intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 20:39:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: submit marker for cpu transfer with event on IOQ queue

Browse Source 
Related-To: NEO-8081

Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
This commit is contained in:
Maciej Plewka
2023-08-04 12:10:12 +00:00
committed by Compute-Runtime-Automation
parent f3b2458a9c
commit 2e249d819e
8 changed files with 197 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
opencl/source/command_queue/cpu_data_transfer_handler.cpp
View File

@@ -27,6 +27,7 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
EventBuilder eventBuilder;
bool eventCompleted = false;
bool mapOperation = transferProperties.cmdType == CL_COMMAND_MAP_BUFFER || transferProperties.cmdType == CL_COMMAND_MAP_IMAGE;
ErrorCodeHelper err(&retVal, CL_SUCCESS);
if (mapOperation) {
@@ -45,8 +46,18 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
}
transferProperties.memObj->removeMappedPtr(unmapInfo.ptr);
}
auto blockQueue = false;
TaskCountType taskLevel = 0u;
TakeOwnershipWrapper<CommandQueue> queueOwnership(*this);
auto commandStreamReceiverOwnership = getGpgpuCommandStreamReceiver().obtainUniqueOwnership();
obtainTaskLevelAndBlockedStatus(taskLevel, eventsRequest.numEventsInWaitList, eventsRequest.eventWaitList, blockQueue, transferProperties.cmdType);
bool isMarkerRequiredForEventSignal = !blockQueue &&
!transferProperties.blocking &&
!transferProperties.finishRequired &&
!isOOQEnabled() &&
eventsRequest.outEvent != nullptr;
if (eventsRequest.outEvent) {
if (eventsRequest.outEvent && !isMarkerRequiredForEventSignal) {
eventBuilder.create<Event>(this, transferProperties.cmdType, CompletionStamp::notReady, CompletionStamp::notReady);
outEventObj = eventBuilder.getEvent();
outEventObj->setQueueTimeStamp();
@@ -54,13 +65,6 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
*eventsRequest.outEvent = outEventObj;
}
TakeOwnershipWrapper<CommandQueue> queueOwnership(*this);
auto commandStreamReceiverOwnership = getGpgpuCommandStreamReceiver().obtainUniqueOwnership();
auto blockQueue = false;
TaskCountType taskLevel = 0u;
obtainTaskLevelAndBlockedStatus(taskLevel, eventsRequest.numEventsInWaitList, eventsRequest.eventWaitList, blockQueue, transferProperties.cmdType);
DBG_LOG(LogTaskCounts, __FUNCTION__, "taskLevel", taskLevel);
if (outEventObj) {
@@ -81,9 +85,10 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
mapOperation ? transferProperties.mapFlags == CL_MAP_READ : unmapInfo.readOnly,
eventBuilder);
}
commandStreamReceiverOwnership.unlock();
queueOwnership.unlock();
if (!isMarkerRequiredForEventSignal) {
commandStreamReceiverOwnership.unlock();
queueOwnership.unlock();
}
// read/write buffers are always blocking
if (!blockQueue || transferProperties.blocking) {
@@ -94,11 +99,9 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
outEventObj->setSubmitTimeStamp();
}
// wait for the completness of previous commands
if (transferProperties.cmdType != CL_COMMAND_UNMAP_MEM_OBJECT) {
if (!transferProperties.memObj->isMemObjZeroCopy() || transferProperties.blocking) {
finish();
eventCompleted = true;
}
if (transferProperties.finishRequired) {
finish();
eventCompleted = true;
}
if (outEventObj) {
@@ -158,6 +161,12 @@ void *CommandQueue::cpuDataTransferHandler(TransferProperties &transferPropertie
} else {
outEventObj->updateExecutionStatus();
}
} else if (isMarkerRequiredForEventSignal) {
enqueueMarkerWithWaitList(0, nullptr, eventsRequest.outEvent);
commandStreamReceiverOwnership.unlock();
queueOwnership.unlock();
outEventObj = castToObject<Event>(*eventsRequest.outEvent);
outEventObj->setCmdType(transferProperties.cmdType);
}
if (modifySimulationFlags) {
auto graphicsAllocation = transferProperties.memObj->getGraphicsAllocation(getDevice().getRootDeviceIndex());