Skip to content

Commit fb4b1cc

Browse files
KamilKoprykIntelCompute-Runtime-Automation
KamilKoprykIntel
authored and
Compute-Runtime-Automation
committed
Use internal blitter for internal memory transfers
Signed-off-by: Kamil Kopryk <[email protected]> Related-To: NEO-6621
1 parent c354b73 commit fb4b1cc

File tree

14 files changed

+106
-27
lines changed

14 files changed

+106
-27
lines changed

‎opencl/source/command_queue/command_queue.cpp‎

Lines changed: 9 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -275,15 +275,19 @@ CommandStreamReceiver &CommandQueue::selectCsrForBuiltinOperation(const CsrSelec
275275

276276
void CommandQueue::initializeBcsEngine(bool internalUsage) {
277277
if (bcsAllowed && !bcsInitialized) {
278+
auto &hwInfo = device->getHardwareInfo();
279+
auto &hwHelper = HwHelper::get(hwInfo.platform.eRenderCoreFamily);
278280
auto &neoDevice = device->getNearestGenericSubDevice(0)->getDevice();
279281
auto &selectorCopyEngine = neoDevice.getSelectorCopyEngine();
280-
auto bcsEngineType = EngineHelpers::getBcsEngineType(device->getHardwareInfo(), device->getDeviceBitfield(), selectorCopyEngine, internalUsage);
281-
bcsEngines[EngineHelpers::getBcsIndex(bcsEngineType)] = neoDevice.tryGetEngine(bcsEngineType, EngineUsage::Regular);
282+
auto bcsEngineType = EngineHelpers::getBcsEngineType(hwInfo, device->getDeviceBitfield(), selectorCopyEngine, internalUsage);
283+
auto bcsIndex = EngineHelpers::getBcsIndex(bcsEngineType);
284+
auto engineUsage = (internalUsage && hwHelper.preferInternalBcsEngine()) ? EngineUsage::Internal : EngineUsage::Regular;
285+
bcsEngines[bcsIndex] = neoDevice.tryGetEngine(bcsEngineType, engineUsage);
282286
bcsEngineTypes.push_back(bcsEngineType);
283287
bcsInitialized = true;
284-
if (bcsEngines[EngineHelpers::getBcsIndex(bcsEngineType)]) {
285-
bcsEngines[EngineHelpers::getBcsIndex(bcsEngineType)]->osContext->ensureContextInitialized();
286-
bcsEngines[EngineHelpers::getBcsIndex(bcsEngineType)]->commandStreamReceiver->initDirectSubmission();
288+
if (bcsEngines[bcsIndex]) {
289+
bcsEngines[bcsIndex]->osContext->ensureContextInitialized();
290+
bcsEngines[bcsIndex]->commandStreamReceiver->initDirectSubmission();
287291
}
288292
}
289293
}

‎opencl/test/unit_test/command_queue/command_queue_tests.cpp‎

Lines changed: 26 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -305,6 +305,32 @@ TEST(CommandQueue, givenDeviceWithSubDevicesSupportingBlitOperationsWhenQueueIsC
305305
EXPECT_EQ(bcsEngine.commandStreamReceiver, cmdQ.getBcsCommandStreamReceiver(aub_stream::EngineType::ENGINE_BCS));
306306
}
307307

308+
TEST(CommandQueue, whenCommandQueueWithInternalUsageIsCreatedThenInternalBcsEngineIsUsed) {
309+
DebugManagerStateRestore restorer;
310+
DebugManager.flags.EnableBlitterForEnqueueOperations.set(1);
311+
DebugManager.flags.DeferCmdQBcsInitialization.set(0);
312+
HardwareInfo hwInfo = *defaultHwInfo;
313+
hwInfo.capabilityTable.blitterOperationsSupported = true;
314+
REQUIRE_FULL_BLITTER_OR_SKIP(&hwInfo);
315+
316+
auto &hwHelper = HwHelper::get(hwInfo.platform.eRenderCoreFamily);
317+
auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&hwInfo));
318+
auto internalUsage = true;
319+
auto expectedEngineType = EngineHelpers::linkCopyEnginesSupported(hwInfo, device->getDeviceBitfield())
320+
? aub_stream::EngineType::ENGINE_BCS2
321+
: aub_stream::EngineType::ENGINE_BCS;
322+
323+
for (auto preferInternalBcsEngine : {0, 1}) {
324+
DebugManager.flags.PreferInternalBcsEngine.set(preferInternalBcsEngine);
325+
auto engineUsage = hwHelper.preferInternalBcsEngine() ? EngineUsage::Internal : EngineUsage::Regular;
326+
MockCommandQueue cmdQ(nullptr, device.get(), 0, internalUsage);
327+
auto &bcsEngine = device->getEngine(expectedEngineType, engineUsage);
328+
329+
EXPECT_NE(nullptr, cmdQ.getBcsCommandStreamReceiver(expectedEngineType));
330+
EXPECT_EQ(bcsEngine.commandStreamReceiver, cmdQ.getBcsCommandStreamReceiver(expectedEngineType));
331+
}
332+
}
333+
308334
INSTANTIATE_TEST_CASE_P(uint32_t,
309335
CommandQueueWithBlitOperationsTests,
310336
::testing::Values(CL_COMMAND_WRITE_BUFFER,

