Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

refactor: add skip event residency flag to compacted event 

Related-To: NEO-10064

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz 2024-02-23 10:25:06 +00:00 committed by Compute-Runtime-Automation
parent a5989eb31b
commit 6011eb147f
9 changed files with 81 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@ -182,7 +182,7 @@ struct CommandListCoreFamily : public CommandListImp {
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingDispatch) override; uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingDispatch) override;
void appendMultiPartitionPrologue(uint32_t partitionDataSize) override; void appendMultiPartitionPrologue(uint32_t partitionDataSize) override;
void appendMultiPartitionEpilogue() override; void appendMultiPartitionEpilogue() override;
void appendEventForProfilingAllWalkers(Event *event, bool beforeWalker, bool singlePacketEvent); void appendEventForProfilingAllWalkers(Event *event, void **syncCmdBuffer, bool beforeWalker, bool singlePacketEvent, bool skipAddingEventToResidency);
ze_result_t addEventsToCmdList(uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingAllowed, bool trackDependencies, bool waitForImplicitInOrderDependency); ze_result_t addEventsToCmdList(uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingAllowed, bool trackDependencies, bool waitForImplicitInOrderDependency);
ze_result_t reserveSpace(size_t size, void **ptr) override; ze_result_t reserveSpace(size_t size, void **ptr) override;
@ -289,7 +289,7 @@ struct CommandListCoreFamily : public CommandListImp {
void adjustWriteKernelTimestamp(uint64_t globalAddress, uint64_t contextAddress, bool maskLsb, uint32_t mask, bool workloadPartition); void adjustWriteKernelTimestamp(uint64_t globalAddress, uint64_t contextAddress, bool maskLsb, uint32_t mask, bool workloadPartition);
void appendEventForProfiling(Event *event, bool beforeWalker, bool skipBarrierForEndProfiling); void appendEventForProfiling(Event *event, bool beforeWalker, bool skipBarrierForEndProfiling);
void appendEventForProfilingCopyCommand(Event *event, bool beforeWalker); void appendEventForProfilingCopyCommand(Event *event, bool beforeWalker);
void appendSignalEventPostWalker(Event *event, bool skipBarrierForEndProfiling); void appendSignalEventPostWalker(Event *event, void **syncCmdBuffer, bool skipBarrierForEndProfiling, bool skipAddingEventToResidency);
virtual void programStateBaseAddress(NEO::CommandContainer &container, bool useSbaProperties); virtual void programStateBaseAddress(NEO::CommandContainer &container, bool useSbaProperties);
void appendComputeBarrierCommand(); void appendComputeBarrierCommand();
NEO::PipeControlArgs createBarrierFlags(); NEO::PipeControlArgs createBarrierFlags();
@ -320,9 +320,9 @@ struct CommandListCoreFamily : public CommandListImp {
CmdListEventOperation estimateEventPostSync(Event *event, uint32_t operations); CmdListEventOperation estimateEventPostSync(Event *event, uint32_t operations);
void dispatchPostSyncCopy(uint64_t gpuAddress, uint32_t value, bool workloadPartition); void dispatchPostSyncCopy(uint64_t gpuAddress, uint32_t value, bool workloadPartition);
void dispatchPostSyncCompute(uint64_t gpuAddress, uint32_t value, bool workloadPartition); void dispatchPostSyncCompute(uint64_t gpuAddress, uint32_t value, bool workloadPartition);
void dispatchPostSyncCommands(const CmdListEventOperation &eventOperations, uint64_t gpuAddress, uint32_t value, bool useLastPipeControl, bool signalScope, bool skipPartitionOffsetProgramming); void dispatchPostSyncCommands(const CmdListEventOperation &eventOperations, uint64_t gpuAddress, void **syncCmdBuffer, uint32_t value, bool useLastPipeControl, bool signalScope, bool skipPartitionOffsetProgramming);
void dispatchEventRemainingPacketsPostSyncOperation(Event *event); void dispatchEventRemainingPacketsPostSyncOperation(Event *event);
void dispatchEventPostSyncOperation(Event *event, uint32_t value, bool omitFirstOperation, bool useMax, bool useLastPipeControl, bool skipPartitionOffsetProgramming); void dispatchEventPostSyncOperation(Event *event, void **syncCmdBuffer, uint32_t value, bool omitFirstOperation, bool useMax, bool useLastPipeControl, bool skipPartitionOffsetProgramming);
bool isKernelUncachedMocsRequired(bool kernelState) { bool isKernelUncachedMocsRequired(bool kernelState) {
this->containsStatelessUncachedResource |= kernelState; this->containsStatelessUncachedResource |= kernelState;
if (this->stateBaseAddressTracking) { if (this->stateBaseAddressTracking) {

