intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-10 15:12:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

Apply heuristics when setting TG dispatch size on XE_HPC_CORE

Browse Source 
The default TG dispatch size can be changed
to a better value based on number of threads in TG or
currently available amount of threads on GPU.
Decision on what TG dispatch size should be are based on
implemented heuristics.

Signed-off-by: Rafal Maziejuk <rafal.maziejuk@intel.com>
Related-To: NEO-6989
This commit is contained in:
Rafal Maziejuk
2022-08-03 12:22:30 +00:00
committed by Compute-Runtime-Automation
parent 52133e61ce
commit ed0c36117e
21 changed files with 188 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -65,6 +65,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
size_t globalOffsets[3] = {dispatchInfo.getOffset().x, dispatchInfo.getOffset().y, dispatchInfo.getOffset().z};
size_t startWorkGroups[3] = {walkerArgs.startOfWorkgroups->x, walkerArgs.startOfWorkgroups->y, walkerArgs.startOfWorkgroups->z};
size_t numWorkGroups[3] = {walkerArgs.numberOfWorkgroups->x, walkerArgs.numberOfWorkgroups->y, walkerArgs.numberOfWorkgroups->z};
auto threadGroupCount = static_cast<uint32_t>(walkerArgs.numberOfWorkgroups->x * walkerArgs.numberOfWorkgroups->y * walkerArgs.numberOfWorkgroups->z);
if (walkerArgs.currentTimestampPacketNodes && commandQueue.getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
auto timestampPacketNode = walkerArgs.currentTimestampPacketNodes->peekNodes().at(walkerArgs.currentDispatchIndex);
@@ -83,6 +84,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
kernel.getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
simd,
walkerArgs.localWorkSizes,
threadGroupCount,
walkerArgs.offsetInterfaceDescriptorTable,
walkerArgs.interfaceDescriptorIndex,
walkerArgs.preemptionMode,

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

@@ -60,6 +60,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
size_t globalOffsets[3] = {dispatchInfo.getOffset().x, dispatchInfo.getOffset().y, dispatchInfo.getOffset().z};
size_t startWorkGroups[3] = {walkerArgs.startOfWorkgroups->x, walkerArgs.startOfWorkgroups->y, walkerArgs.startOfWorkgroups->z};
size_t numWorkGroups[3] = {walkerArgs.numberOfWorkgroups->x, walkerArgs.numberOfWorkgroups->y, walkerArgs.numberOfWorkgroups->z};
auto threadGroupCount = static_cast<uint32_t>(walkerArgs.numberOfWorkgroups->x * walkerArgs.numberOfWorkgroups->y * walkerArgs.numberOfWorkgroups->z);
uint32_t requiredWalkOrder = 0u;
bool localIdsGenerationByRuntime = EncodeDispatchKernel<GfxFamily>::isRuntimeLocalIdsGenerationRequired(
@@ -98,6 +99,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
kernel.getKernelStartAddress(localIdsGenerationByRuntime, kernelUsesLocalIds, isCcsUsed, false),
simd,
walkerArgs.localWorkSizes,
threadGroupCount,
walkerArgs.offsetInterfaceDescriptorTable,
walkerArgs.interfaceDescriptorIndex,
walkerArgs.preemptionMode,

2
opencl/source/helpers/hardware_commands_helper.h
View File

@@ -47,6 +47,7 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
size_t bindingTablePointer,
[[maybe_unused]] size_t offsetSamplerState,
uint32_t numSamplers,
const uint32_t threadGroupCount,
uint32_t numThreadsPerThreadGroup,
const Kernel &kernel,
uint32_t bindingTablePrefetchSize,
@@ -79,6 +80,7 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
uint64_t kernelStartOffset,
uint32_t simd,
const size_t localWorkSize[3],
const uint32_t threadGroupCount,
const uint64_t offsetInterfaceDescriptorTable,
uint32_t &interfaceDescriptorIndex,
PreemptionMode preemptionMode,

