intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-11 08:07:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

Improve TimestampPacket residency flow

Browse Source 
- Dont call makeResident in enqueueHandler for blocked path
- Fill csrDeps for blit enqueue only in unblocked path
- Call makeResident on all dependencies during blocked command flush

Change-Id: I6658e4695483bee63eca205f85687ea5f951b099
Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2019-09-04 11:34:23 +02:00
committed by sys_ocldev
parent 45707269c9
commit 6dae106f07
8 changed files with 72 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
runtime/command_queue/enqueue_common.h
View File

@@ -193,7 +193,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
BlitProperties blitProperties;
if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
csrDeps.fillFromEventsRequestAndMakeResident(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
csrDeps.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
size_t nodesCount = 0u;
if (blitEnqueue || isCacheFlushCommand(commandType)) {
@@ -219,7 +219,8 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
bool flushDependenciesForNonKernelCommand = false;
if (blitEnqueue) {
blitProperties = processDispatchForBlitEnqueue(multiDispatchInfo, previousTimestampPacketNodes, eventsRequest, commandStream, commandType);
blitProperties = processDispatchForBlitEnqueue(multiDispatchInfo, previousTimestampPacketNodes, eventsRequest, commandStream,
commandType, blockQueue);
} else if (multiDispatchInfo.empty() == false) {
processDispatchForKernels<commandType>(multiDispatchInfo, printfHandler, eventBuilder.getEvent(),
hwTimeStamps, blockQueue, devQueueHw, csrDeps, blockedCommandsData.get(),
@@ -251,6 +252,8 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
flushDependenciesForNonKernelCommand, &blitProperties);
if (!blockQueue) {
csrDeps.makeResident(getGpgpuCommandStreamReceiver());
if (parentKernel) {
processDeviceEnqueue(devQueueHw, multiDispatchInfo, hwTimeStamps, blocking);
}
@@ -438,20 +441,21 @@ void CommandQueueHw<GfxFamily>::processDispatchForKernels(const MultiDispatchInf
template <typename GfxFamily>
BlitProperties CommandQueueHw<GfxFamily>::processDispatchForBlitEnqueue(const MultiDispatchInfo &multiDispatchInfo,
TimestampPacketContainer &previousTimestampPacketNodes,
const EventsRequest &eventsRequest,
LinearStream &commandStream,
uint32_t commandType) {
const EventsRequest &eventsRequest, LinearStream &commandStream,
uint32_t commandType, bool queueBlocked) {
auto blitDirection = BlitProperties::obtainBlitDirection(commandType);
auto blitCommandStreamReceiver = getBcsCommandStreamReceiver();
auto blitProperties = BlitProperties::constructPropertiesForReadWriteBuffer(blitDirection, *blitCommandStreamReceiver,
multiDispatchInfo.peekBuiltinOpParams(), false);
if (!queueBlocked) {
blitProperties.csrDependencies.fillFromEventsRequest(eventsRequest, *blitCommandStreamReceiver,
CsrDependencies::DependenciesType::All);
blitProperties.csrDependencies.fillFromEventsRequestAndMakeResident(eventsRequest, *blitCommandStreamReceiver,
CsrDependencies::DependenciesType::All);
blitProperties.csrDependencies.push_back(&previousTimestampPacketNodes);
}
blitProperties.csrDependencies.push_back(&previousTimestampPacketNodes);
blitProperties.outputTimestampPacket = timestampPacketContainer.get();
auto currentTimestampPacketNode = timestampPacketContainer->peekNodes().at(0);
@@ -681,7 +685,8 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
dispatchFlags.preemptionMode = PreemptionHelper::taskPreemptionMode(*device, multiDispatchInfo);
dispatchFlags.outOfOrderExecutionAllowed = !eventBuilder.getEvent() || getGpgpuCommandStreamReceiver().isNTo1SubmissionModelEnabled();
if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
dispatchFlags.csrDependencies.fillFromEventsRequestAndMakeResident(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
dispatchFlags.csrDependencies.makeResident(getGpgpuCommandStreamReceiver());
}
dispatchFlags.numGrfRequired = numGrfRequired;
dispatchFlags.specialPipelineSelectMode = specialPipelineSelectMode;
@@ -841,6 +846,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueCommandWithoutKernel(
UNRECOVERABLE_IF(!enqueueProperties.blitProperties);
auto bcsCsr = getBcsCommandStreamReceiver();
enqueueProperties.blitProperties->csrDependencies.makeResident(*bcsCsr);
previousTimestampPacketNodes->makeResident(*bcsCsr);
timestampPacketContainer->makeResident(*bcsCsr);
bcsCsr->blitBuffer(*enqueueProperties.blitProperties);
@@ -855,7 +861,8 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueCommandWithoutKernel(
dispatchFlags.outOfOrderExecutionAllowed = getGpgpuCommandStreamReceiver().isNTo1SubmissionModelEnabled();
if (getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
dispatchFlags.csrDependencies.fillFromEventsRequestAndMakeResident(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
dispatchFlags.csrDependencies.fillFromEventsRequest(eventsRequest, getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OutOfCsr);
dispatchFlags.csrDependencies.makeResident(getGpgpuCommandStreamReceiver());
}
CompletionStamp completionStamp = getGpgpuCommandStreamReceiver().flushTask(
commandStream,