intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-21 09:14:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: Unify logic calculating threads per work group part 3

Browse Source 
Related-To: NEO-8087
Signed-off-by: Cencelewska, Katarzyna <katarzyna.cencelewska@intel.com>
This commit is contained in:
Cencelewska, Katarzyna
2023-06-30 14:17:13 +00:00
committed by Compute-Runtime-Automation
parent d2f1cf98d7
commit 61f701aba5
17 changed files with 93 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
level_zero/core/source/kernel/kernel_imp.cpp
View File

@@ -345,7 +345,7 @@ ze_result_t KernelImp::setGroupSize(uint32_t groupSizeX, uint32_t groupSizeY,
auto grfSize = this->module->getDevice()->getHwInfo().capabilityTable.grfSize;
uint32_t perThreadDataSizeForWholeThreadGroupNeeded =
static_cast<uint32_t>(NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(
simdSize, grfSize, numChannels, itemsInGroup));
simdSize, grfSize, numChannels, itemsInGroup, !kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper));
if (perThreadDataSizeForWholeThreadGroupNeeded >
perThreadDataSizeForWholeThreadGroupAllocated) {
alignedFree(perThreadDataForWholeThreadGroup);

7
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_2.cpp
View File

@@ -991,7 +991,8 @@ struct CmdlistAppendLaunchKernelWithImplicitArgsTests : CmdlistAppendLaunchKerne
result = commandList->appendLaunchKernel(kernel->toHandle(), &groupCount, nullptr, 0, nullptr, launchParams, false);
EXPECT_EQ(ZE_RESULT_SUCCESS, result);
implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&expectedImplicitArgs, *kernelDescriptor);
const auto &gfxCoreHelper = device->getGfxCoreHelper();
implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&expectedImplicitArgs, *kernelDescriptor, !kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper);
auto sizeCrossThreadData = kernel->getCrossThreadDataSize();
auto sizePerThreadDataForWholeGroup = kernel->getPerThreadDataSizeForWholeThreadGroup();
EXPECT_EQ(indirectHeap->getUsed(), sizeCrossThreadData + sizePerThreadDataForWholeGroup + implicitArgsProgrammingSize);
@@ -1027,7 +1028,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CmdlistAppendLaunchKernelWithImplicitArgsTests, giv
generateLocalIDs(expectedLocalIds, expectedImplicitArgs.simdWidth, localSize, workgroupDimOrder, false, grfSize, gfxCoreHelper);
auto localIdsProgrammingSize = implicitArgsProgrammingSize - sizeof(ImplicitArgs);
size_t sizeForLocalIds = NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize);
size_t sizeForLocalIds = NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize, !kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper);
EXPECT_EQ(0, memcmp(expectedLocalIds, indirectHeapAllocation->getUnderlyingBuffer(), sizeForLocalIds));
alignedFree(expectedLocalIds);
@@ -1073,7 +1074,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CmdlistAppendLaunchKernelWithImplicitArgsTests, giv
generateLocalIDs(expectedLocalIds, expectedImplicitArgs.simdWidth, localSize, expectedDimOrder, false, grfSize, gfxCoreHelper);
auto localIdsProgrammingSize = implicitArgsProgrammingSize - sizeof(ImplicitArgs);
size_t sizeForLocalIds = NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize);
size_t sizeForLocalIds = NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize, !kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper);
EXPECT_EQ(0, memcmp(expectedLocalIds, indirectHeapAllocation->getUnderlyingBuffer(), sizeForLocalIds));
alignedFree(expectedLocalIds);

3
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_3.cpp
View File

@@ -2062,7 +2062,8 @@ struct CommandListAppendLaunchKernelWithImplicitArgs : CommandListAppendLaunchKe
template <typename FamilyType>
uint64_t getIndirectHeapOffsetForImplicitArgsBuffer(const Mock<::L0::KernelImp> &kernel) {
if (FamilyType::supportsCmdSet(IGFX_XE_HP_CORE)) {
auto implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(kernel.pImplicitArgs.get(), kernel.getKernelDescriptor());
const auto &gfxCoreHelper = device->getGfxCoreHelper();
auto implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(kernel.pImplicitArgs.get(), kernel.getKernelDescriptor(), !kernel.kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper);
return implicitArgsProgrammingSize - sizeof(ImplicitArgs);
} else {
return 0u;

