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

SWDEV-526836 - Switch PAL backend to CmdReleaseThenAcquire() (#175)

Browse Source
This commit is contained in:
Andryeyev, German
2025-04-15 11:49:53 -04:00
committed by GitHub
parent fcaefe97b8
commit f6c804edc0
1 changed files with 18 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/pal/palvirtual.hpp
+18 -20
View File
@@ -491,31 +491,29 @@ class VirtualGPU : public device::VirtualDevice {
void addBarrier(RgpSqqtBarrierReason reason = RgpSqqtBarrierReason::MemDependency,
BarrierType type = BarrierType::KernelToKernel) const {
Pal::BarrierInfo barrier = {};
barrier.pipePointWaitCount = 1;
Pal::HwPipePoint point = Pal::HwPipePostCs;
barrier.pPipePoints = &point;
barrier.transitionCount = 1;
Pal::BarrierTransition trans = {};
trans.srcCacheMask = Pal::CoherShader;
trans.dstCacheMask = Pal::CoherShader;
trans.imageInfo.oldLayout.usages = Pal::LayoutShaderRead;
trans.imageInfo.oldLayout.engines = Pal::LayoutComputeEngine;
trans.imageInfo.newLayout.usages = Pal::LayoutShaderRead;
trans.imageInfo.newLayout.engines = Pal::LayoutComputeEngine;
Pal::AcquireReleaseInfo barrier = {
.srcGlobalStageMask = Pal::PipelineStageCs,
.dstGlobalStageMask = Pal::PipelineStageCs,
.srcGlobalAccessMask = Pal::CoherShader,
.dstGlobalAccessMask = Pal::CoherShader,
.memoryBarrierCount = 0,
.pMemoryBarriers = nullptr,
.imageBarrierCount = 0,
.pImageBarriers = nullptr,
.reason = static_cast<uint32_t>(reason)
};
if (type == BarrierType::KernelToCopy) {
trans.dstCacheMask = Pal::CoherCopy;
barrier.dstGlobalAccessMask = Pal::CoherCopy;
} else if (type == BarrierType::CopyToKernel) {
trans.srcCacheMask = Pal::CoherCopy;
barrier.srcGlobalAccessMask = Pal::CoherCopy;
} else if (type == BarrierType::CopyToCopy) {
trans.dstCacheMask = trans.srcCacheMask = Pal::CoherCopy;
barrier.srcGlobalAccessMask = barrier.dstGlobalAccessMask = Pal::CoherCopy;
} else if (type == BarrierType::FlushL2) {
trans.dstCacheMask = trans.srcCacheMask = Pal::CoherCopy | Pal::CoherCpu;
barrier.srcGlobalAccessMask = barrier.dstGlobalAccessMask = Pal::CoherCopy | Pal::CoherCpu;
}
barrier.pTransitions = &trans;
barrier.waitPoint = Pal::HwPipePreCs;
barrier.reason = static_cast<uint32_t>(reason);
iCmd()->CmdBarrier(barrier);
iCmd()->CmdReleaseThenAcquire(barrier);
queues_[engineID_]->submit<true>(false);
}