intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-19 06:24:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: add BlitSyncPropertiesExt to BlitSyncProperties

Browse Source 
Added BlitSyncPropertiesExt to provide additional information for
different platform and/or blitter commands.

Signed-off-by: Aravind Gopalakrishnan <aravind.gopalakrishnan@intel.com>
Signed-off-by: Young Jin Yoon <young.jin.yoon@intel.com>
This commit is contained in:
Young Jin Yoon
2025-03-27 14:02:20 +00:00
committed by Compute-Runtime-Automation
parent 3438a14fc4
commit b5a259aded
41 changed files with 338 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
cmake/fill_core_types.cmake
View File

@@ -17,3 +17,4 @@ list(APPEND DG2_AND_LATER_CORE_TYPES ${PVC_AND_LATER_CORE_TYPES} "XE_HPG_CORE")
list(APPEND MTL_AND_LATER_CORE_TYPES ${DG2_AND_LATER_CORE_TYPES})
list(APPEND XEHP_AND_LATER_CORE_TYPES ${MTL_AND_LATER_CORE_TYPES})
list(APPEND XE_HPC_AND_BEFORE_CORE_TYPES "GEN12LP" "XE_HPG_CORE" "XE_HPC_CORE")
list(APPEND XE3_AND_BEFORE_CORE_TYPES "GEN12LP" "XE_HPG_CORE" "XE_HPC_CORE" "XE2_HPG_CORE" "XE3_CORE")

9
cmake/setup_platform_flags.cmake
View File

@@ -31,6 +31,15 @@ foreach(CORE_TYPE ${XE_HPC_AND_BEFORE_CORE_TYPES})
endif()
endforeach()
foreach(CORE_TYPE ${XE3_AND_BEFORE_CORE_TYPES})
if(TESTS_${CORE_TYPE})
set(TESTS_XE3_AND_BEFORE 1)
endif()
if(SUPPORT_${CORE_TYPE})
set(SUPPORT_XE3_AND_BEFORE 1)
endif()
endforeach()
foreach(CORE_TYPE ${XEHP_AND_LATER_CORE_TYPES})
if(TESTS_${CORE_TYPE})
set(TESTS_XEHP_AND_LATER 1)

14
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@@ -8,6 +8,7 @@
#pragma once
#include "shared/source/command_stream/transfer_direction.h"
#include "shared/source/helpers/blit_properties.h"
#include "shared/source/helpers/hw_mapper.h"
#include "shared/source/helpers/pipe_control_args.h"
#include "shared/source/helpers/vec.h"
@@ -239,7 +240,8 @@ struct CommandListCoreFamily : public CommandListImp {
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size);
uint64_t size,
Event *signalEvent);
MOCKABLE_VIRTUAL ze_result_t appendMemoryCopyBlitRegion(AlignedAllocationData *srcAllocationData,
AlignedAllocationData *dstAllocationData,
@@ -335,6 +337,8 @@ struct CommandListCoreFamily : public CommandListImp {
void addFlushRequiredCommand(bool flushOperationRequired, Event *signalEvent, bool copyOperation, bool flushL3InPipeControl);
void handlePostSubmissionState();
MOCKABLE_VIRTUAL void setAdditionalBlitProperties(NEO::BlitProperties &blitProperties, Event *signalEvent);
void setupFillKernelArguments(size_t baseOffset,
size_t patternSize,
size_t dstSize,
@@ -367,6 +371,13 @@ struct CommandListCoreFamily : public CommandListImp {
}
return this->containsStatelessUncachedResource;
}
bool isUsingSystemAllocation(const NEO::AllocationType &allocType) const {
return ((allocType == NEO::AllocationType::bufferHostMemory) ||
(allocType == NEO::AllocationType::svmCpu) ||
(allocType == NEO::AllocationType::externalHostPtr));
}
void postInitComputeSetup();
NEO::PreemptionMode obtainKernelPreemptionMode(Kernel *kernel);
virtual bool isRelaxedOrderingDispatchAllowed(uint32_t numWaitEvents, bool copyOffload) { return false; }
@@ -407,6 +418,7 @@ struct CommandListCoreFamily : public CommandListImp {
bool allowCbWaitEventsNoopDispatch = false;
bool copyOperationFenceSupported = false;
bool implicitSynchronizedDispatchForCooperativeKernelsAllowed = false;
bool useAdditionalBlitProperties = false;
};
template <PRODUCT_FAMILY gfxProductFamily>

27
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -264,6 +264,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::initialize(Device *device, NEO
this->maxLocalSubRegionSize = productHelper.getMaxLocalSubRegionSize(hwInfo);
this->l3FlushAfterPostSyncRequired = productHelper.isL3FlushAfterPostSyncRequired(heaplessModeEnabled);
this->compactL3FlushEventPacket = L0GfxCoreHelper::useCompactL3FlushEventPacket(hwInfo, this->l3FlushAfterPostSyncRequired);
this->useAdditionalBlitProperties = productHelper.useAdditionalBlitProperties();
if (NEO::debugManager.flags.OverrideThreadArbitrationPolicy.get() != -1) {
this->defaultPipelinedThreadArbitrationPolicy = NEO::debugManager.flags.OverrideThreadArbitrationPolicy.get();
@@ -1413,10 +1414,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyKernelWithGA(v
auto dstAllocationType = dstPtrAlloc->getAllocationType();
launchParams.isBuiltInKernel = true;
launchParams.isDestinationAllocationInSystemMemory =
(dstAllocationType == NEO::AllocationType::bufferHostMemory) ||
(dstAllocationType == NEO::AllocationType::svmCpu) ||
(dstAllocationType == NEO::AllocationType::externalHostPtr);
launchParams.isDestinationAllocationInSystemMemory = this->isUsingSystemAllocation(dstAllocationType);
if constexpr (checkIfAllocationImportedRequired()) {
launchParams.isDestinationAllocationImported = this->isAllocationImported(dstPtrAlloc, device->getDriverHandle()->getSvmAllocsManager());
@@ -1430,17 +1428,21 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyBlit(uintptr_t
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size) {
uint64_t size,
Event *signalEvent) {
dstOffset += ptrDiff<uintptr_t>(dstPtr, dstPtrAlloc->getGpuAddress());
srcOffset += ptrDiff<uintptr_t>(srcPtr, srcPtrAlloc->getGpuAddress());
auto clearColorAllocation = device->getNEODevice()->getDefaultEngine().commandStreamReceiver->getClearColorAllocation();
auto blitProperties = NEO::BlitProperties::constructPropertiesForCopy(dstPtrAlloc, srcPtrAlloc, {dstOffset, 0, 0}, {srcOffset, 0, 0}, {size, 0, 0}, 0, 0, 0, 0, clearColorAllocation);
commandContainer.addToResidencyContainer(dstPtrAlloc);
commandContainer.addToResidencyContainer(srcPtrAlloc);
commandContainer.addToResidencyContainer(clearColorAllocation);
if (useAdditionalBlitProperties) {
setAdditionalBlitProperties(blitProperties, signalEvent);
}
NEO::BlitPropertiesContainer blitPropertiesContainer{blitProperties};
NEO::BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForBufferPerRow(blitProperties, *commandContainer.getCommandStream(), *this->dummyBlitWa.rootDeviceEnvironment);
@@ -1559,7 +1561,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendPageFaultCopy(NEO::Graph
if (isCopyOnly(false)) {
return appendMemoryCopyBlit(dstAddress, dstAllocation, 0u,
srcAddress, srcAllocation, 0u,
size);
size, nullptr);
} else {
CmdListKernelLaunchParams launchParams = {};
launchParams.isKernelSplitOperation = rightSize > 0;
@@ -1712,7 +1714,9 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
launchParams.pipeControlSignalling = (signalEvent && singlePipeControlPacket) || getDcFlushRequired(dstAllocationStruct.needsFlush);
appendEventForProfilingAllWalkers(signalEvent, nullptr, nullptr, true, singlePipeControlPacket, false, isCopyOnlyEnabled);
if (!useAdditionalBlitProperties || !isCopyOnlyEnabled) {
appendEventForProfilingAllWalkers(signalEvent, nullptr, nullptr, true, singlePipeControlPacket, false, isCopyOnlyEnabled);
}
if (isCopyOnlyEnabled) {
if (NEO::debugManager.flags.FlushTlbBeforeCopy.get() == 1) {
@@ -1723,7 +1727,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
ret = appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr,
dstAllocationStruct.alloc, dstAllocationStruct.offset,
srcAllocationStruct.alignedAllocationPtr,
srcAllocationStruct.alloc, srcAllocationStruct.offset, size);
srcAllocationStruct.alloc, srcAllocationStruct.offset, size, signalEvent);
} else {
if (NEO::debugManager.flags.FlushTlbBeforeCopy.get() == 1) {
NEO::PipeControlArgs args;
@@ -1784,7 +1788,9 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
appendCopyOperationFence(signalEvent, srcAllocationStruct.alloc, dstAllocationStruct.alloc, isCopyOnlyEnabled);
appendEventForProfilingAllWalkers(signalEvent, nullptr, nullptr, false, singlePipeControlPacket, false, isCopyOnlyEnabled);
if (!useAdditionalBlitProperties || !isCopyOnlyEnabled) {
appendEventForProfilingAllWalkers(signalEvent, nullptr, nullptr, false, singlePipeControlPacket, false, isCopyOnlyEnabled);
}
bool l3flushInPipeControl = !l3FlushAfterPostSyncRequired || isSplitOperation;
addFlushRequiredCommand(dstAllocationStruct.needsFlush, signalEvent, isCopyOnlyEnabled, l3flushInPipeControl);
@@ -2475,7 +2481,6 @@ void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingCopyCommand(Ev
return;
}
commandContainer.addToResidencyContainer(event->getAllocation(this->device));
if (beforeWalker) {
event->resetKernelCountAndPacketUsedCount();
} else {

