intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-04 15:53:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: move process barrier with post sync to a function

Browse Source 
Related-To: NEO-7824
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
This commit is contained in:
Kamil Kopryk
2024-02-23 04:57:45 +00:00
committed by Compute-Runtime-Automation
parent 01a721df3e
commit 10d610d163
2 changed files with 66 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
shared/source/command_stream/command_stream_receiver_hw.h
View File

@@ -241,6 +241,14 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
bool areMultipleSubDevicesInContext,
bool setGeneralStateBaseAddress);
inline void processBarrierWithPostSync(LinearStream &commandStreamTask,
DispatchFlags &dispatchFlags,
bool &levelClosed,
void *&currentPipeControlForNooping,
void *&epiloguePipeControlLocation,
bool &hasStallingCmdsOnTaskStream,
PipeControlArgs &args);
inline CompletionStamp updateTaskCountAndGetCompletionStamp(bool levelClosed);
inline void programSamplerCacheFlushBetweenRedescribedSurfaceReads(LinearStream &commandStreamCSR);
bool bcsRelaxedOrderingAllowed(const BlitPropertiesContainer &blitPropertiesContainer, bool hasStallingCmds) const;

108
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -366,8 +366,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
using MI_BATCH_BUFFER_END = typename GfxFamily::MI_BATCH_BUFFER_END;
using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
auto &rootDeviceEnvironment = this->peekRootDeviceEnvironment();
DEBUG_BREAK_IF(&commandStreamTask == &commandStream);
DEBUG_BREAK_IF(!(dispatchFlags.preemptionMode == PreemptionMode::Disabled ? device.getPreemptionMode() == PreemptionMode::Disabled : true));
DEBUG_BREAK_IF(taskLevel >= CompletionStamp::notReady);
@@ -393,55 +391,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
bool hasStallingCmdsOnTaskStream = false;
if (dispatchFlags.blocking || dispatchFlags.dcFlush || dispatchFlags.guardCommandBufferWithPipeControl || this->heapStorageRequiresRecyclingTag) {
if (this->dispatchMode == DispatchMode::immediateDispatch) {
// for ImmediateDispatch we will send this right away, therefore this pipe control will close the level
// for BatchedSubmissions it will be nooped and only last ppc in batch will be emitted.
levelClosed = true;
// if we guard with ppc, flush dc as well to speed up completion latency
if (dispatchFlags.guardCommandBufferWithPipeControl || this->heapStorageRequiresRecyclingTag || dispatchFlags.blocking) {
dispatchFlags.dcFlush = this->dcFlushSupport;
}
}
this->heapStorageRequiresRecyclingTag = false;
epiloguePipeControlLocation = ptrOffset(commandStreamTask.getCpuBase(), commandStreamTask.getUsed());
if ((dispatchFlags.outOfOrderExecutionAllowed || timestampPacketWriteEnabled) &&
!dispatchFlags.dcFlush) {
currentPipeControlForNooping = epiloguePipeControlLocation;
}
hasStallingCmdsOnTaskStream = true;
auto address = getTagAllocation()->getGpuAddress();
args.dcFlushEnable = getDcFlushRequired(dispatchFlags.dcFlush);
args.notifyEnable = isUsedNotifyEnableForPostSync();
args.tlbInvalidation |= dispatchFlags.memoryMigrationRequired;
args.textureCacheInvalidationEnable |= dispatchFlags.textureCacheFlush;
args.workloadPartitionOffset = isMultiTileOperationEnabled();
args.stateCacheInvalidationEnable = dispatchFlags.stateCacheInvalidation;
MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(
commandStreamTask,
PostSyncMode::immediateData,
address,
taskCount + 1,
rootDeviceEnvironment,
args);
DBG_LOG(LogTaskCounts, __FUNCTION__, "Line: ", __LINE__, "taskCount", peekTaskCount());