48
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@ -465,7 +465,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelIndirect(ze_
ret = appendLaunchKernelWithParams(Kernel::fromHandle(kernelHandle), pDispatchArgumentsBuffer, ret = appendLaunchKernelWithParams(Kernel::fromHandle(kernelHandle), pDispatchArgumentsBuffer,
nullptr, launchParams); nullptr, launchParams);
addToMappedEventList(event); addToMappedEventList(event);
appendSignalEventPostWalker(event, false); appendSignalEventPostWalker(event, nullptr, false, false);
handleInOrderDependencyCounter(event, isInOrderNonWalkerSignalingRequired(event)); handleInOrderDependencyCounter(event, isInOrderNonWalkerSignalingRequired(event));
@ -516,7 +516,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchMultipleKernelsInd
} }
} }
addToMappedEventList(event); addToMappedEventList(event);
appendSignalEventPostWalker(event, false); appendSignalEventPostWalker(event, nullptr, false, false);
return ret; return ret;
} }
@ -554,7 +554,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendEventReset(ze_event_hand
bool useMaxPackets = event->isEventTimestampFlagSet() || (event->getPacketsInUse() < this->partitionCount); bool useMaxPackets = event->isEventTimestampFlagSet() || (event->getPacketsInUse() < this->partitionCount);
bool appendPipeControlWithPostSync = (!isCopyOnly()) && (event->isSignalScope() || event->isEventTimestampFlagSet()); bool appendPipeControlWithPostSync = (!isCopyOnly()) && (event->isSignalScope() || event->isEventTimestampFlagSet());
dispatchEventPostSyncOperation(event, Event::STATE_CLEARED, false, useMaxPackets, appendPipeControlWithPostSync, false); dispatchEventPostSyncOperation(event, nullptr, Event::STATE_CLEARED, false, useMaxPackets, appendPipeControlWithPostSync, false);
if (!isCopyOnly()) { if (!isCopyOnly()) {
if (this->partitionCount > 1) { if (this->partitionCount > 1) {
@ -602,7 +602,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryRangesBarrier(uint
appendEventForProfiling(signalEvent, true, false); appendEventForProfiling(signalEvent, true, false);
applyMemoryRangesBarrier(numRanges, pRangeSizes, pRanges); applyMemoryRangesBarrier(numRanges, pRangeSizes, pRanges);
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
addToMappedEventList(signalEvent); addToMappedEventList(signalEvent);
if (this->isInOrderExecutionEnabled()) { if (this->isInOrderExecutionEnabled()) {
@ -1272,7 +1272,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyBlitRegion(Ali
} }
makeResidentDummyAllocation(); makeResidentDummyAllocation();
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
return ZE_RESULT_SUCCESS; return ZE_RESULT_SUCCESS;
} }
@ -1301,7 +1301,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendCopyImageBlit(NEO::Graph
NEO::BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForImageRegion(blitProperties, *commandContainer.getCommandStream(), dummyBlitWa); NEO::BlitCommandsHelper<GfxFamily>::dispatchBlitCommandsForImageRegion(blitProperties, *commandContainer.getCommandStream(), dummyBlitWa);
makeResidentDummyAllocation(); makeResidentDummyAllocation();
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
return ZE_RESULT_SUCCESS; return ZE_RESULT_SUCCESS;
} }
@ -1459,7 +1459,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
launchParams.pipeControlSignalling = (signalEvent && singlePipeControlPacket) || dstAllocationStruct.needsFlush; launchParams.pipeControlSignalling = (signalEvent && singlePipeControlPacket) || dstAllocationStruct.needsFlush;
appendEventForProfilingAllWalkers(signalEvent, true, singlePipeControlPacket); appendEventForProfilingAllWalkers(signalEvent, nullptr, true, singlePipeControlPacket, false);
if (isCopyOnly()) { if (isCopyOnly()) {
ret = appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr, ret = appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr,
@ -1524,7 +1524,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
} }
} }
appendEventForProfilingAllWalkers(signalEvent, false, singlePipeControlPacket); appendEventForProfilingAllWalkers(signalEvent, nullptr, false, singlePipeControlPacket, false);
addFlushRequiredCommand(dstAllocationStruct.needsFlush, signalEvent); addFlushRequiredCommand(dstAllocationStruct.needsFlush, signalEvent);
addToMappedEventList(signalEvent); addToMappedEventList(signalEvent);
@ -1922,7 +1922,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(void *ptr,
launchParams.isKernelSplitOperation = (fillArguments.leftRemainingBytes > 0 || fillArguments.rightRemainingBytes > 0); launchParams.isKernelSplitOperation = (fillArguments.leftRemainingBytes > 0 || fillArguments.rightRemainingBytes > 0);
bool singlePipeControlPacket = eventSignalPipeControl(launchParams.isKernelSplitOperation, dcFlush); bool singlePipeControlPacket = eventSignalPipeControl(launchParams.isKernelSplitOperation, dcFlush);