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

@@ -104,7 +104,7 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
const Kernel &kernel);
static size_t getSizeRequiredIOH(
const Kernel &kernel,
size_t localWorkSize = 256);
const size_t localWorkSizes[3]);
static size_t getSizeRequiredSSH(
const Kernel &kernel);

21
opencl/source/helpers/hardware_commands_helper_base.inl
View File

@@ -45,20 +45,33 @@ size_t HardwareCommandsHelper<GfxFamily>::getSizeRequiredDSH(const Kernel &kerne
template <typename GfxFamily>
size_t HardwareCommandsHelper<GfxFamily>::getSizeRequiredIOH(const Kernel &kernel,
size_t localWorkSize) {
const size_t localWorkSizes[3]) {
auto localWorkSize = Math::computeTotalElementsCount(localWorkSizes);
typedef typename GfxFamily::WALKER_TYPE WALKER_TYPE;
const auto &kernelDescriptor = kernel.getDescriptor();
const auto &hwInfo = kernel.getHardwareInfo();
const auto &gfxCoreHelper = kernel.getGfxCoreHelper();
auto numChannels = kernelDescriptor.kernelAttributes.numLocalIdChannels;
uint32_t grfSize = hwInfo.capabilityTable.grfSize;
auto simdSize = kernelDescriptor.kernelAttributes.simdSize;
uint32_t requiredWalkOrder = 0u;
auto isHwLocalIdGeneration = !NEO::EncodeDispatchKernel<GfxFamily>::isRuntimeLocalIdsGenerationRequired(
numChannels,
localWorkSizes,
std::array<uint8_t, 3>{
{kernelDescriptor.kernelAttributes.workgroupWalkOrder[0],
kernelDescriptor.kernelAttributes.workgroupWalkOrder[1],
kernelDescriptor.kernelAttributes.workgroupWalkOrder[2]}},
kernelDescriptor.kernelAttributes.flags.requiresWorkgroupWalkOrder,
requiredWalkOrder,
simdSize);
auto size = kernel.getCrossThreadDataSize() +
getPerThreadDataSizeTotal(simdSize, grfSize, numChannels, localWorkSize);
getPerThreadDataSizeTotal(simdSize, grfSize, numChannels, localWorkSize, isHwLocalIdGeneration, gfxCoreHelper);
auto pImplicitArgs = kernel.getImplicitArgs();
if (pImplicitArgs) {
size += ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor);
size += ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor, isHwLocalIdGeneration, gfxCoreHelper);
}
return alignUp(size, WALKER_TYPE::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
}
@@ -94,7 +107,7 @@ size_t HardwareCommandsHelper<GfxFamily>::getTotalSizeRequiredIOH(
const MultiDispatchInfo &multiDispatchInfo) {
return getSizeRequired(multiDispatchInfo, [](const DispatchInfo &dispatchInfo) { return getSizeRequiredIOH(
*dispatchInfo.getKernel(),
Math::computeTotalElementsCount(dispatchInfo.getLocalWorkgroupSize())); });
dispatchInfo.getLocalWorkgroupSize().values); });
}
template <typename GfxFamily>

