P4 to Git Change 1537141 by gandryey@gera-w8 on 2018/04/05 11:56:20

SWDEV-149190 - [CQE OCL][DTB-BLOCKER][WIN][LNX][QR]observed soft hang while running oclruntime sub-test of ocltst due to faulty CL#1529531
	- Pal::PerfExperiment can't works with extra offsets for allocations. Thus disable suballocation for the thread trace.
	- Minor clean-ups: make sure the profiling end is called for threadtrace commands; call buf release on mem copy in the program loader; add comments to the scratch buffer alignment.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#80 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#60 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palthreadtrace.cpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#85 edit


[ROCm/clr commit: 1d972ed59a]

projects/clr/rocclr/runtime/device/pal/paldevice.cpp

@@ -1814,7 +1814,8 @@ bool Device::allocScratch(uint regNum, const VirtualGPU* vgpu) {
               scratchBuf->regNum_ * numMaxWaves * sizeof(uint32_t);
           scratchBuf->size_ = std::min(scratchBuf->size_, info().maxMemAllocSize_);
           scratchBuf->size_ = std::min(scratchBuf->size_, uint64_t(3 * Gi));
-          scratchBuf->size_ = amd::alignUp(scratchBuf->size_, 0xFFFF);
+          // Note: Generic address space setup in HW requires 64KB alignment for scratch
+          scratchBuf->size_ = amd::alignUp(scratchBuf->size_, 64 * Ki);
           scratchBuf->offset_ = offset;
           size += scratchBuf->size_;
           offset += scratchBuf->size_;

projects/clr/rocclr/runtime/device/pal/palprogram.cpp

@@ -98,6 +98,7 @@ void Segment::copy(size_t offset, const void* src, size_t size) {
       srcOffs += tmpSize;
       tmpSize = std::min(static_cast<size_t>(xferBuf.size()), size);
     }
+    gpuAccess_->dev().xferWrite().release(gpu, xferBuf);
     gpu.waitAllEngines();
   }
 }

projects/clr/rocclr/runtime/device/pal/palthreadtrace.cpp

@@ -58,8 +58,11 @@ bool PalThreadTraceReference::finalize() {
   // Acquire GPU memory for the query from the pool and bind it.
   Pal::GpuMemoryRequirements gpuMemReqs = {};
   iPerf()->GetGpuMemoryRequirements(&gpuMemReqs);
-
-  memory_ = new Memory(gpu().dev(), amd::alignUp(gpuMemReqs.size, gpuMemReqs.alignment));
+  // Note: Pal:::PerfExperiments can't apply the runtime offsets for suballocaiotns,
+  // thus disable suballocation.
+  Pal::gpusize bufSize = std::max(gpuMemReqs.size,
+    gpu_.dev().settings().subAllocationMaxSize_ + gpuMemReqs.alignment);
+  memory_ = new Memory(gpu().dev(), amd::alignUp(bufSize, gpuMemReqs.alignment));
 
   if (nullptr == memory_) {
     return false;

projects/clr/rocclr/runtime/device/pal/palvirtual.cpp

@@ -2427,6 +2427,8 @@ void VirtualGPU::submitThreadTraceMemObjects(amd::ThreadTraceMemObjectsCommand&
       LogError("Unsupported command type for ThreadTraceMemObjects!");
       break;
   }
+
+  profilingEnd(cmd);
 }
 
 void VirtualGPU::submitThreadTrace(amd::ThreadTraceCommand& cmd) {
@@ -2469,6 +2471,8 @@ void VirtualGPU::submitThreadTrace(amd::ThreadTraceCommand& cmd) {
       LogError("Unsupported command type for ThreadTrace!");
       break;
   }
+
+  profilingEnd(cmd);
 }
 
 void VirtualGPU::submitAcquireExtObjects(amd::AcquireExtObjectsCommand& vcmd) {