P4 to Git Change 1973670 by gandryey@gera-win10 on 2019/07/25 11:27:11

SWDEV-79445 - OCL generic changes and code clean-up
	- Don't create an extra queue for DMA transfers when SDMA is disabled. That should allow to avoid useless sync operations.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#13 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#75 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paltimestamp.cpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paltimestamp.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#143 edit


[ROCm/clr commit: b9b798616f]

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

@@ -1332,7 +1332,12 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
     }
   }
 
-  gpu.engineID_ = SdmaEngine;
+  if (dev().settings().disableSdma_) {
+    // Make sure compute is done before CP DMA start
+    gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+  } else {
+    gpu.engineID_ = SdmaEngine;
+  }
 
   // Wait for the resources, since runtime may use async transfers
   wait(gpu, waitOnBusyEngine);
@@ -1423,6 +1428,11 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
     }
   }
 
+  if (dev().settings().disableSdma_) {
+    // Make sure CP dma is done
+    gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+  }
+
   gpu.eventEnd(gpu.engineID_, event);
 
   // Mark source and destination as busy