nullptr,
0,
false,
(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) >> 1),
container.getIndirectObjectHeapBaseAddress(),
container.getInstructionHeapBaseAddress(),
false,
gmmHelper,
false);
auto pCmd = reinterpret_cast<STATE_BASE_ADDRESS *>(container.getCommandStream()->getSpace(sizeof(STATE_BASE_ADDRESS)));
*pCmd = sbaCmd;
EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), false);
}
template <typename Family>
void EncodeL3State<Family>::encode(CommandContainer &container, bool enableSLM) {
auto offset = L3CNTLRegisterOffset<Family>::registerOffset;
auto data = PreambleHelper<Family>::getL3Config(container.getDevice()->getHardwareInfo(), enableSLM);
EncodeSetMMIO<Family>::encodeIMM(container, offset, data, false);
}
template <typename GfxFamily>
void EncodeMiFlushDW<GfxFamily>::appendMiFlushDw(MI_FLUSH_DW *miFlushDwCmd) {}
template <typename GfxFamily>
void EncodeMiFlushDW<GfxFamily>::programMiFlushDwWA(LinearStream &commandStream) {}
template <typename GfxFamily>
size_t EncodeMiFlushDW<GfxFamily>::getMiFlushDwWaSize() {
return 0;
}
template <typename GfxFamily>
inline void EncodeWA<GfxFamily>::encodeAdditionalPipelineSelect(Device &device, LinearStream &stream, bool is3DPipeline) {}
template <typename GfxFamily>
inline size_t EncodeWA<GfxFamily>::getAdditionalPipelineSelectSize(Device &device) {
return 0;
}
template <typename GfxFamily>
void EncodeSurfaceState<GfxFamily>::encodeExtraBufferParams(R_SURFACE_STATE *surfaceState, GraphicsAllocation *allocation, GmmHelper *gmmHelper,
bool isReadOnly, uint32_t numAvailableDevices) {
}
template <typename Family>
void EncodeSempahore<Family>::programMiSemaphoreWait(MI_SEMAPHORE_WAIT *cmd,
uint64_t compareAddress,
uint32_t compareData,
COMPARE_OPERATION compareMode,
bool registerPollMode) {
MI_SEMAPHORE_WAIT localCmd = Family::cmdInitMiSemaphoreWait;
localCmd.setCompareOperation(compareMode);
localCmd.setSemaphoreDataDword(compareData);
localCmd.setSemaphoreGraphicsAddress(compareAddress);
localCmd.setWaitMode(MI_SEMAPHORE_WAIT::WAIT_MODE::WAIT_MODE_POLLING_MODE);
*cmd = localCmd;
}
} // namespace NEO
|