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

@@ -87,11 +87,13 @@ size_t HardwareCommandsHelper<GfxFamily>::sendCrossThreadData(
auto pImplicitArgs = kernel.getImplicitArgs();
if (pImplicitArgs) {
const auto &kernelDescriptor = kernel.getDescriptor();
auto sizeForImplicitArgsProgramming = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor);
const auto &gfxCoreHelper = kernel.getGfxCoreHelper();
auto isHwLocalIdGeneration = false;
auto sizeForImplicitArgsProgramming = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor, isHwLocalIdGeneration, gfxCoreHelper);
auto implicitArgsGpuVA = indirectHeap.getGraphicsAllocation()->getGpuAddress() + indirectHeap.getUsed();
auto ptrToPatchImplicitArgs = indirectHeap.getSpace(sizeForImplicitArgsProgramming);
const auto &gfxCoreHelper = kernel.getGfxCoreHelper();
ImplicitArgsHelper::patchImplicitArgs(ptrToPatchImplicitArgs, *pImplicitArgs, kernelDescriptor, {}, gfxCoreHelper);
auto implicitArgsCrossThreadPtr = ptrOffset(reinterpret_cast<uint64_t *>(kernel.getCrossThreadData()), kernelDescriptor.payloadMappings.implicitArgs.implicitArgsBuffer);

14
opencl/source/helpers/hardware_commands_helper_xehp_and_later.inl
View File

@@ -76,12 +76,6 @@ size_t HardwareCommandsHelper<GfxFamily>::sendCrossThreadData(
pImplicitArgs->localIdTablePtr = indirectHeap.getGraphicsAllocation()->getGpuAddress() + offsetCrossThreadData;
const auto &kernelDescriptor = kernel.getDescriptor();
auto sizeForImplicitArgsProgramming = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor);
auto sizeForLocalIdsProgramming = sizeForImplicitArgsProgramming - sizeof(ImplicitArgs);
offsetCrossThreadData += sizeForLocalIdsProgramming;
auto ptrToPatchImplicitArgs = indirectHeap.getSpace(sizeForImplicitArgsProgramming);
const auto &kernelAttributes = kernelDescriptor.kernelAttributes;
uint32_t requiredWalkOrder = 0u;
@@ -96,7 +90,15 @@ size_t HardwareCommandsHelper<GfxFamily>::sendCrossThreadData(
kernelAttributes.flags.requiresWorkgroupWalkOrder,
requiredWalkOrder,
kernelDescriptor.kernelAttributes.simdSize);
const auto &gfxCoreHelper = kernel.getGfxCoreHelper();
auto sizeForImplicitArgsProgramming = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor, !generationOfLocalIdsByRuntime, gfxCoreHelper);
auto sizeForLocalIdsProgramming = sizeForImplicitArgsProgramming - sizeof(ImplicitArgs);
offsetCrossThreadData += sizeForLocalIdsProgramming;
auto ptrToPatchImplicitArgs = indirectHeap.getSpace(sizeForImplicitArgsProgramming);
ImplicitArgsHelper::patchImplicitArgs(ptrToPatchImplicitArgs, *pImplicitArgs, kernelDescriptor, std::make_pair(generationOfLocalIdsByRuntime, requiredWalkOrder), gfxCoreHelper);
}

