intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 12:23:05 +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) {