intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-04 23:56:39 +08:00
Code Issues Packages Projects Releases Wiki Activity

Feature(OCL) Use tag nodes for root device synchronization

Browse Source 
With this commit events created on multi root device contexts will
synchronize using signaled TagNodes instead of using taskCounts.

Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>

Related-To: NEO-7105
This commit is contained in:
Maciej Plewka
2022-12-07 11:07:43 +00:00
committed by Compute-Runtime-Automation
parent 6fac234655
commit 547d1c37b3
62 changed files with 995 additions and 210 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
opencl/source/helpers/properties_helper.cpp
View File

@@ -20,7 +20,6 @@ namespace NEO {
void flushDependentCsr(CommandStreamReceiver &dependentCsr, CsrDependencies &csrDeps) {
auto csrOwnership = dependentCsr.obtainUniqueOwnership();
dependentCsr.updateTagFromWait();
csrDeps.taskCountContainer.push_back({dependentCsr.peekTaskCount(), reinterpret_cast<uint64_t>(dependentCsr.getTagAddress())});
}
void EventsRequest::fillCsrDependenciesForTimestampPacketContainer(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr, CsrDependencies::DependenciesType depsType) const {
@@ -54,6 +53,7 @@ void EventsRequest::fillCsrDependenciesForTimestampPacketContainer(CsrDependenci
if (productHelper.isDcFlushAllowed()) {
if (!dependentCsr.isLatestTaskCountFlushed()) {
flushDependentCsr(dependentCsr, csrDeps);
//csrDeps.taskCountContainer.push_back({dependentCsr.peekTaskCount(), reinterpret_cast<uint64_t>(dependentCsr.getTagAddress())});
currentCsr.makeResident(*dependentCsr.getTagAllocation());
}
}
@@ -62,23 +62,22 @@ void EventsRequest::fillCsrDependenciesForTimestampPacketContainer(CsrDependenci
}
}
void EventsRequest::fillCsrDependenciesForTaskCountContainer(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr) const {
void EventsRequest::fillCsrDependenciesForRootDevices(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr) const {
for (cl_uint i = 0; i < this->numEventsInWaitList; i++) {
auto event = castToObjectOrAbort<Event>(this->eventWaitList[i]);
if (event->isUserEvent() || CompletionStamp::notReady == event->peekTaskCount()) {
continue;
}
if (event->getCommandQueue() && event->getCommandQueue()->getDevice().getRootDeviceIndex() != currentCsr.getRootDeviceIndex()) {
auto timestampPacketContainer = event->getMultiRootDeviceTimestampPacketNodes();
if (!timestampPacketContainer || timestampPacketContainer->peekNodes().empty()) {
continue;
}
auto &dependentCsr = event->getCommandQueue()->getGpgpuCommandStreamReceiver();
if (!dependentCsr.isLatestTaskCountFlushed()) {
flushDependentCsr(dependentCsr, csrDeps);
} else {
csrDeps.taskCountContainer.push_back({event->peekTaskCount(), reinterpret_cast<uint64_t>(dependentCsr.getTagAddress())});
}
auto graphicsAllocation = event->getCommandQueue()->getGpgpuCommandStreamReceiver().getTagsMultiAllocation()->getGraphicsAllocation(currentCsr.getRootDeviceIndex());
currentCsr.getResidencyAllocations().push_back(graphicsAllocation);
csrDeps.multiRootTimeStampSyncContainer.push_back(timestampPacketContainer);
}
}
}

2
opencl/source/helpers/properties_helper.h
View File

@@ -25,7 +25,7 @@ struct EventsRequest {
: numEventsInWaitList(numEventsInWaitList), eventWaitList(eventWaitList), outEvent(outEvent) {}
void fillCsrDependenciesForTimestampPacketContainer(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr, CsrDependencies::DependenciesType depsType) const;
void fillCsrDependenciesForTaskCountContainer(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr) const;
void fillCsrDependenciesForRootDevices(CsrDependencies &csrDeps, CommandStreamReceiver &currentCsr) const;
void setupBcsCsrForOutputEvent(CommandStreamReceiver &bcsCsr) const;
cl_uint numEventsInWaitList;

