intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-20 00:24:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: change parameter names into more meaningful

Browse Source 
- change additional size into local region size
- change walk order into dispatch walk order to distinguish for local id walk

Related-To: NEO-13350

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2024-11-27 00:13:50 +00:00
committed by Compute-Runtime-Automation
parent b56d4324c5
commit 6b7235cd6c
26 changed files with 168 additions and 168 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@@ -305,7 +305,7 @@ struct CommandListCoreFamily : public CommandListImp {
ze_result_t setGlobalWorkSizeIndirect(NEO::CrossThreadDataOffset offsets[3], uint64_t crossThreadAddress, uint32_t lws[3]); ze_result_t setGlobalWorkSizeIndirect(NEO::CrossThreadDataOffset offsets[3], uint64_t crossThreadAddress, uint32_t lws[3]);
ze_result_t programSyncBuffer(Kernel &kernel, NEO::Device &device, const ze_group_count_t &threadGroupDimensions); ze_result_t programSyncBuffer(Kernel &kernel, NEO::Device &device, const ze_group_count_t &threadGroupDimensions);
void programRegionGroupBarrier(Kernel &kernel, const ze_group_count_t &threadGroupDimensions, size_t additionalSizeParam); void programRegionGroupBarrier(Kernel &kernel, const ze_group_count_t &threadGroupDimensions, size_t localRegionSize);
void appendWriteKernelTimestamp(Event *event, CommandToPatchContainer *outTimeStampSyncCmds, bool beforeWalker, bool maskLsb, bool workloadPartition, bool copyOperation); void appendWriteKernelTimestamp(Event *event, CommandToPatchContainer *outTimeStampSyncCmds, bool beforeWalker, bool maskLsb, bool workloadPartition, bool copyOperation);
void adjustWriteKernelTimestamp(uint64_t globalAddress, uint64_t contextAddress, uint64_t baseAddress, CommandToPatchContainer *outTimeStampSyncCmds, bool maskLsb, uint32_t mask, bool workloadPartition, bool copyOperation); void adjustWriteKernelTimestamp(uint64_t globalAddress, uint64_t contextAddress, uint64_t baseAddress, CommandToPatchContainer *outTimeStampSyncCmds, bool maskLsb, uint32_t mask, bool workloadPartition, bool copyOperation);
void appendEventForProfiling(Event *event, CommandToPatchContainer *outTimeStampSyncCmds, bool beforeWalker, bool skipBarrierForEndProfiling, bool skipAddingEventToResidency, bool copyOperation); void appendEventForProfiling(Event *event, CommandToPatchContainer *outTimeStampSyncCmds, bool beforeWalker, bool skipBarrierForEndProfiling, bool skipAddingEventToResidency, bool copyOperation);

10
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -2821,14 +2821,14 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::programSyncBuffer(Kernel &kern
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::programRegionGroupBarrier(Kernel &kernel, const ze_group_count_t &threadGroupDimensions, size_t additionalSizeParam) { void CommandListCoreFamily<gfxCoreFamily>::programRegionGroupBarrier(Kernel &kernel, const ze_group_count_t &threadGroupDimensions, size_t localRegionSize) {
auto neoDevice = device->getNEODevice(); auto neoDevice = device->getNEODevice();
neoDevice->allocateSyncBufferHandler(); neoDevice->allocateSyncBufferHandler();
const size_t requestedNumberOfWorkgroups = threadGroupDimensions.groupCountX * threadGroupDimensions.groupCountY * threadGroupDimensions.groupCountZ; const size_t requestedNumberOfWorkgroups = threadGroupDimensions.groupCountX * threadGroupDimensions.groupCountY * threadGroupDimensions.groupCountZ;
size_t size = alignUp((requestedNumberOfWorkgroups / additionalSizeParam) * (additionalSizeParam + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize); size_t size = alignUp((requestedNumberOfWorkgroups / localRegionSize) * (localRegionSize + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize);
auto patchData = neoDevice->syncBufferHandler->obtainAllocationAndOffset(size); auto patchData = neoDevice->syncBufferHandler->obtainAllocationAndOffset(size);
@@ -4279,11 +4279,11 @@ template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::setAdditionalKernelLaunchParams(CmdListKernelLaunchParams &launchParams, Kernel &kernel) const { void CommandListCoreFamily<gfxCoreFamily>::setAdditionalKernelLaunchParams(CmdListKernelLaunchParams &launchParams, Kernel &kernel) const {
auto &kernelDescriptor = kernel.getImmutableData()->getDescriptor(); auto &kernelDescriptor = kernel.getImmutableData()->getDescriptor();
if (launchParams.additionalSizeParam == NEO::additionalKernelLaunchSizeParamNotSet) { if (launchParams.localRegionSize == NEO::localRegionSizeParamNotSet) {
launchParams.additionalSizeParam = kernelDescriptor.kernelAttributes.additionalSize; launchParams.localRegionSize = kernelDescriptor.kernelAttributes.localRegionSize;
} }
if (launchParams.requiredDispatchWalkOrder == NEO::RequiredDispatchWalkOrder::none) { if (launchParams.requiredDispatchWalkOrder == NEO::RequiredDispatchWalkOrder::none) {
launchParams.requiredDispatchWalkOrder = kernelDescriptor.kernelAttributes.walkOrder; launchParams.requiredDispatchWalkOrder = kernelDescriptor.kernelAttributes.dispatchWalkOrder;
} }
if (launchParams.requiredPartitionDim == NEO::RequiredPartitionDim::none) { if (launchParams.requiredPartitionDim == NEO::RequiredPartitionDim::none) {
launchParams.requiredPartitionDim = kernelDescriptor.kernelAttributes.partitionDim; launchParams.requiredPartitionDim = kernelDescriptor.kernelAttributes.partitionDim;

