P4 to Git Change 1539091 by vsytchen@vsytchen-ocl-win10 on 2018/04/10 14:57:40

SWDEV-150166 - [OCL][LC][Ellesmere][Windows] oclcompiler.OCLStructOperations[30] causing crash

	1. Set the default amount of command buffer allocated per queue to 4 for pre-gfx9 and 12 for gfx9+ pal devices.
	2. Limit the amount of command buffers allocated for device queue to 2.

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/14594/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#50 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#89 edit

rocclr/runtime/device/pal/palsettings.cpp

@@ -142,8 +142,7 @@ Settings::Settings() {
   subAllocationMaxSize_ =
     std::min(static_cast<uint64_t>(GPU_MAX_SUBALLOC_SIZE) * Ki, subAllocationChunkSize_);
 
-  // Note: More command buffers may cause a HW hang with HWSC on VI family in OCLPerfKernelArguments
-  maxCmdBuffers_ = 8;
+  maxCmdBuffers_ = 12;
 }
 
 bool Settings::create(const Pal::DeviceProperties& palProp,
@@ -293,6 +292,12 @@ bool Settings::create(const Pal::DeviceProperties& palProp,
         maxAllocSize_ = 3ULL * Gi;
       }
 
+      // Note: More than 4 command buffers may cause a HW hang
+      // with HWSC on pre-gfx9 devices in OCLPerfKernelArguments
+      if (!aiPlus_) {
+        maxCmdBuffers_ = 4;
+      }
+
       supportRA_ = false;
       partialDispatch_ = GPU_PARTIAL_DISPATCH;
       numMemDependencies_ = GPU_NUM_MEM_DEPENDENCY;

rocclr/runtime/device/pal/palvirtual.cpp

@@ -784,7 +784,7 @@ bool VirtualGPU::create(bool profiling, uint deviceQueueSize, uint rtCUs,
   uint idx = index() % (dev().numComputeEngines() - firstQueue);
   uint64_t residency_limit = dev().properties().gpuMemoryProperties.flags.supportPerSubmitMemRefs ? 0 :
     (dev().properties().gpuMemoryProperties.maxLocalMemSize >> 2);
-  uint max_cmd_buffers = dev().settings().maxCmdBuffers_;
+  uint max_cmd_buffers = (idx == 0) ? 2 : dev().settings().maxCmdBuffers_;
 
   if (dev().numComputeEngines()) {
     //! @todo There is a hang with a mix of user and non user queues.