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

Correct tg dispatch size heuristic

Browse Source 
Multiply available thread count by tile count
if implicit scaling is used

Related-To: NEO-6989

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2022-10-26 13:49:19 +00:00
committed by Compute-Runtime-Automation
parent e0c7e2e130
commit 6cf8b4daca
11 changed files with 45 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -103,7 +103,7 @@ struct EncodeDispatchKernel {
static void programBarrierEnable(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, uint32_t value, const HardwareInfo &hwInfo);
static void adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, 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);
static void adjustBindingTablePrefetch(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, uint32_t samplerCount, uint32_t bindingTableEntryCount);

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

@@ -714,7 +714,7 @@ void EncodeDispatchKernel<Family>::adjustBindingTablePrefetch(INTERFACE_DESCRIPT
}
template <typename Family>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, 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) {}
template <typename Family>
constexpr bool EncodeDispatchKernel<Family>::shouldUpdateGlobalAtomics(bool &currentVal, bool refVal, bool updateCurrent) { return false; }

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

@@ -234,7 +234,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
}
auto threadGroupCount = cmd.getThreadGroupIdXDimension() * cmd.getThreadGroupIdYDimension() * cmd.getThreadGroupIdZDimension();
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
memcpy_s(iddPtr, sizeof(idd), &idd, sizeof(idd));

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

@@ -282,7 +282,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
walkerCmd.setPredicateEnable(args.isPredicate);
auto threadGroupCount = walkerCmd.getThreadGroupIdXDimension() * walkerCmd.getThreadGroupIdYDimension() * walkerCmd.getThreadGroupIdZDimension();
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(idd, *args.device, hwInfo, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired);
EncodeDispatchKernel<Family>::appendAdditionalIDDFields(&idd, hwInfo, threadsPerThreadGroup,
args.dispatchInterface->getSlmTotalSize(),

2
shared/source/xe_hp_core/command_encoder_xe_hp_core.cpp
View File

@@ -61,7 +61,7 @@ void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DAT
}
template <>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, 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) {
const auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1);

7
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 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) {
const auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
@@ -47,6 +47,10 @@ void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCR
constexpr uint32_t maxThreadsInTGForTGDispatchSize4 = 32u;
auto &hwHelper = HwHelper::get(hwInfo.platform.eRenderCoreFamily);
uint32_t availableThreadCount = hwHelper.calculateAvailableThreadCount(hwInfo, numGrf);
if (ImplicitScalingHelper::isImplicitScalingEnabled(device.getDeviceBitfield(), true)) {
const uint32_t tilesCount = device.getNumSubDevices();
availableThreadCount *= tilesCount;
}
uint32_t numberOfThreadsInThreadGroup = interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup();
uint32_t dispatchedTotalThreadCount = numberOfThreadsInThreadGroup * threadGroupCount;
UNRECOVERABLE_IF(numberOfThreadsInThreadGroup == 0u);
@@ -63,6 +67,7 @@ void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCR
uint32_t exponent = INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1 - interfaceDescriptor.getThreadGroupDispatchSize();
uint32_t threadGroupDispatchSize = 1u << exponent;
if ((dispatchedTotalThreadCount % (numberOfThreadsInThreadGroup * threadGroupDispatchSize)) != 0) {
interfaceDescriptor.setThreadGroupDispatchSize(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1);
}

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

@@ -86,7 +86,7 @@ void EncodeDispatchKernel<Family>::appendAdditionalIDDFields(INTERFACE_DESCRIPTO
}
template <>
void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor, 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) {
const auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
if (interfaceDescriptor.getNumberOfThreadsInGpgpuThreadGroup() == 1) {

2
shared/test/unit_test/encoders/command_encoder_tests_dg2.cpp
View File

@@ -60,7 +60,7 @@ HWTEST2_F(DG2CommandEncoderTest, givenInterfaceDescriptorDataWhenForceThreadGrou
for (auto numberOfThreadsInGroup : {1u, 4u, 16u}) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, 0, 0);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
if (numberOfThreadsInGroup == 1) {

4
shared/test/unit_test/encoders/test_encode_dispatch_kernel_xehp_and_later.cpp
View File

@@ -628,7 +628,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDa
uint32_t revisions[] = {REVISION_A0, REVISION_B};
for (auto revision : revisions) {
hwInfo.platform.usRevId = hwInfoConfig.getHwRevIdFromStepping(revision, hwInfo);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, 0, 0);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
@@ -652,7 +652,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInterfaceDescriptorDa
DebugManagerStateRestore restorer;
DebugManager.flags.ForceThreadGroupDispatchSize.set(forceThreadGroupDispatchSize);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, pDevice->getHardwareInfo(), threadGroupCount, 1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, pDevice->getHardwareInfo(), threadGroupCount, 1);
EXPECT_NE(defaultThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
EXPECT_EQ(forceThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());

35
shared/test/unit_test/xe_hpc_core/test_encode_xe_hpc_core.cpp
View File

@@ -522,7 +522,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenInterfaceDescriptorDataAndNonS
for (const auto &revision : {REVISION_A0, REVISION_B}) {
hwInfo.platform.usRevId = hwInfoConfig.getHwRevIdFromStepping(revision, hwInfo);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, 0, 0);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 0, 0);
if (hwInfoConfig.isDisableOverdispatchAvailable(hwInfo)) {
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
@@ -545,12 +545,35 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDispatchSizeSmallerOrEqualToAv
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
for (const auto threadGroupCount : {1u, 2u}) {
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenMultipleTilesAndImplicitScalingWhenAdjustInterfaceDescriptorDataIsCalledThenThreadGroupDispatchSizeIsCorrectlySet) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
DebugManagerStateRestore restorer;
DebugManager.flags.EnableWalkerPartition.set(0);
const auto &hwInfoConfig = *HwInfoConfig::get(productFamily);
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = hwInfoConfig.getHwRevIdFromStepping(REVISION_B, hwInfo);
hwInfo.gtSystemInfo.EUCount = 32;
hwInfo.gtSystemInfo.DualSubSliceCount = hwInfo.gtSystemInfo.MaxDualSubSlicesSupported;
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
const uint32_t threadGroupCount = 32u;
const uint32_t numGrf = 32;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(64u);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EXPECT_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);
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;
@@ -568,7 +591,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupWh
for (const auto &[numberOfThreadsInThreadGroup, expectedThreadGroupDispatchSize] : testParams) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
if (hwInfo.gtSystemInfo.MaxDualSubSlicesSupported == hwInfo.gtSystemInfo.DualSubSliceCount) {
EXPECT_EQ(expectedThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
@@ -596,7 +619,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAn
for (const auto &[numberOfThreadsInThreadGroup, expectedThreadGroupDispatchSize] : testParams) {
iddArg.setNumberOfThreadsInGpgpuThreadGroup(numberOfThreadsInThreadGroup);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EXPECT_EQ(expectedThreadGroupDispatchSize, iddArg.getThreadGroupDispatchSize());
}
@@ -614,7 +637,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenIndivisibleDispatchSizeWhenCal
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(18u);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}
@@ -631,7 +654,7 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenThreadGroupCountZeroWhenCallin
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, hwInfo, threadGroupCount, numGrf);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
}