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

@@ -207,7 +207,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
commandListPreemptionMode, // preemptionMode commandListPreemptionMode, // preemptionMode
launchParams.requiredPartitionDim, // requiredPartitionDim launchParams.requiredPartitionDim, // requiredPartitionDim
launchParams.requiredDispatchWalkOrder, // requiredDispatchWalkOrder launchParams.requiredDispatchWalkOrder, // requiredDispatchWalkOrder
launchParams.additionalSizeParam, // additionalSizeParam launchParams.localRegionSize, // localRegionSize
0, // partitionCount 0, // partitionCount
launchParams.reserveExtraPayloadSpace, // reserveExtraPayloadSpace launchParams.reserveExtraPayloadSpace, // reserveExtraPayloadSpace
maxWgCountPerTile, // maxWgCountPerTile maxWgCountPerTile, // maxWgCountPerTile

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

@@ -256,7 +256,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
} }
if (kernel->usesRegionGroupBarrier()) { if (kernel->usesRegionGroupBarrier()) {
programRegionGroupBarrier(*kernel, threadGroupDimensions, launchParams.additionalSizeParam); programRegionGroupBarrier(*kernel, threadGroupDimensions, launchParams.localRegionSize);
} }
bool uncachedMocsKernel = isKernelUncachedMocsRequired(kernelImp->getKernelRequiresUncachedMocs()); bool uncachedMocsKernel = isKernelUncachedMocsRequired(kernelImp->getKernelRequiresUncachedMocs());
@@ -331,7 +331,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
kernelPreemptionMode, // preemptionMode kernelPreemptionMode, // preemptionMode
launchParams.requiredPartitionDim, // requiredPartitionDim launchParams.requiredPartitionDim, // requiredPartitionDim
launchParams.requiredDispatchWalkOrder, // requiredDispatchWalkOrder launchParams.requiredDispatchWalkOrder, // requiredDispatchWalkOrder
launchParams.additionalSizeParam, // additionalSizeParam launchParams.localRegionSize, // localRegionSize
this->partitionCount, // partitionCount this->partitionCount, // partitionCount
launchParams.reserveExtraPayloadSpace, // reserveExtraPayloadSpace launchParams.reserveExtraPayloadSpace, // reserveExtraPayloadSpace
maxWgCountPerTile, // maxWgCountPerTile maxWgCountPerTile, // maxWgCountPerTile

2
level_zero/core/source/cmdlist/cmdlist_launch_params.h
View File

@@ -54,7 +54,7 @@ struct CmdListKernelLaunchParams {
uint32_t externalPerThreadScratchSize[2] = {0U, 0U}; uint32_t externalPerThreadScratchSize[2] = {0U, 0U};
NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none; NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none;
NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none; NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
uint32_t additionalSizeParam = NEO::additionalKernelLaunchSizeParamNotSet; uint32_t localRegionSize = NEO::localRegionSizeParamNotSet;
uint32_t numKernelsInSplitLaunch = 0; uint32_t numKernelsInSplitLaunch = 0;
uint32_t numKernelsExecutedInSplitLaunch = 0; uint32_t numKernelsExecutedInSplitLaunch = 0;
uint32_t reserveExtraPayloadSpace = 0; uint32_t reserveExtraPayloadSpace = 0;

6
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_6.cpp
View File

@@ -65,13 +65,13 @@ HWTEST2_F(MultiTileImmediateCommandListTest, givenMultipleTilesWhenAllocatingBar
size_t requestedNumberOfWorkgroups = threadGroupDimensions.groupCountX * threadGroupDimensions.groupCountY * threadGroupDimensions.groupCountZ; size_t requestedNumberOfWorkgroups = threadGroupDimensions.groupCountX * threadGroupDimensions.groupCountY * threadGroupDimensions.groupCountZ;
size_t additionalSizeParam = 4; size_t localRegionSize = 4;
whiteBoxCmdList->programRegionGroupBarrier(mockKernel, threadGroupDimensions, additionalSizeParam); whiteBoxCmdList->programRegionGroupBarrier(mockKernel, threadGroupDimensions, localRegionSize);
auto patchData = neoDevice->syncBufferHandler->obtainAllocationAndOffset(1); auto patchData = neoDevice->syncBufferHandler->obtainAllocationAndOffset(1);
size_t expectedOffset = alignUp((requestedNumberOfWorkgroups / additionalSizeParam) * (additionalSizeParam + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize); size_t expectedOffset = alignUp((requestedNumberOfWorkgroups / localRegionSize) * (localRegionSize + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize);
EXPECT_EQ(patchData.second, expectedOffset); EXPECT_EQ(patchData.second, expectedOffset);
} }

