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

refactor: move prefetch to dedicated method

Browse Source 
Signed-off-by: Bartosz Dunajski <bartosz.dunajski@intel.com>
This commit is contained in:
Bartosz Dunajski
2025-06-06 15:48:00 +00:00
committed by Compute-Runtime-Automation
parent b6018010c4
commit 6d7148ee5e
5 changed files with 89 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@@ -414,6 +414,11 @@ struct CommandListCoreFamily : public CommandListImp {
void dispatchInOrderPostOperationBarrier(Event *signalOperation, bool dcFlushRequired, bool copyOperation);
NEO::GraphicsAllocation *getDeviceCounterAllocForResidency(NEO::GraphicsAllocation *counterDeviceAlloc);
bool isHighPriorityImmediateCmdList() const;
void prefetchKernelMemory(NEO::LinearStream &cmdStream, const Kernel &kernel, const NEO::GraphicsAllocation &ioh, size_t iohOffset, CommandToPatchContainer *outListCommands, uint64_t cmdId);
virtual void addKernelIsaMemoryPrefetchPadding(NEO::LinearStream &cmdStream, const Kernel &kernel, uint64_t cmdId) {}
virtual void addKernelIndirectDataMemoryPrefetchPadding(NEO::LinearStream &cmdStream, const Kernel &kernel, uint64_t cmdId) {}
virtual uint64_t getPrefetchCmdId() const { return std::numeric_limits<uint64_t>::max(); }
virtual uint32_t getIohSizeForPrefetch(const Kernel &kernel, uint32_t reserveExtraSpace) const;
NEO::InOrderPatchCommandsContainer<GfxFamily> inOrderPatchCmds;

45
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -394,6 +394,37 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::close() {
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::programL3(bool isSLMused) {}
template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::prefetchKernelMemory(NEO::LinearStream &cmdStream, const Kernel &kernel, const NEO::GraphicsAllocation &ioh, size_t iohOffset, CommandToPatchContainer *outListCommands, uint64_t cmdId) {
if (NEO::debugManager.flags.EnableMemoryPrefetch.get() != 1) {
return;
}
auto &rootExecEnv = device->getNEODevice()->getRootDeviceEnvironment();
auto currentCmdStreamPtr = cmdStream.getSpace(0);
auto cmdStreamOffset = cmdStream.getUsed();
NEO::EncodeMemoryPrefetch<GfxFamily>::programMemoryPrefetch(cmdStream, ioh, kernel.getIndirectSize(), iohOffset, rootExecEnv);
addKernelIndirectDataMemoryPrefetchPadding(cmdStream, kernel, cmdId);
NEO::EncodeMemoryPrefetch<GfxFamily>::programMemoryPrefetch(cmdStream, *kernel.getIsaAllocation(), kernel.getImmutableData()->getIsaSize(), kernel.getIsaOffsetInParentAllocation(), rootExecEnv);
addKernelIsaMemoryPrefetchPadding(cmdStream, kernel, cmdId);
if (outListCommands) {
auto &prefetchToPatch = outListCommands->emplace_back();
prefetchToPatch.type = CommandToPatch::PrefetchKernelMemory;
prefetchToPatch.pDestination = currentCmdStreamPtr;
prefetchToPatch.patchSize = cmdStream.getUsed() - cmdStreamOffset;
prefetchToPatch.offset = iohOffset;
}
}
template <GFXCORE_FAMILY gfxCoreFamily>
uint32_t CommandListCoreFamily<gfxCoreFamily>::getIohSizeForPrefetch(const Kernel &kernel, uint32_t reserveExtraSpace) const {
return kernel.getIndirectSize() + reserveExtraSpace;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernel(ze_kernel_handle_t kernelHandle,
const ze_group_count_t &threadGroupDimensions,
@@ -413,13 +444,10 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernel(ze_kernel_h
callId);
}
if (NEO::debugManager.flags.EnableMemoryPrefetch.get() == 1) {
auto kernel = Kernel::fromHandle(kernelHandle);
NEO::LinearStream &cmdStream = *commandContainer.getCommandStream();
auto heap = commandContainer.getIndirectHeap(NEO::IndirectHeapType::indirectObject);
NEO::EncodeMemoryPrefetch<GfxFamily>::programMemoryPrefetch(cmdStream, *heap->getGraphicsAllocation(), kernel->getIndirectSize(), heap->getUsed(), device->getNEODevice()->getRootDeviceEnvironment());
NEO::EncodeMemoryPrefetch<GfxFamily>::programMemoryPrefetch(cmdStream, *kernel->getIsaAllocation(), static_cast<uint32_t>(kernel->getImmutableData()->getIsaSize()), kernel->getIsaOffsetInParentAllocation(), device->getNEODevice()->getRootDeviceEnvironment());
}
auto kernel = Kernel::fromHandle(kernelHandle);
auto ioh = commandContainer.getHeapWithRequiredSizeAndAlignment(NEO::IndirectHeapType::indirectObject, getIohSizeForPrefetch(*kernel, launchParams.reserveExtraPayloadSpace), GfxFamily::indirectDataAlignment);
prefetchKernelMemory(*commandContainer.getCommandStream(), *kernel, *ioh->getGraphicsAllocation(), ioh->getUsed(), launchParams.outListCommands, getPrefetchCmdId());
ze_result_t ret = addEventsToCmdList(numWaitEvents, phWaitEvents, launchParams.outListCommands, launchParams.relaxedOrderingDispatch, true, true, launchParams.omitAddingWaitEventsResidency, false);
if (ret) {
@@ -445,8 +473,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernel(ze_kernel_h
return ZE_RESULT_ERROR_INVALID_ARGUMENT;
}
auto res = appendLaunchKernelWithParams(Kernel::fromHandle(kernelHandle), threadGroupDimensions,
event, launchParams);
auto res = appendLaunchKernelWithParams(kernel, threadGroupDimensions, event, launchParams);
if (!launchParams.skipInOrderNonWalkerSignaling) {
handleInOrderDependencyCounter(event, isInOrderNonWalkerSignalingRequired(event) && !(event && event->isCounterBased() && event->isUsingContextEndOffset()), false);

1
level_zero/core/source/cmdlist/command_to_patch.h
View File

@@ -31,6 +31,7 @@ struct CommandToPatch {
ComputeWalkerInlineDataScratch,
ComputeWalkerImplicitArgsScratch,
NoopSpace,
PrefetchKernelMemory,
Invalid
};
void *pDestination = nullptr;