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
411
412
|
/*
* Copyright (C) 2019-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "shared/source/built_ins/built_ins.h"
#include "shared/source/built_ins/sip.h"
#include "shared/source/command_container/command_encoder.h"
#include "shared/source/command_stream/command_stream_receiver_hw.h"
#include "shared/source/command_stream/linear_stream.h"
#include "shared/source/command_stream/preemption.h"
#include "shared/source/command_stream/thread_arbitration_policy.h"
#include "shared/source/device/device.h"
#include "shared/source/helpers/hw_helper.h"
#include "shared/source/helpers/hw_info.h"
#include "shared/source/helpers/interlocked_max.h"
#include "shared/source/helpers/preamble.h"
#include "shared/source/memory_manager/memory_manager.h"
#include "shared/source/memory_manager/residency_container.h"
#include "shared/source/os_interface/os_context.h"
#include "shared/source/page_fault_manager/cpu_page_fault_manager.h"
#include "shared/source/unified_memory/unified_memory.h"
#include "level_zero/core/source/cmdlist/cmdlist.h"
#include "level_zero/core/source/cmdlist/cmdlist_hw.h"
#include "level_zero/core/source/cmdqueue/cmdqueue_hw.h"
#include "level_zero/core/source/device/device.h"
#include "level_zero/core/source/fence/fence.h"
#include "level_zero/tools/source/metrics/metric.h"
#include "pipe_control_args.h"
#include <limits>
#include <thread>
namespace L0 {
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandQueueHw<gfxCoreFamily>::createFence(const ze_fence_desc_t *desc,
ze_fence_handle_t *phFence) {
*phFence = Fence::create(this, desc);
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandQueueHw<gfxCoreFamily>::destroy() {
this->printFunctionsPrintfOutput();
delete commandStream;
buffers.destroy(this->getDevice()->getNEODevice()->getMemoryManager());
delete this;
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
uint32_t numCommandLists,
ze_command_list_handle_t *phCommandLists,
ze_fence_handle_t hFence,
bool performMigration) {
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
using MI_BATCH_BUFFER_END = typename GfxFamily::MI_BATCH_BUFFER_END;
using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
using POST_SYNC_OPERATION = typename PIPE_CONTROL::POST_SYNC_OPERATION;
auto lockCSR = csr->obtainUniqueOwnership();
for (auto i = 0u; i < numCommandLists; i++) {
auto commandList = CommandList::fromHandle(phCommandLists[i]);
if (isCopyOnlyCommandQueue != commandList->isCopyOnly()) {
return ZE_RESULT_ERROR_INVALID_COMMAND_LIST_TYPE;
}
}
size_t spaceForResidency = 0;
size_t preemptionSize = 0u;
size_t debuggerCmdsSize = 0;
constexpr size_t residencyContainerSpaceForPreemption = 2;
constexpr size_t residencyContainerSpaceForFence = 1;
constexpr size_t residencyContainerSpaceForTagWrite = 1;
NEO::Device *neoDevice = device->getNEODevice();
NEO::PreemptionMode statePreemption = commandQueuePreemptionMode;
auto devicePreemption = device->getDevicePreemptionMode();
if (commandQueuePreemptionMode == NEO::PreemptionMode::Initial) {
preemptionSize += NEO::PreemptionHelper::getRequiredCmdStreamSize<GfxFamily>(commandQueuePreemptionMode,
devicePreemption) +
NEO::PreemptionHelper::getRequiredPreambleSize<GfxFamily>(*neoDevice) +
NEO::PreemptionHelper::getRequiredStateSipCmdSize<GfxFamily>(*neoDevice);
statePreemption = devicePreemption;
}
if (!commandQueueDebugCmdsProgrammed) {
debuggerCmdsSize += NEO::PreambleHelper<GfxFamily>::getKernelDebuggingCommandsSize(neoDevice->isDebuggerActive());
}
if (devicePreemption == NEO::PreemptionMode::MidThread) {
spaceForResidency += residencyContainerSpaceForPreemption;