appendEventForProfilingAllWalkers(signalEvent, true, singlePipeControlPacket); appendEventForProfilingAllWalkers(signalEvent, nullptr, true, singlePipeControlPacket, false);
if (fillArguments.leftRemainingBytes > 0) { if (fillArguments.leftRemainingBytes > 0) {
launchParams.numKernelsInSplitLaunch++; launchParams.numKernelsInSplitLaunch++;
@ -2067,7 +2067,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(void *ptr,
} }
addToMappedEventList(signalEvent); addToMappedEventList(signalEvent);
appendEventForProfilingAllWalkers(signalEvent, false, singlePipeControlPacket); appendEventForProfilingAllWalkers(signalEvent, nullptr, false, singlePipeControlPacket, false);
addFlushRequiredCommand(hostPointerNeedsFlush, signalEvent); addFlushRequiredCommand(hostPointerNeedsFlush, signalEvent);
bool nonWalkerInOrderCmdChaining = false; bool nonWalkerInOrderCmdChaining = false;
@ -2144,7 +2144,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendBlitFill(void *ptr,
this->dummyBlitWa); this->dummyBlitWa);
makeResidentDummyAllocation(); makeResidentDummyAllocation();
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
if (isInOrderExecutionEnabled()) { if (isInOrderExecutionEnabled()) {
appendSignalInOrderDependencyCounter(signalEvent); appendSignalInOrderDependencyCounter(signalEvent);
@ -2155,7 +2155,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendBlitFill(void *ptr,
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::appendSignalEventPostWalker(Event *event, bool skipBarrierForEndProfiling) { void CommandListCoreFamily<gfxCoreFamily>::appendSignalEventPostWalker(Event *event, void **syncCmdBuffer, bool skipBarrierForEndProfiling, bool skipAddingEventToResidency) {
if (event == nullptr || !event->getPoolAllocation(this->device)) { if (event == nullptr || !event->getPoolAllocation(this->device)) {
return; return;
} }
@ -2163,10 +2163,12 @@ void CommandListCoreFamily<gfxCoreFamily>::appendSignalEventPostWalker(Event *ev
appendEventForProfiling(event, false, skipBarrierForEndProfiling); appendEventForProfiling(event, false, skipBarrierForEndProfiling);
} else { } else {
event->resetKernelCountAndPacketUsedCount(); event->resetKernelCountAndPacketUsedCount();
if (!skipAddingEventToResidency) {
commandContainer.addToResidencyContainer(event->getPoolAllocation(this->device)); commandContainer.addToResidencyContainer(event->getPoolAllocation(this->device));
}
event->setPacketsInUse(this->partitionCount); event->setPacketsInUse(this->partitionCount);
dispatchEventPostSyncOperation(event, Event::STATE_SIGNALED, false, false, !isCopyOnly(), false); dispatchEventPostSyncOperation(event, syncCmdBuffer, Event::STATE_SIGNALED, false, false, !isCopyOnly(), false);
} }
} }
@ -2183,7 +2185,7 @@ void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingCopyCommand(Ev
NEO::MiFlushArgs args{this->dummyBlitWa}; NEO::MiFlushArgs args{this->dummyBlitWa};
NEO::EncodeMiFlushDW<GfxFamily>::programWithWa(*commandContainer.getCommandStream(), 0, 0, args); NEO::EncodeMiFlushDW<GfxFamily>::programWithWa(*commandContainer.getCommandStream(), 0, 0, args);
makeResidentDummyAllocation(); makeResidentDummyAllocation();
dispatchEventPostSyncOperation(event, Event::STATE_SIGNALED, true, false, false, false); dispatchEventPostSyncOperation(event, nullptr, Event::STATE_SIGNALED, true, false, false, false);
} }
appendWriteKernelTimestamp(event, beforeWalker, false, false); appendWriteKernelTimestamp(event, beforeWalker, false, false);
} }
@ -2376,7 +2378,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(ze_event_han
event->setPacketsInUse(this->partitionCount); event->setPacketsInUse(this->partitionCount);
bool appendPipeControlWithPostSync = (!isCopyOnly()) && (event->isSignalScope() || event->isEventTimestampFlagSet()); bool appendPipeControlWithPostSync = (!isCopyOnly()) && (event->isSignalScope() || event->isEventTimestampFlagSet());
dispatchEventPostSyncOperation(event, Event::STATE_SIGNALED, false, false, appendPipeControlWithPostSync, false); dispatchEventPostSyncOperation(event, nullptr, Event::STATE_SIGNALED, false, false, appendPipeControlWithPostSync, false);
if (this->isInOrderExecutionEnabled()) { if (this->isInOrderExecutionEnabled()) {
appendSignalInOrderDependencyCounter(event); appendSignalInOrderDependencyCounter(event);
@ -2643,7 +2645,7 @@ void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfiling(Event *event,
bool workloadPartition = setupTimestampEventForMultiTile(event); bool workloadPartition = setupTimestampEventForMultiTile(event);
appendWriteKernelTimestamp(event, beforeWalker, true, workloadPartition); appendWriteKernelTimestamp(event, beforeWalker, true, workloadPartition);
} else { } else {
dispatchEventPostSyncOperation(event, Event::STATE_SIGNALED, true, false, false, true); dispatchEventPostSyncOperation(event, nullptr, Event::STATE_SIGNALED, true, false, false, true);
