intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 21:18:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: cache flush dependency for queue blocked

Browse Source 
Related-to: NEO-9872, HSD-18038461954
Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
This commit is contained in:
Maciej Plewka
2024-05-20 10:39:06 +00:00
committed by Compute-Runtime-Automation
parent e01d34741d
commit 90df4b298b
12 changed files with 215 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
opencl/source/command_queue/command_queue_hw.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2023 Intel Corporation
* Copyright (C) 2018-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -9,6 +9,7 @@
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/command_stream/preemption.h"
#include "shared/source/device/device.h"
#include "shared/source/helpers/bcs_ccs_dependency_pair_container.h"
#include "shared/source/helpers/engine_control.h"
#include "shared/source/helpers/hw_info.h"
#include "shared/source/os_interface/os_context.h"
@@ -401,7 +402,8 @@ class CommandQueueHw : public CommandQueue {
EventBuilder &externalEventBuilder,
std::unique_ptr<PrintfHandler> &&printfHandler,
CommandStreamReceiver *bcsCsr,
TagNodeBase *multiRootDeviceSyncNode);
TagNodeBase *multiRootDeviceSyncNode,
CsrDependencyContainer *csrDependencies);
CompletionStamp enqueueCommandWithoutKernel(Surface **surfaces,
size_t surfaceCount,
@@ -449,6 +451,7 @@ class CommandQueueHw : public CommandQueue {
protected:
MOCKABLE_VIRTUAL void enqueueHandlerHook(const unsigned int commandType, const MultiDispatchInfo &dispatchInfo){};
MOCKABLE_VIRTUAL bool prepareCsrDependency(CsrDependencies &csrDeps, CsrDependencyContainer &dependencyTags, TimestampPacketDependencies &timestampPacketDependencies, TagAllocatorBase *allocator, bool blockQueue);
size_t calculateHostPtrSizeForImage(const size_t *region, size_t rowPitch, size_t slicePitch, Image *image);
cl_int enqueueReadWriteBufferOnCpuWithMemoryTransfer(cl_command_type commandType, Buffer *buffer,