15
opencl/test/unit_test/command_queue/dispatch_walker_tests.cpp
View File

@@ -741,10 +741,8 @@ HWTEST_F(DispatchWalkerTest, GivenBlockedQueueWhenDispatchingWalkerThenRequiredH
CsrDependencies(),
walkerArgs);
Vec3<size_t> localWorkgroupSize(workGroupSize);
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernel, Math::computeTotalElementsCount(localWorkgroupSize));
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernel, workGroupSize);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(kernel);
EXPECT_LE(expectedSizeDSH, blockedCommandsData->dsh->getMaxAvailableSpace());
@@ -1433,7 +1431,7 @@ HWTEST_F(DispatchWalkerTest, WhenKernelRequiresImplicitArgsThenIohRequiresMoreSp
size_t workItems[3] = {1, 1, 1};
size_t workGroupSize[3] = {2, 5, 10};
cl_uint dimensions = 1;
Vec3<size_t> localWorkgroupSize(workGroupSize);
auto blockedCommandsData = createBlockedCommandsData(*pCmdQ);
kernelInfo.kernelDescriptor.kernelAttributes.simdSize = 1u;
@@ -1458,7 +1456,7 @@ HWTEST_F(DispatchWalkerTest, WhenKernelRequiresImplicitArgsThenIohRequiresMoreSp
CsrDependencies(),
walkerArgsWithoutImplicitArgs);
auto iohSizeWithoutImplicitArgs = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernelWithoutImplicitArgs, Math::computeTotalElementsCount(localWorkgroupSize));
auto iohSizeWithoutImplicitArgs = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernelWithoutImplicitArgs, workGroupSize);
DispatchInfo dispatchInfoWithImplicitArgs(pClDevice, const_cast<MockKernel *>(&kernelWithImplicitArgs), dimensions, workItems, workGroupSize, globalOffsets);
dispatchInfoWithImplicitArgs.setNumberOfWorkgroups({1, 1, 1});
@@ -1473,7 +1471,7 @@ HWTEST_F(DispatchWalkerTest, WhenKernelRequiresImplicitArgsThenIohRequiresMoreSp
CsrDependencies(),
walkerArgsWithImplicitArgs);
auto iohSizeWithImplicitArgs = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernelWithImplicitArgs, Math::computeTotalElementsCount(localWorkgroupSize));
auto iohSizeWithImplicitArgs = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(kernelWithImplicitArgs, workGroupSize);
EXPECT_LE(iohSizeWithoutImplicitArgs, iohSizeWithImplicitArgs);
@@ -1481,9 +1479,10 @@ HWTEST_F(DispatchWalkerTest, WhenKernelRequiresImplicitArgsThenIohRequiresMoreSp
auto numChannels = kernelInfo.kernelDescriptor.kernelAttributes.numLocalIdChannels;
auto simdSize = kernelInfo.getMaxSimdSize();
uint32_t grfSize = sizeof(typename FamilyType::GRF);
const auto &gfxCoreHelper = getHelper<GfxCoreHelper>();
auto size = kernelWithImplicitArgs.getCrossThreadDataSize() +
HardwareCommandsHelper<FamilyType>::getPerThreadDataSizeTotal(simdSize, grfSize, numChannels, Math::computeTotalElementsCount(localWorkgroupSize)) +
ImplicitArgsHelper::getSizeForImplicitArgsPatching(kernelWithImplicitArgs.getImplicitArgs(), kernelWithImplicitArgs.getDescriptor());
HardwareCommandsHelper<FamilyType>::getPerThreadDataSizeTotal(simdSize, grfSize, numChannels, Math::computeTotalElementsCount(workGroupSize), false, gfxCoreHelper) +
ImplicitArgsHelper::getSizeForImplicitArgsPatching(kernelWithImplicitArgs.getImplicitArgs(), kernelWithImplicitArgs.getDescriptor(), false, gfxCoreHelper);
size = alignUp(size, MemoryConstants::cacheLineSize);
EXPECT_EQ(size, iohSizeWithImplicitArgs);

6
opencl/test/unit_test/command_queue/get_size_required_buffer_tests.cpp
View File

@@ -499,7 +499,8 @@ HWTEST_F(GetSizeRequiredBufferTest, GivenHelloWorldKernelWhenEnqueingKernelThenH
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *pCmdQ, KernelFixture::pKernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*KernelFixture::pKernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*KernelFixture::pKernel, workSize[0]);
size_t localWorkSizes[] = {64, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*KernelFixture::pKernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*KernelFixture::pKernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -538,7 +539,8 @@ HWTEST_F(GetSizeRequiredBufferTest, GivenKernelWithSimpleArgWhenEnqueingKernelTh
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *pCmdQ, KernelFixture::pKernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*KernelFixture::pKernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*KernelFixture::pKernel, workSize[0]);
size_t localWorkSizes[] = {64, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*KernelFixture::pKernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*KernelFixture::pKernel);
EXPECT_EQ(0u, expectedSizeIOH % GPGPU_WALKER::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);

18
opencl/test/unit_test/command_queue/get_size_required_image_tests.cpp
View File

