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
|
/*
* Copyright (C) 2019-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/command_container/command_encoder.h"
#include "opencl/source/cl_device/cl_device.h"
#include "opencl/source/device_queue/device_queue_hw_base.inl"
#include "opencl/source/program/block_kernel_manager.h"
namespace NEO {
template <typename GfxFamily>
size_t DeviceQueueHw<GfxFamily>::getMinimumSlbSize() {
using MEDIA_STATE_FLUSH = typename GfxFamily::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename GfxFamily::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
return sizeof(MEDIA_STATE_FLUSH) +
sizeof(MEDIA_INTERFACE_DESCRIPTOR_LOAD) +
sizeof(PIPE_CONTROL) +
sizeof(GPGPU_WALKER) +
sizeof(MEDIA_STATE_FLUSH) +
sizeof(PIPE_CONTROL) +
DeviceQueueHw<GfxFamily>::getCSPrefetchSize();
}
template <typename GfxFamily>
void DeviceQueueHw<GfxFamily>::buildSlbDummyCommands() {
using MEDIA_STATE_FLUSH = typename GfxFamily::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename GfxFamily::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
auto igilCmdQueue = reinterpret_cast<IGIL_CommandQueue *>(queueBuffer->getUnderlyingBuffer());
auto slbEndOffset = igilCmdQueue->m_controls.m_SLBENDoffsetInBytes;
size_t commandsSize = getMinimumSlbSize() + getWaCommandsSize();
size_t numEnqueues = numberOfDeviceEnqueues;
// buildSlbDummyCommands is called from resetDeviceQueue() - reset slbCS each time
slbCS.replaceBuffer(slbBuffer->getUnderlyingBuffer(), slbBuffer->getUnderlyingBufferSize());
if (slbEndOffset >= 0) {
DEBUG_BREAK_IF(slbEndOffset % commandsSize != 0);
//We always overwrite at most one enqueue space with BB_START command pointing to cleanup section
//if SLBENDoffset is the at the end then BB_START added after scheduler did not corrupt anything so no need to regenerate
numEnqueues = (slbEndOffset == static_cast<int>(commandsSize)) ? 0 : 1;
slbCS.getSpace(slbEndOffset);
}
for (size_t i = 0; i < numEnqueues; i++) {
auto mediaStateFlush = slbCS.getSpaceForCmd<MEDIA_STATE_FLUSH>();
*mediaStateFlush = GfxFamily::cmdInitMediaStateFlush;
addArbCheckCmdWa();
addMiAtomicCmdWa((uint64_t)&igilCmdQueue->m_controls.m_DummyAtomicOperationPlaceholder);
auto mediaIdLoad = slbCS.getSpaceForCmd<MEDIA_INTERFACE_DESCRIPTOR_LOAD>();
*mediaIdLoad = GfxFamily::cmdInitMediaInterfaceDescriptorLoad;
mediaIdLoad->setInterfaceDescriptorTotalLength(2048);
auto dataStartAddress = colorCalcStateSize;
mediaIdLoad->setInterfaceDescriptorDataStartAddress(dataStartAddress + sizeof(INTERFACE_DESCRIPTOR_DATA) * schedulerIDIndex);
addLriCmdWa(true);
if (isProfilingEnabled()) {
addPipeControlCmdWa();
auto pipeControl = slbCS.getSpaceForCmd<PIPE_CONTROL>();
initPipeControl(pipeControl);
} else {
auto noop = slbCS.getSpace(sizeof(PIPE_CONTROL));
memset(noop, 0x0, sizeof(PIPE_CONTROL));
addPipeControlCmdWa(true);
}
auto gpgpuWalker = slbCS.getSpaceForCmd<GPGPU_WALKER>();
*gpgpuWalker = GfxFamily::cmdInitGpgpuWalker;
gpgpuWalker->setSimdSize(GPGPU_WALKER::SIMD_SIZE::SIMD_SIZE_SIMD16);
gpgpuWalker->setThreadGroupIdXDimension(1);
gpgpuWalker->setThreadGroupIdYDimension(1);
gpgpuWalker->setThreadGroupIdZDimension(1);
gpgpuWalker->setRightExecutionMask(0xFFFFFFFF);
gpgpuWalker->setBottomExecutionMask(0xFFFFFFFF);
