AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-467102 - Hidden heap init for graph capture

Browse Source 
If the graph has kernels that does device side allocation,  during packet capture, heap is
allocated because heap pointer has to be added to the AQL packet, and initialized during
graph launch.

Handle race with wait when 2 kernels with device heap are enqueued on multiple streams.

Change-Id: I45933b77fcaf7bc8fdf1bc906462e32b5d8d3688


[ROCm/clr commit: 57156c524d]
This commit is contained in:
Anusha GodavarthySurya
2024-06-11 11:59:05 +00:00
committed by Anusha Godavarthy Surya
parent c8bc9e3f2e
commit 291f079669
9 changed files with 57 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/device/rocm/rocvirtual.cpp
+7 -2
View File
@@ -2962,6 +2962,8 @@ static inline void nontemporalMemcpy(
#endif
}
void VirtualGPU::HiddenHeapInit() { const_cast<Device&>(dev()).HiddenHeapInit(*this); }
// ================================================================================================
bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
const amd::Kernel& kernel, const_address parameters, void* eventHandle,
@@ -3016,7 +3018,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
amd::Memory* const* memories =
reinterpret_cast<amd::Memory* const*>(parameters + kernelParams.memoryObjOffset());
bool isGraphCapture = vcmd != nullptr && vcmd->getCapturingState();
for (int j = 0; j < iteration; j++) {
// Reset global size for dimension dim if split is needed
if (dim != -1) {
@@ -3143,6 +3145,10 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
const_cast<Device&>(dev()).HiddenHeapAlloc(*this);
}
if (dev().HeapBuffer() != nullptr) {
// Initialize hidden heap buffer
if (!isGraphCapture) {
const_cast<Device&>(dev()).HiddenHeapInit(*this);
}
// Add heap pointer to the code
size_t heap_ptr = static_cast<size_t>(dev().HeapBuffer()->virtualAddress());
WriteAqlArgAt(hidden_arguments, heap_ptr, it.size_, it.offset_);
@@ -3225,7 +3231,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
}
address argBuffer = hidden_arguments;
bool isGraphCapture = vcmd != nullptr && vcmd->getCapturingState();
size_t argSize = std::min(gpuKernel.KernargSegmentByteSize(), signature.paramsSize());
// Find all parameters for the current kernel