@@ -96,7 +96,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenCopyingImageThenHeapsAndCommandBufferCons
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_COPY_IMAGE, false, false, *pCmdQ, kernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel);
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -143,7 +144,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenCopyingReadWriteImageThenHeapsAndCommandB
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_COPY_IMAGE, false, false, *pCmdQ, kernel.get(), {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel.get());
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel.get());
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel.get(), localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel.get());
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -200,7 +202,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenReadingImageNonBlockingThenHeapsAndComman
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_READ_IMAGE, false, false, *pCmdQ, kernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel);
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -255,7 +258,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenReadingImageBlockingThenHeapsAndCommandBu
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_READ_IMAGE, false, false, *pCmdQ, kernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel);
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -310,7 +314,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenWritingImageNonBlockingThenHeapsAndComman
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_WRITE_IMAGE, false, false, *pCmdQ, kernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel);
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.
@@ -365,7 +370,8 @@ HWTEST_F(GetSizeRequiredImageTest, WhenWritingImageBlockingThenHeapsAndCommandBu
auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_WRITE_IMAGE, false, false, *pCmdQ, kernel, {});
auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel);
size_t localWorkSizes[] = {256, 1, 1};
auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
// Since each enqueue* may flush, we may see a MI_BATCH_BUFFER_END appended.

14
opencl/test/unit_test/helpers/hardware_commands_helper_tests.cpp
View File

@@ -365,7 +365,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, WhenAllocatingIndirectStateRes
auto usedAfterIOH = ioh.getUsed();
auto usedAfterSSH = ssh.getUsed();
auto sizeRequiredDSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredDSH(*kernel);
auto sizeRequiredIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSize);
auto sizeRequiredIOH = HardwareCommandsHelper<FamilyType>::getSizeRequiredIOH(*kernel, localWorkSizes);
auto sizeRequiredSSH = HardwareCommandsHelper<FamilyType>::getSizeRequiredSSH(*kernel);
EXPECT_GE(sizeRequiredDSH, usedAfterDSH - usedBeforeDSH);
@@ -548,11 +548,11 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, whenSendingIndirectStateThenKe
constexpr uint32_t grfSize = sizeof(typename FamilyType::GRF);
size_t localWorkSize = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
auto numChannels = modifiedKernelInfo.kernelDescriptor.kernelAttributes.numLocalIdChannels;
size_t expectedIohSize = PerThreadDataHelper::getPerThreadDataSizeTotal(modifiedKernelInfo.getMaxSimdSize(), grfSize, numChannels, localWorkSize);
const auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
size_t expectedIohSize = PerThreadDataHelper::getPerThreadDataSizeTotal(modifiedKernelInfo.getMaxSimdSize(), grfSize, numChannels, localWorkSize, !kernelUsesLocalIds, gfxCoreHelper);
ASSERT_LE(expectedIohSize, ioh.getUsed());
auto expectedLocalIds = alignedMalloc(expectedIohSize, 64);
const auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
generateLocalIDs(expectedLocalIds, modifiedKernelInfo.getMaxSimdSize(),
std::array<uint16_t, 3>{{localWorkSizeX, localWorkSizeY, localWorkSizeZ}},
std::array<uint8_t, 3>{{modifiedKernelInfo.kernelDescriptor.kernelAttributes.workgroupDimensionsOrder[0],
@@ -1315,8 +1315,8 @@ struct HardwareCommandsImplicitArgsTests : Test<ClDeviceFixture> {
kernel.setGlobalWorkSizeValues(static_cast<uint32_t>(expectedImplicitArgs.globalSizeX), static_cast<uint32_t>(expectedImplicitArgs.globalSizeY), static_cast<uint32_t>(expectedImplicitArgs.globalSizeZ));
kernel.setGlobalWorkOffsetValues(static_cast<uint32_t>(expectedImplicitArgs.globalOffsetX), static_cast<uint32_t>(expectedImplicitArgs.globalOffsetY), static_cast<uint32_t>(expectedImplicitArgs.globalOffsetZ));
kernel.setNumWorkGroupsValues(expectedImplicitArgs.groupCountX, expectedImplicitArgs.groupCountY, expectedImplicitArgs.groupCountZ);
implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernel.getDescriptor());
const auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernel.getDescriptor(), false, gfxCoreHelper);
auto sizeCrossThreadData = kernel.getCrossThreadDataSize();
HardwareCommandsHelper<FamilyType>::sendCrossThreadData(
@@ -1382,7 +1382,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, HardwareCommandsImplicitArgsTests, givenKernelWithI
generateLocalIDs(expectedLocalIds, expectedImplicitArgs.simdWidth, localSize, workgroupDimOrder, false, grfSize, gfxCoreHelper);
auto localIdsProgrammingSize = implicitArgsProgrammingSize - sizeof(ImplicitArgs);
size_t sizeForLocalIds = PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize);
size_t sizeForLocalIds = PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize, false, gfxCoreHelper);
EXPECT_EQ(0, memcmp(expectedLocalIds, indirectHeapAllocation->getUnderlyingBuffer(), sizeForLocalIds));
alignedFree(expectedLocalIds);
@@ -1416,7 +1416,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, HardwareCommandsImplicitArgsTests, givenKernelWithI
generateLocalIDs(expectedLocalIds, expectedImplicitArgs.simdWidth, localSize, expectedDimOrder, false, grfSize, gfxCoreHelper);
auto localIdsProgrammingSize = implicitArgsProgrammingSize - sizeof(ImplicitArgs);
size_t sizeForLocalIds = PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize);
size_t sizeForLocalIds = PerThreadDataHelper::getPerThreadDataSizeTotal(expectedImplicitArgs.simdWidth, grfSize, 3u, totalLocalSize, false, gfxCoreHelper);
EXPECT_EQ(0, memcmp(expectedLocalIds, indirectHeapAllocation->getUnderlyingBuffer(), sizeForLocalIds));
alignedFree(expectedLocalIds);