76
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_1.cpp
View File

@@ -196,44 +196,44 @@ HWTEST2_F(CommandListAppendLaunchKernel, givenNotEnoughSpaceInCommandStreamWhenA
const uint32_t threadGroupDimensions[3] = {1, 1, 1}; const uint32_t threadGroupDimensions[3] = {1, 1, 1};
NEO::EncodeDispatchKernelArgs dispatchKernelArgs{ NEO::EncodeDispatchKernelArgs dispatchKernelArgs{
0, // eventAddress 0, // eventAddress
0, // postSyncImmValue 0, // postSyncImmValue
0, // inOrderCounterValue 0, // inOrderCounterValue
device->getNEODevice(), // device device->getNEODevice(), // device
nullptr, // inOrderExecInfo nullptr, // inOrderExecInfo
kernel.get(), // dispatchInterface kernel.get(), // dispatchInterface
nullptr, // surfaceStateHeap nullptr, // surfaceStateHeap
nullptr, // dynamicStateHeap nullptr, // dynamicStateHeap
threadGroupDimensions, // threadGroupDimensions threadGroupDimensions, // threadGroupDimensions
nullptr, // outWalkerPtr nullptr, // outWalkerPtr
nullptr, // cpuWalkerBuffer nullptr, // cpuWalkerBuffer
nullptr, // cpuPayloadBuffer nullptr, // cpuPayloadBuffer
nullptr, // outImplicitArgsPtr nullptr, // outImplicitArgsPtr
nullptr, // additionalCommands nullptr, // additionalCommands
PreemptionMode::MidBatch, // preemptionMode PreemptionMode::MidBatch, // preemptionMode
NEO::RequiredPartitionDim::none, // requiredPartitionDim NEO::RequiredPartitionDim::none, // requiredPartitionDim
NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
NEO::additionalKernelLaunchSizeParamNotSet, // additionalSizeParam NEO::localRegionSizeParamNotSet, // localRegionSize
0, // partitionCount 0, // partitionCount
0, // reserveExtraPayloadSpace 0, // reserveExtraPayloadSpace
1, // maxWgCountPerTile 1, // maxWgCountPerTile
NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy
false, // isIndirect false, // isIndirect
false, // isPredicate false, // isPredicate
false, // isTimestampEvent false, // isTimestampEvent
false, // requiresUncachedMocs false, // requiresUncachedMocs
false, // isInternal false, // isInternal
false, // isCooperative false, // isCooperative
false, // isHostScopeSignalEvent false, // isHostScopeSignalEvent
false, // isKernelUsingSystemAllocation false, // isKernelUsingSystemAllocation
false, // isKernelDispatchedFromImmediateCmdList false, // isKernelDispatchedFromImmediateCmdList
false, // isRcs false, // isRcs
commandList->getDcFlushRequired(true), // dcFlushEnable commandList->getDcFlushRequired(true), // dcFlushEnable
false, // isHeaplessModeEnabled false, // isHeaplessModeEnabled
false, // isHeaplessStateInitEnabled false, // isHeaplessStateInitEnabled
false, // interruptEvent false, // interruptEvent
false, // immediateScratchAddressPatching false, // immediateScratchAddressPatching
false, // makeCommandView false, // makeCommandView
}; };
NEO::EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(commandContainer, dispatchKernelArgs); NEO::EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(commandContainer, dispatchKernelArgs);

80
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_3.cpp
View File

@@ -541,7 +541,7 @@ HWTEST2_F(CommandListAppendLaunchKernel, givenKernelUsingRegionGroupBarrierWhenA
std::unique_ptr<L0::CommandList> cmdList(CommandList::createImmediate(productFamily, device, &desc, false, NEO::EngineGroupType::renderCompute, result)); std::unique_ptr<L0::CommandList> cmdList(CommandList::createImmediate(productFamily, device, &desc, false, NEO::EngineGroupType::renderCompute, result));
CmdListKernelLaunchParams launchParams = {}; CmdListKernelLaunchParams launchParams = {};
launchParams.additionalSizeParam = 4; launchParams.localRegionSize = 4;
EXPECT_EQ(ZE_RESULT_SUCCESS, cmdList->appendLaunchKernel(kernel.toHandle(), groupCount, nullptr, 0, nullptr, launchParams, false)); EXPECT_EQ(ZE_RESULT_SUCCESS, cmdList->appendLaunchKernel(kernel.toHandle(), groupCount, nullptr, 0, nullptr, launchParams, false));
auto patchPtr = *reinterpret_cast<uint64_t *>(ptrOffset(kernel.crossThreadData.get(), regionGroupBarrier.stateless)); auto patchPtr = *reinterpret_cast<uint64_t *>(ptrOffset(kernel.crossThreadData.get(), regionGroupBarrier.stateless));
@@ -572,7 +572,7 @@ HWTEST2_F(CommandListAppendLaunchKernel, givenKernelUsingRegionGroupBarrierWhenA
size_t requestedNumberOfWorkgroups = groupCount.groupCountX * groupCount.groupCountY * groupCount.groupCountZ; size_t requestedNumberOfWorkgroups = groupCount.groupCountX * groupCount.groupCountY * groupCount.groupCountZ;
auto offset = alignUp((requestedNumberOfWorkgroups / launchParams.additionalSizeParam) * (launchParams.additionalSizeParam + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize); auto offset = alignUp((requestedNumberOfWorkgroups / launchParams.localRegionSize) * (launchParams.localRegionSize + 1) * 2 * sizeof(uint32_t), MemoryConstants::cacheLineSize);
EXPECT_EQ(patchPtr2, patchPtr + offset); EXPECT_EQ(patchPtr2, patchPtr + offset);
} }
@@ -778,44 +778,44 @@ HWTEST2_F(CommandListAppendLaunchKernel, givenNotEnoughSpaceInCommandStreamWhenA
const uint32_t threadGroupDimensions[3] = {1, 1, 1}; const uint32_t threadGroupDimensions[3] = {1, 1, 1};
NEO::EncodeDispatchKernelArgs dispatchKernelArgs{ NEO::EncodeDispatchKernelArgs dispatchKernelArgs{
0, // eventAddress 0, // eventAddress
0, // postSyncImmValue 0, // postSyncImmValue
0, // inOrderCounterValue 0, // inOrderCounterValue
device->getNEODevice(), // device device->getNEODevice(), // device
nullptr, // inOrderExecInfo nullptr, // inOrderExecInfo
kernel.get(), // dispatchInterface kernel.get(), // dispatchInterface
nullptr, // surfaceStateHeap nullptr, // surfaceStateHeap
nullptr, // dynamicStateHeap nullptr, // dynamicStateHeap
threadGroupDimensions, // threadGroupDimensions threadGroupDimensions, // threadGroupDimensions
nullptr, // outWalkerPtr nullptr, // outWalkerPtr
nullptr, // cpuWalkerBuffer nullptr, // cpuWalkerBuffer
nullptr, // cpuPayloadBuffer nullptr, // cpuPayloadBuffer
nullptr, // outImplicitArgsPtr nullptr, // outImplicitArgsPtr
nullptr, // additionalCommands nullptr, // additionalCommands
PreemptionMode::MidBatch, // preemptionMode PreemptionMode::MidBatch, // preemptionMode
NEO::RequiredPartitionDim::none, // requiredPartitionDim NEO::RequiredPartitionDim::none, // requiredPartitionDim
NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
NEO::additionalKernelLaunchSizeParamNotSet, // additionalSizeParam NEO::localRegionSizeParamNotSet, // localRegionSize
0, // partitionCount 0, // partitionCount
0, // reserveExtraPayloadSpace 0, // reserveExtraPayloadSpace
1, // maxWgCountPerTile 1, // maxWgCountPerTile
NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy
false, // isIndirect false, // isIndirect
false, // isPredicate false, // isPredicate
false, // isTimestampEvent false, // isTimestampEvent
false, // requiresUncachedMocs false, // requiresUncachedMocs
false, // isInternal false, // isInternal
false, // isCooperative false, // isCooperative
false, // isHostScopeSignalEvent false, // isHostScopeSignalEvent
false, // isKernelUsingSystemAllocation false, // isKernelUsingSystemAllocation
false, // isKernelDispatchedFromImmediateCmdList false, // isKernelDispatchedFromImmediateCmdList
false, // isRcs false, // isRcs
commandList->getDcFlushRequired(true), // dcFlushEnable commandList->getDcFlushRequired(true), // dcFlushEnable
false, // isHeaplessModeEnabled false, // isHeaplessModeEnabled
false, // isHeaplessStateInitEnabled false, // isHeaplessStateInitEnabled
false, // interruptEvent false, // interruptEvent
false, // immediateScratchAddressPatching false, // immediateScratchAddressPatching
false, // makeCommandView false, // makeCommandView
}; };
EXPECT_THROW(NEO::EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(commandContainer, dispatchKernelArgs), std::exception); EXPECT_THROW(NEO::EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(commandContainer, dispatchKernelArgs), std::exception);
} }

