intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 09:09:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Disable EUFusion for odd work groups with DPAS on DG2

Browse Source 
Related-To: NEO-7495, HSD-14017007475

Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
This commit is contained in:
Maciej Plewka
2023-01-26 09:55:59 +00:00
committed by Compute-Runtime-Automation
parent 4c58eda90d
commit 429be6b4cb
26 changed files with 495 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@@ -259,9 +259,9 @@ struct CommandListCoreFamily : CommandListImp {
const CmdListKernelLaunchParams &launchParams);
ze_result_t prepareIndirectParams(const ze_group_count_t *threadGroupDimensions);
void updateStreamProperties(Kernel &kernel, bool isCooperative);
void updateStreamPropertiesForRegularCommandLists(Kernel &kernel, bool isCooperative);
void updateStreamPropertiesForFlushTaskDispatchFlags(Kernel &kernel, bool isCooperative);
void updateStreamPropertiesForRegularCommandLists(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect);
void updateStreamPropertiesForFlushTaskDispatchFlags(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect);
void updateStreamProperties(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect);
void clearCommandsToPatch();
size_t getTotalSizeForCopyRegion(const ze_copy_region_t *region, uint32_t pitch, uint32_t slicePitch);

47
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -2344,32 +2344,48 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::prepareIndirectParams(const ze
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel, bool isCooperative) {
void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect) {
if (this->isFlushTaskSubmissionEnabled) {
updateStreamPropertiesForFlushTaskDispatchFlags(kernel, isCooperative);
updateStreamPropertiesForFlushTaskDispatchFlags(kernel, isCooperative, threadGroupDimensions, isIndirect);
} else {
updateStreamPropertiesForRegularCommandLists(kernel, isCooperative);
updateStreamPropertiesForRegularCommandLists(kernel, isCooperative, threadGroupDimensions, isIndirect);
}
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForFlushTaskDispatchFlags(Kernel &kernel, bool isCooperative) {
void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForFlushTaskDispatchFlags(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect) {
auto &rootDeviceEnvironment = device->getNEODevice()->getRootDeviceEnvironment();
auto &kernelAttributes = kernel.getKernelDescriptor().kernelAttributes;
bool fusedEuDisabled = kernelAttributes.flags.requiresDisabledEUFusion;
auto &productHelper = device->getProductHelper();
if (productHelper.isCalculationForDisablingEuFusionWithDpasNeeded()) {
if (threadGroupDimensions) {
uint32_t *groupCountPtr = nullptr;
uint32_t groupCount[3] = {};
if (!isIndirect) {
groupCount[0] = threadGroupDimensions->groupCountX;
groupCount[1] = threadGroupDimensions->groupCountY;
groupCount[2] = threadGroupDimensions->groupCountZ;
groupCountPtr = groupCount;
}
fusedEuDisabled |= productHelper.isFusedEuDisabledForDpas(kernelAttributes.flags.usesSystolicPipelineSelectMode, kernel.getGroupSize(), groupCountPtr);
}
}
requiredStreamState.stateComputeMode.setPropertiesGrfNumberThreadArbitration(kernelAttributes.numGrfRequired, kernelAttributes.threadArbitrationPolicy, rootDeviceEnvironment);
requiredStreamState.frontEndState.setPropertiesComputeDispatchAllWalkerEnableDisableEuFusion(isCooperative, kernelAttributes.flags.requiresDisabledEUFusion, rootDeviceEnvironment);
requiredStreamState.frontEndState.setPropertiesComputeDispatchAllWalkerEnableDisableEuFusion(isCooperative, fusedEuDisabled, rootDeviceEnvironment);
requiredStreamState.pipelineSelect.setPropertySystolicMode(kernelAttributes.flags.usesSystolicPipelineSelectMode, rootDeviceEnvironment);
}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForRegularCommandLists(Kernel &kernel, bool isCooperative) {
void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForRegularCommandLists(Kernel &kernel, bool isCooperative, const ze_group_count_t *threadGroupDimensions, bool isIndirect) {
using VFE_STATE_TYPE = typename GfxFamily::VFE_STATE_TYPE;
auto &rootDeviceEnvironment = device->getNEODevice()->getRootDeviceEnvironment();
auto &kernelAttributes = kernel.getKernelDescriptor().kernelAttributes;
KernelImp &kernelImp = static_cast<KernelImp &>(kernel);
currentMocsState = static_cast<int32_t>(device->getMOCS(!kernelImp.getKernelRequiresUncachedMocs(), false) >> 1);
@@ -2391,8 +2407,23 @@ void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForRegularComma
currentIndirectObjectBaseAddress = ioh->getHeapGpuBase();
currentIndirectObjectSize = ioh->getHeapSizeInPages();
bool fusedEuDisabled = kernelAttributes.flags.requiresDisabledEUFusion;
auto &productHelper = device->getProductHelper();
if (productHelper.isCalculationForDisablingEuFusionWithDpasNeeded()) {
if (threadGroupDimensions) {
uint32_t *groupCountPtr = nullptr;
uint32_t groupCount[3] = {};
if (!isIndirect) {
groupCount[0] = threadGroupDimensions->groupCountX;
groupCount[1] = threadGroupDimensions->groupCountY;
groupCount[2] = threadGroupDimensions->groupCountZ;
groupCountPtr = groupCount;
}
fusedEuDisabled |= productHelper.isFusedEuDisabledForDpas(kernelAttributes.flags.usesSystolicPipelineSelectMode, kernel.getGroupSize(), groupCountPtr);
}
}
if (!containsAnyKernel) {
requiredStreamState.frontEndState.setProperties(isCooperative, kernelAttributes.flags.requiresDisabledEUFusion, cmdListDefaultDisableOverdispatch, -1, rootDeviceEnvironment);
requiredStreamState.frontEndState.setProperties(isCooperative, fusedEuDisabled, cmdListDefaultDisableOverdispatch, -1, rootDeviceEnvironment);
requiredStreamState.pipelineSelect.setProperties(true, false, kernelAttributes.flags.usesSystolicPipelineSelectMode, rootDeviceEnvironment);
requiredStreamState.stateBaseAddress.setProperties(kernelImp.getKernelDescriptor().kernelAttributes.flags.useGlobalAtomics, currentMocsState,
@@ -2425,7 +2456,7 @@ void CommandListCoreFamily<gfxCoreFamily>::updateStreamPropertiesForRegularComma
rootDeviceEnvironment);
}
finalStreamState.frontEndState.setProperties(isCooperative, kernelAttributes.flags.requiresDisabledEUFusion, cmdListDefaultDisableOverdispatch, -1, rootDeviceEnvironment);
finalStreamState.frontEndState.setProperties(isCooperative, fusedEuDisabled, cmdListDefaultDisableOverdispatch, -1, rootDeviceEnvironment);
bool isPatchingVfeStateAllowed = NEO::DebugManager.flags.AllowPatchingVfeStateInCommandLists.get();
if (finalStreamState.frontEndState.isDirty() && logicalStateHelperBlock) {
if (isPatchingVfeStateAllowed) {