3
shared/source/command_container/command_encoder_bdw_and_later.inl
View File

@@ -161,7 +161,8 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
sizePerThreadData, hwInfo);
uint32_t sizeThreadData = sizePerThreadDataForWholeGroup + sizeCrossThreadData;
uint32_t sizeForImplicitArgsPatching = NEO::ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor);
bool isHwLocalIdGeneration = false;
uint32_t sizeForImplicitArgsPatching = NEO::ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor, isHwLocalIdGeneration, gfxCoreHelper);
uint32_t iohRequiredSize = sizeThreadData + sizeForImplicitArgsPatching;
uint64_t offsetThreadData = 0u;
{

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

@@ -200,7 +200,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
}
uint32_t sizeThreadData = sizePerThreadDataForWholeGroup + sizeCrossThreadData;
uint32_t sizeForImplicitArgsPatching = NEO::ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor);
uint32_t sizeForImplicitArgsPatching = NEO::ImplicitArgsHelper::getSizeForImplicitArgsPatching(pImplicitArgs, kernelDescriptor, !localIdsGenerationByRuntime, gfxCoreHelper);
uint32_t iohRequiredSize = sizeThreadData + sizeForImplicitArgsPatching;
{
auto heap = container.getIndirectHeap(HeapType::INDIRECT_OBJECT);

7
shared/source/helpers/per_thread_data.h
View File

@@ -6,6 +6,7 @@
*/
#pragma once
#include "shared/source/helpers/gfx_core_helper.h"
#include "shared/source/helpers/local_id_gen.h"
#include <cstddef>
@@ -19,8 +20,10 @@ struct PerThreadDataHelper {
uint32_t simd,
uint32_t grfSize,
uint32_t numChannels,
size_t localWorkSize) {
return getThreadsPerWG(simd, static_cast<uint32_t>(localWorkSize)) * getPerThreadSizeLocalIDs(simd, grfSize, numChannels);
size_t localWorkSize,
bool isHwLocalIdGeneration,
const GfxCoreHelper &gfxCoreHelper) {
return gfxCoreHelper.calculateNumThreadsPerThreadGroup(simd, static_cast<uint32_t>(localWorkSize), grfSize, isHwLocalIdGeneration) * getPerThreadSizeLocalIDs(simd, grfSize, numChannels);
}
}; // namespace PerThreadDataHelper
} // namespace NEO