‎opencl/test/unit_test/helpers/hw_helper_tests_xehp_and_later.cpp‎

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -84,6 +84,22 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, HwHelperTestXeHPAndLater, givenXeHPAndLaterPlatform
8484
EXPECT_TRUE(hwHelper.timestampPacketWriteSupported());
8585
}
8686

87+
HWCMDTEST_F(IGFX_XE_HP_CORE, HwHelperTestXeHPAndLater, givenXeHPAndLaterPlatformWhenPreferInternalBcsEngineThenReturnTrue) {
88+
auto &hwHelper = HwHelperHw<FamilyType>::get();
89+
EXPECT_TRUE(hwHelper.preferInternalBcsEngine());
90+
}
91+
92+
HWCMDTEST_F(IGFX_XE_HP_CORE, HwHelperTestXeHPAndLater, givenXeHPAndLaterPlatformAndDebugFlagsWhenPreferInternalBcsEngineThenReturnsCorrectResult) {
93+
DebugManagerStateRestore restore;
94+
auto &hwHelper = HwHelperHw<FamilyType>::get();
95+
96+
DebugManager.flags.PreferInternalBcsEngine.set(1);
97+
EXPECT_TRUE(hwHelper.preferInternalBcsEngine());
98+
99+
DebugManager.flags.PreferInternalBcsEngine.set(0);
100+
EXPECT_FALSE(hwHelper.preferInternalBcsEngine());
101+
}
102+
87103
HWCMDTEST_F(IGFX_XE_HP_CORE, HwHelperTestXeHPAndLater, givenAllFlagsSetWhenGetGpgpuEnginesThenReturnThreeRcsEnginesFourCcsEnginesAndOneBcsEngine) {
88104
HardwareInfo hwInfo = *defaultHwInfo;
89105
hwInfo.featureTable.flags.ftrCCSNode = true;

‎opencl/test/unit_test/test_files/igdrcl.config‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -386,6 +386,7 @@ MakeEachAllocationResident = -1
386386
AssignBCSAtEnqueue = -1
387387
DeferCmdQGpgpuInitialization = -1
388388
DeferCmdQBcsInitialization = -1
389+
PreferInternalBcsEngine = -1
389390
ReuseKernelBinaries = -1
390391
EnableChipsetUniqueUUID = -1
391392
ForceSimdMessageSizeInWalker = -1

‎opencl/test/unit_test/xe_hpc_core/hw_helper_tests_xe_hpc_core.cpp‎

Lines changed: 8 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -208,7 +208,7 @@ HWTEST2_F(HwHelperTestsXeHpcCore, givenRevisionEnumAndPlatformFamilyTypeThenProp
208208
}
209209

210210
XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEnginesAndFourCcsEnginesAndLinkCopyEngines) {
211-
const size_t numEngines = 17;
211+
const size_t numEngines = 18;
212212

213213
HardwareInfo hwInfo = *defaultHwInfo;
214214
hwInfo.featureTable.flags.ftrCCSNode = true;
@@ -241,6 +241,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEn
241241
{aub_stream::ENGINE_BCS, false, true},
242242
{aub_stream::ENGINE_BCS1, false, true},
243243
{aub_stream::ENGINE_BCS2, false, true},
244+
{aub_stream::ENGINE_BCS2, false, true},
244245
{aub_stream::ENGINE_BCS3, false, true},
245246
{aub_stream::ENGINE_BCS4, false, true},
246247
{aub_stream::ENGINE_BCS5, false, true},
@@ -257,7 +258,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEn
257258
}
258259

259260
XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEnginesAndFourCcsEnginesAndEightLinkCopyEngines) {
260-
const size_t numEngines = 17;
261+
const size_t numEngines = 18;
261262

