intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 14:55:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: adjust thread group dispatch size

Browse Source 
adjust thread group dispatch size on pvc if chosen size does not evenly
divide dimension

this is to avoid leftover thread groups

Related-To: NEO-7927

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2023-04-27 16:58:48 +00:00
committed by Compute-Runtime-Automation
parent 4f297cf971
commit c84c7a0c91
13 changed files with 125 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -129,7 +129,7 @@ struct EncodeDispatchKernel {
static void programBarrierEnable(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, uint32_t value, const HardwareInfo &hwInfo);
static void adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf);
static void adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WALKER_TYPE &walkerCmd);
static void adjustBindingTablePrefetch(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, uint32_t samplerCount, uint32_t bindingTableEntryCount);

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

@@ -723,7 +723,7 @@ void EncodeDispatchKernel<Family>::adjustBindingTablePrefetch(INTERFACE_DESCRIPT
}
template <typename Family>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf) {}
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WALKER_TYPE &walkerCmd) {}
template <typename Family>
size_t EncodeDispatchKernel<Family>::getSizeRequiredDsh(const KernelDescriptor &kernelDescriptor, uint32_t iddCount) {

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

@@ -260,7 +260,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
}
auto threadGroupCount = cmd.getThreadGroupIdXDimension() * cmd.getThreadGroupIdYDimension() * cmd.getThreadGroupIdZDimension();
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired, cmd);
memcpy_s(iddPtr, sizeof(idd), &idd, sizeof(idd));

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

@@ -298,7 +298,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
walkerCmd.setPredicateEnable(args.isPredicate);
auto threadGroupCount = walkerCmd.getThreadGroupIdXDimension() * walkerCmd.getThreadGroupIdYDimension() * walkerCmd.getThreadGroupIdZDimension();
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired, walkerCmd);
EncodeDispatchKernel<Family>::appendAdditionalIDDFields(&idd, rootDeviceEnvironment, threadsPerThreadGroup,
args.dispatchInterface->getSlmTotalSize(),

23
shared/source/xe_hpc_core/command_encoder_xe_hpc_core.cpp
View File

@@ -31,7 +31,7 @@ void EncodeDispatchKernel<Family>::adjustTimestampPacket(WALKER_TYPE &walkerCmd,
}
template <>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf) {
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WALKER_TYPE &walkerCmd) {
const auto &productHelper = device.getProductHelper();
if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
@@ -54,15 +54,30 @@ void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCR
uint32_t numberOfThreadsInThreadGroup = interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup();
uint32_t dispatchedTotalThreadCount = numberOfThreadsInThreadGroup * threadGroupCount;
UNRECOVERABLE_IF(numberOfThreadsInThreadGroup == 0u);
auto tgDispatchSizeSelected = 1u;
if (dispatchedTotalThreadCount <= availableThreadCount) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1);
tgDispatchSizeSelected = 1;
} else if (numberOfThreadsInThreadGroup <= maxThreadsInTGForTGDispatchSize8) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8);
tgDispatchSizeSelected = 8;
} else if (numberOfThreadsInThreadGroup <= maxThreadsInTGForTGDispatchSize4) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4);
tgDispatchSizeSelected = 4;
} else {
tgDispatchSizeSelected = 2;
}
if (walkerCmd.getThreadGroupIdYDimension() > 1 || walkerCmd.getThreadGroupIdZDimension() > 1) {
while (walkerCmd.getThreadGroupIdXDimension() % tgDispatchSizeSelected != 0) {
tgDispatchSizeSelected /= 2;
}
}
if (tgDispatchSizeSelected == 8) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8);
} else if (tgDispatchSizeSelected == 1) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1);
} else if (tgDispatchSizeSelected == 2) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2);
} else {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4);
}
}
}

2
shared/source/xe_hpg_core/command_encoder_xe_hpg_core.cpp
View File

@@ -87,7 +87,7 @@ void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTO
}
template <>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf) {
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, const Device &device, const HardwareInfo &hwInfo, const uint32_t threadGroupCount, const uint32_t numGrf, WALKER_TYPE &walkerCmd) {
const auto &productHelper = device.getProductHelper();
if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
if (interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup() == 1) {

5
shared/test/unit_test/encoders/command_encoder_tests_dg2.cpp
View File

@@ -67,8 +67,9 @@ HWTEST2_F(DG2CommandEncoderTest, givenDG2AndCommandContainerWithDirtyHeapWhenGet
HWTEST2_F(DG2CommandEncoderTest, givenInterfaceDescriptorDataWhenForceThreadGroupDispatchSizeVariableIsDefaultThenThreadGroupDispatchSizeIsNotChanged, IsDG2) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
INTERFACE_DESCRIPTOR_DATA iddArg;
WALKER_TYPE walkerCmd{};
iddArg = FamilyType::cmdInitInterfaceDescriptorData;
const uint32_t forceThreadGroupDispatchSize = -1;
auto hwInfo = pDevice->getHardwareInfo();
@@ -83,7 +84,7 @@ HWTEST2_F(DG2CommandEncoderTest, givenInterfaceDescriptorDataWhenForceThreadGrou
for (auto numberOfThreadsInGroup : {1u, 4u, 16u}) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0, walkerCmd);
if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
if (numberOfThreadsInGroup == 1) {

10
shared/test/unit_test/encoders/test_encode_dispatch_kernel_xehp_and_later.cpp
View File