}
bool directSubmissionEnabled = csr->isDirectSubmissionEnabled();
L0::Fence *fence = nullptr;
device->activateMetricGroups();
size_t totalCmdBuffers = 0;
for (auto i = 0u; i < numCommandLists; i++) {
auto commandList = CommandList::fromHandle(phCommandLists[i]);
bool indirectAllocationsAllowed = commandList->hasIndirectAllocationsAllowed();
if (indirectAllocationsAllowed) {
UnifiedMemoryControls unifiedMemoryControls = commandList->getUnifiedMemoryControls();
auto svmAllocsManager = device->getDriverHandle()->getSvmAllocsManager();
svmAllocsManager->addInternalAllocationsToResidencyContainer(neoDevice->getRootDeviceIndex(),
commandList->commandContainer.getResidencyContainer(),
unifiedMemoryControls.generateMask());
}
totalCmdBuffers += commandList->commandContainer.getCmdBufferAllocations().size();
spaceForResidency += commandList->commandContainer.getResidencyContainer().size();
auto commandListPreemption = commandList->getCommandListPreemptionMode();
if (statePreemption != commandListPreemption) {
preemptionSize += sizeof(PIPE_CONTROL);
preemptionSize += NEO::PreemptionHelper::getRequiredCmdStreamSize<GfxFamily>(commandListPreemption, statePreemption);
statePreemption = commandListPreemption;
}
interlockedMax(commandQueuePerThreadScratchSize, commandList->getCommandListPerThreadScratchSize());
if (commandList->getCommandListPerThreadScratchSize() != 0) {
heapContainer.push_back(commandList->commandContainer.getIndirectHeap(NEO::HeapType::SURFACE_STATE)->getGraphicsAllocation());
for (auto element : commandList->commandContainer.sshAllocations) {
heapContainer.push_back(element);
}
}
}
size_t linearStreamSizeEstimate = totalCmdBuffers * sizeof(MI_BATCH_BUFFER_START);
linearStreamSizeEstimate += csr->getCmdsSizeForHardwareContext();
if (directSubmissionEnabled) {
linearStreamSizeEstimate += sizeof(MI_BATCH_BUFFER_START);
} else {
linearStreamSizeEstimate += sizeof(MI_BATCH_BUFFER_END);
}
if (hFence) {
fence = Fence::fromHandle(hFence);
spaceForResidency += residencyContainerSpaceForFence;
linearStreamSizeEstimate += isCopyOnlyCommandQueue ? NEO::EncodeMiFlushDW<GfxFamily>::getMiFlushDwCmdSizeForDataWrite() : NEO::MemorySynchronizationCommands<GfxFamily>::getSizeForPipeControlWithPostSyncOperation(device->getHwInfo());
}
spaceForResidency += residencyContainerSpaceForTagWrite;
residencyContainer.reserve(spaceForResidency);
auto scratchSpaceController = csr->getScratchSpaceController();
bool gsbaStateDirty = false;
bool frontEndStateDirty = false;
handleScratchSpace(residencyContainer,
heapContainer,
scratchSpaceController,
gsbaStateDirty, frontEndStateDirty);
gsbaStateDirty |= !gsbaInit;
frontEndStateDirty |= !frontEndInit;
if (!isCopyOnlyCommandQueue) {
if (!gpgpuEnabled) {
linearStreamSizeEstimate += estimatePipelineSelect();
}
if (frontEndStateDirty) {
linearStreamSizeEstimate += estimateFrontEndCmdSize();
}
if (gsbaStateDirty) {
linearStreamSizeEstimate += estimateStateBaseAddressCmdSize();
}
linearStreamSizeEstimate += preemptionSize + debuggerCmdsSize;
}
linearStreamSizeEstimate += isCopyOnlyCommandQueue ? NEO::EncodeMiFlushDW<GfxFamily>::getMiFlushDwCmdSizeForDataWrite() : NEO::MemorySynchronizationCommands<GfxFamily>::getSizeForPipeControlWithPostSyncOperation(device->getHwInfo());
size_t alignedSize = alignUp<size_t>(linearStreamSizeEstimate, minCmdBufferPtrAlign);
size_t padding = alignedSize - linearStreamSizeEstimate;
reserveLinearStreamSize(alignedSize);
NEO::LinearStream child(commandStream->getSpace(alignedSize), alignedSize);
const auto globalFenceAllocation = csr->getGlobalFenceAllocation();
if (globalFenceAllocation) {