2
opencl/source/command_queue/hardware_interface_bdw_and_later.inl
View File

@@ -87,7 +87,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
kernel.getKernelInfo().kernelDescriptor, // kernelDescriptor kernel.getKernelInfo().kernelDescriptor, // kernelDescriptor
kernel.getExecutionType(), // kernelExecutionType kernel.getExecutionType(), // kernelExecutionType
RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
0, // additionalSizeParam 0, // localRegionSize
0, // maxFrontEndThreads 0, // maxFrontEndThreads
false}; // requiredSystemFence false}; // requiredSystemFence
EncodeDispatchKernel<GfxFamily>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs); EncodeDispatchKernel<GfxFamily>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs);

14
opencl/source/command_queue/hardware_interface_xehp_and_later.inl
View File

@@ -146,12 +146,12 @@ inline void HardwareInterface<GfxFamily>::programWalker(
auto maxFrontEndThreads = device.getDeviceInfo().maxFrontEndThreads; auto maxFrontEndThreads = device.getDeviceInfo().maxFrontEndThreads;
EncodeWalkerArgs encodeWalkerArgs{ EncodeWalkerArgs encodeWalkerArgs{
kernelInfo.kernelDescriptor, // kernelDescriptor kernelInfo.kernelDescriptor, // kernelDescriptor
kernel.getExecutionType(), // kernelExecutionType kernel.getExecutionType(), // kernelExecutionType
kernelAttributes.walkOrder, // requiredDispatchWalkOrder kernelAttributes.dispatchWalkOrder, // requiredDispatchWalkOrder
kernelAttributes.additionalSize, // additionalSizeParam kernelAttributes.localRegionSize, // localRegionSize
maxFrontEndThreads, // maxFrontEndThreads maxFrontEndThreads, // maxFrontEndThreads
requiredSystemFence}; // requiredSystemFence requiredSystemFence}; // requiredSystemFence
EncodeDispatchKernel<GfxFamily>::template encodeAdditionalWalkerFields<WalkerType>(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs); EncodeDispatchKernel<GfxFamily>::template encodeAdditionalWalkerFields<WalkerType>(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs);
EncodeDispatchKernel<GfxFamily>::template encodeWalkerPostSyncFields<WalkerType>(walkerCmd, encodeWalkerArgs); EncodeDispatchKernel<GfxFamily>::template encodeWalkerPostSyncFields<WalkerType>(walkerCmd, encodeWalkerArgs);
@@ -192,7 +192,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
queueCsr.getDcFlushSupport(), // dcFlush queueCsr.getDcFlushSupport(), // dcFlush
kernel.isSingleSubdevicePreferred(), // forceExecutionOnSingleTile kernel.isSingleSubdevicePreferred(), // forceExecutionOnSingleTile
false, // blockDispatchToCommandBuffer false, // blockDispatchToCommandBuffer
requiredWalkOrder != 0}; // isRequiredWorkGroupOrder requiredWalkOrder != 0}; // isRequiredDispatchWorkGroupOrder
ImplicitScalingDispatch<GfxFamily>::template dispatchCommands<WalkerType>(commandStream, ImplicitScalingDispatch<GfxFamily>::template dispatchCommands<WalkerType>(commandStream,
walkerCmd, walkerCmd,

2
opencl/test/unit_test/xe2_hpg_core/dispatch_walker_tests_xe2_hpg_core.cpp
View File

@@ -30,7 +30,7 @@ XE2_HPG_CORETEST_F(WalkerDispatchTestsXe2HpGCore, whenEncodeAdditionalWalkerFiel
kernelDescriptor, // kernelDescriptor kernelDescriptor, // kernelDescriptor
KernelExecutionType::concurrent, // kernelExecutionType KernelExecutionType::concurrent, // kernelExecutionType
NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
0, // additionalSizeParam 0, // localRegionSize
0, // maxFrontEndThreads 0, // maxFrontEndThreads
true}; // requiredSystemFence true}; // requiredSystemFence
{ {