2
shared/source/kernel/implicit_args.h
View File

@@ -51,7 +51,7 @@ inline constexpr const char *implicitArgsRelocationSymbolName = "__INTEL_PATCH_C
namespace ImplicitArgsHelper {
std::array<uint8_t, 3> getDimensionOrderForLocalIds(const uint8_t *workgroupDimensionsOrder, std::optional<std::pair<bool /* localIdsGeneratedByRuntime */, uint32_t /* walkOrderForHwGenerationOfLocalIds */>> hwGenerationOfLocalIdsParams);
uint32_t getGrfSize(uint32_t simd);
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor);
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool localIdsGeneratedByRuntime, const GfxCoreHelper &gfxCoreHelper);
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool /* localIdsGeneratedByRuntime */, uint32_t /* walkOrderForHwGenerationOfLocalIds */>> hwGenerationOfLocalIdsParams, const GfxCoreHelper &gfxCoreHelper);
} // namespace ImplicitArgsHelper
} // namespace NEO

8
shared/source/kernel/implicit_args_helper.cpp
View File

@@ -41,7 +41,7 @@ uint32_t getGrfSize(uint32_t simd) {
return 32u;
}
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor) {
uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const KernelDescriptor &kernelDescriptor, bool isHwLocalIdGeneration, const GfxCoreHelper &gfxCoreHelper) {
if (!pImplicitArgs) {
return 0;
}
@@ -56,15 +56,15 @@ uint32_t getSizeForImplicitArgsPatching(const ImplicitArgs *pImplicitArgs, const
auto itemsInGroup = Math::computeTotalElementsCount(localWorkSize);
uint32_t localIdsSizeNeeded =
alignUp(static_cast<uint32_t>(NEO::PerThreadDataHelper::getPerThreadDataSizeTotal(
simdSize, grfSize, 3u, itemsInGroup)),
simdSize, grfSize, 3u, itemsInGroup, isHwLocalIdGeneration, gfxCoreHelper)),
MemoryConstants::cacheLineSize);
return implicitArgsSize + localIdsSizeNeeded;
}
}
void *patchImplicitArgs(void *ptrToPatch, const ImplicitArgs &implicitArgs, const KernelDescriptor &kernelDescriptor, std::optional<std::pair<bool, uint32_t>> hwGenerationOfLocalIdsParams, const GfxCoreHelper &gfxCoreHelper) {
auto totalSizeToProgram = getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor);
auto localIdsGeneratedByHw = hwGenerationOfLocalIdsParams.has_value() ? hwGenerationOfLocalIdsParams.value().first : false;
auto totalSizeToProgram = getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, localIdsGeneratedByHw, gfxCoreHelper);
auto retVal = ptrOffset(ptrToPatch, totalSizeToProgram);
auto patchImplicitArgsBufferInCrossThread = NEO::isValidOffset<>(kernelDescriptor.payloadMappings.implicitArgs.implicitArgsBuffer);