const auto &rootDeviceEnvironment = this->device->getNEODevice()->getRootDeviceEnvironment(); const auto &rootDeviceEnvironment = this->device->getNEODevice()->getRootDeviceEnvironment();
@ -2713,7 +2715,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWriteGlobalTimestamp(
args); args);
} }
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
if (this->isInOrderExecutionEnabled()) { if (this->isInOrderExecutionEnabled()) {
appendSignalInOrderDependencyCounter(signalEvent); appendSignalInOrderDependencyCounter(signalEvent);
@ -3260,7 +3262,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendBarrier(ze_event_handle_
} }
addToMappedEventList(signalEvent); addToMappedEventList(signalEvent);
appendSignalEventPostWalker(signalEvent, this->isInOrderExecutionEnabled()); appendSignalEventPostWalker(signalEvent, nullptr, this->isInOrderExecutionEnabled(), false);
if (isInOrderExecutionEnabled()) { if (isInOrderExecutionEnabled()) {
appendSignalInOrderDependencyCounter(signalEvent); appendSignalInOrderDependencyCounter(signalEvent);
@ -3425,7 +3427,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWaitOnMemory(void *desc,
NEO::MemorySynchronizationCommands<GfxFamily>::addAdditionalSynchronization(*commandContainer.getCommandStream(), gpuAddress, true, rootDeviceEnvironment); NEO::MemorySynchronizationCommands<GfxFamily>::addAdditionalSynchronization(*commandContainer.getCommandStream(), gpuAddress, true, rootDeviceEnvironment);
} }
appendSignalEventPostWalker(signalEvent, false); appendSignalEventPostWalker(signalEvent, nullptr, false, false);
if (this->isInOrderExecutionEnabled()) { if (this->isInOrderExecutionEnabled()) {
appendSignalInOrderDependencyCounter(signalEvent); appendSignalInOrderDependencyCounter(signalEvent);
@ -3553,7 +3555,7 @@ void CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCompute(uint64_t gpuA
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCommands(const CmdListEventOperation &eventOperations, uint64_t gpuAddress, uint32_t value, bool useLastPipeControl, bool signalScope, bool skipPartitionOffsetProgramming) { void CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCommands(const CmdListEventOperation &eventOperations, uint64_t gpuAddress, void **syncCmdBuffer, uint32_t value, bool useLastPipeControl, bool signalScope, bool skipPartitionOffsetProgramming) {
decltype(&CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCompute) dispatchFunction = &CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCompute; decltype(&CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCompute) dispatchFunction = &CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCompute;
if (isCopyOnly()) { if (isCopyOnly()) {
dispatchFunction = &CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCopy; dispatchFunction = &CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCopy;
@ -3603,7 +3605,7 @@ void CommandListCoreFamily<gfxCoreFamily>::dispatchPostSyncCommands(const CmdLis
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::dispatchEventPostSyncOperation(Event *event, uint32_t value, bool omitFirstOperation, bool useMax, bool useLastPipeControl, bool skipPartitionOffsetProgramming) { void CommandListCoreFamily<gfxCoreFamily>::dispatchEventPostSyncOperation(Event *event, void **syncCmdBuffer, uint32_t value, bool omitFirstOperation, bool useMax, bool useLastPipeControl, bool skipPartitionOffsetProgramming) {
uint32_t packets = event->getPacketsInUse(); uint32_t packets = event->getPacketsInUse();
if (this->signalAllEventPackets || useMax) { if (this->signalAllEventPackets || useMax) {
packets = event->getMaxPacketsCount(); packets = event->getMaxPacketsCount();
@ -3616,7 +3618,7 @@ void CommandListCoreFamily<gfxCoreFamily>::dispatchEventPostSyncOperation(Event
eventPostSync.operationCount--; eventPostSync.operationCount--;
} }
dispatchPostSyncCommands(eventPostSync, gpuAddress, value, useLastPipeControl, event->isSignalScope(), skipPartitionOffsetProgramming); dispatchPostSyncCommands(eventPostSync, gpuAddress, syncCmdBuffer, value, useLastPipeControl, event->isSignalScope(), skipPartitionOffsetProgramming);
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
@ -3629,7 +3631,7 @@ void CommandListCoreFamily<gfxCoreFamily>::dispatchEventRemainingPacketsPostSync
eventAddress += event->getSinglePacketSize() * event->getPacketsInUse(); eventAddress += event->getSinglePacketSize() * event->getPacketsInUse();
constexpr bool appendLastPipeControl = false; constexpr bool appendLastPipeControl = false;
dispatchPostSyncCommands(remainingPacketsOperation, eventAddress, Event::STATE_SIGNALED, appendLastPipeControl, event->isSignalScope(), false); dispatchPostSyncCommands(remainingPacketsOperation, eventAddress, nullptr, Event::STATE_SIGNALED, appendLastPipeControl, event->isSignalScope(), false);
} }
} }

