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: 922e14c46d]

projects/clr/rocclr/runtime/device/gpu/gpublit.cpp

@@ -938,6 +938,8 @@ KernelBlitManager::copyBufferToImage(
     static const bool CopyRect = false;
     // Flush DMA for ASYNC copy
     static const bool FlushDMA = true;
+    size_t imgRowPitch = size[0] * gpuMem(dstMemory).elementSize();
+    size_t imgSlicePitch = imgRowPitch * size[1];
 
     if (setup_.disableCopyBufferToImage_) {
         result = DmaBlitManager::copyBufferToImage(
@@ -948,7 +950,9 @@ KernelBlitManager::copyBufferToImage(
     }
     // Check if buffer is in system memory with direct access
     else if (gpuMem(srcMemory).isHostMemDirectAccess() &&
-             (rowPitch == 0) && (slicePitch == 0)) {
+             (((rowPitch == 0) && (slicePitch == 0)) ||
+              ((rowPitch == imgRowPitch) &&
+               ((slicePitch == 0) || (slicePitch == imgSlicePitch))))) {
         // First attempt to do this all with DMA,
         // but there are restriciton with older hardware
         if (dev().settings().imageDMA_) {
@@ -1330,6 +1334,8 @@ KernelBlitManager::copyImageToBuffer(
     static const bool CopyRect = false;
     // Flush DMA for ASYNC copy
     static const bool FlushDMA = true;
+    size_t imgRowPitch = size[0] * gpuMem(srcMemory).elementSize();
+    size_t imgSlicePitch = imgRowPitch * size[1];
 
     if (setup_.disableCopyImageToBuffer_) {
         result = HostBlitManager::copyImageToBuffer(
@@ -1340,7 +1346,9 @@ KernelBlitManager::copyImageToBuffer(
     }
     // Check if buffer is in system memory with direct access
     else if (gpuMem(dstMemory).isHostMemDirectAccess() &&
-             (rowPitch == 0) && (slicePitch == 0)) {
+             (((rowPitch == 0) && (slicePitch == 0)) ||
+              ((rowPitch == imgRowPitch) &&
+                ((slicePitch == 0) || (slicePitch == imgSlicePitch))))) {
         // First attempt to do this all with DMA,
         // but there are restriciton with older hardware
         if (dev().settings().imageDMA_) {

projects/clr/rocclr/runtime/platform/memory.cpp

@@ -1226,17 +1226,21 @@ Image::validateRegion(const Coord3D& origin, const Coord3D& region) const
 }
 
 bool
-Image::isSliceValid(
-    const size_t&   rowPitch,
-    const size_t&   slice,
-    const size_t&   height) const
+Image::isRowSliceValid(
+    size_t  rowPitch,
+    size_t  slice,
+    size_t  width,
+    size_t  height) const
 {
     size_t  tmpHeight =
         (getType() == CL_MEM_OBJECT_IMAGE1D_ARRAY) ? 1 : height;
 
+    bool valid = (rowPitch == 0) || ((rowPitch != 0) &&
+             (rowPitch >= width * getImageFormat().getElementSize()));
+
     return ((slice == 0) ||
             ((slice != 0) &&
-             (slice >= rowPitch * tmpHeight))) ? true : false;
+             (slice >= rowPitch * tmpHeight))) ? valid : false;
 }
 
 void

projects/clr/rocclr/runtime/platform/memory.hpp

@@ -577,10 +577,11 @@ public:
         ) const;
 
     //! Returns true if the slice value for the image is valid
-    bool isSliceValid(
-        const size_t& rowPitch,     //!< The row pitch value
-        const size_t& slicePitch,   //!< The slice pitch value
-        const size_t& height        //!< The height of the copy region
+    bool isRowSliceValid(
+        size_t  rowPitch,       //!< The row pitch value
+        size_t  slicePitch,     //!< The slice pitch value
+        size_t  width,          //!< The width of the copy region
+        size_t  height          //!< The height of the copy region
         ) const;
 
     //! Creates a view memory object