14
opencl/source/helpers/task_information.cpp
View File

@@ -117,10 +117,11 @@ CompletionStamp &CommandMapUnmap::submit(TaskCountType taskLevel, bool terminate
CommandComputeKernel::CommandComputeKernel(CommandQueue &commandQueue, std::unique_ptr<KernelOperation> &kernelOperation, std::vector<Surface *> surfaces,
bool flushDC, bool usesSLM, uint32_t commandType, std::unique_ptr<PrintfHandler> &&printfHandler,
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount)
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount,
TagNodeBase *multiRootDeviceSyncNode)
: Command(commandQueue, kernelOperation), surfaces(std::move(surfaces)), flushDC(flushDC), slmUsed(usesSLM),
commandType(commandType), printfHandler(std::move(printfHandler)), kernel(kernel),
kernelCount(kernelCount), preemptionMode(preemptionMode) {
kernelCount(kernelCount), preemptionMode(preemptionMode), multiRootDeviceSyncNode(multiRootDeviceSyncNode) {
UNRECOVERABLE_IF(nullptr == this->kernel);
kernel->incRefInternal();
}
@@ -162,6 +163,9 @@ CompletionStamp &CommandComputeKernel::submit(TaskCountType taskLevel, bool term
printfHandler->makeResident(commandStreamReceiver);
}
makeTimestampPacketsResident(commandStreamReceiver);
if (multiRootDeviceSyncNode != nullptr) {
commandStreamReceiver.makeResident(*multiRootDeviceSyncNode->getBaseGraphicsAllocation());
}
if (kernelOperation->blitPropertiesContainer.size() > 0) {
CsrDependencies csrDeps;
@@ -213,7 +217,7 @@ CompletionStamp &CommandComputeKernel::submit(TaskCountType taskLevel, bool term
false); // hasRelaxedOrderingDependencies
if (commandQueue.getContext().getRootDeviceIndices().size() > 1) {
eventsRequest.fillCsrDependenciesForTaskCountContainer(dispatchFlags.csrDependencies, commandStreamReceiver);
eventsRequest.fillCsrDependenciesForRootDevices(dispatchFlags.csrDependencies, commandStreamReceiver);
}
const bool isHandlingBarrier = commandQueue.getGpgpuCommandStreamReceiver().isStallingCommandsOnNextFlushRequired();
@@ -306,7 +310,7 @@ TaskCountType CommandWithoutKernel::dispatchBlitOperation() {
blitProperties.outputTimestampPacket = currentTimestampPacketNodes->peekNodes()[0];
if (commandQueue.getContext().getRootDeviceIndices().size() > 1) {
eventsRequest.fillCsrDependenciesForTaskCountContainer(blitProperties.csrDependencies, *bcsCsr);
eventsRequest.fillCsrDependenciesForRootDevices(blitProperties.csrDependencies, *bcsCsr);
}
const auto newTaskCount = bcsCsr->flushBcsTask(kernelOperation->blitPropertiesContainer, false, commandQueue.isProfilingEnabled(), commandQueue.getDevice());
@@ -388,7 +392,7 @@ CompletionStamp &CommandWithoutKernel::submit(TaskCountType taskLevel, bool term
false); // hasRelaxedOrderingDependencies
if (commandQueue.getContext().getRootDeviceIndices().size() > 1) {
eventsRequest.fillCsrDependenciesForTaskCountContainer(dispatchFlags.csrDependencies, commandStreamReceiver);
eventsRequest.fillCsrDependenciesForRootDevices(dispatchFlags.csrDependencies, commandStreamReceiver);
}
const bool isHandlingBarrier = commandQueue.getGpgpuCommandStreamReceiver().isStallingCommandsOnNextFlushRequired();

3
opencl/source/helpers/task_information.h
View File

@@ -131,7 +131,7 @@ class CommandComputeKernel : public Command {
public:
CommandComputeKernel(CommandQueue &commandQueue, std::unique_ptr<KernelOperation> &kernelOperation, std::vector<Surface *> surfaces,
bool flushDC, bool usesSLM, uint32_t commandType, std::unique_ptr<PrintfHandler> &&printfHandler,
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount);
PreemptionMode preemptionMode, Kernel *kernel, uint32_t kernelCount, TagNodeBase *multiRootDeviceSyncNode);
~CommandComputeKernel() override;
@@ -150,6 +150,7 @@ class CommandComputeKernel : public Command {
Kernel *kernel;
uint32_t kernelCount;
PreemptionMode preemptionMode;
TagNodeBase *multiRootDeviceSyncNode;
};
class CommandWithoutKernel : public Command {