6
shared/source/command_container/command_encoder.h
View File

@@ -63,7 +63,7 @@ struct EncodeDispatchKernelArgs {
PreemptionMode preemptionMode = PreemptionMode::Initial; PreemptionMode preemptionMode = PreemptionMode::Initial;
NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none; NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none;
NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none; NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
uint32_t additionalSizeParam = NEO::additionalKernelLaunchSizeParamNotSet; uint32_t localRegionSize = NEO::localRegionSizeParamNotSet;
uint32_t partitionCount = 0u; uint32_t partitionCount = 0u;
uint32_t reserveExtraPayloadSpace = 0; uint32_t reserveExtraPayloadSpace = 0;
uint32_t maxWgCountPerTile = 0; uint32_t maxWgCountPerTile = 0;
@@ -109,7 +109,7 @@ struct EncodeWalkerArgs {
const KernelDescriptor &kernelDescriptor; const KernelDescriptor &kernelDescriptor;
KernelExecutionType kernelExecutionType = KernelExecutionType::defaultType; KernelExecutionType kernelExecutionType = KernelExecutionType::defaultType;
NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none; NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
uint32_t additionalSizeParam = NEO::additionalKernelLaunchSizeParamNotSet; uint32_t localRegionSize = NEO::localRegionSizeParamNotSet;
uint32_t maxFrontEndThreads = 0; uint32_t maxFrontEndThreads = 0;
bool requiredSystemFence = false; bool requiredSystemFence = false;
}; };
@@ -188,7 +188,7 @@ struct EncodeDispatchKernel {
static void setupPostSyncForRegularEvent(WalkerType &walkerCmd, const EncodeDispatchKernelArgs &args); static void setupPostSyncForRegularEvent(WalkerType &walkerCmd, const EncodeDispatchKernelArgs &args);
template <typename WalkerType> template <typename WalkerType>
static void setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredWalkOrder); static void setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredDispatchWalkOrder);
template <typename WalkerType> template <typename WalkerType>
static void setupPostSyncForInOrderExec(WalkerType &walkerCmd, const EncodeDispatchKernelArgs &args); static void setupPostSyncForInOrderExec(WalkerType &walkerCmd, const EncodeDispatchKernelArgs &args);

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

@@ -283,7 +283,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
kernelDescriptor, // kernelDescriptor kernelDescriptor, // kernelDescriptor
KernelExecutionType::defaultType, // kernelExecutionType KernelExecutionType::defaultType, // kernelExecutionType
args.requiredDispatchWalkOrder, // requiredDispatchWalkOrder args.requiredDispatchWalkOrder, // requiredDispatchWalkOrder
args.additionalSizeParam, // additionalSizeParam args.localRegionSize, // localRegionSize
args.device->getDeviceInfo().maxFrontEndThreads, // maxFrontEndThreads args.device->getDeviceInfo().maxFrontEndThreads, // maxFrontEndThreads
args.requiresSystemMemoryFence()}; // requiredSystemFence args.requiresSystemMemoryFence()}; // requiredSystemFence
EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, cmd, walkerArgs); EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, cmd, walkerArgs);
@@ -684,7 +684,7 @@ void EncodeDispatchKernel<Family>::encodeEuSchedulingPolicy(InterfaceDescriptorT
template <typename Family> template <typename Family>
template <typename WalkerType> template <typename WalkerType>
void EncodeDispatchKernel<Family>::setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredWalkOrder) {} void EncodeDispatchKernel<Family>::setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredDispatchWalkOrder) {}
template <typename Family> template <typename Family>
template <typename WalkerType> template <typename WalkerType>

2
shared/source/command_container/command_encoder_enablers.inl
View File

@@ -30,7 +30,7 @@ template uint64_t NEO::EncodeDispatchKernel<Family>::getScratchAddressForImmedia
template void NEO::EncodeDispatchKernel<Family>::patchScratchAddressInImplicitArgs<false>(ImplicitArgs &implicitArgs, uint64_t scratchAddress, bool scratchPtrPatchingRequired); template void NEO::EncodeDispatchKernel<Family>::patchScratchAddressInImplicitArgs<false>(ImplicitArgs &implicitArgs, uint64_t scratchAddress, bool scratchPtrPatchingRequired);
template void NEO::EncodeDispatchKernel<Family>::forceComputeWalkerPostSyncFlushWithWrite<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd); template void NEO::EncodeDispatchKernel<Family>::forceComputeWalkerPostSyncFlushWithWrite<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd);
template void NEO::EncodeDispatchKernel<Family>::setWalkerRegionSettings<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, template void NEO::EncodeDispatchKernel<Family>::setWalkerRegionSettings<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount,
uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredWalkOrder); uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredDispatchWalkOrder);
template void NEO::EncodeDispatchKernel<Family>::overrideDefaultValues<Family::DefaultWalkerType, Family::DefaultWalkerType::InterfaceDescriptorType>(Family::DefaultWalkerType &walkerCmd, Family::DefaultWalkerType::InterfaceDescriptorType &interfaceDescriptor); template void NEO::EncodeDispatchKernel<Family>::overrideDefaultValues<Family::DefaultWalkerType, Family::DefaultWalkerType::InterfaceDescriptorType>(Family::DefaultWalkerType &walkerCmd, Family::DefaultWalkerType::InterfaceDescriptorType &interfaceDescriptor);
template void NEO::EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs); template void NEO::EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs);
template void NEO::EncodeDispatchKernel<Family>::encodeComputeDispatchAllWalker<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs); template void NEO::EncodeDispatchKernel<Family>::encodeComputeDispatchAllWalker<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs);

2
shared/source/command_container/command_encoder_from_xe_hpg_core_to_xe3_core.inl
View File

@@ -90,6 +90,6 @@ void EncodeDispatchKernel<Family>::adjustTimestampPacket(WalkerType &walkerCmd,
template <typename Family> template <typename Family>
template <typename WalkerType> template <typename WalkerType>
void EncodeDispatchKernel<Family>::setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredWalkOrder) {} void EncodeDispatchKernel<Family>::setWalkerRegionSettings(WalkerType &walkerCmd, const NEO::Device &device, uint32_t partitionCount, uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredDispatchWalkOrder) {}
} // namespace NEO } // namespace NEO