4
level_zero/core/source/cmdlist/cmdlist_hw_gen12lp_to_xe3.inl
View File

@@ -14,4 +14,8 @@ constexpr bool CommandListCoreFamily<gfxCoreFamily>::checkIfAllocationImportedRe
return false;
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::setAdditionalBlitProperties(NEO::BlitProperties &blitProperties, Event *signalEvent) {
}
} // namespace L0

2
level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
View File

@@ -812,7 +812,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendPageFaultCopy(N
ret = static_cast<DeviceImp *>(this->device)->bcsSplit.appendSplitCall<gfxCoreFamily, uintptr_t, uintptr_t>(this, dstAddress, srcAddress, size, nullptr, 0u, nullptr, false, relaxedOrdering, direction, [&](uintptr_t dstAddressParam, uintptr_t srcAddressParam, size_t sizeParam, ze_event_handle_t hSignalEventParam) {
this->appendMemoryCopyBlit(dstAddressParam, dstAllocation, 0u,
srcAddressParam, srcAllocation, 0u,
sizeParam);
sizeParam, nullptr);
return CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(hSignalEventParam, false);
});
} else {

3
level_zero/core/source/cmdlist/cmdlist_hw_skl_to_tgllp.inl
View File

@@ -207,9 +207,10 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
.inOrderIncrementValue = 0,
.device = neoDevice,
.inOrderExecInfo = nullptr,
.isCounterBasedEvent = false,
.isTimestampEvent = false,
.isHostScopeSignalEvent = false,
.isKernelUsingSystemAllocation = false,
.isUsingSystemAllocation = false,
.dcFlushEnable = this->dcFlushSupport,
.interruptEvent = false,
.isFlushL3ForExternalAllocationRequired = false,

5
level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
View File

@@ -319,6 +319,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
bool inOrderExecSignalRequired = false;
bool inOrderNonWalkerSignalling = false;
bool isCounterBasedEvent = false;
uint64_t inOrderCounterValue = 0;
uint64_t inOrderIncrementValue = 0;
@@ -348,6 +349,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
inOrderCounterValue = this->inOrderExecInfo->getCounterValue() + getInOrderIncrementValue();
inOrderExecInfo = this->inOrderExecInfo.get();
if (eventForInOrderExec && eventForInOrderExec->isCounterBased()) {
isCounterBasedEvent = true;
if (eventForInOrderExec->getInOrderIncrementValue() > 0) {
inOrderIncrementGpuAddress = eventForInOrderExec->getInOrderExecInfo()->getBaseDeviceAddress();
inOrderIncrementValue = eventForInOrderExec->getInOrderIncrementValue();
@@ -382,9 +384,10 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
.inOrderIncrementValue = inOrderIncrementValue,
.device = neoDevice,
.inOrderExecInfo = inOrderExecInfo,
.isCounterBasedEvent = isCounterBasedEvent,
.isTimestampEvent = isTimestampEvent,
.isHostScopeSignalEvent = isHostSignalScopeEvent,
.isKernelUsingSystemAllocation = isKernelUsingSystemAllocation,
.isUsingSystemAllocation = isKernelUsingSystemAllocation,
.dcFlushEnable = this->dcFlushSupport,
.interruptEvent = interruptEvent,
.isFlushL3ForExternalAllocationRequired = isFlushL3AfterPostSync && isKernelUsingExternalAllocation,

4
level_zero/core/test/unit_tests/mocks/mock_cmdlist.h
View File

@@ -114,6 +114,7 @@ struct WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>
using BaseClass::requiresDcFlushForDcMitigation;
using BaseClass::requiresQueueUncachedMocs;
using BaseClass::scratchAddressPatchingEnabled;
using BaseClass::setAdditionalBlitProperties;
using BaseClass::setupTimestampEventForMultiTile;
using BaseClass::signalAllEventPackets;
using BaseClass::stateBaseAddressTracking;
@@ -628,6 +629,7 @@ class MockCommandListCoreFamily : public CommandListCoreFamily<gfxCoreFamily> {
using BaseClass::encodeMiFlush;
using BaseClass::getDeviceCounterAllocForResidency;
using BaseClass::ownedPrivateAllocations;
using BaseClass::setAdditionalBlitProperties;
using BaseClass::taskCountUpdateFenceRequired;
ze_result_t executeMemAdvise(ze_device_handle_t hDevice,
@@ -654,7 +656,7 @@ class MockCommandListCoreFamily : public CommandListCoreFamily<gfxCoreFamily> {
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size));
uint64_t size, Event *signalEvent));
ADDMETHOD_VOIDRETURN(allocateOrReuseKernelPrivateMemory,
false,

