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

RelaxedOrdering: Optimize GPU Queue stall by adding early return

Browse Source 
Related-To: NEO-7458

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2022-11-24 15:00:51 +00:00
committed by Compute-Runtime-Automation
parent a104636b31
commit a969636b6a
6 changed files with 39 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
shared/source/command_stream/linear_stream.h
View File

@@ -37,6 +37,8 @@ class LinearStream {
uint64_t getGpuBase() const;
void setGpuBase(uint64_t gpuAddress);
uint64_t getCurrentGpuAddressPosition() const;
void overrideMaxSize(size_t newMaxSize);
void replaceBuffer(void *buffer, size_t bufferSize);
GraphicsAllocation *getGraphicsAllocation() const;
@@ -112,4 +114,9 @@ inline GraphicsAllocation *LinearStream::getGraphicsAllocation() const {
inline void LinearStream::replaceGraphicsAllocation(GraphicsAllocation *gfxAllocation) {
graphicsAllocation = gfxAllocation;
}
inline uint64_t LinearStream::getCurrentGpuAddressPosition() const {
return (getGpuBase() + getUsed());
}
} // namespace NEO

2
shared/source/direct_submission/direct_submission_hw.h
View File

@@ -148,8 +148,6 @@ class DirectSubmissionHw {
size_t getSizeEnd();
uint64_t getCommandBufferPositionGpuAddress(void *position);
void dispatchPartitionRegisterConfiguration();
size_t getSizePartitionRegisterConfigurationSection();

30
shared/source/direct_submission/direct_submission_hw.inl
View File

@@ -82,7 +82,7 @@ DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmis
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchRelaxedOrderingSchedulerSection(uint32_t value) {
uint64_t schedulerStartAddress = getCommandBufferPositionGpuAddress(ringCommandStream.getSpace(0));
uint64_t schedulerStartAddress = ringCommandStream.getCurrentGpuAddressPosition();
uint64_t deferredTasksListGpuVa = deferredTasksListAllocation->getGpuAddress();
// 1. Init section
@@ -409,7 +409,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::startRingBuffer() {
if (ringCommandStream.getAvailableSpace() < requiredSize) {
switchRingBuffers();
}
uint64_t gpuStartVa = getCommandBufferPositionGpuAddress(ringCommandStream.getSpace(0));
uint64_t gpuStartVa = ringCommandStream.getCurrentGpuAddressPosition();
if (!this->partitionConfigSet) {
dispatchPartitionRegisterConfiguration();
@@ -558,14 +558,6 @@ inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeEnd() {
return size;
}
template <typename GfxFamily, typename Dispatcher>
inline uint64_t DirectSubmissionHw<GfxFamily, Dispatcher>::getCommandBufferPositionGpuAddress(void *position) {
void *currentBase = ringCommandStream.getCpuBase();
size_t offset = ptrDiff(position, currentBase);
return ringCommandStream.getGraphicsAllocation()->getGpuAddress() + static_cast<uint64_t>(offset);
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatch() {
size_t size = getSizeSemaphoreSection(false);
@@ -624,8 +616,8 @@ void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBu
}
dispatchStartSection(commandStreamAddress);
void *returnPosition = ringCommandStream.getSpace(0);
uint64_t returnGpuPointer = getCommandBufferPositionGpuAddress(returnPosition);
uint64_t returnGpuPointer = ringCommandStream.getCurrentGpuAddressPosition();
if (this->relaxedOrderingEnabled) {
dispatchRelaxedOrderingReturnPtrRegs(relaxedOrderingReturnPtrCmdStream, returnGpuPointer);
@@ -659,13 +651,21 @@ void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBu
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchRelaxedOrderingQueueStall() {
LinearStream bbStartStream(ringCommandStream.getSpace(EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart()),
EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart());
LriHelper<GfxFamily>::program(&ringCommandStream, CS_GPR_R5, 1, true);
dispatchSemaphoreSection(currentQueueWorkCount, false);
// patch conditional bb_start with current GPU address
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(bbStartStream, ringCommandStream.getCurrentGpuAddressPosition(),
CS_GPR_R1, 0, CompareOperation::Equal, false);
}
template <typename GfxFamily, typename Dispatcher>
size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatchRelaxedOrderingQueueStall() {
return getSizeSemaphoreSection(false) + sizeof(typename GfxFamily::MI_LOAD_REGISTER_IMM);
return getSizeSemaphoreSection(false) + sizeof(typename GfxFamily::MI_LOAD_REGISTER_IMM) +
EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart();
}
template <typename GfxFamily, typename Dispatcher>
@@ -762,8 +762,6 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffe
requiredMinimalSize += RelaxedOrderingHelper::getSizeTaskStoreSection<GfxFamily>() + RelaxedOrderingHelper::getSizeReturnPtrRegs<GfxFamily>();
}
getCommandBufferPositionGpuAddress(ringCommandStream.getSpace(0));
if (ringCommandStream.getAvailableSpace() < requiredMinimalSize) {
switchRingBuffers();
}
@@ -828,7 +826,7 @@ template <typename GfxFamily, typename Dispatcher>
inline uint64_t DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffers() {
GraphicsAllocation *nextRingBuffer = switchRingBuffersAllocations();
void *flushPtr = ringCommandStream.getSpace(0);
uint64_t currentBufferGpuVa = getCommandBufferPositionGpuAddress(flushPtr);
uint64_t currentBufferGpuVa = ringCommandStream.getCurrentGpuAddressPosition();
if (ringStart) {
dispatchSwitchRingBufferSection(nextRingBuffer->getGpuAddress());

5
shared/source/direct_submission/direct_submission_prefetch_mitigation_xe_hp_core_and_later.inl
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2021 Intel Corporation
* Copyright (C) 2021-2022 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -11,8 +11,7 @@ namespace NEO {
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchPrefetchMitigation() {
auto addressToJump = ptrOffset(ringCommandStream.getSpace(0u), getSizeStartSection());
dispatchStartSection(getCommandBufferPositionGpuAddress(addressToJump));
dispatchStartSection(ringCommandStream.getCurrentGpuAddressPosition() + getSizeStartSection());
}
template <typename GfxFamily, typename Dispatcher>