intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-04 15:53:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: Implement V2 version of tg dispatch size algorithm.

Browse Source 
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
Related-To: NEO-6989

-Prevent imbalance in multi dimensional dispatches
-Make sure to utilize as much Eus as possible
-Prefer highest possible tg dspatch count possible
-Make sure that xe_core doesn't have uneven workgroups
This commit is contained in:
Michal Mrozek
2023-11-03 11:56:01 +00:00
committed by Compute-Runtime-Automation
parent 098ea3a997
commit ed897c302d
4 changed files with 215 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
shared/test/common/test_files/igdrcl.config
View File

@@ -558,4 +558,5 @@ PrintGmmCompressionParams = 0
SkipInOrderNonWalkerSignalingAllowed = 0
PrintKernelDispatchParameters = 0
SetAmountOfReusableAllocationsPerCmdQueue = -1
ForceThreadGroupDispatchSizeAlgorithm = -1
# Please don't edit below this line

144
shared/test/unit_test/xe_hpc_core/test_encode_xe_hpc_core.cpp
View File

@@ -535,6 +535,8 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenMultipleTilesAndImplicitScalin
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsCorrectlySet) {
DebugManagerStateRestore restorer;
DebugManager.flags.ForceThreadGroupDispatchSizeAlgorithm.set(1u);
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
@@ -559,6 +561,8 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupWh
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAndDimensionsWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsCorrectlySet) {
DebugManagerStateRestore restorer;
DebugManager.flags.ForceThreadGroupDispatchSizeAlgorithm.set(1u);
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
WALKER_TYPE walkerCmd{};
@@ -675,6 +679,146 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDifferentNumGrfWhenCallingAdju
}
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenVariousDispatchParamtersWhenAlogrithmV2IsUsedThenProperThreadGroupDispatchSizeIsChoosen) {
DebugManagerStateRestore restorer;
DebugManager.flags.ForceThreadGroupDispatchSizeAlgorithm.set(2u);
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 mutableHwInfo = pDevice->getRootDeviceEnvironment().getMutableHardwareInfo();
mutableHwInfo->gtSystemInfo.MaxSubSlicesSupported = 64u;
mutableHwInfo->gtSystemInfo.ThreadCount = 4096u;
auto hwInfo = pDevice->getHardwareInfo();
hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
uint32_t numGrf = GrfConfig::DefaultGrfNumber;
const uint32_t threadGroupCount = gfxCoreHelper.calculateAvailableThreadCount(hwInfo, numGrf);
INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(1);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(256);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(64u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(64);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(512);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(512);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(8u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(512);
walkerCmd.setThreadGroupIdYDimension(1);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(510);
walkerCmd.setThreadGroupIdYDimension(512);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(509);
walkerCmd.setThreadGroupIdYDimension(512);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(508);
walkerCmd.setThreadGroupIdYDimension(512);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(16u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(508);
walkerCmd.setThreadGroupIdYDimension(512);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(16u);
numGrf = GrfConfig::LargeGrfNumber;
walkerCmd.setThreadGroupIdXDimension(508);
walkerCmd.setThreadGroupIdYDimension(512);
walkerCmd.setThreadGroupIdZDimension(1);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(1);
walkerCmd.setThreadGroupIdYDimension(510);
walkerCmd.setThreadGroupIdZDimension(512);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(1);
walkerCmd.setThreadGroupIdYDimension(509);
walkerCmd.setThreadGroupIdZDimension(512);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(16u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(1);
walkerCmd.setThreadGroupIdYDimension(508);
walkerCmd.setThreadGroupIdZDimension(512);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_4, iddArg.getThreadGroupDispatchSize());
iddArg.setNumberOfThreadsInGpgpuThreadGroup(32u);
numGrf = GrfConfig::DefaultGrfNumber;
walkerCmd.setThreadGroupIdXDimension(1);
walkerCmd.setThreadGroupIdYDimension(508);
walkerCmd.setThreadGroupIdZDimension(512);
EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, threadGroupCount, numGrf, walkerCmd);
EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
}
XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAndDebugFlagDisabledWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsDefault) {
using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
using WALKER_TYPE = typename FamilyType::WALKER_TYPE;