3
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_2.cpp
View File

@@ -82,7 +82,7 @@ class MockCommandListHw : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFam
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size) override {
uint64_t size, Event *signalEvent) override {
appendMemoryCopyBlitCalledTimes++;
if (failOnFirstCopy && appendMemoryCopyBlitCalledTimes == 1) {
return ZE_RESULT_ERROR_UNKNOWN;
@@ -108,6 +108,7 @@ class MockCommandListHw : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFam
return ZE_RESULT_SUCCESS;
}
void setAdditionalBlitProperties(NEO::BlitProperties &blitProperties, Event *signalEvent) override {}
ze_result_t appendMemoryCopyKernel2d(AlignedAllocationData *dstAlignedAllocation, AlignedAllocationData *srcAlignedAllocation,
Builtin builtin, const ze_copy_region_t *dstRegion,
uint32_t dstPitch, size_t dstOffset,

3
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_memory.cpp
View File

@@ -683,6 +683,9 @@ HWTEST2_F(AppendMemoryCopyTests, givenCommandListAndHostPointersWhenMemoryCopyCa
HWTEST2_F(AppendMemoryCopyTests, givenCopyCommandListWhenTimestampPassedToMemoryCopyThenAppendProfilingCalledOnceBeforeAndAfterCommand, MatchAny) {
if (device->getProductHelper().useAdditionalBlitProperties()) {
GTEST_SKIP();
}
using MI_STORE_REGISTER_MEM = typename FamilyType::MI_STORE_REGISTER_MEM;
using MI_FLUSH_DW = typename FamilyType::MI_FLUSH_DW;

71
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_blit.cpp
View File

@@ -42,7 +42,7 @@ class MockCommandListForMemFill : public WhiteBox<::L0::CommandListCoreFamily<gf
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size) override {
uint64_t size, Event *signalEvent) override {
appendMemoryCopyBlitCalledTimes++;
return ZE_RESULT_SUCCESS;
}
@@ -213,7 +213,7 @@ HWTEST2_F(AppendMemoryCopyTests, givenCopyOnlyCommandListThenDcFlushIsNotAddedAf
NEO::MockGraphicsAllocation mockAllocationDst(0, 1u /*num gmms*/, NEO::AllocationType::internalHostMemory,
reinterpret_cast<void *>(dstPtr), 0x1000, 0, sizeof(uint32_t),
MemoryPool::system4KBPages, MemoryManager::maxOsContextCount);
commandList->appendMemoryCopyBlit(ptrOffset(dstPtr, dstOffset), &mockAllocationDst, 0, ptrOffset(srcPtr, srcOffset), &mockAllocationSrc, 0, copySize);
commandList->appendMemoryCopyBlit(ptrOffset(dstPtr, dstOffset), &mockAllocationDst, 0, ptrOffset(srcPtr, srcOffset), &mockAllocationSrc, 0, copySize, nullptr);
auto &commandContainer = commandList->getCmdContainer();
GenCmdList genCmdList;
@@ -429,5 +429,72 @@ HWTEST2_F(AppendMemoryCopyTests, givenCopyCommandListWhenNotTiled1DArrayImagePas
EXPECT_EQ(cmd->getSourceSurfaceHeight(), arrayLevels);
}
template <GFXCORE_FAMILY gfxCoreFamily>
class MockCommandListForAdditionalBlitProperties : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>> {
public:
using BaseClass = WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>;
using BaseClass::setAdditionalBlitProperties;
using BaseClass::useAdditionalBlitProperties;
};
HWTEST_F(AppendMemoryCopyTests, givenBlitPropertiesWhenCallingSetAdditionalBlitPropertiesThenSyncPropertiesExtRemainsUnchanged) {
NEO::BlitProperties blitProperties{}, blitProperties2{}, blitPropertiesExpected{};
EncodePostSyncArgs &postSyncArgs = blitProperties.blitSyncProperties.postSyncArgs;
EncodePostSyncArgs &postSyncArgs2 = blitProperties2.blitSyncProperties.postSyncArgs;
EncodePostSyncArgs &postSyncArgsExpected = blitPropertiesExpected.blitSyncProperties.postSyncArgs;
auto commandList = std::make_unique<MockCommandListForAdditionalBlitProperties<FamilyType::gfxCoreFamily>>();
EXPECT_FALSE(commandList->useAdditionalBlitProperties);
commandList->setAdditionalBlitProperties(blitProperties, nullptr);
EXPECT_EQ(postSyncArgs.isTimestampEvent, postSyncArgsExpected.isTimestampEvent);
EXPECT_EQ(postSyncArgs.postSyncImmValue, postSyncArgsExpected.postSyncImmValue);
EXPECT_EQ(postSyncArgs.interruptEvent, postSyncArgsExpected.interruptEvent);
EXPECT_EQ(postSyncArgs.eventAddress, postSyncArgsExpected.eventAddress);
commandList->useAdditionalBlitProperties = true;
commandList->setAdditionalBlitProperties(blitProperties2, nullptr);
EXPECT_EQ(postSyncArgs2.isTimestampEvent, postSyncArgsExpected.isTimestampEvent);
EXPECT_EQ(postSyncArgs2.postSyncImmValue, postSyncArgsExpected.postSyncImmValue);
EXPECT_EQ(postSyncArgs2.interruptEvent, postSyncArgsExpected.interruptEvent);
EXPECT_EQ(postSyncArgs2.eventAddress, postSyncArgsExpected.eventAddress);
EXPECT_EQ(nullptr, postSyncArgs2.inOrderExecInfo);
}
template <GFXCORE_FAMILY gfxCoreFamily>
class MockCommandListForAdditionalBlitProperties2 : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>> {
public:
using BaseClass = WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>;
using BaseClass::useAdditionalBlitProperties;
void setAdditionalBlitProperties(NEO::BlitProperties &blitProperties, Event *signalEvent) override {
additionalBlitPropertiesCalled++;
BaseClass::setAdditionalBlitProperties(blitProperties, signalEvent);
}
uint32_t additionalBlitPropertiesCalled = 0;
};
HWTEST_F(AppendMemoryCopyTests, givenCopyOnlyCommandListWithUseAdditionalBlitPropertiesWhenCallingAppendMemoryCopyBlitThenAdditionalBlitPropertiesCalled) {
auto commandList = std::make_unique<MockCommandListForAdditionalBlitProperties2<FamilyType::gfxCoreFamily>>();
commandList->initialize(device, NEO::EngineGroupType::copy, 0u);
uintptr_t srcPtr = 0x5001;
uintptr_t dstPtr = 0x7001;
uint64_t srcOffset = 0x101;
uint64_t dstOffset = 0x201;
uint64_t copySize = 0x301;
NEO::MockGraphicsAllocation mockAllocationSrc(0, 1u /*num gmms*/, NEO::AllocationType::internalHostMemory,
reinterpret_cast<void *>(srcPtr), 0x1000, 0, sizeof(uint32_t),
MemoryPool::system4KBPages, MemoryManager::maxOsContextCount);
NEO::MockGraphicsAllocation mockAllocationDst(0, 1u /*num gmms*/, NEO::AllocationType::internalHostMemory,
reinterpret_cast<void *>(dstPtr), 0x1000, 0, sizeof(uint32_t),
MemoryPool::system4KBPages, MemoryManager::maxOsContextCount);
commandList->useAdditionalBlitProperties = false;
EXPECT_EQ(0u, commandList->additionalBlitPropertiesCalled);
commandList->appendMemoryCopyBlit(ptrOffset(dstPtr, dstOffset), &mockAllocationDst, 0, ptrOffset(srcPtr, srcOffset), &mockAllocationSrc, 0, copySize, nullptr);
EXPECT_EQ(0u, commandList->additionalBlitPropertiesCalled);
commandList->useAdditionalBlitProperties = true;
commandList->appendMemoryCopyBlit(ptrOffset(dstPtr, dstOffset), &mockAllocationDst, 0, ptrOffset(srcPtr, srcOffset), &mockAllocationSrc, 0, copySize, nullptr);
EXPECT_EQ(1u, commandList->additionalBlitPropertiesCalled);
}
} // namespace ult
} // namespace L0