@@ -655,8 +655,9 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInlineDataRequiredAnd
HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDataWhenForceThreadGroupDispatchSizeVariableIsDefaultThenThreadGroupDispatchSizeIsNotChanged) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
INTERFACE_DESCRIPTOR_DATA iddArg;
WALKER_TYPE walkerCmd{};
iddArg = FamilyType::cmdInitInterfaceDescriptorData;
const uint32_t forceThreadGroupDispatchSize = -1;
auto hwInfo = pDevice->getHardwareInfo();
@@ -668,7 +669,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDa
uint32_t revisions[] = {REVISION_A0, REVISION_B};
for (auto revision : revisions) {
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(revision, hwInfo);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0, walkerCmd);
if (productHelper.isDisableOverdispatchAvailable(hwInfo)) {
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
@@ -680,8 +681,9 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDa
HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDataWhenForceThreadGroupDispatchSizeVariableIsSetThenThreadGroupDispatchSizeIsChanged) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
INTERFACE_DESCRIPTOR_DATA iddArg;
WALKER_TYPE walkerCmd{};
iddArg = FamilyType::cmdInitInterfaceDescriptorData;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
@@ -692,7 +694,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDa
DebugManagerStateRestore restorer;
DebugManager.flags.ForceThreadGroupDispatchSize.set(forceThreadGroupDispatchSize);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, pDevice->getHardwareInfo(), threadGroupCount, 1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, pDevice->getHardwareInfo(), threadGroupCount, 1, walkerCmd);
EXPECT_NE(defaultThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
EXPECT_EQ(forceThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());

86
shared/test/unit_test/xe_hpc_core/test_encode_xe_hpc_core.cpp
View File

@@ -491,7 +491,8 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenCleanHeapsAndSlmNotChangedAndU
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDispatchSizeSmallerOrEqualToAvailableThreadCountWhenAdjustInterfaceDescriptorDataIsCalledThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
const auto &productHelper = pDevice->getProductHelper();
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
@@ -502,7 +503,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDispatchSizeSmallerOrEqualToAv
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
for (const auto threadGroupCount : {1u, 2u}) {
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
@@ -510,6 +511,8 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDispatchSizeSmallerOrEqualToAv
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenMultipleTilesAndImplicitScalingWhenAdjustInterfaceDescriptorDataIsCalledThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
DebugManagerStateRestore restorer;
DebugManager.flags.EnableWalkerPartition.set(0);
const auto &productHelper = pDevice->getProductHelper();
@@ -523,18 +526,19 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenMultipleTilesAndImplicitScalin
const uint32_t threadGroupCount = gfxCoreHelper.calculateAvailableThreadCount(hwInfo, numGrf) / 32u;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(64u);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
ASSERT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
DebugManager.flags.EnableWalkerPartition.set(1);
pDevice->numSubDevices = 2;
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
const auto &productHelper = pDevice->getProductHelper();
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
@@ -549,15 +553,70 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupWh
for (const auto &[numberOfThreadsInThreadGroup, expectedThreadGroupDispatchSize] : testParams) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(expectedThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
}
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAndDimensionsWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
const auto &productHelper = pDevice->getProductHelper();
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
const uint32_t threadGroupCount = 512u;
const uint32_t numGrf = GrfConfig::DefaultGrfNumber;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(16u);
{
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8, iddArg.getThreadGroupDispatchSize());
}
walkerCmd.setThreadGroupIdYDimension(2);
walkerCmd.setThreadGroupIdZDimension(1);
{
walkerCmd.setThreadGroupIdXDimension(4);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4, iddArg.getThreadGroupDispatchSize());
}
{
walkerCmd.setThreadGroupIdXDimension(2);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
}
{
walkerCmd.setThreadGroupIdXDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(2);
{
walkerCmd.setThreadGroupIdXDimension(4);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4, iddArg.getThreadGroupDispatchSize());
}
{
walkerCmd.setThreadGroupIdXDimension(2);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
}
{
walkerCmd.setThreadGroupIdXDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDifferentNumGrfWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
const auto &productHelper = pDevice->getProductHelper();
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
@@ -570,7 +629,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDifferentNumGrfWhenCallingAdju
const uint32_t numGrf = GrfConfig::DefaultGrfNumber;
const uint32_t threadGroupCount = gfxCoreHelper.calculateAvailableThreadCount(hwInfo, numGrf);
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
ASSERT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
@@ -578,13 +637,15 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDifferentNumGrfWhenCallingAdju
const uint32_t numGrf = GrfConfig::LargeGrfNumber;
const uint32_t threadGroupCount = gfxCoreHelper.calculateAvailableThreadCount(hwInfo, numGrf);
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAndDebugFlagDisabledWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsDefault) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
DebugManagerStateRestore restorer;
DebugManager.flags.AdjustThreadGroupDispatchSize.set(0);
const auto &productHelper = pDevice->getProductHelper();
@@ -601,7 +662,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAn
for (const auto &[numberOfThreadsInThreadGroup, expectedThreadGroupDispatchSize] : testParams) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(expectedThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
}
@@ -609,7 +670,8 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAn
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenThreadGroupCountZeroWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsSetToDefault) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
const auto &productHelper = pDevice->getProductHelper();
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
@@ -619,7 +681,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenThreadGroupCountZeroWhenCallin
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}