intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-06 02:18:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: program in-order barrier as non stalling instruction

Browse Source 
Related-To: NEO-7966

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2023-06-29 11:23:19 +00:00
committed by Compute-Runtime-Automation
parent c5d675570a
commit f9a58b5544
4 changed files with 133 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -2921,7 +2921,9 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendBarrier(ze_event_handle_
appendEventForProfiling(signalEvent, true);
if (isCopyOnly()) {
if (this->inOrderExecutionEnabled) {
appendSignalInOrderDependencyCounter();
} else if (isCopyOnly()) {
NEO::MiFlushArgs args{this->dummyBlitWa};
uint64_t gpuAddress = 0u;
TaskCountType value = 0u;

17
level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
View File

@@ -385,6 +385,23 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendBarrier(
ze_event_handle_t *phWaitEvents) {
ze_result_t ret = ZE_RESULT_SUCCESS;
if (isInOrderExecutionEnabled() && numWaitEvents == 0) {
auto signalEvent = Event::fromHandle(hSignalEvent);
bool earlyReturn = true;
if (signalEvent) {
if (signalEvent->isEventTimestampFlagSet()) {
earlyReturn = false;
} else {
signalEvent->enableInOrderExecMode(*this->inOrderDependencyCounterAllocation, this->inOrderDependencyCounter);
}
}
if (earlyReturn) {
return ZE_RESULT_SUCCESS;
}
}
if (this->isFlushTaskSubmissionEnabled) {
checkAvailableSpace(numWaitEvents, false);
checkWaitEventsState(numWaitEvents, phWaitEvents);

6
level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
View File

@@ -395,12 +395,6 @@ void CommandListCoreFamily<gfxCoreFamily>::appendComputeBarrierCommand() {
uint64_t gpuWriteAddress = 0;
uint64_t writeValue = 0;
if (this->inOrderExecutionEnabled) {
postSyncMode = NEO::PostSyncMode::ImmediateData;
gpuWriteAddress = this->inOrderDependencyCounterAllocation->getGpuAddress();
writeValue = this->inOrderDependencyCounter + 1;
}
NEO::MemorySynchronizationCommands<GfxFamily>::addSingleBarrier(*commandContainer.getCommandStream(), postSyncMode, gpuWriteAddress, writeValue, args);
}
}