262263
HardwareInfo hwInfo = *defaultHwInfo;
263264
hwInfo.featureTable.flags.ftrCCSNode = true;
@@ -290,6 +291,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEn
290291
{aub_stream::ENGINE_BCS, false, true},
291292
{aub_stream::ENGINE_BCS1, false, true},
292293
{aub_stream::ENGINE_BCS2, false, true},
294+
{aub_stream::ENGINE_BCS2, false, true},
293295
{aub_stream::ENGINE_BCS3, false, true},
294296
{aub_stream::ENGINE_BCS4, false, true},
295297
{aub_stream::ENGINE_BCS5, false, true},
@@ -306,7 +308,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, whenGetGpgpuEnginesThenReturnTwoCccsEn
306308
}
307309

308310
XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, givenCccsAsDefaultEngineWhenGetEnginesCalledThenChangeDefaultEngine) {
309-
const size_t numEngines = 17;
311+
const size_t numEngines = 18;
310312

311313
HardwareInfo hwInfo = *defaultHwInfo;
312314
hwInfo.featureTable.flags.ftrCCSNode = true;
@@ -339,6 +341,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, givenCccsAsDefaultEngineWhenGetEngines
339341
{aub_stream::ENGINE_BCS, false, true},
340342
{aub_stream::ENGINE_BCS1, false, true},
341343
{aub_stream::ENGINE_BCS2, false, true},
344+
{aub_stream::ENGINE_BCS2, false, true},
342345
{aub_stream::ENGINE_BCS3, false, true},
343346
{aub_stream::ENGINE_BCS4, false, true},
344347
{aub_stream::ENGINE_BCS5, false, true},
@@ -355,7 +358,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, givenCccsAsDefaultEngineWhenGetEngines
355358
}
356359

357360
XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, givenOneCcsEnabledWhenGetEnginesCalledThenCreateOnlyOneCcs) {
358-
const size_t numEngines = 14;
361+
const size_t numEngines = 15;
359362

360363
HardwareInfo hwInfo = *defaultHwInfo;
361364
hwInfo.featureTable.flags.ftrCCSNode = true;
@@ -385,6 +388,7 @@ XE_HPC_CORETEST_F(HwHelperTestsXeHpcCore, givenOneCcsEnabledWhenGetEnginesCalled
385388
{aub_stream::ENGINE_BCS, false, true},
386389
{aub_stream::ENGINE_BCS1, false, true},
387390
{aub_stream::ENGINE_BCS2, false, true},
391+
{aub_stream::ENGINE_BCS2, false, true},
388392
{aub_stream::ENGINE_BCS3, false, true},
389393
{aub_stream::ENGINE_BCS4, false, true},
390394
{aub_stream::ENGINE_BCS5, false, true},

‎opencl/test/unit_test/xe_hpc_core/pvc/engine_node_helper_tests_pvc.cpp‎

Lines changed: 13 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,5 @@
11
/*
2-
* Copyright (C) 2021 Intel Corporation
2+
* Copyright (C) 2021-2022 Intel Corporation
33
*
44
* SPDX-License-Identifier: MIT
55
*
@@ -47,8 +47,16 @@ PVCTEST_F(EngineNodeHelperPvcTests, givenPvcBaseDieA0AndTile1WhenGettingBcsEngin
4747
auto deviceBitfield = 0b10;
4848
auto &selectorCopyEngine = pDevice->getNearestGenericSubDevice(0)->getSelectorCopyEngine();
4949

50-
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, true));
51-
EXPECT_EQ(ENGINE_BCS4, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, true));
52-
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, true));
53-
EXPECT_EQ(ENGINE_BCS4, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, true));
50+
{
51+
auto internalUsage = true;
52+
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
53+
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
54+
}
55+
{
56+
auto internalUsage = false;
57+
EXPECT_EQ(ENGINE_BCS, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
58+
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
59+
EXPECT_EQ(ENGINE_BCS4, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
60+
EXPECT_EQ(ENGINE_BCS2, EngineHelpers::getBcsEngineType(*pHwInfo, deviceBitfield, selectorCopyEngine, internalUsage));
61+
}
5462
}

‎shared/source/debug_settings/debug_variables_base.inl‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -271,6 +271,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, MakeEachAllocationResident, -1, "-1: default, 0:
271271
DECLARE_DEBUG_VARIABLE(int32_t, AssignBCSAtEnqueue, -1, "-1: default, 0:disabled, 1: enabled.")
272272
DECLARE_DEBUG_VARIABLE(int32_t, DeferCmdQGpgpuInitialization, -1, "-1: default, 0:disabled, 1: enabled.")
273273
DECLARE_DEBUG_VARIABLE(int32_t, DeferCmdQBcsInitialization, -1, "-1: default, 0:disabled, 1: enabled.")
274+
DECLARE_DEBUG_VARIABLE(int32_t, PreferInternalBcsEngine, -1, "-1: default, 0:disabled, 1: enabled. When enabled use internal bcs engine for internal transfers, when disabled use regular engine")
274275
DECLARE_DEBUG_VARIABLE(int32_t, ReuseKernelBinaries, -1, "-1: default, 0:disabled, 1: enabled. If enabled, driver reuses kernel binaries.")
275276

