intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 14:55:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: stall RelaxedOrdering scheduler when programming Semaphores

Browse Source 
Related-To: NEO-7458

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2023-06-27 15:48:18 +00:00
committed by Compute-Runtime-Automation
parent 87eb609958
commit 2383896dbb
3 changed files with 140 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
level_zero/core/source/cmdlist/cmdlist_hw_immediate.h
View File

@@ -182,6 +182,7 @@ struct CommandListCoreFamilyImmediate : public CommandListCoreFamily<gfxCoreFami
void printKernelsPrintfOutput(bool hangDetected);
ze_result_t synchronizeInOrderExecution(uint64_t timeout) const;
bool hasStallingCmdsForRelaxedOrdering(uint32_t numWaitEvents, bool relaxedOrderingDispatch);
MOCKABLE_VIRTUAL void checkAssert();
ComputeFlushMethodType computeFlushMethod = nullptr;

36
level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
View File

@@ -331,6 +331,11 @@ bool CommandListCoreFamilyImmediate<gfxCoreFamily>::waitForEventsFromHost() {
return true;
}
template <GFXCORE_FAMILY gfxCoreFamily>
bool CommandListCoreFamilyImmediate<gfxCoreFamily>::hasStallingCmdsForRelaxedOrdering(uint32_t numWaitEvents, bool relaxedOrderingDispatch) {
return (!relaxedOrderingDispatch && (numWaitEvents > 0 || this->inOrderDependencyCounter > 0));
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendLaunchKernel(
ze_kernel_handle_t kernelHandle, const ze_group_count_t *threadGroupDimensions,
@@ -354,7 +359,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendLaunchKernel(
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernel(kernelHandle, threadGroupDimensions,
hSignalEvent, numWaitEvents, phWaitEvents,
launchParams, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -370,7 +376,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendLaunchKernelInd
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelIndirect(kernelHandle, pDispatchArgumentsBuffer,
hSignalEvent, numWaitEvents, phWaitEvents, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -405,6 +412,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryCopy(
checkWaitEventsState(numWaitEvents, phWaitEvents);
}
bool hasStallindCmds = hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch);
ze_result_t ret;
CpuMemCopyInfo cpuMemCopyInfo(dstptr, srcptr, size);
this->device->getDriverHandle()->findAllocationDataForRange(const_cast<void *>(srcptr), size, &cpuMemCopyInfo.srcAllocData);
@@ -420,6 +429,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryCopy(
auto isSplitNeeded = this->isAppendSplitNeeded(dstptr, srcptr, size, direction);
if (isSplitNeeded) {
relaxedOrderingDispatch = isRelaxedOrderingDispatchAllowed(1); // split generates more than 1 event
hasStallindCmds = !relaxedOrderingDispatch;
ret = static_cast<DeviceImp *>(this->device)->bcsSplit.appendSplitCall<gfxCoreFamily, void *, const void *>(this, dstptr, srcptr, size, hSignalEvent, numWaitEvents, phWaitEvents, true, relaxedOrderingDispatch, direction, [&](void *dstptrParam, const void *srcptrParam, size_t sizeParam, ze_event_handle_t hSignalEventParam) {
return CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(dstptrParam, srcptrParam, sizeParam, hSignalEventParam, 0u, nullptr, relaxedOrderingDispatch, true);
});
@@ -427,7 +438,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryCopy(
ret = CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(dstptr, srcptr, size, hSignalEvent,
numWaitEvents, phWaitEvents, relaxedOrderingDispatch, forceDisableCopyOnlyInOrderSignaling);
}
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallindCmds, relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -450,12 +462,16 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryCopyRegio
checkWaitEventsState(numWaitEvents, phWaitEvents);
}
bool hasStallindCmds = hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch);
ze_result_t ret;
NEO::TransferDirection direction;
auto isSplitNeeded = this->isAppendSplitNeeded(dstPtr, srcPtr, this->getTotalSizeForCopyRegion(dstRegion, dstPitch, dstSlicePitch), direction);
if (isSplitNeeded) {
relaxedOrderingDispatch = isRelaxedOrderingDispatchAllowed(1); // split generates more than 1 event
hasStallindCmds = !relaxedOrderingDispatch;
ret = static_cast<DeviceImp *>(this->device)->bcsSplit.appendSplitCall<gfxCoreFamily, uint32_t, uint32_t>(this, dstRegion->originX, srcRegion->originX, dstRegion->width, hSignalEvent, numWaitEvents, phWaitEvents, true, relaxedOrderingDispatch, direction, [&](uint32_t dstOriginXParam, uint32_t srcOriginXParam, size_t sizeParam, ze_event_handle_t hSignalEventParam) {
ze_copy_region_t dstRegionLocal = {};
ze_copy_region_t srcRegionLocal = {};
@@ -475,7 +491,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryCopyRegio
hSignalEvent, numWaitEvents, phWaitEvents, relaxedOrderingDispatch, forceDisableCopyOnlyInOrderSignaling);
}
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallindCmds, relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -493,7 +509,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendMemoryFill(void
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(ptr, pattern, patternSize, size, hSignalEvent, numWaitEvents, phWaitEvents, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -620,7 +636,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendImageCopyRegion
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendImageCopyRegion(hDstImage, hSrcImage, pDstRegion, pSrcRegion, hSignalEvent,
numWaitEvents, phWaitEvents, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -641,7 +658,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendImageCopyFromMe
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendImageCopyFromMemory(hDstImage, srcPtr, pDstRegion, hSignalEvent,
numWaitEvents, phWaitEvents, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -662,7 +679,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendImageCopyToMemo
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendImageCopyToMemory(dstPtr, hSrcImage, pSrcRegion, hSignalEvent,
numWaitEvents, phWaitEvents, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>
@@ -694,7 +711,8 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::appendLaunchCooperati
}
auto ret = CommandListCoreFamily<gfxCoreFamily>::appendLaunchCooperativeKernel(kernelHandle, launchKernelArgs, hSignalEvent, numWaitEvents, waitEventHandles, relaxedOrderingDispatch);
return flushImmediate(ret, true, false, relaxedOrderingDispatch, hSignalEvent);
return flushImmediate(ret, true, hasStallingCmdsForRelaxedOrdering(numWaitEvents, relaxedOrderingDispatch), relaxedOrderingDispatch, hSignalEvent);
}
template <GFXCORE_FAMILY gfxCoreFamily>