8
shared/source/command_container/command_encoder_xehp_and_later.inl
View File

@@ -402,7 +402,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
kernelDescriptor, // kernelDescriptor kernelDescriptor, // kernelDescriptor
kernelExecutionType, // kernelExecutionType kernelExecutionType, // kernelExecutionType
args.requiredDispatchWalkOrder, // requiredDispatchWalkOrder args.requiredDispatchWalkOrder, // requiredDispatchWalkOrder
args.additionalSizeParam, // additionalSizeParam args.localRegionSize, // localRegionSize
args.device->getDeviceInfo().maxFrontEndThreads, // maxFrontEndThreads args.device->getDeviceInfo().maxFrontEndThreads, // maxFrontEndThreads
args.requiresSystemMemoryFence()}; // requiresMemoryFence args.requiresSystemMemoryFence()}; // requiresMemoryFence
EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, walkerArgs); EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, walkerArgs);
@@ -412,7 +412,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
EncodeDispatchKernel<Family>::overrideDefaultValues(walkerCmd, idd); EncodeDispatchKernel<Family>::overrideDefaultValues(walkerCmd, idd);
uint32_t workgroupSize = args.dispatchInterface->getGroupSize()[0] * args.dispatchInterface->getGroupSize()[1] * args.dispatchInterface->getGroupSize()[2]; uint32_t workgroupSize = args.dispatchInterface->getGroupSize()[0] * args.dispatchInterface->getGroupSize()[1] * args.dispatchInterface->getGroupSize()[2];
bool isRequiredWorkGroupOrder = args.requiredDispatchWalkOrder != NEO::RequiredDispatchWalkOrder::none; bool isRequiredDispatchWorkGroupOrder = args.requiredDispatchWalkOrder != NEO::RequiredDispatchWalkOrder::none;
if (args.partitionCount > 1 && !args.isInternal) { if (args.partitionCount > 1 && !args.isInternal) {
const uint64_t workPartitionAllocationGpuVa = args.device->getDefaultEngine().commandStreamReceiver->getWorkPartitionAllocationGpuAddress(); const uint64_t workPartitionAllocationGpuVa = args.device->getDefaultEngine().commandStreamReceiver->getWorkPartitionAllocationGpuAddress();
@@ -429,7 +429,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
args.dcFlushEnable, // dcFlush args.dcFlushEnable, // dcFlush
EncodeDispatchKernel<Family>::singleTileExecImplicitScalingRequired(args.isCooperative), // forceExecutionOnSingleTile EncodeDispatchKernel<Family>::singleTileExecImplicitScalingRequired(args.isCooperative), // forceExecutionOnSingleTile
args.makeCommandView, // blockDispatchToCommandBuffer args.makeCommandView, // blockDispatchToCommandBuffer
isRequiredWorkGroupOrder}; // isRequiredWorkGroupOrder isRequiredDispatchWorkGroupOrder}; // isRequiredDispatchWorkGroupOrder
ImplicitScalingDispatch<Family>::dispatchCommands(*listCmdBufferStream, ImplicitScalingDispatch<Family>::dispatchCommands(*listCmdBufferStream,
walkerCmd, walkerCmd,
@@ -438,7 +438,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
args.partitionCount = implicitScalingArgs.partitionCount; args.partitionCount = implicitScalingArgs.partitionCount;
} else { } else {
args.partitionCount = 1; args.partitionCount = 1;
EncodeDispatchKernel<Family>::setWalkerRegionSettings(walkerCmd, *args.device, args.partitionCount, workgroupSize, args.maxWgCountPerTile, isRequiredWorkGroupOrder); EncodeDispatchKernel<Family>::setWalkerRegionSettings(walkerCmd, *args.device, args.partitionCount, workgroupSize, args.maxWgCountPerTile, isRequiredDispatchWorkGroupOrder);
if (!args.makeCommandView) { if (!args.makeCommandView) {
auto buffer = listCmdBufferStream->getSpaceForCmd<WalkerType>(); auto buffer = listCmdBufferStream->getSpaceForCmd<WalkerType>();

2
shared/source/command_container/implicit_scaling.h
View File

@@ -53,7 +53,7 @@ struct ImplicitScalingDispatchCommandArgs {
bool dcFlush = false; bool dcFlush = false;
bool forceExecutionOnSingleTile = false; bool forceExecutionOnSingleTile = false;
bool blockDispatchToCommandBuffer = false; bool blockDispatchToCommandBuffer = false;
bool isRequiredWorkGroupOrder = false; bool isRequiredDispatchWorkGroupOrder = false;
}; };
template <typename GfxFamily> template <typename GfxFamily>

2
shared/source/command_container/implicit_scaling_xehp_and_later.inl
View File

@@ -51,7 +51,7 @@ WalkerPartition::WalkerPartitionArgs prepareWalkerPartitionArgs(ImplicitScalingD
args.workgroupSize = dispatchCommandArgs.workgroupSize; args.workgroupSize = dispatchCommandArgs.workgroupSize;
args.maxWgCountPerTile = dispatchCommandArgs.maxWgCountPerTile; args.maxWgCountPerTile = dispatchCommandArgs.maxWgCountPerTile;
args.isRequiredWorkGroupOrder = dispatchCommandArgs.isRequiredWorkGroupOrder; args.isRequiredDispatchWorkGroupOrder = dispatchCommandArgs.isRequiredDispatchWorkGroupOrder;
return args; return args;
} }

2
shared/source/command_container/walker_partition_interface.h
View File

@@ -35,7 +35,7 @@ struct WalkerPartitionArgs {
bool dcFlushEnable = false; bool dcFlushEnable = false;
bool forceExecutionOnSingleTile = false; bool forceExecutionOnSingleTile = false;
bool blockDispatchToCommandBuffer = false; bool blockDispatchToCommandBuffer = false;
bool isRequiredWorkGroupOrder = false; bool isRequiredDispatchWorkGroupOrder = false;
}; };
inline constexpr uint32_t wparidCCSOffset = 0x221C; inline constexpr uint32_t wparidCCSOffset = 0x221C;