276277
/*DIRECT SUBMISSION FLAGS*/

‎shared/source/helpers/blit_commands_helper.cpp‎

Lines changed: 5 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -177,6 +177,7 @@ BlitOperationResult BlitHelper::blitMemoryToAllocationBanks(const Device &device
177177
if (!hwInfo.capabilityTable.blitterOperationsSupported) {
178178
return BlitOperationResult::Unsupported;
179179
}
180+
auto &hwHelper = HwHelper::get(hwInfo.platform.eRenderCoreFamily);
180181

181182
UNRECOVERABLE_IF(memoryBanks.none());
182183

@@ -189,10 +190,12 @@ BlitOperationResult BlitHelper::blitMemoryToAllocationBanks(const Device &device
189190

190191
UNRECOVERABLE_IF(!pRootDevice->getDeviceBitfield().test(tileId));
191192
auto pDeviceForBlit = pRootDevice->getNearestGenericSubDevice(tileId);
192-
193193
auto &selectorCopyEngine = pDeviceForBlit->getSelectorCopyEngine();
194194
auto deviceBitfield = pDeviceForBlit->getDeviceBitfield();
195-
auto bcsEngine = pDeviceForBlit->tryGetEngine(EngineHelpers::getBcsEngineType(hwInfo, deviceBitfield, selectorCopyEngine, true), EngineUsage::Regular);
195+
auto internalUsage = true;
196+
auto bcsEngineType = EngineHelpers::getBcsEngineType(hwInfo, deviceBitfield, selectorCopyEngine, internalUsage);
197+
auto bcsEngineUsage = hwHelper.preferInternalBcsEngine() ? EngineUsage::Internal : EngineUsage::Regular;
198+
auto bcsEngine = pDeviceForBlit->tryGetEngine(bcsEngineType, bcsEngineUsage);
196199
if (!bcsEngine) {
197200
return BlitOperationResult::Unsupported;
198201
}

‎shared/source/helpers/engine_node_helper.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -103,7 +103,7 @@ aub_stream::EngineType getBcsEngineType(const HardwareInfo &hwInfo, const Device
103103
return DebugManager.flags.ForceBCSForInternalCopyEngine.get() == 0 ? aub_stream::EngineType::ENGINE_BCS
104104
: static_cast<aub_stream::EngineType>(aub_stream::EngineType::ENGINE_BCS1 + DebugManager.flags.ForceBCSForInternalCopyEngine.get() - 1);
105105
}
106-
return selectLinkCopyEngine(hwInfo, deviceBitfield, selectorCopyEngine.selector);
106+
return aub_stream::ENGINE_BCS2;
107107
}
108108

109109
const bool isMainCopyEngineAlreadyUsed = selectorCopyEngine.isMainUsed.exchange(true);

‎shared/source/helpers/hw_helper.h‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -143,7 +143,7 @@ class HwHelper {
143143
virtual void setSipKernelData(uint32_t *&sipKernelBinary, size_t &kernelBinarySize) const = 0;
144144
virtual void adjustPreemptionSurfaceSize(size_t &csrSize) const = 0;
145145
virtual size_t getSamplerStateSize() const = 0;
146-
146+
virtual bool preferInternalBcsEngine() const = 0;
147147
virtual bool isScratchSpaceSurfaceStateAccessible() const = 0;
148148
virtual uint64_t getRenderSurfaceStateBaseAddress(void *renderSurfaceState) const = 0;
149149
virtual uint32_t getRenderSurfaceStatePitch(void *renderSurfaceState) const = 0;
@@ -391,7 +391,7 @@ class HwHelperHw : public HwHelper {
391391
void adjustPreemptionSurfaceSize(size_t &csrSize) const override;
392392

393393
bool isScratchSpaceSurfaceStateAccessible() const override;
394-
394+
bool preferInternalBcsEngine() const override;
395395
size_t getMax3dImageWidthOrHeight() const override;
396396
uint64_t getMaxMemAllocSize() const override;
397397
uint64_t getPatIndex(CacheRegion cacheRegion, CachePolicy cachePolicy) const override;

0 commit comments

Comments
 (0)