if (debugManager.flags.AddPatchInfoCommentsForAUBDump.get()) {
flatBatchBufferHelper->setPatchInfoData(PatchInfoData(address, 0u,
PatchInfoAllocationType::tagAddress,
commandStreamTask.getGraphicsAllocation()->getGpuAddress(),
commandStreamTask.getUsed() - 2 * sizeof(uint64_t),
PatchInfoAllocationType::defaultType));
flatBatchBufferHelper->setPatchInfoData(PatchInfoData(address, 0u,
PatchInfoAllocationType::tagValue,
commandStreamTask.getGraphicsAllocation()->getGpuAddress(),
commandStreamTask.getUsed() - sizeof(uint64_t),
PatchInfoAllocationType::defaultType));
}
processBarrierWithPostSync(commandStreamTask, dispatchFlags, levelClosed, currentPipeControlForNooping,
epiloguePipeControlLocation, hasStallingCmdsOnTaskStream, args);
}
this->latestSentTaskCount = taskCount + 1;
@@ -1829,6 +1781,62 @@ inline void CommandStreamReceiverHw<GfxFamily>::programStateBaseAddressCommon(
}
}
template <typename GfxFamily>
inline void CommandStreamReceiverHw<GfxFamily>::processBarrierWithPostSync(LinearStream &commandStreamTask, DispatchFlags &dispatchFlags, bool &levelClosed, void *&currentPipeControlForNooping, void *&epiloguePipeControlLocation, bool &hasStallingCmdsOnTaskStream, PipeControlArgs &args) {
if (this->dispatchMode == DispatchMode::immediateDispatch) {
// for ImmediateDispatch we will send this right away, therefore this pipe control will close the level
// for BatchedSubmissions it will be nooped and only last ppc in batch will be emitted.
levelClosed = true;
// if we guard with ppc, flush dc as well to speed up completion latency
if (dispatchFlags.guardCommandBufferWithPipeControl || this->heapStorageRequiresRecyclingTag || dispatchFlags.blocking) {
dispatchFlags.dcFlush = this->dcFlushSupport;
}
}
this->heapStorageRequiresRecyclingTag = false;
epiloguePipeControlLocation = ptrOffset(commandStreamTask.getCpuBase(), commandStreamTask.getUsed());
if ((dispatchFlags.outOfOrderExecutionAllowed || timestampPacketWriteEnabled) &&
!dispatchFlags.dcFlush) {
currentPipeControlForNooping = epiloguePipeControlLocation;
}
hasStallingCmdsOnTaskStream = true;
auto address = getTagAllocation()->getGpuAddress();
auto &rootDeviceEnvironment = this->peekRootDeviceEnvironment();
args.dcFlushEnable = getDcFlushRequired(dispatchFlags.dcFlush);
args.notifyEnable = isUsedNotifyEnableForPostSync();
args.tlbInvalidation |= dispatchFlags.memoryMigrationRequired;
args.textureCacheInvalidationEnable |= dispatchFlags.textureCacheFlush;
args.workloadPartitionOffset = isMultiTileOperationEnabled();
args.stateCacheInvalidationEnable = dispatchFlags.stateCacheInvalidation;
MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(
commandStreamTask,
PostSyncMode::immediateData,
address,
taskCount + 1,
rootDeviceEnvironment,
args);
DBG_LOG(LogTaskCounts, __FUNCTION__, "Line: ", __LINE__, "taskCount", peekTaskCount());
if (debugManager.flags.AddPatchInfoCommentsForAUBDump.get()) {
flatBatchBufferHelper->setPatchInfoData(PatchInfoData(address, 0u,
PatchInfoAllocationType::tagAddress,
commandStreamTask.getGraphicsAllocation()->getGpuAddress(),
commandStreamTask.getUsed() - 2 * sizeof(uint64_t),
PatchInfoAllocationType::defaultType));
flatBatchBufferHelper->setPatchInfoData(PatchInfoData(address, 0u,
PatchInfoAllocationType::tagValue,
commandStreamTask.getGraphicsAllocation()->getGpuAddress(),
commandStreamTask.getUsed() - sizeof(uint64_t),
PatchInfoAllocationType::defaultType));
}
}
template <typename GfxFamily>
inline CompletionStamp CommandStreamReceiverHw<GfxFamily>::updateTaskCountAndGetCompletionStamp(bool levelClosed) {