25
shared/test/unit_test/kernel/implicit_args_helper_tests.cpp
View File

@@ -57,8 +57,8 @@ TEST(ImplicitArgsHelperTest, givenSimdGreaterThanOneWhenGettingGrfSizeThenGrfSiz
TEST(ImplicitArgsHelperTest, givenNoImplicitArgsWhenGettingSizeForImplicitArgsProgrammingThenZeroIsReturned) {
KernelDescriptor kernelDescriptor{};
EXPECT_EQ(0u, ImplicitArgsHelper::getSizeForImplicitArgsPatching(nullptr, kernelDescriptor));
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
EXPECT_EQ(0u, ImplicitArgsHelper::getSizeForImplicitArgsPatching(nullptr, kernelDescriptor, false, *gfxCoreHelper.get()));
}
TEST(ImplicitArgsHelperTest, givenImplicitArgsWithoutImplicitArgsBufferOffsetInPayloadMappingWhenGettingSizeForImplicitArgsProgrammingThenCorrectSizeIsReturned) {
@@ -75,8 +75,9 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithoutImplicitArgsBufferOffsetInP
auto totalWorkgroupSize = implicitArgs.localSizeX * implicitArgs.localSizeY * implicitArgs.localSizeZ;
auto localIdsSize = alignUp(PerThreadDataHelper::getPerThreadDataSizeTotal(implicitArgs.simdWidth, 32u /* grfSize */, 3u /* num channels */, totalWorkgroupSize), MemoryConstants::cacheLineSize);
EXPECT_EQ(localIdsSize + implicitArgs.structSize, ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor));
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
auto localIdsSize = alignUp(PerThreadDataHelper::getPerThreadDataSizeTotal(implicitArgs.simdWidth, 32u /* grfSize */, 3u /* num channels */, totalWorkgroupSize, false, *gfxCoreHelper.get()), MemoryConstants::cacheLineSize);
EXPECT_EQ(localIdsSize + implicitArgs.structSize, ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, false, *gfxCoreHelper.get()));
}
TEST(ImplicitArgsHelperTest, givenImplicitArgsWithImplicitArgsBufferOffsetInPayloadMappingWhenGettingSizeForImplicitArgsProgrammingThenCorrectSizeIsReturned) {
@@ -90,8 +91,8 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithImplicitArgsBufferOffsetInPayl
implicitArgs.localSizeX = 2;
implicitArgs.localSizeY = 3;
implicitArgs.localSizeZ = 4;
EXPECT_EQ(alignUp(implicitArgs.structSize, MemoryConstants::cacheLineSize), ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor));
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
EXPECT_EQ(alignUp(implicitArgs.structSize, MemoryConstants::cacheLineSize), ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, false, *gfxCoreHelper.get()));
}
TEST(ImplicitArgsHelperTest, givenImplicitArgsWithoutImplicitArgsBufferOffsetInPayloadMappingWhenPatchingImplicitArgsThenOnlyProperRegionIsPatched) {
@@ -108,8 +109,8 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithoutImplicitArgsBufferOffsetInP
implicitArgs.localSizeX = 2;
implicitArgs.localSizeY = 3;
implicitArgs.localSizeZ = 4;
auto totalSizeForPatching = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor);
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
auto totalSizeForPatching = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, false, *gfxCoreHelper.get());
auto totalWorkgroupSize = implicitArgs.localSizeX * implicitArgs.localSizeY * implicitArgs.localSizeZ;
auto localIdsPatchingSize = totalWorkgroupSize * 3 * sizeof(uint16_t);
@@ -119,7 +120,7 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithoutImplicitArgsBufferOffsetInP
uint8_t pattern = 0xcd;
memset(memoryToPatch.get(), pattern, totalSizeForPatching);
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
auto retVal = ImplicitArgsHelper::patchImplicitArgs(memoryToPatch.get(), implicitArgs, kernelDescriptor, {}, *gfxCoreHelper.get());
EXPECT_EQ(retVal, ptrOffset(memoryToPatch.get(), totalSizeForPatching));
@@ -150,8 +151,8 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithImplicitArgsBufferOffsetInPayl
implicitArgs.localSizeX = 2;
implicitArgs.localSizeY = 3;
implicitArgs.localSizeZ = 4;
auto totalSizeForPatching = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor);
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
auto totalSizeForPatching = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&implicitArgs, kernelDescriptor, false, *gfxCoreHelper.get());
EXPECT_EQ(0x80u, totalSizeForPatching);
@@ -160,7 +161,7 @@ TEST(ImplicitArgsHelperTest, givenImplicitArgsWithImplicitArgsBufferOffsetInPayl
uint8_t pattern = 0xcd;
memset(memoryToPatch.get(), pattern, totalSizeForPatching);
auto gfxCoreHelper = GfxCoreHelper::create(defaultHwInfo->platform.eRenderCoreFamily);
auto retVal = ImplicitArgsHelper::patchImplicitArgs(memoryToPatch.get(), implicitArgs, kernelDescriptor, {}, *gfxCoreHelper.get());
EXPECT_EQ(retVal, ptrOffset(memoryToPatch.get(), totalSizeForPatching));