P4 to Git Change 1058007 by rili@rili_opencl_stg_01 on 2014/07/22 17:28:41

EPR #399808 - Fixed wrong conversion of sRGBA when using host copy instead of blit kernel transfer

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_memobj.cpp#68 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/blit.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/blit.hpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#107 edit

rocclr/runtime/device/blit.cpp

@@ -683,6 +683,20 @@ HostBlitManager::fillImage(
     // rowPitch and slicePitch in bytes
     size_t  devRowPitch;
     size_t  devSlicePitch;
+
+    void *newpattern = const_cast<void *>(pattern);
+    cl_float4 fFillColor;
+
+    // Converting a linear RGB floating-point color value to a normalized 8-bit unsigned integer sRGB value so that the cpu path can treat sRGB as RGB for host transfer.
+    if (memory.owner()->asImage()->getImageFormat().image_channel_order == CL_sRGBA) {
+        float *fColor = static_cast<float *>(newpattern);
+        fFillColor.s[0] = sRGBmap(fColor[0]) / 255.0f;
+        fFillColor.s[1] = sRGBmap(fColor[1]) / 255.0f;
+        fFillColor.s[2] = sRGBmap(fColor[2]) / 255.0f;
+        fFillColor.s[3] = fColor[3];
+        newpattern = static_cast<void *>(&fFillColor);
+    }
+
     // Map memory
     void* fillMem = memory.cpuMap(vDev_, (entire) ? Memory::CpuWriteOnly : 0,
         startLayer, numLayers, &devRowPitch, &devSlicePitch);
@@ -693,7 +707,7 @@ HostBlitManager::fillImage(
 
     float fillValue[4];
     memset(fillValue, 0, sizeof(fillValue));
-    memory.owner()->asImage()->getImageFormat().formatColor(pattern, fillValue);
+    memory.owner()->asImage()->getImageFormat().formatColor(newpattern, fillValue);
 
     size_t  elementSize = memory.owner()->asImage()->getImageFormat().getElementSize();
     size_t  offset  = origin[0] * elementSize;
@@ -734,4 +748,28 @@ HostBlitManager::fillImage(
     return true;
 }
 
+cl_uint
+HostBlitManager::sRGBmap(float fc) const
+{
+    double c = (double)fc;
+
+#ifdef ATI_OS_LINUX
+    if (isnan(c))
+        c = 0.0;
+#else
+    if (_isnan(c))
+        c = 0.0;
+#endif
+
+    if (c > 1.0)
+        c = 1.0;
+    else if (c < 0.0)
+        c = 0.0;
+    else if (c < 0.0031308)
+        c = 12.92 * c;
+    else
+        c = (1055.0/1000.0) * pow(c, 5.0/12.0) - (55.0/1000.0);
+
+    return (cl_uint)(c * 255.0 + 0.5);
+}
 } // namespace gpu

rocclr/runtime/device/blit.hpp

@@ -352,6 +352,8 @@ public:
         bool        entire = false      //!< Entire buffer will be updated
         ) const;
 
+    cl_uint sRGBmap(float fc) const;
+
 protected:
     VirtualDevice&      vDev_;  //!< Virtual device object
     const amd::Device&  dev_;   //!< Physical device

rocclr/runtime/device/gpu/gpublit.cpp

@@ -2583,6 +2583,9 @@ KernelBlitManager::fillImage(
     fillType = FillImage;
     dim = 3;
 
+    void *newpattern = const_cast<void *>(pattern);
+    cl_uint4  iFillColor;
+
     bool rejected = false;
     bool    releaseView = false;
     // For depth, we need to create a view
@@ -2605,6 +2608,13 @@ KernelBlitManager::fillImage(
         }
 
         if (gpuMem(memory).cal()->format_ == CM_SURF_FMT_RGBA8_SRGB) {
+            // Converting a linear RGB floating-point color value to a 8-bit unsigned integer sRGB value because hw is not support write_imagef for sRGB.
+            float *fColor = static_cast<float *>(newpattern);
+            iFillColor.s[0] = sRGBmap(fColor[0]);
+            iFillColor.s[1] = sRGBmap(fColor[1]);
+            iFillColor.s[2] = sRGBmap(fColor[2]);
+            iFillColor.s[3] = (cl_uint)(fColor[3]*255.0f);
+            newpattern = static_cast<void*>(&iFillColor);
             for (uint i = 0; i < RejectedFormatChannelTotal; ++i) {
                 if (RejectedOrder[i].clOldType_ == newFormat.image_channel_order) {
                     newFormat.image_channel_order = RejectedOrder[i].clNewType_;
@@ -2657,9 +2667,9 @@ KernelBlitManager::fillImage(
     // Program kernels arguments for the blit operation
     Memory*  mem = memView;
     setArgument(kernels_[fillType], 0, sizeof(cl_mem), &mem);
-    setArgument(kernels_[fillType], 1, sizeof(cl_float4), pattern);
-    setArgument(kernels_[fillType], 2, sizeof(cl_int4), pattern);
-    setArgument(kernels_[fillType], 3, sizeof(cl_uint4), pattern);
+    setArgument(kernels_[fillType], 1, sizeof(cl_float4), newpattern);
+    setArgument(kernels_[fillType], 2, sizeof(cl_int4), newpattern);
+    setArgument(kernels_[fillType], 3, sizeof(cl_uint4), newpattern);
 
     cl_int fillOrigin[4] = { (cl_int)origin[0],
                              (cl_int)origin[1],