2
shared/source/command_container/walker_partition_xehp_and_later.h
View File

@@ -531,7 +531,7 @@ void *programPartitionedWalker(void *&inputAddress, uint32_t &totalBytesProgramm
args.partitionCount, args.partitionCount,
args.workgroupSize, args.workgroupSize,
args.maxWgCountPerTile, args.maxWgCountPerTile,
args.isRequiredWorkGroupOrder); args.isRequiredDispatchWorkGroupOrder);
appendWalkerFields<GfxFamily, WalkerType>(*inputWalker, args.tileCount, workgroupCount); appendWalkerFields<GfxFamily, WalkerType>(*inputWalker, args.tileCount, workgroupCount);
} }

10
shared/source/device_binary_format/zebin/zeinfo.h
View File

@@ -350,7 +350,7 @@ using IndirectStatelessCountT = int32_t;
using HasSampleT = bool; using HasSampleT = bool;
using PrivateSizeT = int32_t; using PrivateSizeT = int32_t;
using SpillSizeT = int32_t; using SpillSizeT = int32_t;
using AdditionalSizeT = int32_t; using LocalRegionSizeT = int32_t;
using WalkOrderT = int32_t; using WalkOrderT = int32_t;
using PartitionDimT = int32_t; using PartitionDimT = int32_t;
@@ -384,8 +384,8 @@ inline constexpr IndirectStatelessCountT indirectStatelessCount = 0;
inline constexpr HasSampleT hasSample = false; inline constexpr HasSampleT hasSample = false;
inline constexpr PrivateSizeT privateSize = 0; inline constexpr PrivateSizeT privateSize = 0;
inline constexpr SpillSizeT spillSize = 0; inline constexpr SpillSizeT spillSize = 0;
inline constexpr AdditionalSizeT additionalSize = -1; inline constexpr LocalRegionSizeT localRegionSize = -1;
inline constexpr WalkOrderT walkOrder = -1; inline constexpr WalkOrderT dispatchWalkOrder = -1;
inline constexpr PartitionDimT partitionDim = -1; inline constexpr PartitionDimT partitionDim = -1;
} // namespace Defaults } // namespace Defaults
@@ -422,8 +422,8 @@ struct ExecutionEnvBaseT {
HasSampleT hasSample = Defaults::hasSample; HasSampleT hasSample = Defaults::hasSample;
PrivateSizeT privateSize = Defaults::privateSize; PrivateSizeT privateSize = Defaults::privateSize;
SpillSizeT spillSize = Defaults::spillSize; SpillSizeT spillSize = Defaults::spillSize;
AdditionalSizeT additionalSize = Defaults::additionalSize; LocalRegionSizeT localRegionSize = Defaults::localRegionSize;
WalkOrderT walkOrder = Defaults::walkOrder; WalkOrderT dispatchWalkOrder = Defaults::dispatchWalkOrder;
PartitionDimT partitionDim = Defaults::partitionDim; PartitionDimT partitionDim = Defaults::partitionDim;
}; };

8
shared/source/device_binary_format/zebin/zeinfo_decoder.cpp
View File

@@ -718,11 +718,11 @@ void populateKernelExecutionEnvironment(KernelDescriptor &dst, const KernelExecu
dst.kernelAttributes.hasIndirectStatelessAccess = (execEnv.indirectStatelessCount > 0); dst.kernelAttributes.hasIndirectStatelessAccess = (execEnv.indirectStatelessCount > 0);
dst.kernelAttributes.numThreadsRequired = static_cast<uint32_t>(execEnv.euThreadCount); dst.kernelAttributes.numThreadsRequired = static_cast<uint32_t>(execEnv.euThreadCount);
if (execEnv.additionalSize != Types::Kernel::ExecutionEnv::Defaults::additionalSize) { if (execEnv.localRegionSize != Types::Kernel::ExecutionEnv::Defaults::localRegionSize) {
dst.kernelAttributes.additionalSize = static_cast<uint32_t>(execEnv.additionalSize); dst.kernelAttributes.localRegionSize = static_cast<uint32_t>(execEnv.localRegionSize);
} }
if (execEnv.walkOrder != Types::Kernel::ExecutionEnv::Defaults::walkOrder) { if (execEnv.dispatchWalkOrder != Types::Kernel::ExecutionEnv::Defaults::dispatchWalkOrder) {
dst.kernelAttributes.walkOrder = EncodeParamsApiMappings::walkOrder[execEnv.walkOrder]; dst.kernelAttributes.dispatchWalkOrder = EncodeParamsApiMappings::dispatchWalkOrder[execEnv.dispatchWalkOrder];
} }
if (execEnv.partitionDim != Types::Kernel::ExecutionEnv::Defaults::partitionDim) { if (execEnv.partitionDim != Types::Kernel::ExecutionEnv::Defaults::partitionDim) {
dst.kernelAttributes.partitionDim = EncodeParamsApiMappings::partitionDim[execEnv.partitionDim]; dst.kernelAttributes.partitionDim = EncodeParamsApiMappings::partitionDim[execEnv.partitionDim];

4
shared/source/helpers/definitions/command_encoder_args.h
View File

@@ -43,11 +43,11 @@ enum class RequiredDispatchWalkOrder : uint32_t {
additional additional
}; };
static constexpr uint32_t additionalKernelLaunchSizeParamNotSet = 0; static constexpr uint32_t localRegionSizeParamNotSet = 0;
namespace EncodeParamsApiMappings { namespace EncodeParamsApiMappings {
static constexpr std::array<NEO::RequiredPartitionDim, 3> partitionDim = {{RequiredPartitionDim::x, NEO::RequiredPartitionDim::y, NEO::RequiredPartitionDim::z}}; static constexpr std::array<NEO::RequiredPartitionDim, 3> partitionDim = {{RequiredPartitionDim::x, NEO::RequiredPartitionDim::y, NEO::RequiredPartitionDim::z}};
static constexpr std::array<NEO::RequiredDispatchWalkOrder, 3> walkOrder = {{NEO::RequiredDispatchWalkOrder::x, NEO::RequiredDispatchWalkOrder::y, NEO::RequiredDispatchWalkOrder::additional}}; static constexpr std::array<NEO::RequiredDispatchWalkOrder, 3> dispatchWalkOrder = {{NEO::RequiredDispatchWalkOrder::x, NEO::RequiredDispatchWalkOrder::y, NEO::RequiredDispatchWalkOrder::additional}};
} // namespace EncodeParamsApiMappings } // namespace EncodeParamsApiMappings
} // namespace NEO } // namespace NEO