mediaStateFlush = slbCS.getSpaceForCmd<MEDIA_STATE_FLUSH>();
*mediaStateFlush = GfxFamily::cmdInitMediaStateFlush;
addArbCheckCmdWa();
addPipeControlCmdWa();
auto pipeControl2 = slbCS.getSpaceForCmd<PIPE_CONTROL>();
initPipeControl(pipeControl2);
addLriCmdWa(false);
auto prefetch = slbCS.getSpace(getCSPrefetchSize());
memset(prefetch, 0x0, getCSPrefetchSize());
}
// always the same BBStart position (after 128 enqueues)
auto bbStartOffset = (commandsSize * 128) - slbCS.getUsed();
slbCS.getSpace(bbStartOffset);
auto bbStart = slbCS.getSpaceForCmd<MI_BATCH_BUFFER_START>();
*bbStart = GfxFamily::cmdInitBatchBufferStart;
auto slbPtr = reinterpret_cast<uintptr_t>(slbBuffer->getUnderlyingBuffer());
bbStart->setBatchBufferStartAddressGraphicsaddress472(slbPtr);
igilCmdQueue->m_controls.m_CleanupSectionSize = 0;
igilQueue->m_controls.m_CleanupSectionAddress = 0;
}
template <typename GfxFamily>
void DeviceQueueHw<GfxFamily>::addMediaStateClearCmds() {
typedef typename GfxFamily::MEDIA_VFE_STATE MEDIA_VFE_STATE;
addPipeControlCmdWa();
auto pipeControl = slbCS.getSpaceForCmd<PIPE_CONTROL>();
*pipeControl = GfxFamily::cmdInitPipeControl;
pipeControl->setGenericMediaStateClear(true);
pipeControl->setCommandStreamerStallEnable(true);
addDcFlushToPipeControlWa(pipeControl);
PreambleHelper<GfxFamily>::programVFEState(&slbCS, device->getHardwareInfo(), 0u, 0, device->getSharedDeviceInfo().maxFrontEndThreads, aub_stream::EngineType::ENGINE_RCS, AdditionalKernelExecInfo::NotApplicable);
}
template <typename GfxFamily>
size_t DeviceQueueHw<GfxFamily>::getMediaStateClearCmdsSize() {
using MEDIA_VFE_STATE = typename GfxFamily::MEDIA_VFE_STATE;
// PC with GenreicMediaStateClear + WA PC
size_t size = 2 * sizeof(PIPE_CONTROL);
// VFE state cmds
size += sizeof(PIPE_CONTROL);
size += sizeof(MEDIA_VFE_STATE);
return size;
}
template <typename GfxFamily>
void DeviceQueueHw<GfxFamily>::setupIndirectState(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentIDCount, bool isCcsUsed) {
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
void *pDSH = dynamicStateHeap.getCpuBase();
// Set scheduler ID to last entry in first table, it will have ID == 0, blocks will have following entries.
auto igilCmdQueue = reinterpret_cast<IGIL_CommandQueue *>(queueBuffer->getUnderlyingBuffer());
igilCmdQueue->m_controls.m_IDTstart = colorCalcStateSize + sizeof(INTERFACE_DESCRIPTOR_DATA) * (interfaceDescriptorEntries - 2);
// Parent's dsh is located after ColorCalcState and 2 ID tables
igilCmdQueue->m_controls.m_DynamicHeapStart = offsetDsh + alignUp((uint32_t)parentKernel->getDynamicStateHeapSize(), GPGPU_WALKER::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
igilCmdQueue->m_controls.m_DynamicHeapSizeInBytes = (uint32_t)dshBuffer->getUnderlyingBufferSize();
igilCmdQueue->m_controls.m_CurrentDSHoffset = igilCmdQueue->m_controls.m_DynamicHeapStart;
igilCmdQueue->m_controls.m_ParentDSHOffset = offsetDsh;
uint32_t blockIndex = parentIDCount;
pDSH = ptrOffset(pDSH, colorCalcStateSize);
INTERFACE_DESCRIPTOR_DATA *pIDDestination = static_cast<INTERFACE_DESCRIPTOR_DATA *>(pDSH);
BlockKernelManager *blockManager = parentKernel->getProgram()->getBlockKernelManager();
uint32_t blockCount = static_cast<uint32_t>(blockManager->getCount());
