intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 09:09:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: Experimental support of immediate cmd list in-order execution [6/n]

Browse Source 
Related-To: LOCI-4332

- Signal appendWaitOnEvents API call
- Signal appendBarrier call
- Handle sync allocation residency

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2023-05-10 17:13:49 +00:00
committed by Compute-Runtime-Automation
parent 5c988e8a76
commit 41478c5972
18 changed files with 136 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
level_zero/core/source/cmdlist/cmdlist.h
View File

@@ -139,7 +139,7 @@ struct CommandList : _ze_command_list_handle_t {
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingDispatch) = 0;
virtual ze_result_t appendMemoryPrefetch(const void *ptr, size_t count) = 0;
virtual ze_result_t appendSignalEvent(ze_event_handle_t hEvent) = 0;
virtual ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies) = 0;
virtual ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies, bool signalInOrderCompletion) = 0;
virtual ze_result_t appendWriteGlobalTimestamp(uint64_t *dstptr, ze_event_handle_t hSignalEvent,
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents) = 0;
virtual ze_result_t appendMemoryCopyFromContext(void *dstptr, ze_context_handle_t hContextSrc,

2
level_zero/core/source/cmdlist/cmdlist_hw.h
View File

@@ -167,7 +167,7 @@ struct CommandListCoreFamily : CommandListImp {
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents) override;
ze_result_t appendSignalEvent(ze_event_handle_t hEvent) override;
ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies) override;
ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies, bool signalInOrderCompletion) override;
void appendWaitOnInOrderDependency(NEO::GraphicsAllocation *dependencyCounterAllocation, uint32_t waitValue, bool relaxedOrderingAllowed);
ze_result_t appendWriteGlobalTimestamp(uint64_t *dstptr, ze_event_handle_t hSignalEvent,
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents) override;

13
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -2033,7 +2033,7 @@ inline ze_result_t CommandListCoreFamily<gfxCoreFamily>::addEventsToCmdList(uint
if (numWaitEvents > 0) {
if (phWaitEvents) {
CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numWaitEvents, phWaitEvents, relaxedOrderingAllowed, trackDependencies);
CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numWaitEvents, phWaitEvents, relaxedOrderingAllowed, trackDependencies, false);
} else {
return ZE_RESULT_ERROR_INVALID_ARGUMENT;
}
@@ -2095,9 +2095,11 @@ void CommandListCoreFamily<gfxCoreFamily>::appendWaitOnInOrderDependency(NEO::Gr
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies) {
ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies, bool signalInOrderCompletion) {
using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;
signalInOrderCompletion &= this->inOrderExecutionEnabled;
NEO::Device *neoDevice = device->getNEODevice();
uint32_t callId = 0;
if (NEO::DebugManager.flags.EnableSWTags.get()) {
@@ -2166,6 +2168,11 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(uint32_t nu
commandContainer.addToResidencyContainer(this->csr->getTagAllocation());
}
if (signalInOrderCompletion) {
NEO::EncodeStoreMemory<GfxFamily>::programStoreDataImm(*commandContainer.getCommandStream(), this->inOrderDependencyCounterAllocation->getGpuAddress(),
this->inOrderDependencyCounter + 1, 0, false, false);
}
makeResidentDummyAllocation();
if (NEO::DebugManager.flags.EnableSWTags.get()) {
@@ -2266,7 +2273,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWriteGlobalTimestamp(
if (numWaitEvents > 0) {
if (phWaitEvents) {
CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numWaitEvents, phWaitEvents, false, true);
CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numWaitEvents, phWaitEvents, false, true, false);
} else {
return ZE_RESULT_ERROR_INVALID_ARGUMENT;
}

2
level_zero/core/source/cmdlist/cmdlist_hw_immediate.h
View File

@@ -92,7 +92,7 @@ struct CommandListCoreFamilyImmediate : public CommandListCoreFamily<gfxCoreFami
NEO::GraphicsAllocation *srcAllocation,
size_t size, bool flushHost) override;
ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies) override;
ze_result_t appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phEvent, bool relaxedOrderingAllowed, bool trackDependencies, bool signalInOrderCompletion) override;
ze_result_t appendWriteGlobalTimestamp(uint64_t *dstptr, ze_event_handle_t hSignalEvent,
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents) override;

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

@@ -543,7 +543,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendPageFaultCopy(N
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingAllowed, bool trackDependencies) {
ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendWaitOnEvents(uint32_t numEvents, ze_event_handle_t *phWaitEvents, bool relaxedOrderingAllowed, bool trackDependencies, bool signalInOrderCompletion) {
bool allSignaled = true;
for (auto i = 0u; i < numEvents; i++) {
allSignaled &= (!this->dcFlushSupport && Event::fromHandle(phWaitEvents[i])->isAlreadyCompleted());
@@ -555,7 +555,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendWaitOnEvents(ui
checkAvailableSpace(numEvents, false);
checkWaitEventsState(numEvents, phWaitEvents);
}
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numEvents, phWaitEvents, relaxedOrderingAllowed, trackDependencies);
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(numEvents, phWaitEvents, relaxedOrderingAllowed, trackDependencies, signalInOrderCompletion);
this->dependenciesPresent = true;
return flushImmediate(ret, true, true, false, nullptr);
}
@@ -691,6 +691,8 @@ template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::flushImmediate(ze_result_t inputRet, bool performMigration, bool hasStallingCmds,
bool hasRelaxedOrderingDependencies, ze_event_handle_t hSignalEvent) {
if (inputRet == ZE_RESULT_SUCCESS) {
this->commandContainer.addToResidencyContainer(this->inOrderDependencyCounterAllocation);
if (this->isFlushTaskSubmissionEnabled) {
inputRet = executeCommandListImmediateWithFlushTask(performMigration, hasStallingCmds, hasRelaxedOrderingDependencies);
} else {

12
level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
View File

@@ -385,7 +385,17 @@ void CommandListCoreFamily<gfxCoreFamily>::appendComputeBarrierCommand() {
appendMultiTileBarrier(*neoDevice);
} else {
NEO::PipeControlArgs args = createBarrierFlags();
NEO::MemorySynchronizationCommands<GfxFamily>::addSingleBarrier(*commandContainer.getCommandStream(), args);
NEO::PostSyncMode postSyncMode = NEO::PostSyncMode::NoWrite;
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);
}
}