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

@ -247,7 +247,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
*reinterpret_cast<typename GfxFamily::RENDER_SURFACE_STATE *>(surfaceStateSpace) = surfaceState; *reinterpret_cast<typename GfxFamily::RENDER_SURFACE_STATE *>(surfaceStateSpace) = surfaceState;
} }
appendSignalEventPostWalker(event, false); appendSignalEventPostWalker(event, nullptr, false, false);
commandContainer.addToResidencyContainer(kernelImmutableData->getIsaGraphicsAllocation()); commandContainer.addToResidencyContainer(kernelImmutableData->getIsaGraphicsAllocation());
auto &residencyContainer = kernel->getResidencyContainer(); auto &residencyContainer = kernel->getResidencyContainer();
@ -329,11 +329,11 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelSplit(Kernel
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingAllWalkers(Event *event, bool beforeWalker, bool singlePacketEvent) { void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingAllWalkers(Event *event, void **syncCmdBuffer, bool beforeWalker, bool singlePacketEvent, bool skipAddingEventToResidency) {
if (beforeWalker) { if (beforeWalker) {
appendEventForProfiling(event, true, false); appendEventForProfiling(event, true, false);
} else { } else {
appendSignalEventPostWalker(event, false); appendSignalEventPostWalker(event, syncCmdBuffer, false, skipAddingEventToResidency);
} }
} }

