P4 to Git Change 1309866 by gandryey@gera-w8 on 2016/09/01 13:50:12

SWDEV-79445 - OCL generic changes and code clean-up
	- Improve image fill performance with multiple writes in a single thread. The current split has 3 regions

Affected files ...

... //depot/stg/opencl/drivers/opencl/library/common.hsa/src/blitKernels.cl#4 edit
... //depot/stg/opencl/drivers/opencl/library/common/src/blitKernels.cl#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#123 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.hpp#40 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.hpp#4 edit

rocclr/runtime/device/gpu/gpublit.cpp

@@ -2592,23 +2592,25 @@ KernelBlitManager::fillImage(
         }
     }
 
+    // Perform workload split to allow multiple operations in a single thread
+    globalWorkSize[0] = (size[0] + TransferSplitSize - 1) / TransferSplitSize;
     // Find the current blit type
     if (memView->cal()->dimSize_ == 1) {
-        globalWorkSize[0] = amd::alignUp(size[0], 256);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 256);
         globalWorkSize[1] = amd::alignUp(size[1], 1);
         globalWorkSize[2] = amd::alignUp(size[2], 1);
         localWorkSize[0] = 256;
         localWorkSize[1] = localWorkSize[2] = 1;
     }
     else if (memView->cal()->dimSize_ == 2) {
-        globalWorkSize[0] = amd::alignUp(size[0], 16);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 16);
         globalWorkSize[1] = amd::alignUp(size[1], 16);
         globalWorkSize[2] = amd::alignUp(size[2], 1);
         localWorkSize[0] = localWorkSize[1] = 16;
         localWorkSize[2] = 1;
     }
     else {
-        globalWorkSize[0] = amd::alignUp(size[0], 8);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 8);
         globalWorkSize[1] = amd::alignUp(size[1], 8);
         globalWorkSize[2] = amd::alignUp(size[2], 4);
         localWorkSize[0] = localWorkSize[1] = 8;

rocclr/runtime/device/gpu/gpublit.hpp

@@ -383,6 +383,7 @@ public:
 
 private:
     static const size_t MaxXferBuffers = 2;
+    static const uint   TransferSplitSize = 3;
 
     //! Copies a buffer object to an image object
     bool copyBufferToImageKernel(

rocclr/runtime/device/pal/palblit.cpp

@@ -2564,23 +2564,25 @@ KernelBlitManager::fillImage(
         }
     }
 
+    // Perform workload split to allow multiple operations in a single thread
+    globalWorkSize[0] = (size[0] + TransferSplitSize - 1) / TransferSplitSize;
     // Find the current blit type
     if (memView->desc().dimSize_ == 1) {
-        globalWorkSize[0] = amd::alignUp(size[0], 256);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 256);
         globalWorkSize[1] = amd::alignUp(size[1], 1);
         globalWorkSize[2] = amd::alignUp(size[2], 1);
         localWorkSize[0] = 256;
         localWorkSize[1] = localWorkSize[2] = 1;
     }
     else if (memView->desc().dimSize_ == 2) {
-        globalWorkSize[0] = amd::alignUp(size[0], 16);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 16);
         globalWorkSize[1] = amd::alignUp(size[1], 16);
         globalWorkSize[2] = amd::alignUp(size[2], 1);
         localWorkSize[0] = localWorkSize[1] = 16;
         localWorkSize[2] = 1;
     }
     else {
-        globalWorkSize[0] = amd::alignUp(size[0], 8);
+        globalWorkSize[0] = amd::alignUp(globalWorkSize[0], 8);
         globalWorkSize[1] = amd::alignUp(size[1], 8);
         globalWorkSize[2] = amd::alignUp(size[2], 4);
         localWorkSize[0] = localWorkSize[1] = 8;

rocclr/runtime/device/pal/palblit.hpp

@@ -388,6 +388,7 @@ public:
 
 private:
     static const size_t MaxXferBuffers = 2;
+    static const uint   TransferSplitSize = 3;
 
     //! Copies a buffer object to an image object
     bool copyBufferToImageKernel(