P4 to Git Change 1057445 by rili@rili_opencl_stg on 2014/07/21 14:11:34

EPR #399808 - Add CL_RGB, CL_UNORM_INT_101010 support

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#105 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudefs.hpp#111 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#186 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.cpp#106 edit


[ROCm/clr commit: 1b9e65b27b]

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

@@ -917,6 +917,7 @@ static const FormatConvertion RejectedData[] =
     { CL_FLOAT,                 CL_UNSIGNED_INT32 },
     { CL_SIGNED_INT8,           CL_UNSIGNED_INT8  },
     { CL_SIGNED_INT16,          CL_UNSIGNED_INT16 },
+    { CL_UNORM_INT_101010,      CL_UNSIGNED_INT8 },
     { CL_SIGNED_INT32,          CL_UNSIGNED_INT32 }
 };
 
@@ -927,6 +928,7 @@ static const FormatConvertion RejectedOrder[] =
     { CL_RA,                    CL_RG },
     { CL_LUMINANCE,             CL_R  },
     { CL_INTENSITY,             CL_R },
+    { CL_RGB,                   CL_RGBA },
     { CL_BGRA,                  CL_RGBA },
     { CL_ARGB,                  CL_RGBA },
     { CL_sRGB,                  CL_RGBA },

projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp

@@ -259,6 +259,12 @@ MemoryFormatMap[] = {
     { { CL_RA,                      CL_FLOAT },
       { GSL_CHANNEL_ORDER_RA,       CM_SURF_FMT_RG32F } },
 
+    // RGB
+    { { CL_RGB,                     CL_UNORM_INT_101010 },
+      { GSL_CHANNEL_ORDER_RGB,      CM_SURF_FMT_BGR10_X2 } },
+    { { CL_RGB,                     CL_UNSIGNED_INT8 },             // This is used only by blit kernel
+      { GSL_CHANNEL_ORDER_RGBA,     CM_SURF_FMT_RGBA8UI } },
+
     // RGBA
     { { CL_RGBA,                    CL_UNORM_INT8 },
       { GSL_CHANNEL_ORDER_RGBA,     CM_SURF_FMT_RGBA8 } },
@@ -427,6 +433,7 @@ MemoryFormatSize[] = {
     { CM_SURF_FMT_RG16F,    4,   2 }, /**< A 2 component, 16-bit float value per component */
     { CM_SURF_FMT_RGBA16F,  8,   4 }, /**< A 4 component, 16-bit float value per component */
 
+    { CM_SURF_FMT_BGR10_X2, 4,   4 }, /**< 4 component, unnormalized signed 10-bit integer value per component packed as (@c XXRRRRRRRRRRGGGGGGGGGGBBBBBBBBBB)*/
     { CM_SURF_FMT_DEPTH32F, 4,   1 }, /**< A one component, 32 float value per component */
     { CM_SURF_FMT_DEPTH16 , 2,   1 }, /**< A one component, 16 unsigned int value per component */
     { CM_SURF_FMT_DEPTH24_STEN8 , 4 ,1}, /**< A one component, 32 float value per component */

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

@@ -197,6 +197,9 @@ static uint32_t GetHSAILImageFormatType(cmSurfFmt format)
         formatType = 6;
         break;
 */
+    case CM_SURF_FMT_BGR10_X2:
+        formatType = 7;
+        break;
     case CM_SURF_FMT_sR8:
     case CM_SURF_FMT_sRG8:
     case CM_SURF_FMT_sRGBA8:
@@ -285,6 +288,9 @@ static uint32_t GetHSAILImageOrderType(gslChannelOrder chOrder)
     case HSA_IMAGE_FMT_R10G10B10_UNORM:
         orderType = 6;
         break;*/
+    case GSL_CHANNEL_ORDER_RGB:
+        orderType = 7;
+        break;
     case GSL_CHANNEL_ORDER_RGBA:
         orderType = 8;
         break;

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

@@ -760,6 +760,8 @@ Image::Format::getElementSize() const
         break;
 
     case CL_UNORM_INT_101010:
+        bytesPerPixel = 4;
+        break;
     case CL_SIGNED_INT32:
     case CL_UNSIGNED_INT32:
     case CL_FLOAT:
@@ -916,6 +918,9 @@ Image::supportedFormats[] = {
 
     {CL_RG, CL_HALF_FLOAT},         {CL_RG, CL_FLOAT},
 
+    // RGB
+    {CL_RGB, CL_UNORM_INT_101010},
+
     // RGBA
     {CL_RGBA, CL_SNORM_INT8},       {CL_RGBA, CL_SNORM_INT16},
     {CL_RGBA, CL_UNORM_INT8},       {CL_RGBA, CL_UNORM_INT16},
@@ -1334,9 +1339,9 @@ Image::Format::formatColor(const void* colorRGBA, void* colorFormat) const
 
     union t101010 {
         struct {
-            uint32_t    r_: 10;
-            uint32_t    g_: 10;
             uint32_t    b_: 10;
+            uint32_t    g_: 10;
+            uint32_t    r_: 10;
             uint32_t    a_: 2;
         };
         uint32_t    rgba_;