10
level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
View File

@ -286,7 +286,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
std::list<void *> additionalCommands; std::list<void *> additionalCommands;
if (compactEvent) { if (compactEvent) {
appendEventForProfilingAllWalkers(compactEvent, true, true); appendEventForProfilingAllWalkers(compactEvent, nullptr, true, true, false);
} }
bool inOrderExecSignalRequired = (this->isInOrderExecutionEnabled() && !launchParams.isKernelSplitOperation && !launchParams.pipeControlSignalling); bool inOrderExecSignalRequired = (this->isInOrderExecutionEnabled() && !launchParams.isKernelSplitOperation && !launchParams.pipeControlSignalling);
@ -297,7 +297,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
if (inOrderExecSignalRequired) { if (inOrderExecSignalRequired) {
if (inOrderNonWalkerSignalling) { if (inOrderNonWalkerSignalling) {
dispatchEventPostSyncOperation(eventForInOrderExec, Event::STATE_CLEARED, false, false, false, false); dispatchEventPostSyncOperation(eventForInOrderExec, nullptr, Event::STATE_CLEARED, false, false, false, false);
} else { } else {
inOrderCounterValue = this->inOrderExecInfo->getCounterValue() + getInOrderIncrementValue(); inOrderCounterValue = this->inOrderExecInfo->getCounterValue() + getInOrderIncrementValue();
inOrderExecInfo = this->inOrderExecInfo.get(); inOrderExecInfo = this->inOrderExecInfo.get();
@ -348,7 +348,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
} }
if (compactEvent) { if (compactEvent) {
appendEventForProfilingAllWalkers(compactEvent, false, true); appendEventForProfilingAllWalkers(compactEvent, nullptr, false, true, launchParams.omitAddingEventResidency);
} else if (event) { } else if (event) {
event->setPacketsInUse(partitionCount); event->setPacketsInUse(partitionCount);
if (l3FlushEnable) { if (l3FlushEnable) {
@ -506,12 +506,12 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelSplit(Kernel
} }
template <GFXCORE_FAMILY gfxCoreFamily> template <GFXCORE_FAMILY gfxCoreFamily>
void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingAllWalkers(Event *event, bool beforeWalker, bool singlePacketEvent) { void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfilingAllWalkers(Event *event, void **syncCmdBuffer, bool beforeWalker, bool singlePacketEvent, bool skipAddingEventToResidency) {
if (isCopyOnly() || singlePacketEvent) { if (isCopyOnly() || singlePacketEvent) {
if (beforeWalker) { if (beforeWalker) {
appendEventForProfiling(event, true, false); appendEventForProfiling(event, true, false);
} else { } else {
appendSignalEventPostWalker(event, false); appendSignalEventPostWalker(event, syncCmdBuffer, false, skipAddingEventToResidency);
} }
} else { } else {
if (event) { if (event) {