4
shared/source/kernel/kernel_descriptor.h
View File

@@ -57,8 +57,8 @@ struct KernelDescriptor {
uint32_t numThreadsRequired = 0u; uint32_t numThreadsRequired = 0u;
uint32_t spillFillScratchMemorySize = 0u; uint32_t spillFillScratchMemorySize = 0u;
uint32_t privateScratchMemorySize = 0u; uint32_t privateScratchMemorySize = 0u;
uint32_t additionalSize = NEO::additionalKernelLaunchSizeParamNotSet; uint32_t localRegionSize = NEO::localRegionSizeParamNotSet;
NEO::RequiredDispatchWalkOrder walkOrder = NEO::RequiredDispatchWalkOrder::none; NEO::RequiredDispatchWalkOrder dispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
NEO::RequiredPartitionDim partitionDim = NEO::RequiredPartitionDim::none; NEO::RequiredPartitionDim partitionDim = NEO::RequiredPartitionDim::none;
ThreadArbitrationPolicy threadArbitrationPolicy = NotPresent; ThreadArbitrationPolicy threadArbitrationPolicy = NotPresent;
uint16_t requiredWorkgroupSize[3] = {0U, 0U, 0U}; uint16_t requiredWorkgroupSize[3] = {0U, 0U, 0U};

78
shared/test/unit_test/fixtures/command_container_fixture.cpp
View File

@@ -39,44 +39,44 @@ EncodeDispatchKernelArgs CommandEncodeStatesFixture::createDefaultDispatchKernel
bool requiresUncachedMocs) { bool requiresUncachedMocs) {
EncodeDispatchKernelArgs args{ EncodeDispatchKernelArgs args{
0, // eventAddress 0, // eventAddress
0, // postSyncImmValue 0, // postSyncImmValue
0, // inOrderCounterValue 0, // inOrderCounterValue
device, // device device, // device
nullptr, // inOrderExecInfo nullptr, // inOrderExecInfo
dispatchInterface, // dispatchInterface dispatchInterface, // dispatchInterface
nullptr, // surfaceStateHeap nullptr, // surfaceStateHeap
nullptr, // dynamicStateHeap nullptr, // dynamicStateHeap
threadGroupDimensions, // threadGroupDimensions threadGroupDimensions, // threadGroupDimensions
nullptr, // outWalkerPtr nullptr, // outWalkerPtr
nullptr, // cpuWalkerBuffer nullptr, // cpuWalkerBuffer
nullptr, // cpuPayloadBuffer nullptr, // cpuPayloadBuffer
nullptr, // outImplicitArgsPtr nullptr, // outImplicitArgsPtr
nullptr, // additionalCommands nullptr, // additionalCommands
PreemptionMode::Disabled, // preemptionMode PreemptionMode::Disabled, // preemptionMode
NEO::RequiredPartitionDim::none, // requiredPartitionDim NEO::RequiredPartitionDim::none, // requiredPartitionDim
NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
NEO::additionalKernelLaunchSizeParamNotSet, // additionalSizeParam NEO::localRegionSizeParamNotSet, // localRegionSize
1, // partitionCount 1, // partitionCount
0, // reserveExtraPayloadSpace 0, // reserveExtraPayloadSpace
1, // maxWgCountPerTile 1, // maxWgCountPerTile
NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy NEO::ThreadArbitrationPolicy::NotPresent, // defaultPipelinedThreadArbitrationPolicy
false, // isIndirect false, // isIndirect
false, // isPredicate false, // isPredicate
false, // isTimestampEvent false, // isTimestampEvent
requiresUncachedMocs, // requiresUncachedMocs requiresUncachedMocs, // requiresUncachedMocs
false, // isInternal false, // isInternal
false, // isCooperative false, // isCooperative
false, // isHostScopeSignalEvent false, // isHostScopeSignalEvent
false, // isKernelUsingSystemAllocation false, // isKernelUsingSystemAllocation
false, // isKernelDispatchedFromImmediateCmdList false, // isKernelDispatchedFromImmediateCmdList
false, // isRcs false, // isRcs
false, // dcFlushEnable false, // dcFlushEnable
false, // isHeaplessModeEnabled false, // isHeaplessModeEnabled
false, // isHeaplessStateInitEnabled false, // isHeaplessStateInitEnabled
false, // interruptEvent false, // interruptEvent
false, // immediateScratchAddressPatching false, // immediateScratchAddressPatching
false, // makeCommandView false, // makeCommandView
}; };
return args; return args;
@@ -87,7 +87,7 @@ EncodeWalkerArgs CommandEncodeStatesFixture::createDefaultEncodeWalkerArgs(const
kernelDescriptor, // kernelDescriptor kernelDescriptor, // kernelDescriptor
NEO::KernelExecutionType::defaultType, // kernelExecutionType NEO::KernelExecutionType::defaultType, // kernelExecutionType
NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder NEO::RequiredDispatchWalkOrder::none, // requiredDispatchWalkOrder
0, // additionalSizeParam 0, // localRegionSize
0, // maxFrontEndThreads 0, // maxFrontEndThreads
false}; // requiredSystemFence false}; // requiredSystemFence

2
shared/test/unit_test/fixtures/implicit_scaling_fixture.cpp
View File

@@ -54,7 +54,7 @@ ImplicitScalingDispatchCommandArgs ImplicitScalingFixture::createDispatchCommand
dcFlushFlag, // dcFlush dcFlushFlag, // dcFlush
forceExecutionOnSingleTileFlag, // forceExecutionOnSingleTile forceExecutionOnSingleTileFlag, // forceExecutionOnSingleTile
false, // blockDispatchToCommandBuffer false, // blockDispatchToCommandBuffer
false}; // isRequiredWorkGroupOrder false}; // isRequiredDispatchWorkGroupOrder
return args; return args;
} }