4
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_memory_extension.cpp
View File

@@ -120,7 +120,7 @@ class MockCommandListExtensionHw : public WhiteBox<::L0::CommandListCoreFamily<g
uint64_t dstOffset, uintptr_t srcPtr,
NEO::GraphicsAllocation *srcPtrAlloc,
uint64_t srcOffset,
uint64_t size) override {
uint64_t size, Event *signalEvent) override {
appendMemoryCopyBlitCalledTimes++;
if (failOnFirstCopy && appendMemoryCopyBlitCalledTimes == 1) {
return ZE_RESULT_ERROR_UNKNOWN;
@@ -128,6 +128,8 @@ class MockCommandListExtensionHw : public WhiteBox<::L0::CommandListCoreFamily<g
return ZE_RESULT_SUCCESS;
}
void setAdditionalBlitProperties(NEO::BlitProperties &blitProperties, Event *signalEvent) override {}
ze_result_t appendMemoryCopyBlitRegion(AlignedAllocationData *srcAllocationData,
AlignedAllocationData *dstAllocationData,
ze_copy_region_t srcRegion,

23
level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp
View File

@@ -3436,21 +3436,22 @@ void BcsSplitInOrderCmdListTests::verifySplitCmds(LinearStream &cmdStream, size_
ASSERT_NE(cmdList.end(), itor);
ASSERT_NE(nullptr, genCmdCast<XY_COPY_BLT *>(*itor));
auto flushDwItor = find<MI_FLUSH_DW *>(++itor, cmdList.end());
ASSERT_NE(cmdList.end(), flushDwItor);
auto signalSubCopyEvent = genCmdCast<MI_FLUSH_DW *>(*flushDwItor);
ASSERT_NE(nullptr, signalSubCopyEvent);
while (signalSubCopyEvent->getDestinationAddress() != signalSubCopyEventGpuVa) {
flushDwItor = find<MI_FLUSH_DW *>(++flushDwItor, cmdList.end());
if (!device->getProductHelper().useAdditionalBlitProperties()) {
auto flushDwItor = find<MI_FLUSH_DW *>(++itor, cmdList.end());
ASSERT_NE(cmdList.end(), flushDwItor);
signalSubCopyEvent = genCmdCast<MI_FLUSH_DW *>(*flushDwItor);
auto signalSubCopyEvent = genCmdCast<MI_FLUSH_DW *>(*flushDwItor);
ASSERT_NE(nullptr, signalSubCopyEvent);
}
itor = ++flushDwItor;
while (signalSubCopyEvent->getDestinationAddress() != signalSubCopyEventGpuVa) {
flushDwItor = find<MI_FLUSH_DW *>(++flushDwItor, cmdList.end());
ASSERT_NE(cmdList.end(), flushDwItor);
signalSubCopyEvent = genCmdCast<MI_FLUSH_DW *>(*flushDwItor);
ASSERT_NE(nullptr, signalSubCopyEvent);
}
itor = ++flushDwItor;
}
auto semaphoreCmds = findAll<MI_SEMAPHORE_WAIT *>(beginItor, itor);
EXPECT_EQ(numExpectedSemaphores, semaphoreCmds.size());

2
opencl/test/unit_test/command_stream/command_stream_receiver_hw_2_tests.cpp
View File

@@ -2259,4 +2259,4 @@ HWTEST_F(BcsTests, given1DNotTiledArrayImageWhenConstructPropertiesThenImageNotT
builtinOpParams);
EXPECT_EQ(blitProperties.copySize.y, builtinOpParams.size.y);
EXPECT_EQ(blitProperties.copySize.z, builtinOpParams.size.z);
}
}

8
shared/source/command_container/command_encoder.h
View File

@@ -57,16 +57,20 @@ struct EncodePostSyncArgs {
uint64_t inOrderIncrementValue = 0;
Device *device = nullptr;
NEO::InOrderExecInfo *inOrderExecInfo = nullptr;
bool isCounterBasedEvent = false;
bool isTimestampEvent = false;
bool isHostScopeSignalEvent = false;
bool isKernelUsingSystemAllocation = false;
bool isUsingSystemAllocation = false;
bool dcFlushEnable = false;
bool interruptEvent = false;
bool isFlushL3ForExternalAllocationRequired = false;
bool isFlushL3ForHostUsmRequired = false;
bool requiresSystemMemoryFence() const {
return (isHostScopeSignalEvent && isKernelUsingSystemAllocation && this->device->getProductHelper().isGlobalFenceInPostSyncRequired(this->device->getHardwareInfo()));
return (isHostScopeSignalEvent && isUsingSystemAllocation && this->device->getProductHelper().isGlobalFenceInPostSyncRequired(this->device->getHardwareInfo()));
}
bool isValidEvent() const {
return (eventAddress != 0) || (isCounterBasedEvent && !isTimestampEvent);
}
};

9
shared/source/command_container/command_encoder_from_xe_hpg_core_to_xe3_core.inl
View File

@@ -42,12 +42,19 @@ bool EncodeDispatchKernel<Family>::singleTileExecImplicitScalingRequired(bool co
return cooperativeKernel;
}
template <typename Family>
template <typename CommandType>
inline auto &EncodePostSync<Family>::getPostSync(CommandType &cmd, size_t index) {
UNRECOVERABLE_IF(index != 0);
return cmd.getPostSync();
}
template <typename Family>
template <typename CommandType>
void EncodePostSync<Family>::setupPostSyncForInOrderExec(CommandType &cmd, const EncodePostSyncArgs &args) {
using POSTSYNC_DATA = decltype(Family::template getPostSyncType<CommandType>());
auto &postSync = cmd.getPostSync();
auto &postSync = getPostSync(cmd, 0);
uint64_t gpuVa = args.inOrderExecInfo->getBaseDeviceAddress() + args.inOrderExecInfo->getAllocationOffset();
UNRECOVERABLE_IF(!(isAligned<immWriteDestinationAddressAlignment>(gpuVa)));

