Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-15 13:01:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

Optimize gpgpu csr's mutex lock in the enqueue blit

Browse Source 
Signed-off-by: Jobczyk, Lukasz <lukasz.jobczyk@intel.com>
This commit is contained in:
Jobczyk, Lukasz
2022-04-20 09:13:26 +00:00
committed by Compute-Runtime-Automation
parent 7f08a9d993
commit 93fe8a4794
1 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
opencl/source/command_queue/enqueue_common.h
View File

@ -1087,6 +1087,7 @@ template <typename GfxFamily>
template <uint32_t cmdType>
cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDispatchInfo, cl_uint numEventsInWaitList, const cl_event *eventWaitList, cl_event *event, bool blocking, CommandStreamReceiver &bcsCsr) {
auto bcsCommandStreamReceiverOwnership = bcsCsr.obtainUniqueOwnership();
std::unique_lock<NEO::CommandStreamReceiver::MutexType> commandStreamReceiverOwnership;
EventsRequest eventsRequest(numEventsInWaitList, eventWaitList, event);
EventBuilder eventBuilder;
@ -1096,7 +1097,6 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
std::unique_ptr<KernelOperation> blockedCommandsData;
TakeOwnershipWrapper<CommandQueueHw<GfxFamily>> queueOwnership(*this);
auto commandStreamReceiverOwnership = getGpgpuCommandStreamReceiver().obtainUniqueOwnership();
auto blockQueue = false;
auto taskLevel = 0u;
@ -1145,6 +1145,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
LinearStream *gpgpuCommandStream = {};
size_t gpgpuCommandStreamStart = {};
if (gpgpuSubmission) {
commandStreamReceiverOwnership = getGpgpuCommandStreamReceiver().obtainUniqueOwnership();
gpgpuCommandStream = obtainCommandStream<cmdType>(csrDeps, true, blockQueue, multiDispatchInfo, eventsRequest, blockedCommandsData, nullptr, 0, false);
gpgpuCommandStreamStart = gpgpuCommandStream->getUsed();
}
@ -1156,6 +1157,10 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
completionStamp = enqueueCommandWithoutKernel(nullptr, 0, gpgpuCommandStream, gpgpuCommandStreamStart, blocking,
enqueueProperties, timestampPacketDependencies, eventsRequest,
eventBuilder, taskLevel, csrDeps, &bcsCsr);
if (gpgpuSubmission) {
commandStreamReceiverOwnership.unlock();
}
if (eventBuilder.getEvent()) {
eventBuilder.getEvent()->flushStamp->replaceStampObject(this->flushStamp->getStampReference());
}
@ -1168,11 +1173,14 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
if (blockQueue) {
enqueueBlocked(cmdType, nullptr, 0, multiDispatchInfo, timestampPacketDependencies, blockedCommandsData, enqueueProperties, eventsRequest, eventBuilder, nullptr, &bcsCsr);
if (gpgpuSubmission) {
commandStreamReceiverOwnership.unlock();
}
}
timestampPacketDependencies.moveNodesToNewContainer(*deferredTimestampPackets);
commandStreamReceiverOwnership.unlock();
queueOwnership.unlock();
bcsCommandStreamReceiverOwnership.unlock();