intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-31 20:13:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add multi tile barrier to marker command on multi tile device

Browse Source 
Related-To: NEO-6262

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2021-12-10 21:31:34 +00:00
committed by Compute-Runtime-Automation
parent d8ea439dab
commit 01348451db
12 changed files with 414 additions and 313 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
opencl/source/command_queue/enqueue_common.h
View File

@@ -168,8 +168,8 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
DeviceQueueHw<GfxFamily> *devQueueHw = castToObject<DeviceQueueHw<GfxFamily>>(devQueue);
TagNodeBase *hwTimeStamps = nullptr;
auto commandStreamReceiverOwnership = getGpgpuCommandStreamReceiver().obtainUniqueOwnership();
CommandStreamReceiver &computeCommandStreamReceiver = getGpgpuCommandStreamReceiver();
auto commandStreamReceiverOwnership = computeCommandStreamReceiver.obtainUniqueOwnership();
EventBuilder eventBuilder;
setupEvent(eventBuilder, event, commandType);
@@ -206,17 +206,17 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
BlitPropertiesContainer blitPropertiesContainer;
if (this->context->getRootDeviceIndices().size() > 1) {
eventsRequest.fillCsrDependenciesForTaskCountContainer(csrDeps, getGpgpuCommandStreamReceiver());
eventsRequest.fillCsrDependenciesForTaskCountContainer(csrDeps, computeCommandStreamReceiver);
}
bool enqueueWithBlitAuxTranslation = isBlitAuxTranslationRequired(multiDispatchInfo);
if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
if (computeCommandStreamReceiver.peekTimestampPacketWriteEnabled()) {
if (!clearDependenciesForSubCapture) {
eventsRequest.fillCsrDependenciesForTimestampPacketContainer(csrDeps, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
eventsRequest.fillCsrDependenciesForTimestampPacketContainer(csrDeps, computeCommandStreamReceiver, CsrDependencies::DependenciesType::OnCsr);
}
auto allocator = getGpgpuCommandStreamReceiver().getTimestampPacketAllocator();
auto allocator = computeCommandStreamReceiver.getTimestampPacketAllocator();
size_t nodesCount = 0u;
if (isCacheFlushCommand(commandType) || isMarkerWithProfiling) {
@@ -231,7 +231,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
}
if (nodesCount > 0) {
obtainNewTimestampPacketNodes(nodesCount, timestampPacketDependencies.previousEnqueueNodes, clearAllDependencies, getGpgpuCommandStreamReceiver());
obtainNewTimestampPacketNodes(nodesCount, timestampPacketDependencies.previousEnqueueNodes, clearAllDependencies, computeCommandStreamReceiver);
csrDeps.timestampPacketContainer.push_back(&timestampPacketDependencies.previousEnqueueNodes);
}
}
@@ -250,10 +250,10 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
}
if (!blockQueue && isOOQEnabled()) {
setupBarrierTimestampForBcsEngines(getGpgpuCommandStreamReceiver().getOsContext().getEngineType(), timestampPacketDependencies);
setupBarrierTimestampForBcsEngines(computeCommandStreamReceiver.getOsContext().getEngineType(), timestampPacketDependencies);
}
if (eventBuilder.getEvent() && getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
if (eventBuilder.getEvent() && computeCommandStreamReceiver.peekTimestampPacketWriteEnabled()) {
eventBuilder.getEvent()->addTimestampPacketNodes(*timestampPacketContainer);
eventBuilder.getEvent()->addTimestampPacketNodes(timestampPacketDependencies.nonAuxToAuxNodes);
eventBuilder.getEvent()->addTimestampPacketNodes(timestampPacketDependencies.auxToNonAuxNodes);
@@ -267,9 +267,9 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
timestampPacketDependencies);
} else if (isCacheFlushCommand(commandType)) {
processDispatchForCacheFlush(surfacesForResidency, numSurfaceForResidency, &commandStream, csrDeps);
} else if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
} else if (computeCommandStreamReceiver.peekTimestampPacketWriteEnabled()) {
if (CL_COMMAND_BARRIER == commandType) {
getGpgpuCommandStreamReceiver().requestStallingCommandsOnNextFlush();
computeCommandStreamReceiver.requestStallingCommandsOnNextFlush();
}
for (size_t i = 0; i < eventsRequest.numEventsInWaitList; i++) {
@@ -288,8 +288,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
if (isMarkerWithProfiling) {
if (numEventsInWaitList == 0) {
PipeControlArgs args(false);
MemorySynchronizationCommands<GfxFamily>::addPipeControl(commandStream, args);
computeCommandStreamReceiver.programComputeBarrierCommand(commandStream);
}
processDispatchForMarkerWithTimestampPacket(*this, &commandStream, eventsRequest, csrDeps);
}
@@ -305,7 +304,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
if (!blockQueue && multiDispatchInfo.peekMainKernel() && multiDispatchInfo.peekMainKernel()->requiresMemoryMigration()) {
for (auto &arg : multiDispatchInfo.peekMainKernel()->getMemObjectsToMigrate()) {
MigrationController::handleMigration(*this->context, getGpgpuCommandStreamReceiver(), arg.second);
MigrationController::handleMigration(*this->context, computeCommandStreamReceiver, arg.second);
migratedMemory = true;
}
}
@@ -315,7 +314,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
}
if (enqueueProperties.operation == EnqueueProperties::Operation::GpuKernel) {
csrDeps.makeResident(getGpgpuCommandStreamReceiver());
csrDeps.makeResident(computeCommandStreamReceiver);
completionStamp = enqueueNonBlocked<commandType>(
surfacesForResidency,
@@ -334,7 +333,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
getBcsForAuxTranslation());
if (parentKernel) {
getGpgpuCommandStreamReceiver().setMediaVFEStateDirty(true);
computeCommandStreamReceiver.setMediaVFEStateDirty(true);
if (devQueueHw->getSchedulerReturnInstance() > 0) {
waitUntilComplete(completionStamp.taskCount, {}, completionStamp.flushStamp, false);
@@ -427,7 +426,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
}
}
if (migratedMemory) {
getGpgpuCommandStreamReceiver().flushBatchedSubmissions();
computeCommandStreamReceiver.flushBatchedSubmissions();
}
}

2
opencl/source/command_queue/gpgpu_walker_base.inl
View File

@@ -208,7 +208,7 @@ size_t EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(uint32_t eventType, c
expectedSizeCS += EnqueueOperation<GfxFamily>::getSizeRequiredForTimestampPacketWrite();
if (isMarkerWithProfiling) {
if (!eventsInWaitlist) {
expectedSizeCS += MemorySynchronizationCommands<GfxFamily>::getSizeForSinglePipeControl();
expectedSizeCS += commandQueue.getGpgpuCommandStreamReceiver().getCmdsSizeForComputeBarrierCommand();
}
expectedSizeCS += 4 * EncodeStoreMMIO<GfxFamily>::size;
}