6
opencl/source/helpers/hardware_commands_helper_base.inl
View File

@@ -113,6 +113,7 @@ size_t HardwareCommandsHelper<GfxFamily>::sendInterfaceDescriptorData(
size_t bindingTablePointer,
[[maybe_unused]] size_t offsetSamplerState,
uint32_t numSamplers,
const uint32_t threadGroupCount,
uint32_t threadsPerThreadGroup,
const Kernel &kernel,
uint32_t bindingTablePrefetchSize,
@@ -169,7 +170,8 @@ size_t HardwareCommandsHelper<GfxFamily>::sendInterfaceDescriptorData(
hardwareInfo);
PreemptionHelper::programInterfaceDescriptorDataPreemption<GfxFamily>(&interfaceDescriptor, preemptionMode);
EncodeDispatchKernel<GfxFamily>::adjustInterfaceDescriptorData(interfaceDescriptor, hardwareInfo);
EncodeDispatchKernel<GfxFamily>::adjustInterfaceDescriptorData(interfaceDescriptor, hardwareInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
*pInterfaceDescriptor = interfaceDescriptor;
return (size_t)offsetInterfaceDescriptor;
@@ -185,6 +187,7 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
uint64_t kernelStartOffset,
uint32_t simd,
const size_t localWorkSize[3],
const uint32_t threadGroupCount,
const uint64_t offsetInterfaceDescriptorTable,
uint32_t &interfaceDescriptorIndex,
PreemptionMode preemptionMode,
@@ -263,6 +266,7 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
dstBindingTablePointer,
samplerStateOffset,
samplerCount,
threadGroupCount,
threadsPerThreadGroup,
kernel,
bindingTablePrefetchSize,

20
opencl/test/unit_test/helpers/hardware_commands_helper_tests.cpp
View File

@@ -88,9 +88,10 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenProgramInterfaceDescriptor
auto usedIndirectHeapBefore = indirectHeap.getUsed();
indirectHeap.getSpace(sizeof(INTERFACE_DESCRIPTOR_DATA));
const uint32_t threadGroupCount = 1u;
size_t crossThreadDataSize = kernel->getCrossThreadDataSize();
HardwareCommandsHelper<FamilyType>::sendInterfaceDescriptorData(
indirectHeap, 0, 0, crossThreadDataSize, 64, 0, 0, 0, 1, *kernel, 0, pDevice->getPreemptionMode(), nullptr, *pDevice);
indirectHeap, 0, 0, crossThreadDataSize, 64, 0, 0, 0, threadGroupCount, 1, *kernel, 0, pDevice->getPreemptionMode(), nullptr, *pDevice);
auto usedIndirectHeapAfter = indirectHeap.getUsed();
EXPECT_EQ(sizeof(INTERFACE_DESCRIPTOR_DATA), usedIndirectHeapAfter - usedIndirectHeapBefore);
@@ -309,6 +310,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenAllocatingIndirectStateRes
const size_t localWorkSize = 256;
const size_t localWorkSizes[3]{localWorkSize, 1, 1};
const uint32_t threadGroupCount = 1u;
auto &commandStream = cmdQ.getCS(1024);
auto pWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
@@ -343,6 +345,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenAllocatingIndirectStateRes
kernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
kernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
idToffset,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -385,6 +388,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWithFourBindingTabl
auto &ssh = cmdQ.getIndirectHeap(IndirectHeap::Type::SURFACE_STATE, 8192);
const size_t localWorkSize = 256;
const size_t localWorkSizes[3]{localWorkSize, 1, 1};
const uint32_t threadGroupCount = 1u;
uint32_t interfaceDescriptorIndex = 0;
auto isCcsUsed = EngineHelpers::isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
auto kernelUsesLocalIds = HardwareCommandsHelper<FamilyType>::kernelUsesLocalIds(*mockKernelWithInternal->mockKernel);
@@ -398,6 +402,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWithFourBindingTabl
mockKernelWithInternal->mockKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
mockKernelWithInternal->mockKernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -431,6 +436,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWith100BindingTable
auto &ssh = cmdQ.getIndirectHeap(IndirectHeap::Type::SURFACE_STATE, 8192);
const size_t localWorkSize = 256;
const size_t localWorkSizes[3]{localWorkSize, 1, 1};
const uint32_t threadGroupCount = 1u;
uint32_t interfaceDescriptorIndex = 0;
auto isCcsUsed = EngineHelpers::isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
auto kernelUsesLocalIds = HardwareCommandsHelper<FamilyType>::kernelUsesLocalIds(*mockKernelWithInternal->mockKernel);
@@ -444,6 +450,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWith100BindingTable
mockKernelWithInternal->mockKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
mockKernelWithInternal->mockKernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -487,6 +494,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, whenSendingIndirectStateThenKe
const size_t localWorkSizeY = 3;
const size_t localWorkSizeZ = 4;
const size_t localWorkSizes[3]{localWorkSizeX, localWorkSizeY, localWorkSizeZ};
const uint32_t threadGroupCount = 1u;
auto &commandStream = cmdQ.getCS(1024);
auto pWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
@@ -523,6 +531,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, whenSendingIndirectStateThenKe
mockKernel.getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
modifiedKernelInfo.getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
idToffset,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -578,6 +587,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenSendingIndirectStateThenBi
ASSERT_NE(nullptr, kernel);
const size_t localWorkSizes[3]{256, 1, 1};
const uint32_t threadGroupCount = 1u;
auto &commandStream = cmdQ.getCS(1024);
auto pWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
@@ -613,6 +623,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenSendingIndirectStateThenBi
kernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
kernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -701,6 +712,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenGettingBindingTableStateTh
EXPECT_EQ(numSurfaces, pKernel->getNumberOfBindingTableStates());
const size_t localWorkSizes[3]{256, 1, 1};
const uint32_t threadGroupCount = 1u;
dsh.getSpace(sizeof(INTERFACE_DESCRIPTOR_DATA));
@@ -722,6 +734,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenGettingBindingTableStateTh
pKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
pKernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -847,6 +860,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithInvalidSamplerS
auto &ssh = cmdQ.getIndirectHeap(IndirectHeap::Type::SURFACE_STATE, 8192);
const size_t localWorkSize = 256;
const size_t localWorkSizes[3]{localWorkSize, 1, 1};
const uint32_t threadGroupCount = 1u;
uint32_t interfaceDescriptorIndex = 0;
auto isCcsUsed = EngineHelpers::isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
auto kernelUsesLocalIds = HardwareCommandsHelper<FamilyType>::kernelUsesLocalIds(*mockKernelWithInternal->mockKernel);
@@ -863,6 +877,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithInvalidSamplerS
mockKernelWithInternal->mockKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
mockKernelWithInternal->mockKernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -887,6 +902,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithInvalidSamplerS
mockKernelWithInternal->mockKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
mockKernelWithInternal->mockKernel->getKernelInfo().getMaxSimdSize(),
localWorkSizes,
threadGroupCount,
0,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),
@@ -909,6 +925,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithSamplersWhenInd
CommandQueueHw<FamilyType> cmdQ(nullptr, pClDevice, 0, false);
const size_t localWorkSizes[3]{1, 1, 1};
const uint32_t threadGroupCount = 1u;
auto &commandStream = cmdQ.getCS(1024);
auto pWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
@@ -957,6 +974,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithSamplersWhenInd
mockKernelWithInternal->mockKernel->getKernelStartAddress(true, kernelUsesLocalIds, isCcsUsed, false),
8,
localWorkSizes,
threadGroupCount,
interfaceDescriptorTableOffset,
interfaceDescriptorIndex,
pDevice->getPreemptionMode(),