uint32_t maxBindingTableCount = 0;
uint32_t totalBlockSSHSize = 0;
igilCmdQueue->m_controls.m_StartBlockID = blockIndex;
for (uint32_t i = 0; i < blockCount; i++) {
const KernelInfo *pBlockInfo = blockManager->getBlockKernelInfo(i);
auto blockKernelStartPointer = getBlockKernelStartPointer(getDevice(), pBlockInfo, isCcsUsed);
auto bindingTableCount = pBlockInfo->patchInfo.bindingTableState->Count;
maxBindingTableCount = std::max(maxBindingTableCount, bindingTableCount);
totalBlockSSHSize += alignUp(pBlockInfo->heapInfo.SurfaceStateHeapSize, BINDING_TABLE_STATE::SURFACESTATEPOINTER_ALIGN_SIZE);
surfaceStateHeap.align(BINDING_TABLE_STATE::SURFACESTATEPOINTER_ALIGN_SIZE);
auto btOffset = EncodeSurfaceState<GfxFamily>::pushBindingTableAndSurfaceStates(surfaceStateHeap, bindingTableCount,
pBlockInfo->heapInfo.pSsh,
pBlockInfo->heapInfo.SurfaceStateHeapSize,
bindingTableCount,
pBlockInfo->patchInfo.bindingTableState->Offset);
parentKernel->setReflectionSurfaceBlockBtOffset(i, static_cast<uint32_t>(btOffset));
// Determine SIMD size
uint32_t simd = pBlockInfo->getMaxSimdSize();
DEBUG_BREAK_IF(pBlockInfo->patchInfo.interfaceDescriptorData == nullptr);
uint32_t idOffset = pBlockInfo->patchInfo.interfaceDescriptorData->Offset;
const INTERFACE_DESCRIPTOR_DATA *pBlockID = static_cast<const INTERFACE_DESCRIPTOR_DATA *>(ptrOffset(pBlockInfo->heapInfo.pDsh, idOffset));
pIDDestination[blockIndex + i] = *pBlockID;
pIDDestination[blockIndex + i].setKernelStartPointerHigh(blockKernelStartPointer >> 32);
pIDDestination[blockIndex + i].setKernelStartPointer(static_cast<uint32_t>(blockKernelStartPointer));
pIDDestination[blockIndex + i].setDenormMode(INTERFACE_DESCRIPTOR_DATA::DENORM_MODE_SETBYKERNEL);
EncodeDispatchKernel<GfxFamily>::programBarrierEnable(pIDDestination[blockIndex + i],
pBlockInfo->patchInfo.executionEnvironment->HasBarriers,
parentKernel->getDevice().getHardwareInfo());
// Set offset to sampler states, block's DHSOffset is added by scheduler
pIDDestination[blockIndex + i].setSamplerStatePointer(static_cast<uint32_t>(pBlockInfo->getBorderColorStateSize()));
auto threadPayload = pBlockInfo->patchInfo.threadPayload;
DEBUG_BREAK_IF(nullptr == threadPayload);
auto numChannels = PerThreadDataHelper::getNumLocalIdChannels(*threadPayload);
auto grfSize = device->getDeviceInfo().grfSize;
auto sizePerThreadData = getPerThreadSizeLocalIDs(simd, grfSize, numChannels);
auto numGrfPerThreadData = static_cast<uint32_t>(sizePerThreadData / grfSize);
// HW requires a minimum of 1 GRF of perThreadData for each thread in a thread group
// when sizeCrossThreadData != 0
numGrfPerThreadData = std::max(numGrfPerThreadData, 1u);
pIDDestination[blockIndex + i].setConstantIndirectUrbEntryReadLength(numGrfPerThreadData);
}
igilCmdQueue->m_controls.m_BTmaxSize = alignUp(maxBindingTableCount * (uint32_t)sizeof(BINDING_TABLE_STATE), INTERFACE_DESCRIPTOR_DATA::BINDINGTABLEPOINTER::BINDINGTABLEPOINTER_ALIGN_SIZE);
igilCmdQueue->m_controls.m_BTbaseOffset = alignUp((uint32_t)surfaceStateHeap.getUsed(), INTERFACE_DESCRIPTOR_DATA::BINDINGTABLEPOINTER::BINDINGTABLEPOINTER_ALIGN_SIZE);
igilCmdQueue->m_controls.m_CurrentSSHoffset = igilCmdQueue->m_controls.m_BTbaseOffset;
}
} // namespace NEO
|