4
shared/source/command_container/command_encoder_xehp_and_later.inl
View File

@@ -372,7 +372,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
if (args.postSyncArgs.inOrderExecInfo) {
EncodePostSync<Family>::setupPostSyncForInOrderExec(walkerCmd, args.postSyncArgs);
} else if (args.postSyncArgs.eventAddress) {
} else if (args.postSyncArgs.isValidEvent()) {
EncodePostSync<Family>::setupPostSyncForRegularEvent(walkerCmd, args.postSyncArgs);
} else {
EncodeDispatchKernel<Family>::forceComputeWalkerPostSyncFlushWithWrite(walkerCmd);
@@ -487,7 +487,7 @@ template <typename CommandType>
void EncodePostSync<Family>::setupPostSyncForRegularEvent(CommandType &cmd, const EncodePostSyncArgs &args) {
using POSTSYNC_DATA = decltype(Family::template getPostSyncType<CommandType>());
auto &postSync = cmd.getPostSync();
auto &postSync = getPostSync(cmd, 0);
auto operationType = POSTSYNC_DATA::OPERATION_WRITE_IMMEDIATE_DATA;
uint64_t gpuVa = args.eventAddress;

3
shared/source/gen12lp/command_stream_receiver_hw_gen12lp.cpp
View File

@@ -16,6 +16,7 @@ using Family = NEO::Gen12LpFamily;
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/helpers/address_patch.h"
#include "shared/source/helpers/blit_commands_helper_base.inl"
#include "shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl"
#include "shared/source/helpers/populate_factory.h"
#include "shared/source/helpers/state_base_address_tgllp_and_later.inl"
@@ -332,6 +333,8 @@ bool BlitCommandsHelper<Family>::preBlitCommandWARequired() {
template class CommandStreamReceiverHw<Family>;
template struct BlitCommandsHelper<Family>;
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
const Family::GPGPU_WALKER Family::cmdInitGpgpuWalker = Family::GPGPU_WALKER::sInit();
const Family::INTERFACE_DESCRIPTOR_DATA Family::cmdInitInterfaceDescriptorData = Family::INTERFACE_DESCRIPTOR_DATA::sInit();
const Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD Family::cmdInitMediaInterfaceDescriptorLoad = Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD::sInit();

6
shared/source/helpers/CMakeLists.txt
View File

@@ -234,6 +234,12 @@ if(SUPPORT_XE3_AND_LATER)
)
endif()
if(SUPPORT_XE3_AND_BEFORE)
list(APPEND NEO_CORE_HELPERS
${CMAKE_CURRENT_SOURCE_DIR}/blit_commands_helper_from_gen12lp_to_xe3.inl
)
endif()
set(NEO_CORE_HELPERS_GMM_CALLBACKS_WINDOWS
${CMAKE_CURRENT_SOURCE_DIR}/windows/gmm_callbacks.cpp
${CMAKE_CURRENT_SOURCE_DIR}/windows/gmm_callbacks.h

3
shared/source/helpers/blit_commands_helper.h
View File

@@ -96,5 +96,8 @@ struct BlitCommandsHelper {
static void encodeWa(LinearStream &cmdStream, const BlitProperties &blitProperties, uint32_t &latestSentBcsWaValue);
static size_t getWaCmdsSize(const BlitPropertiesContainer &blitPropertiesContainer);
template <typename CommandType>
static void applyAdditionalBlitProperties(const BlitProperties &blitProperties, CommandType &cmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
};
} // namespace NEO

9
shared/source/helpers/blit_commands_helper_base.inl
View File

@@ -209,10 +209,13 @@ void BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForBufferPerRow(const Bl
appendColorDepth(blitProperties, bltCmd);
const bool useAdditionalBlitProperties = rootDeviceEnvironment.getHelper<ProductHelper>().useAdditionalBlitProperties();
for (uint64_t slice = 0; slice < blitProperties.copySize.z; slice++) {
for (uint64_t row = 0; row < blitProperties.copySize.y; row++) {
uint64_t offset = 0;
uint64_t sizeToBlit = blitProperties.copySize.x;
bool lastIteration = (slice == blitProperties.copySize.z - 1) && (row == blitProperties.copySize.y - 1);
while (sizeToBlit != 0) {
if (sizeToBlit > maxWidth) {
// dispatch 2D blit: maxBlitWidth x (1 .. maxBlitHeight)
@@ -223,6 +226,8 @@ void BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForBufferPerRow(const Bl
width = sizeToBlit;
height = 1;
}
auto blitSize = width * height;
auto lastCommand = lastIteration && (sizeToBlit - blitSize == 0);
bltCmd.setDestinationX2CoordinateRight(static_cast<uint32_t>(width));
bltCmd.setDestinationY2CoordinateBottom(static_cast<uint32_t>(height));
@@ -237,6 +242,9 @@ void BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForBufferPerRow(const Bl
bltCmd.setDestinationBaseAddress(dstAddr);
bltCmd.setSourceBaseAddress(srcAddr);
if (useAdditionalBlitProperties && lastCommand) {
applyAdditionalBlitProperties(blitProperties, bltCmd, rootDeviceEnvironment, lastCommand);
}
appendBlitCommandsForBuffer(blitProperties, bltCmd, rootDeviceEnvironment);
@@ -245,7 +253,6 @@ void BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForBufferPerRow(const Bl
dispatchPostBlitCommand(linearStream, rootDeviceEnvironment);
auto blitSize = width * height;
sizeToBlit -= blitSize;
offset += blitSize;
}

18
shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl Normal file
View File

@@ -0,0 +1,18 @@
/*
* Copyright (C) 2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/helpers/blit_properties.h"
namespace NEO {
template <typename GfxFamily>
template <typename CommandType>
void BlitCommandsHelper<GfxFamily>::applyAdditionalBlitProperties(const BlitProperties &blitProperties,
CommandType &cmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last) {
}
} // namespace NEO

2
shared/source/helpers/blit_properties.cpp
View File

@@ -220,4 +220,4 @@ void BlitProperties::transform1DArrayTo2DArrayIfNeeded() {
this->copySize.y = 1;
}
}
} // namespace NEO
} // namespace NEO

6
shared/source/helpers/blit_properties.h
View File

@@ -6,6 +6,7 @@
*/
#pragma once
#include "shared/source/command_container/command_encoder.h"
#include "shared/source/command_stream/csr_deps.h"
#include "shared/source/gmm_helper/gmm_lib.h"
#include "shared/source/helpers/aux_translation.h"
@@ -17,6 +18,7 @@
namespace NEO {
struct TimestampPacketDependencies;
struct BlitSyncPropertiesExt;
class TagNodeBase;
class TimestampPacketContainer;
class GraphicsAllocation;
@@ -31,13 +33,13 @@ enum class BlitSyncMode {
};
struct BlitSyncProperties {
EncodePostSyncArgs postSyncArgs{};
TagNodeBase *outputTimestampPacket = nullptr;
BlitSyncMode syncMode = BlitSyncMode::none;
uint64_t deviceGpuWriteAddress = 0;
uint64_t hostGpuWriteAddress = 0;
uint64_t timestampGpuWriteAddress = 0;
uint64_t writeValue = 0;
bool isTimestampMode() const {
return (syncMode == BlitSyncMode::timestamp) || (syncMode == BlitSyncMode::timestampAndImmediate);
}
@@ -104,4 +106,4 @@ struct BlitProperties {
bool isSystemMemoryPoolUsed = false;
};
} // namespace NEO
} // namespace NEO