2
level_zero/core/source/cmdlist/cmdlist_hw_base.inl
View File

@@ -152,7 +152,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
std::list<void *> additionalCommands;
updateStreamProperties(*kernel, launchParams.isCooperative);
updateStreamProperties(*kernel, launchParams.isCooperative, threadGroupDimensions, launchParams.isIndirect);
NEO::EncodeDispatchKernelArgs dispatchKernelArgs{
0, // eventAddress
neoDevice, // device

2
level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
View File

@@ -265,7 +265,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
this->containsStatelessUncachedResource |= kernelImp->getKernelRequiresUncachedMocs();
this->requiresQueueUncachedMocs |= kernelImp->getKernelRequiresQueueUncachedMocs();
updateStreamProperties(*kernel, launchParams.isCooperative);
updateStreamProperties(*kernel, launchParams.isCooperative, threadGroupDimensions, launchParams.isIndirect);
auto localMemSize = static_cast<uint32_t>(neoDevice->getDeviceInfo().localMemSize);
auto slmTotalSize = kernelImp->getSlmTotalSize();

1
level_zero/core/source/kernel/kernel_imp.cpp
View File

@@ -384,7 +384,6 @@ ze_result_t KernelImp::suggestGroupSize(uint32_t globalSizeX, uint32_t globalSiz
NEO::computeWorkgroupSize2D(maxWorkGroupSize, retGroupSize, workItems, simd);
}
}
*groupSizeX = static_cast<uint32_t>(retGroupSize[0]);
*groupSizeY = static_cast<uint32_t>(retGroupSize[1]);
*groupSizeZ = static_cast<uint32_t>(retGroupSize[2]);