6
level_zero/core/source/device/bcs_split.h
View File

@ -109,7 +109,7 @@ struct BcsSplit {
cmdList->addEventsToCmdList(numWaitEvents, phWaitEvents, hasRelaxedOrderingDependencies, false, true); cmdList->addEventsToCmdList(numWaitEvents, phWaitEvents, hasRelaxedOrderingDependencies, false, true);
if (signalEvent && i == 0u) { if (signalEvent && i == 0u) {
cmdList->appendEventForProfilingAllWalkers(signalEvent, true, true); cmdList->appendEventForProfilingAllWalkers(signalEvent, nullptr, true, true, false);
} }
auto localSize = totalSize / engineCount; auto localSize = totalSize / engineCount;
@ -137,9 +137,9 @@ struct BcsSplit {
cmdList->addEventsToCmdList(static_cast<uint32_t>(cmdQsForSplit.size()), eventHandles.data(), hasRelaxedOrderingDependencies, false, true); cmdList->addEventsToCmdList(static_cast<uint32_t>(cmdQsForSplit.size()), eventHandles.data(), hasRelaxedOrderingDependencies, false, true);
if (signalEvent) { if (signalEvent) {
cmdList->appendEventForProfilingAllWalkers(signalEvent, false, true); cmdList->appendEventForProfilingAllWalkers(signalEvent, nullptr, false, true, false);
} }
cmdList->appendEventForProfilingAllWalkers(this->events.marker[markerEventIndex], false, true); cmdList->appendEventForProfilingAllWalkers(this->events.marker[markerEventIndex], nullptr, false, true, false);
if (cmdList->isInOrderExecutionEnabled()) { if (cmdList->isInOrderExecutionEnabled()) {
cmdList->appendSignalInOrderDependencyCounter(signalEvent); cmdList->appendSignalInOrderDependencyCounter(signalEvent);

2
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_1.cpp
View File

@ -2945,7 +2945,7 @@ HWTEST2_F(CommandListCreate, givenNullEventWhenAppendEventAfterWalkerThenNothing
auto usedBefore = commandList->getCmdContainer().getCommandStream()->getUsed(); auto usedBefore = commandList->getCmdContainer().getCommandStream()->getUsed();
commandList->appendSignalEventPostWalker(nullptr, false); commandList->appendSignalEventPostWalker(nullptr, nullptr, false, false);
EXPECT_EQ(commandList->getCmdContainer().getCommandStream()->getUsed(), usedBefore); EXPECT_EQ(commandList->getCmdContainer().getCommandStream()->getUsed(), usedBefore);
} }

6
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_signal_event.cpp
View File

@ -412,7 +412,7 @@ HWTEST2_F(CommandListAppendUsedPacketSignalEvent,
event->signalScope = ZE_EVENT_SCOPE_FLAG_HOST; event->signalScope = ZE_EVENT_SCOPE_FLAG_HOST;
commandList->partitionCount = packets; commandList->partitionCount = packets;
commandList->appendSignalEventPostWalker(event.get(), false); commandList->appendSignalEventPostWalker(event.get(), nullptr, false, false);
EXPECT_EQ(packets, event->getPacketsInUse()); EXPECT_EQ(packets, event->getPacketsInUse());
auto gpuAddress = event->getCompletionFieldGpuAddress(device); auto gpuAddress = event->getCompletionFieldGpuAddress(device);
@ -472,7 +472,7 @@ HWTEST2_F(CommandListAppendUsedPacketSignalEvent,
event->signalScope = ZE_EVENT_SCOPE_FLAG_HOST; event->signalScope = ZE_EVENT_SCOPE_FLAG_HOST;
commandList->partitionCount = packets; commandList->partitionCount = packets;
commandList->appendSignalEventPostWalker(event.get(), false); commandList->appendSignalEventPostWalker(event.get(), nullptr, false, false);
EXPECT_EQ(packets, event->getPacketsInUse()); EXPECT_EQ(packets, event->getPacketsInUse());
auto gpuAddress = event->getCompletionFieldGpuAddress(device); auto gpuAddress = event->getCompletionFieldGpuAddress(device);
@ -645,7 +645,7 @@ HWTEST2_F(CommandListAppendUsedPacketSignalEvent,
event->setEventTimestampFlag(false); event->setEventTimestampFlag(false);
commandList->appendSignalEventPostWalker(event.get(), false); commandList->appendSignalEventPostWalker(event.get(), nullptr, false, false);
size_t usedAfterSize = cmdStream->getUsed(); size_t usedAfterSize = cmdStream->getUsed();
GenCmdList cmdList; GenCmdList cmdList;

36
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_xehp_and_later.cpp
View File