1
shared/source/os_interface/product_helper.h
View File

@@ -271,6 +271,7 @@ class ProductHelper {
virtual void setRenderCompressedFlags(HardwareInfo &hwInfo) const = 0;
virtual bool isCompressionForbidden(const HardwareInfo &hwInfo) const = 0;
virtual bool isExposingSubdevicesAllowed() const = 0;
virtual bool useAdditionalBlitProperties() const = 0;
virtual ~ProductHelper() = default;

5
shared/source/os_interface/product_helper.inl
View File

@@ -1121,4 +1121,9 @@ bool ProductHelperHw<gfxProduct>::isExposingSubdevicesAllowed() const {
return true;
}
template <PRODUCT_FAMILY gfxProduct>
bool ProductHelperHw<gfxProduct>::useAdditionalBlitProperties() const {
return false;
}
} // namespace NEO

1
shared/source/os_interface/product_helper_hw.h
View File

@@ -209,6 +209,7 @@ class ProductHelperHw : public ProductHelper {
void setRenderCompressedFlags(HardwareInfo &hwInfo) const override;
bool isCompressionForbidden(const HardwareInfo &hwInfo) const override;
bool isExposingSubdevicesAllowed() const override;
bool useAdditionalBlitProperties() const override;
~ProductHelperHw() override = default;

4
shared/source/xe2_hpg_core/command_stream_receiver_hw_xe2_hpg_core.cpp
View File

@@ -15,6 +15,7 @@ using Family = NEO::Xe2HpgCoreFamily;
#include "shared/source/command_stream/command_stream_receiver_hw_heap_addressing.inl"
#include "shared/source/command_stream/command_stream_receiver_hw_xehp_and_later.inl"
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl"
#include "shared/source/helpers/blit_commands_helper_pvc_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_xe2_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_xehp_and_later.inl"
@@ -271,6 +272,9 @@ template void BlitCommandsHelper<Family>::appendColorDepth<typename Family::XY_B
template void BlitCommandsHelper<Family>::appendColorDepth<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd);
template void BlitCommandsHelper<Family>::appendBlitCommandsForBuffer<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
const Family::COMPUTE_WALKER Family::cmdInitGpgpuWalker = Family::COMPUTE_WALKER::sInit();
const Family::CFE_STATE Family::cmdInitCfeState = Family::CFE_STATE::sInit();
const Family::INTERFACE_DESCRIPTOR_DATA Family::cmdInitInterfaceDescriptorData = Family::INTERFACE_DESCRIPTOR_DATA::sInit();

4
shared/source/xe3_core/command_stream_receiver_hw_xe3_core.cpp
View File

@@ -14,6 +14,7 @@ using Family = NEO::Xe3CoreFamily;
#include "shared/source/command_stream/command_stream_receiver_hw_heap_addressing.inl"
#include "shared/source/command_stream/command_stream_receiver_hw_xehp_and_later.inl"
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl"
#include "shared/source/helpers/blit_commands_helper_pvc_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_xe2_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_xehp_and_later.inl"
@@ -225,6 +226,9 @@ template void BlitCommandsHelper<Family>::appendColorDepth<typename Family::XY_B
template void BlitCommandsHelper<Family>::appendColorDepth<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd);
template void BlitCommandsHelper<Family>::appendBlitCommandsForBuffer<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
const Family::COMPUTE_WALKER Family::cmdInitGpgpuWalker = Family::COMPUTE_WALKER::sInit();
const Family::CFE_STATE Family::cmdInitCfeState = Family::CFE_STATE::sInit();
const Family::INTERFACE_DESCRIPTOR_DATA Family::cmdInitInterfaceDescriptorData = Family::INTERFACE_DESCRIPTOR_DATA::sInit();

3
shared/source/xe_hpc_core/command_stream_receiver_hw_xe_hpc_core.cpp
View File

@@ -16,6 +16,7 @@ using Family = NEO::XeHpcCoreFamily;
#include "shared/source/command_stream/command_stream_receiver_hw_heap_addressing.inl"
#include "shared/source/command_stream/command_stream_receiver_hw_xehp_and_later.inl"
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl"
#include "shared/source/helpers/blit_commands_helper_pvc_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_xehp_and_later.inl"
#include "shared/source/helpers/blit_properties.h"
@@ -273,6 +274,8 @@ template class CommandStreamReceiverHw<Family>;
template struct BlitCommandsHelper<Family>;
template void BlitCommandsHelper<Family>::appendBlitCommandsForBuffer<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
const Family::COMPUTE_WALKER Family::cmdInitGpgpuWalker = Family::COMPUTE_WALKER::sInit();
const Family::CFE_STATE Family::cmdInitCfeState = Family::CFE_STATE::sInit();
const Family::INTERFACE_DESCRIPTOR_DATA Family::cmdInitInterfaceDescriptorData = Family::INTERFACE_DESCRIPTOR_DATA::sInit();

3
shared/source/xe_hpg_core/command_stream_receiver_hw_xe_hpg_core.cpp
View File

@@ -14,6 +14,7 @@ using Family = NEO::XeHpgCoreFamily;
#include "shared/source/command_stream/command_stream_receiver_hw_dg2_and_later.inl"
#include "shared/source/command_stream/command_stream_receiver_hw_heap_addressing.inl"
#include "shared/source/command_stream/command_stream_receiver_hw_xehp_and_later.inl"
#include "shared/source/helpers/blit_commands_helper_from_gen12lp_to_xe3.inl"
#include "shared/source/helpers/blit_commands_helper_xehp_and_later.inl"
#include "shared/source/helpers/populate_factory.h"
#include "shared/source/helpers/state_base_address_xehp_and_later.inl"
@@ -162,6 +163,8 @@ template struct BlitCommandsHelper<Family>;
template void BlitCommandsHelper<Family>::appendColorDepth<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd);
template void BlitCommandsHelper<Family>::appendBlitCommandsForBuffer<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment);
template void BlitCommandsHelper<Family>::applyAdditionalBlitProperties<typename Family::XY_BLOCK_COPY_BLT>(const BlitProperties &blitProperties, typename Family::XY_BLOCK_COPY_BLT &blitCmd, const RootDeviceEnvironment &rootDeviceEnvironment, bool last);
const Family::COMPUTE_WALKER Family::cmdInitGpgpuWalker = Family::COMPUTE_WALKER::sInit();
const Family::CFE_STATE Family::cmdInitCfeState = Family::CFE_STATE::sInit();
const Family::INTERFACE_DESCRIPTOR_DATA Family::cmdInitInterfaceDescriptorData = Family::INTERFACE_DESCRIPTOR_DATA::sInit();