residencyContainer.push_back(globalFenceAllocation);
}
csr->programHardwareContext(child);
if (device->getL0Debugger()) {
residencyContainer.push_back(device->getL0Debugger()->getSbaTrackingBuffer(csr->getOsContext().getContextId()));
}
if (!isCopyOnlyCommandQueue) {
if (!gpgpuEnabled) {
programPipelineSelect(child);
}
if (!commandQueueDebugCmdsProgrammed && neoDevice->isDebuggerActive()) {
NEO::PreambleHelper<GfxFamily>::programKernelDebugging(&child);
commandQueueDebugCmdsProgrammed = true;
}
if (frontEndStateDirty) {
programFrontEnd(scratchSpaceController->getScratchPatchAddress(), child);
}
if (gsbaStateDirty) {
auto indirectHeap = CommandList::fromHandle(phCommandLists[0])->commandContainer.getIndirectHeap(NEO::HeapType::INDIRECT_OBJECT);
programGeneralStateBaseAddress(scratchSpaceController->calculateNewGSH(), indirectHeap->getGraphicsAllocation()->isAllocatedInLocalMemoryPool(), child);
}
if (commandQueuePreemptionMode == NEO::PreemptionMode::Initial) {
NEO::PreemptionHelper::programCsrBaseAddress<GfxFamily>(child, *neoDevice, csr->getPreemptionAllocation());
NEO::PreemptionHelper::programStateSip<GfxFamily>(child, *neoDevice);
NEO::PreemptionHelper::programCmdStream<GfxFamily>(child,
devicePreemption,
commandQueuePreemptionMode,
csr->getPreemptionAllocation());
commandQueuePreemptionMode = devicePreemption;
statePreemption = commandQueuePreemptionMode;
}
const bool sipKernelUsed = devicePreemption == NEO::PreemptionMode::MidThread ||
neoDevice->isDebuggerActive();
if (devicePreemption == NEO::PreemptionMode::MidThread) {
residencyContainer.push_back(csr->getPreemptionAllocation());
}
if (sipKernelUsed) {
auto sipIsa = NEO::SipKernel::getSipKernelAllocation(*neoDevice);
residencyContainer.push_back(sipIsa);
}
if (neoDevice->getDebugger()) {
UNRECOVERABLE_IF(device->getDebugSurface() == nullptr);
residencyContainer.push_back(device->getDebugSurface());
}
}
for (auto i = 0u; i < numCommandLists; ++i) {
auto commandList = CommandList::fromHandle(phCommandLists[i]);
auto cmdBufferAllocations = commandList->commandContainer.getCmdBufferAllocations();
auto cmdBufferCount = cmdBufferAllocations.size();
auto commandListPreemption = commandList->getCommandListPreemptionMode();
if (statePreemption != commandListPreemption) {
NEO::PipeControlArgs args;
NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControl(child, args);
NEO::PreemptionHelper::programCmdStream<GfxFamily>(child,
commandListPreemption,
statePreemption,
csr->getPreemptionAllocation());
statePreemption = commandListPreemption;
}
for (size_t iter = 0; iter < cmdBufferCount; iter++) {
auto allocation = cmdBufferAllocations[iter];
NEO::EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&child, allocation->getGpuAddress(), true);
}
printfFunctionContainer.insert(printfFunctionContainer.end(),
commandList->getPrintfFunctionContainer().begin(),
commandList->getPrintfFunctionContainer().end());
NEO::PageFaultManager *pageFaultManager = nullptr;
if (performMigration) {
pageFaultManager = device->getDriverHandle()->getMemoryManager()->getPageFaultManager();
if (pageFaultManager == nullptr) {
performMigration = false;
}
}
for (auto alloc : commandList->commandContainer.getResidencyContainer()) {
if (residencyContainer.end() ==
std::find(residencyContainer.begin(), residencyContainer.end(), alloc)) {
residencyContainer.push_back(alloc);
if (performMigration) {
if (alloc &&
(alloc->getAllocationType() == NEO::GraphicsAllocation::AllocationType::SVM_GPU ||
alloc->getAllocationType() == NEO::GraphicsAllocation::AllocationType::SVM_CPU)) {
pageFaultManager->moveAllocationToGpuDomain(reinterpret_cast<void *>(alloc->getGpuAddress()));
