P4 to Git Change 1201783 by gandryey@gera-w8 on 2015/10/20 18:03:34

SWDEV-79151 - clenqueuereadImage is slow when using a pinned buffer and a row_picth!0
	- Add a check if the provided rowPitch is equal to the actual transfer width. SDMA doesn't support row/slice pitches, thus runtime still has to fall back to compute in other cases

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_memobj.cpp#78 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#120 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.cpp#122 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.hpp#92 edit


[ROCm/clr commit: 69374d1cd9]

projects/clr/opencl/api/opencl/amdocl/cl_memobj.cpp

@@ -2363,7 +2363,7 @@ RUNTIME_ENTRY(cl_int, clEnqueueReadImage, (
     }
 
     if (!srcImage->validateRegion(srcOrigin, srcRegion) ||
-        !srcImage->isSliceValid(row_pitch, slice_pitch, region[1])) {
+        !srcImage->isRowSliceValid(row_pitch, slice_pitch, region[0], region[1])) {
         return CL_INVALID_VALUE;
     }
 
@@ -2561,7 +2561,7 @@ RUNTIME_ENTRY(cl_int, clEnqueueWriteImage, (
     }
 
     if (!dstImage->validateRegion(dstOrigin, dstRegion) ||
-        !dstImage->isSliceValid(input_row_pitch, input_slice_pitch, region[1])) {
+        !dstImage->isRowSliceValid(input_row_pitch, input_slice_pitch, region[0], region[1])) {
         return CL_INVALID_VALUE;
     }