6
shared/test/common/helpers/mock_product_helper_hw.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2024 Intel Corporation
* Copyright (C) 2022-2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -27,6 +27,7 @@ struct MockProductHelperHw : NEO::ProductHelperHw<productFamily> {
std::vector<uint32_t> getSupportedNumGrfs(const ReleaseHelper *releaseHelper) const override;
aub_stream::EngineType getDefaultCopyEngine() const override;
bool isBufferPoolAllocatorSupported() const override;
bool useAdditionalBlitProperties() const override { return enableAdditionalBlitProperties; }
bool use128MbEdram = false;
bool enableMidThreadPreemption = false;
@@ -36,6 +37,7 @@ struct MockProductHelperHw : NEO::ProductHelperHw<productFamily> {
bool isCooperativeEngineSupportedValue = true;
bool returnedIsUnlockingLockedPtrNecessary = false;
bool isBufferPoolAllocatorSupportedValue = true;
bool enableAdditionalBlitProperties = false;
uint32_t returnedStepping = 0;
uint32_t returnedL1CachePolicy = 0;
uint32_t returnedL1CachePolicyIfDebugger = 0;
@@ -43,4 +45,4 @@ struct MockProductHelperHw : NEO::ProductHelperHw<productFamily> {
std::vector<int32_t> threadArbPolicies = {};
aub_stream::EngineType mockDefaultCopyEngine = aub_stream::EngineType::ENGINE_BCS;
};
} // namespace NEO
} // namespace NEO

1
shared/test/common/test_macros/header/common_matchers.h
View File

@@ -30,6 +30,7 @@ using IsBeforeXeHpgCore = IsBeforeGfxCore<IGFX_XE_HPG_CORE>;
using IsAtLeastXeHpcCore = IsAtLeastGfxCore<IGFX_XE_HPC_CORE>;
using IsAtMostXeHpcCore = IsAtMostGfxCore<IGFX_XE_HPC_CORE>;
using IsBeforeXeHpcCore = IsBeforeGfxCore<IGFX_XE_HPC_CORE>;
using IsWithinXeHpcCoreAndXe3Core = IsWithinGfxCore<IGFX_XE_HPC_CORE, IGFX_XE3_CORE>;
using IsAtLeastXe2HpgCore = IsAtLeastGfxCore<IGFX_XE2_HPG_CORE>;
using IsAtMostXe2HpgCore = IsAtMostGfxCore<IGFX_XE2_HPG_CORE>;

22
shared/test/unit_test/encoders/test_command_encoder.cpp
View File

@@ -8,10 +8,12 @@
#include "shared/source/command_container/cmdcontainer.h"
#include "shared/source/command_container/command_encoder.h"
#include "shared/source/command_container/encode_surface_state.h"
#include "shared/source/helpers/in_order_cmd_helpers.h"
#include "shared/test/common/cmd_parse/gen_cmd_parse.h"
#include "shared/test/common/fixtures/device_fixture.h"
#include "shared/test/common/mocks/mock_device.h"
#include "shared/test/common/mocks/mock_graphics_allocation.h"
#include "shared/test/common/mocks/mock_timestamp_container.h"
#include "shared/test/common/test_macros/hw_test.h"
using namespace NEO;
@@ -220,7 +222,7 @@ HWTEST_F(CommandEncoderTest, givenEncodePostSyncArgsWhenCallingRequiresSystemMem
for (bool kernelUsingSystemAllocation : {true, false}) {
args.device = pDevice;
args.isHostScopeSignalEvent = hostScopeSignalEvent;
args.isKernelUsingSystemAllocation = kernelUsingSystemAllocation;
args.isUsingSystemAllocation = kernelUsingSystemAllocation;
if (hostScopeSignalEvent && kernelUsingSystemAllocation && pDevice->getProductHelper().isGlobalFenceInPostSyncRequired(pDevice->getHardwareInfo())) {
EXPECT_TRUE(args.requiresSystemMemoryFence());
@@ -230,3 +232,21 @@ HWTEST_F(CommandEncoderTest, givenEncodePostSyncArgsWhenCallingRequiresSystemMem
}
}
}
HWTEST_F(CommandEncoderTest, givenEncodePostSyncArgsWhenCallingIsRegularEventThenCorrectValuesAreReturned) {
EncodePostSyncArgs args{};
MockTagAllocator<DeviceAllocNodeType<true>> deviceTagAllocator(0, pDevice->getMemoryManager());
auto inOrderExecInfo = InOrderExecInfo::create(deviceTagAllocator.getTag(), deviceTagAllocator.getTag(), *pDevice, 1, false); // setting duplicateStorage = true;
for (bool counterBasedEvent : {true, false}) {
for (bool timestampEvent : {true, false}) {
for (uint64_t eventAddress : {0, 0x1010}) {
args.device = pDevice;
args.isCounterBasedEvent = counterBasedEvent;
args.isTimestampEvent = timestampEvent;
args.eventAddress = eventAddress;
bool expectedValidEvent = (eventAddress != 0) || (counterBasedEvent && !timestampEvent);
EXPECT_EQ(expectedValidEvent, args.isValidEvent());
}
}
}
}

3
shared/test/unit_test/fixtures/command_container_fixture.h
View File

@@ -54,9 +54,10 @@ class CommandEncodeStatesFixture : public DeviceFixture {
.inOrderIncrementValue = 0,
.device = device,
.inOrderExecInfo = nullptr,
.isCounterBasedEvent = false,
.isTimestampEvent = false,
.isHostScopeSignalEvent = false,
.isKernelUsingSystemAllocation = false,
.isUsingSystemAllocation = false,
.dcFlushEnable = false,
.interruptEvent = false,
.isFlushL3ForExternalAllocationRequired = false,