@ -950,7 +950,7 @@ struct CommandListSignalAllEventPacketFixture : public ModuleFixture {
commandList->setupTimestampEventForMultiTile(event.get()); commandList->setupTimestampEventForMultiTile(event.get());
size_t sizeBefore = cmdStream->getUsed(); size_t sizeBefore = cmdStream->getUsed();
commandList->appendSignalEventPostWalker(event.get(), false); commandList->appendSignalEventPostWalker(event.get(), nullptr, false, false);
size_t sizeAfter = cmdStream->getUsed(); size_t sizeAfter = cmdStream->getUsed();
EXPECT_EQ(ZE_RESULT_SUCCESS, result); EXPECT_EQ(ZE_RESULT_SUCCESS, result);
@ -2397,5 +2397,39 @@ HWTEST2_F(CommandListCreate, givenPlatformSupportsHdcUntypedCacheFlushWhenAppend
EXPECT_TRUE(timestampPostSyncFound); EXPECT_TRUE(timestampPostSyncFound);
} }
HWTEST2_F(CommandListCreate, givenAppendSignalEventWhenSkipAddToResidencyTrueThenEventAllocationNotAddedToResidency, IsAtLeastXeHpCore) {
auto commandList = std::make_unique<WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>>();
auto result = commandList->initialize(device, NEO::EngineGroupType::compute, 0u);
ASSERT_EQ(ZE_RESULT_SUCCESS, result);
ze_event_pool_desc_t eventPoolDesc = {};
eventPoolDesc.count = 1;
eventPoolDesc.flags = 0;
ze_event_desc_t eventDesc = {};
eventDesc.index = 0;
eventDesc.signal = 0;
auto eventPool = std::unique_ptr<L0::EventPool>(L0::EventPool::create(driverHandle.get(), context, 0, nullptr, &eventPoolDesc, result));
EXPECT_EQ(ZE_RESULT_SUCCESS, result);
auto event = std::unique_ptr<L0::Event>(L0::Event::create<typename FamilyType::TimestampPacketType>(eventPool.get(), &eventDesc, device));
ASSERT_NE(nullptr, event.get());
auto &residencyContainer = commandList->getCmdContainer().getResidencyContainer();
auto eventAllocation = event->getPoolAllocation(device);
bool skipAdd = true;
commandList->appendEventForProfilingAllWalkers(event.get(), nullptr, false, true, skipAdd);
auto eventAllocIt = std::find(residencyContainer.begin(), residencyContainer.end(), eventAllocation);
EXPECT_EQ(residencyContainer.end(), eventAllocIt);
skipAdd = false;
commandList->appendEventForProfilingAllWalkers(event.get(), nullptr, false, true, skipAdd);
eventAllocIt = std::find(residencyContainer.begin(), residencyContainer.end(), eventAllocation);
EXPECT_NE(residencyContainer.end(), eventAllocIt);
}
} // namespace ult } // namespace ult
} // namespace L0 } // namespace L0

2
level_zero/core/test/unit_tests/xe_hpc_core/test_cmdlist_xe_hpc_core.cpp
View File

@ -686,7 +686,7 @@ HWTEST2_F(CommandListEventFenceTestsXeHpcCore, givenCommandListWithRegularEventA
auto eventPool = std::unique_ptr<L0::EventPool>(L0::EventPool::create(driverHandle.get(), context, 0, nullptr, &eventPoolDesc, result)); auto eventPool = std::unique_ptr<L0::EventPool>(L0::EventPool::create(driverHandle.get(), context, 0, nullptr, &eventPoolDesc, result));
EXPECT_EQ(ZE_RESULT_SUCCESS, result); EXPECT_EQ(ZE_RESULT_SUCCESS, result);
auto event = std::unique_ptr<L0::Event>(L0::Event::create<typename FamilyType::TimestampPacketType>(eventPool.get(), &eventDesc, device)); auto event = std::unique_ptr<L0::Event>(L0::Event::create<typename FamilyType::TimestampPacketType>(eventPool.get(), &eventDesc, device));
commandList->appendSignalEventPostWalker(event.get(), false); commandList->appendSignalEventPostWalker(event.get(), nullptr, false, false);
GenCmdList cmdList; GenCmdList cmdList;
ASSERT_TRUE(FamilyType::Parse::parseCommandBuffer( ASSERT_TRUE(FamilyType::Parse::parseCommandBuffer(