45
opencl/source/command_queue/enqueue_common.h
View File

@@ -9,6 +9,7 @@
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/command_stream/wait_status.h"
#include "shared/source/direct_submission/relaxed_ordering_helper.h"
#include "shared/source/helpers/bcs_ccs_dependency_pair_container.h"
#include "shared/source/helpers/engine_node_helper.h"
#include "shared/source/helpers/flat_batch_buffer_helper.h"
#include "shared/source/helpers/flush_stamp.h"
@@ -457,7 +458,8 @@ cl_int CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
eventBuilder,
std::move(printfHandler),
nullptr,
multiRootEventSyncStamp);
multiRootEventSyncStamp,
nullptr);
}
if (deferredTimestampPackets.get()) {
@@ -994,7 +996,8 @@ void CommandQueueHw<GfxFamily>::enqueueBlocked(
EventBuilder &externalEventBuilder,
std::unique_ptr<PrintfHandler> &&printfHandler,
CommandStreamReceiver *bcsCsr,
TagNodeBase *multiRootDeviceSyncNode) {
TagNodeBase *multiRootDeviceSyncNode,
CsrDependencyContainer *dependencyTags) {
TakeOwnershipWrapper<CommandQueueHw<GfxFamily>> queueOwnership(*this);
@@ -1033,9 +1036,8 @@ void CommandQueueHw<GfxFamily>::enqueueBlocked(
storeTimestampPackets = (timestampPacketContainer != nullptr);
}
if (enqueueProperties.operation != EnqueueProperties::Operation::gpuKernel) {
command = std::make_unique<CommandWithoutKernel>(*this, blockedCommandsData);
command = std::make_unique<CommandWithoutKernel>(*this, blockedCommandsData, dependencyTags);
} else {
// store task data in event
std::vector<Surface *> allSurfaces;
@@ -1244,6 +1246,23 @@ size_t CommandQueueHw<GfxFamily>::calculateHostPtrSizeForImage(const size_t *reg
return Image::calculateHostPtrSize(region, dstRowPitch, dstSlicePitch, bytesPerPixel, image->getImageDesc().image_type);
}
template <typename GfxFamily>
bool CommandQueueHw<GfxFamily>::prepareCsrDependency(CsrDependencies &csrDeps, CsrDependencyContainer &dependencyTags, TimestampPacketDependencies &timestampPacketDependencies, TagAllocatorBase *allocator, bool blockQueue) {
for (auto &dependentCsr : csrDeps.csrWithMultiEngineDependencies) {
auto tag = allocator->getTag();
timestampPacketDependencies.multiCsrDependencies.add(tag);
if (!blockQueue) {
bool submitStatus = dependentCsr->submitDependencyUpdate(tag);
if (!submitStatus) {
return submitStatus;
}
} else {
dependencyTags.push_back(std::make_pair(dependentCsr, tag));
}
}
return true;
}
template <typename GfxFamily>
bool CommandQueueHw<GfxFamily>::isSplitEnqueueBlitNeeded(TransferDirection transferDirection, size_t transferSize, CommandStreamReceiver &csr) {
auto bcsSplit = getDevice().isBcsSplitSupported() &&
@@ -1438,14 +1457,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
if (isCacheFlushForBcsRequired() && gpgpuSubmission) {
timestampPacketDependencies.cacheFlushNodes.add(allocator->getTag());
}
for (auto &dependentCsr : csrDeps.csrWithMultiEngineDependencies) {
auto tag = allocator->getTag();
timestampPacketDependencies.multiCsrDependencies.add(tag);
bool submitStatus = dependentCsr->submitDependencyUpdate(tag);
if (!submitStatus) {
return CL_OUT_OF_RESOURCES;
}
}
obtainNewTimestampPacketNodes(1, timestampPacketDependencies.previousEnqueueNodes, clearAllDependencies, bcsCsr);
csrDeps.timestampPacketContainer.push_back(&timestampPacketDependencies.previousEnqueueNodes);
@@ -1472,6 +1484,13 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
gpgpuCommandStream = obtainCommandStream<cmdType>(csrDeps, true, blockQueue, multiDispatchInfo, eventsRequest, blockedCommandsData, nullptr, 0, false, false);
gpgpuCommandStreamStart = gpgpuCommandStream->getUsed();
}
CsrDependencyContainer dependencyTags;
if (csrDeps.csrWithMultiEngineDependencies.size() > 0) {
bool submitStatus = prepareCsrDependency(csrDeps, dependencyTags, timestampPacketDependencies, allocator, blockQueue);
if (!submitStatus) {
return CL_OUT_OF_RESOURCES;
}
}
blitPropertiesContainer.push_back(processDispatchForBlitEnqueue(bcsCsr, multiDispatchInfo, timestampPacketDependencies,
eventsRequest, gpgpuCommandStream, cmdType, blockQueue, multiRootEventSyncStamp));
@@ -1501,7 +1520,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
updateFromCompletionStamp(completionStamp, pEventBuilder->getEvent());
if (blockQueue) {
enqueueBlocked(cmdType, nullptr, 0, multiDispatchInfo, timestampPacketDependencies, blockedCommandsData, enqueueProperties, eventsRequest, *pEventBuilder, nullptr, &bcsCsr, multiRootEventSyncStamp);
enqueueBlocked(cmdType, nullptr, 0, multiDispatchInfo, timestampPacketDependencies, blockedCommandsData, enqueueProperties, eventsRequest, *pEventBuilder, nullptr, &bcsCsr, multiRootEventSyncStamp, &dependencyTags);
if (gpgpuSubmission) {
if (debugManager.flags.ForceCsrLockInBcsEnqueueOnlyForGpgpuSubmission.get() == 1) {