77
shared/test/unit_test/helpers/blit_commands_helper_tests.cpp
View File

@@ -15,6 +15,7 @@
#include "shared/test/common/fixtures/device_fixture.h"
#include "shared/test/common/helpers/debug_manager_state_restore.h"
#include "shared/test/common/helpers/default_hw_info.h"
#include "shared/test/common/helpers/mock_product_helper_hw.h"
#include "shared/test/common/mocks/mock_device.h"
#include "shared/test/common/mocks/mock_graphics_allocation.h"
#include "shared/test/common/mocks/mock_timestamp_container.h"
@@ -710,3 +711,79 @@ HWTEST_F(BlitTests, givenPlatformWhenCallingDispatchPreBlitCommandThenNoneMiFlus
blitPropertiesContainer2, false, true, false, false, pDevice->getRootDeviceEnvironment());
EXPECT_NE(estimatedSizeWithoutNode, estimatedSizeWithNode);
}
HWTEST2_F(BlitTests, givenPlatformWithBlitSyncPropertiesWithAndWithoutTimestampModeWhenCallingDispatchBlitCommandForBufferThenTheResultsAreTheSame, MatchAny) {
uint32_t src[] = {1, 2, 3, 4};
uint32_t dst[] = {4, 3, 2, 1};
uint32_t clear[] = {5, 6, 7, 8};
uint64_t srcGpuAddr = 0x12345;
uint64_t dstGpuAddr = 0x54321;
uint64_t clearGpuAddr = 0x5678;
std::unique_ptr<MockGraphicsAllocation> srcAlloc(new MockGraphicsAllocation(src, srcGpuAddr, sizeof(src)));
std::unique_ptr<MockGraphicsAllocation> dstAlloc(new MockGraphicsAllocation(dst, dstGpuAddr, sizeof(dst)));
std::unique_ptr<GraphicsAllocation> clearColorAllocation(new MockGraphicsAllocation(clear, clearGpuAddr, sizeof(clear)));
Vec3<size_t> srcOffsets{1, 0, 0};
Vec3<size_t> dstOffsets{1, 0, 0};
Vec3<size_t> copySize{(BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight) + 1, 2, 2};
size_t srcRowPitch = 0;
size_t srcSlicePitch = 0;
size_t dstRowPitch = 0;
size_t dstSlicePitch = 0;
auto blitProperties = NEO::BlitProperties::constructPropertiesForCopy(dstAlloc.get(), srcAlloc.get(),
dstOffsets, srcOffsets, copySize, srcRowPitch, srcSlicePitch,
dstRowPitch, dstSlicePitch, clearColorAllocation.get());
uint32_t streamBuffer[400] = {};
LinearStream stream(streamBuffer, sizeof(streamBuffer));
NEO::BlitCommandsHelper<FamilyType>::dispatchBlitCommandsForBufferPerRow(blitProperties, stream, pDevice->getRootDeviceEnvironmentRef());
// change it into timestamp mode
blitProperties.blitSyncProperties.syncMode = NEO::BlitSyncMode::timestamp;
uint32_t streamBuffer2[400] = {};
LinearStream stream2(streamBuffer2, sizeof(streamBuffer2));
NEO::BlitCommandsHelper<FamilyType>::dispatchBlitCommandsForBufferPerRow(blitProperties, stream2, pDevice->getRootDeviceEnvironmentRef());
EXPECT_EQ(stream.getUsed(), stream2.getUsed());
EXPECT_EQ(0, memcmp(ptrOffset(stream.getCpuBase(), 0), ptrOffset(stream2.getCpuBase(), 0), std::min(stream.getUsed(), stream2.getUsed())));
// change productHelper to return true
pDevice->getRootDeviceEnvironmentRef().productHelper.reset(new MockProductHelperHw<productFamily>);
auto *mockProductHelper = static_cast<MockProductHelperHw<productFamily> *>(pDevice->getRootDeviceEnvironmentRef().productHelper.get());
mockProductHelper->enableAdditionalBlitProperties = true;
uint32_t streamBuffer3[400] = {};
LinearStream stream3(streamBuffer3, sizeof(streamBuffer2));
NEO::BlitCommandsHelper<FamilyType>::dispatchBlitCommandsForBufferPerRow(blitProperties, stream3, pDevice->getRootDeviceEnvironmentRef());
EXPECT_EQ(stream.getUsed(), stream3.getUsed());
EXPECT_EQ(0, memcmp(ptrOffset(stream.getCpuBase(), 0), ptrOffset(stream3.getCpuBase(), 0), std::min(stream.getUsed(), stream3.getUsed())));
}
HWTEST_F(BlitTests, givenBlitPropertieswithImageOperationWhenCallingEstimateBlitCommandSizeThenBlockCopySizeIsReturned) {
Vec3<size_t> copySize{BlitterConstants::maxBlitWidth - 1, 1, 1};
NEO::CsrDependencies csrDependencies{};
size_t totalSize = NEO::BlitCommandsHelper<FamilyType>::estimateBlitCommandSize(copySize, csrDependencies, false, false, true, pDevice->getRootDeviceEnvironmentRef(), false, false);
size_t expectedSize = sizeof(typename FamilyType::XY_BLOCK_COPY_BLT);
expectedSize += NEO::BlitCommandsHelper<FamilyType>::estimatePostBlitCommandSize();
expectedSize += NEO::BlitCommandsHelper<FamilyType>::estimatePreBlitCommandSize();
EXPECT_EQ(expectedSize, totalSize);
}
using IsWithinGen12LPAndXe3Core = IsWithinGfxCore<IGFX_GEN12LP_CORE, IGFX_XE3_CORE>;
HWTEST2_F(BlitTests, givenXyCopyBltCommandWhenApplyBlitPropertiesIsCalledThenNothingChanged, IsWithinGen12LPAndXe3Core) {
using XY_COPY_BLT = typename FamilyType::XY_COPY_BLT;
auto bltCmd = FamilyType::cmdInitXyCopyBlt;
auto bltCmdBefore = bltCmd;
BlitProperties properties = {};
NEO::BlitCommandsHelper<FamilyType>::applyAdditionalBlitProperties(properties, bltCmd, pDevice->getRootDeviceEnvironment(), false);
EXPECT_EQ(memcmp(&bltCmd, &bltCmdBefore, sizeof(XY_COPY_BLT)), 0);
NEO::BlitCommandsHelper<FamilyType>::applyAdditionalBlitProperties(properties, bltCmd, pDevice->getRootDeviceEnvironment(), true);
EXPECT_EQ(memcmp(&bltCmd, &bltCmdBefore, sizeof(XY_COPY_BLT)), 0);
}

4
shared/test/unit_test/os_interface/product_helper_tests.cpp
View File

@@ -1176,3 +1176,7 @@ HWTEST_F(ProductHelperTest, givenProductHelperWhenIsExposingSubdevicesAllowedThe
HWTEST_F(ProductHelperTest, givenProductHelperWhenGettingIsPrimaryContextsAggregationSupportedThenReturnCorrectValue) {
EXPECT_FALSE(productHelper->isPrimaryContextsAggregationSupported());
}
HWTEST_F(ProductHelperTest, givenProductHelperWhenCallingUseAdditionalBlitPropertiesThenFalseReturned) {
EXPECT_FALSE(productHelper->useAdditionalBlitProperties());
}

4
shared/test/unit_test/xe2_hpg_core/test_encode_xe2_hpg_core.cpp
View File

@@ -373,7 +373,7 @@ XE2_HPG_CORETEST_F(EncodeKernelXe2HpgCoreTest, givenDefaultSettingForFenceWhenKe
dispatchInterface->getCrossThreadDataSizeResult = 0;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, false);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);
@@ -425,7 +425,7 @@ XE2_HPG_CORETEST_F(EncodeKernelXe2HpgCoreTest, givenDefaultSettingForFenceWhenKe
dispatchInterface->getCrossThreadDataSizeResult = 0;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, false);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isHostScopeSignalEvent = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);

4
shared/test/unit_test/xe3_core/test_encode_xe3_core.cpp
View File

@@ -314,7 +314,7 @@ XE3_CORETEST_F(EncodeKernelXe3CoreTest, givenDefaultSettingForFenceWhenKernelUse
dispatchInterface->getCrossThreadDataSizeResult = 0;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, false);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);
@@ -366,7 +366,7 @@ XE3_CORETEST_F(EncodeKernelXe3CoreTest, givenDefaultSettingForFenceWhenKernelUse
dispatchInterface->getCrossThreadDataSizeResult = 0;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, false);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isHostScopeSignalEvent = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);

4
shared/test/unit_test/xe_hpc_core/test_encode_xe_hpc_core.cpp
View File

@@ -374,7 +374,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDefaultSettingForFenceWhenKern
bool requiresUncachedMocs = false;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, requiresUncachedMocs);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);
@@ -445,7 +445,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDefaultSettingForFenceWhenKern
dispatchInterface->getCrossThreadDataSizeResult = 0;
EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, false);
dispatchArgs.postSyncArgs.isKernelUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isUsingSystemAllocation = true;
dispatchArgs.postSyncArgs.isHostScopeSignalEvent = true;
EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);