4
opencl/test/unit_test/kernel/kernel_slm_tests.cpp
View File

@@ -67,6 +67,7 @@ HWCMDTEST_P(IGFX_GEN8_CORE, KernelSLMAndBarrierTest, GivenStaticSlmSizeWhenProgr
// After creating Mock Kernel now create Indirect Heap
auto &indirectHeap = cmdQ.getIndirectHeap(IndirectHeap::Type::DYNAMIC_STATE, 8192);
const uint32_t threadGroupCount = 1u;
uint64_t interfaceDescriptorOffset = indirectHeap.getUsed();
size_t offsetInterfaceDescriptorData = HardwareCommandsHelper<FamilyType>::sendInterfaceDescriptorData(
@@ -78,6 +79,7 @@ HWCMDTEST_P(IGFX_GEN8_CORE, KernelSLMAndBarrierTest, GivenStaticSlmSizeWhenProgr
0,
0,
0,
threadGroupCount,
1,
kernel,
4u,
@@ -154,6 +156,7 @@ HWTEST_F(KernelSLMAndBarrierTest, GivenInterfaceDescriptorProgrammedWhenOverride
CommandQueueHw<FamilyType> cmdQ(nullptr, pClDevice, 0, false);
auto &indirectHeap = cmdQ.getIndirectHeap(IndirectHeap::Type::DYNAMIC_STATE, 8192);
const uint32_t threadGroupCount = 1u;
uint64_t interfaceDescriptorOffset = indirectHeap.getUsed();
INTERFACE_DESCRIPTOR_DATA interfaceDescriptorData;
@@ -166,6 +169,7 @@ HWTEST_F(KernelSLMAndBarrierTest, GivenInterfaceDescriptorProgrammedWhenOverride
0,
0,
0,
threadGroupCount,
1,
kernel,
4u,