}
}
}
}
}
commandQueuePreemptionMode = statePreemption;
if (hFence) {
residencyContainer.push_back(&fence->getAllocation());
if (isCopyOnlyCommandQueue) {
NEO::EncodeMiFlushDW<GfxFamily>::programMiFlushDw(child, fence->getGpuAddress(), Fence::STATE_SIGNALED, false, true);
} else {
NEO::PipeControlArgs args(true);
NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
child, POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
fence->getGpuAddress(),
Fence::STATE_SIGNALED,
device->getHwInfo(),
args);
}
}
dispatchTaskCountWrite(child, true);
residencyContainer.push_back(csr->getTagAllocation());
void *endingCmd = nullptr;
if (directSubmissionEnabled) {
endingCmd = child.getSpace(0);
NEO::EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&child, 0ull, false);
} else {
MI_BATCH_BUFFER_END cmd = GfxFamily::cmdInitBatchBufferEnd;
auto buffer = child.getSpaceForCmd<MI_BATCH_BUFFER_END>();
*(MI_BATCH_BUFFER_END *)buffer = cmd;
}
if (padding) {
void *paddingPtr = child.getSpace(padding);
memset(paddingPtr, 0, padding);
}
submitBatchBuffer(ptrDiff(child.getCpuBase(), commandStream->getCpuBase()), residencyContainer, endingCmd);
this->taskCount = csr->peekTaskCount();
csr->makeSurfacePackNonResident(residencyContainer);
if (getSynchronousMode() == ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS) {
this->synchronize(std::numeric_limits<uint64_t>::max());
}
this->residencyContainer.clear();
this->heapContainer.clear();
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandQueueHw<gfxCoreFamily>::programFrontEnd(uint64_t scratchAddress, NEO::LinearStream &commandStream) {
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
UNRECOVERABLE_IF(csr == nullptr);
NEO::PreambleHelper<GfxFamily>::programVFEState(&commandStream,
device->getHwInfo(),
commandQueuePerThreadScratchSize,
scratchAddress,
device->getMaxNumHwThreads(),
csr->getOsContext().getEngineType(),
NEO::AdditionalKernelExecInfo::NotApplicable);
frontEndInit = true;
}
template <GFXCORE_FAMILY gfxCoreFamily>
size_t CommandQueueHw<gfxCoreFamily>::estimateFrontEndCmdSize() {
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
return NEO::PreambleHelper<GfxFamily>::getVFECommandsSize();
}
template <GFXCORE_FAMILY gfxCoreFamily>
size_t CommandQueueHw<gfxCoreFamily>::estimatePipelineSelect() {
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
return NEO::PreambleHelper<GfxFamily>::getCmdSizeForPipelineSelect(device->getHwInfo());
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandQueueHw<gfxCoreFamily>::programPipelineSelect(NEO::LinearStream &commandStream) {
NEO::PipelineSelectArgs args = {0, 0};
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
NEO::PreambleHelper<GfxFamily>::programPipelineSelect(&commandStream, args, device->getHwInfo());
gpgpuEnabled = true;
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandQueueHw<gfxCoreFamily>::dispatchTaskCountWrite(NEO::LinearStream &commandStream, bool flushDataCache) {
using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
using POST_SYNC_OPERATION = typename PIPE_CONTROL::POST_SYNC_OPERATION;
UNRECOVERABLE_IF(csr == nullptr);
auto taskCountToWrite = csr->peekTaskCount() + 1;
auto gpuAddress = static_cast<uint64_t>(csr->getTagAllocation()->getGpuAddress());
if (isCopyOnlyCommandQueue) {
NEO::EncodeMiFlushDW<GfxFamily>::programMiFlushDw(commandStream, gpuAddress, taskCountToWrite, false, true);
} else {
NEO::PipeControlArgs args(true);
NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
commandStream,
POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
gpuAddress,
taskCountToWrite,
device->getHwInfo(),
args);
}
}
} // namespace L0
|
