intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-30 01:35:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

RelaxedOrdering: Optimize return ptr programming

Browse Source 
Related-To: NEO-7458

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2022-11-28 16:57:36 +00:00
committed by Compute-Runtime-Automation
parent 50e43069ba
commit 2c08f2ca8c
8 changed files with 57 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
shared/source/command_stream/command_stream_receiver_hw.h
View File

@@ -52,7 +52,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
size_t getCmdsSizeForHardwareContext() const override;
static void addBatchBufferEnd(LinearStream &commandStream, void **patchLocation);
void programEndingCmd(LinearStream &commandStream, void **patchLocation, bool directSubmissionEnabled, bool sipWaAllowed);
void programEndingCmd(LinearStream &commandStream, void **patchLocation, bool directSubmissionEnabled, bool hasRelaxedOrderingDependencies, bool sipWaAllowed);
void addBatchBufferStart(MI_BATCH_BUFFER_START *commandBufferMemory, uint64_t startAddress, bool secondary);
size_t getRequiredStateBaseAddressSize(const Device &device) const;

15
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -101,7 +101,8 @@ inline void CommandStreamReceiverHw<GfxFamily>::addBatchBufferEnd(LinearStream &
}
template <typename GfxFamily>
inline void CommandStreamReceiverHw<GfxFamily>::programEndingCmd(LinearStream &commandStream, void **patchLocation, bool directSubmissionEnabled, bool sipWaAllowed) {
inline void CommandStreamReceiverHw<GfxFamily>::programEndingCmd(LinearStream &commandStream, void **patchLocation, bool directSubmissionEnabled,
bool hasRelaxedOrderingDependencies, bool sipWaAllowed) {
if (directSubmissionEnabled) {
uint64_t startAddress = commandStream.getGraphicsAllocation()->getGpuAddress() + commandStream.getUsed();
if (DebugManager.flags.BatchBufferStartPrepatchingWaEnabled.get() == 0) {
@@ -109,7 +110,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::programEndingCmd(LinearStream &c
}
bool indirect = false;
if (DebugManager.flags.DirectSubmissionRelaxedOrdering.get() == 1) {
if (DebugManager.flags.DirectSubmissionRelaxedOrdering.get() == 1 && hasRelaxedOrderingDependencies) {
NEO::EncodeSetMMIO<GfxFamily>::encodeREG(commandStream, CS_GPR_R0, CS_GPR_R3);
NEO::EncodeSetMMIO<GfxFamily>::encodeREG(commandStream, CS_GPR_R0 + 4, CS_GPR_R3 + 4);
@@ -571,7 +572,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
GraphicsAllocation *chainedBatchBuffer = nullptr;
bool directSubmissionEnabled = isDirectSubmissionEnabled();
if (submitTask) {
programEndingCmd(commandStreamTask, &bbEndLocation, directSubmissionEnabled, true);
programEndingCmd(commandStreamTask, &bbEndLocation, directSubmissionEnabled, dispatchFlags.hasRelaxedOrderingDependencies, true);
EncodeNoop<GfxFamily>::emitNoop(commandStreamTask, bbEndPaddingSize);
EncodeNoop<GfxFamily>::alignToCacheLine(commandStreamTask);
@@ -602,7 +603,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
this->programEpilogue(commandStreamCSR, device, &bbEndLocation, dispatchFlags);
} else if (submitCSR) {
programEndingCmd(commandStreamCSR, &bbEndLocation, directSubmissionEnabled, true);
programEndingCmd(commandStreamCSR, &bbEndLocation, directSubmissionEnabled, dispatchFlags.hasRelaxedOrderingDependencies, true);
EncodeNoop<GfxFamily>::emitNoop(commandStreamCSR, bbEndPaddingSize);
EncodeNoop<GfxFamily>::alignToCacheLine(commandStreamCSR);
DEBUG_BREAK_IF(commandStreamCSR.getUsed() > commandStreamCSR.getMaxAvailableSpace());
@@ -1167,7 +1168,7 @@ TaskCountType CommandStreamReceiverHw<GfxFamily>::flushBcsTask(const BlitPropert
}
void *endingCmdPtr = nullptr;
programEndingCmd(commandStream, &endingCmdPtr, blitterDirectSubmission, false);
programEndingCmd(commandStream, &endingCmdPtr, blitterDirectSubmission, false, false);
EncodeNoop<GfxFamily>::alignToCacheLine(commandStream);
@@ -1276,7 +1277,7 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushSmallTask(LinearStream
using MI_BATCH_BUFFER_END = typename GfxFamily::MI_BATCH_BUFFER_END;
void *endingCmdPtr = nullptr;
programEndingCmd(commandStreamTask, &endingCmdPtr, isAnyDirectSubmissionEnabled(), false);
programEndingCmd(commandStreamTask, &endingCmdPtr, isAnyDirectSubmissionEnabled(), false, false);
auto bytesToPad = EncodeBatchBufferStartOrEnd<GfxFamily>::getBatchBufferStartSize() -
EncodeBatchBufferStartOrEnd<GfxFamily>::getBatchBufferEndSize();
@@ -1357,7 +1358,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::programEpilogue(LinearStream &cs
addBatchBufferStart(reinterpret_cast<typename GfxFamily::MI_BATCH_BUFFER_START *>(*batchBufferEndLocation), gpuAddress, false);
this->programEpliogueCommands(csr, dispatchFlags);
programEndingCmd(csr, batchBufferEndLocation, isDirectSubmissionEnabled(), !EngineHelpers::isBcs(osContext->getEngineType()));
programEndingCmd(csr, batchBufferEndLocation, isDirectSubmissionEnabled(), false, !EngineHelpers::isBcs(osContext->getEngineType()));
EncodeNoop<GfxFamily>::alignToCacheLine(csr);
}
}