Created new helper functions for converting texture formats, loading images and…

src/libGLESv2/formatutils.h

+//
+// Copyright (c) 2013 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// formatutils.h: Queries for GL image formats.
+
+#ifndef LIBGLESV2_FORMATUTILS_H_
+#define LIBGLESV2_FORMATUTILS_H_
+
+#define GL_APICALL
+#include <GLES3/gl3.h>
+#include <GLES2/gl2.h>
+#include <GLES2/gl2ext.h>
+
+typedef void (*MipGenerationFunction)(unsigned int sourceWidth, unsigned int sourceHeight, unsigned int sourceDepth,
+                                      const unsigned char *sourceData, int sourceRowPitch, int sourceDepthPitch,
+                                      unsigned char *destData, int destRowPitch, int destDepthPitch);
+
+typedef void (*LoadImageFunction)(int width, int height, int depth,
+                                  const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                  void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+namespace rx
+{
+
+class Renderer;
+
+}
+
+namespace gl
+{
+
+class Context;
+
+bool IsValidInternalFormat(GLint internalFormat, const Context *context);
+bool IsValidFormat(GLenum format, GLuint clientVersion);
+bool IsValidType(GLenum type, GLuint clientVersion);
+
+bool IsValidFormatCombination(GLint internalFormat, GLenum format, GLenum type, GLuint clientVersion);
+bool IsValidCopyTexImageCombination(GLenum textureFormat, GLenum frameBufferFormat, GLuint clientVersion);
+
+bool IsSizedInternalFormat(GLint internalFormat, GLuint clientVersion);
+GLint GetSizedInternalFormat(GLenum format, GLenum type, GLuint clientVersion);
+
+GLuint GetPixelBytes(GLint internalFormat, GLuint clientVersion);
+GLuint GetAlphaBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetRedBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetGreenBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetBlueBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetLuminanceBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetDepthBits(GLint internalFormat, GLuint clientVersion);
+GLuint GetStencilBits(GLint internalFormat, GLuint clientVersion);
+
+GLenum GetFormat(GLint internalFormat, GLuint clientVersion);
+GLenum GetType(GLint internalFormat, GLuint clientVersion);
+
+bool IsColorRenderingSupported(GLint internalFormat, const rx::Renderer *renderer);
+bool IsColorRenderingSupported(GLint internalFormat, const Context *context);
+bool IsTextureFilteringSupported(GLint internalFormat, const rx::Renderer *renderer);
+bool IsTextureFilteringSupported(GLint internalFormat, const Context *context);
+bool IsDepthRenderingSupported(GLint internalFormat, const rx::Renderer *renderer);
+bool IsDepthRenderingSupported(GLint internalFormat, const Context *context);
+bool IsStencilRenderingSupported(GLint internalFormat, const rx::Renderer *renderer);
+bool IsStencilRenderingSupported(GLint internalFormat, const Context *context);
+
+GLuint GetRowPitch(GLint internalFormat, GLenum type, GLuint clientVersion, GLsizei width, GLint alignment);
+GLuint GetDepthPitch(GLint internalFormat, GLenum type, GLuint clientVersion, GLsizei width, GLsizei height, GLint alignment);
+GLuint GetBlockSize(GLint internalFormat, GLenum type, GLuint clientVersion, GLsizei width, GLsizei height);
+
+bool IsFormatCompressed(GLint internalFormat, GLuint clientVersion);
+GLuint GetCompressedBlockWidth(GLint internalFormat, GLuint clientVersion);
+GLuint GetCompressedBlockHeight(GLint internalFormat, GLuint clientVersion);
+
+}
+
+#endif LIBGLESV2_FORMATUTILS_H_

src/libGLESv2/libGLESv2.vcxproj

@@ -244,6 +244,7 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     </ClCompile>
     <ClCompile Include="Fence.cpp" />
     <ClCompile Include="Float16ToFloat32.cpp" />
+    <ClCompile Include="formatutils.cpp" />
     <ClCompile Include="Framebuffer.cpp" />
     <ClCompile Include="HandleAllocator.cpp" />
     <ClCompile Include="libGLESv2.cpp" />
@@ -265,6 +266,8 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     <ClCompile Include="renderer\BufferStorage.cpp" />
     <ClCompile Include="renderer\BufferStorage11.cpp" />
     <ClCompile Include="renderer\BufferStorage9.cpp" />
+    <ClCompile Include="renderer\formatutils11.cpp" />
+    <ClCompile Include="renderer\formatutils9.cpp" />
     <ClCompile Include="renderer\Image.cpp" />
     <ClCompile Include="renderer\Image9.cpp" />
     <ClCompile Include="renderer\IndexBuffer.cpp" />
@@ -274,6 +277,8 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     <ClCompile Include="renderer\ImageSSE2.cpp" />
     <ClCompile Include="renderer\Image11.cpp" />
     <ClCompile Include="renderer\InputLayoutCache.cpp" />
+    <ClCompile Include="renderer\loadimage.cpp" />
+    <ClCompile Include="renderer\loadimageSSE2.cpp" />
     <ClCompile Include="renderer\Query11.cpp" />
     <ClCompile Include="renderer\Query9.cpp" />
     <ClCompile Include="renderer\Renderer.cpp" />
@@ -318,6 +323,7 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     <ClInclude Include="constants.h" />
     <ClInclude Include="Context.h" />
     <ClInclude Include="Fence.h" />
+    <ClInclude Include="formatutils.h" />
     <ClInclude Include="Framebuffer.h" />
     <ClInclude Include="HandleAllocator.h" />
     <ClInclude Include="main.h" />
@@ -335,6 +341,8 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     <ClInclude Include="renderer\BufferStorage.h" />
     <ClInclude Include="renderer\BufferStorage11.h" />
     <ClInclude Include="renderer\BufferStorage9.h" />
+    <ClInclude Include="renderer\formatutils11.h" />
+    <ClInclude Include="renderer\formatutils9.h" />
     <ClInclude Include="renderer\generatemip.h" />
     <ClInclude Include="renderer\Image.h" />
     <ClInclude Include="renderer\Image11.h" />
@@ -344,6 +352,7 @@ copy "$(OutDir)libGLESv2.lib" "$(ProjectDir)..\..\lib\$(Configuration)\"
     <ClInclude Include="renderer\IndexBuffer9.h" />
     <ClInclude Include="renderer\IndexDataManager.h" />
     <ClInclude Include="renderer\InputLayoutCache.h" />
+    <ClInclude Include="renderer\loadimage.h" />
     <ClInclude Include="renderer\Query11.h" />
     <ClInclude Include="renderer\QueryImpl.h" />
     <ClInclude Include="renderer\Query9.h" />

src/libGLESv2/libGLESv2.vcxproj.filters

@@ -218,6 +218,21 @@
     <ClCompile Include="precompiled.cpp">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="renderer\loadimage.cpp">
+      <Filter>Source Files\Renderer</Filter>
+    </ClCompile>
+    <ClCompile Include="renderer\loadimageSSE2.cpp">
+      <Filter>Source Files\Renderer</Filter>
+    </ClCompile>
+    <ClCompile Include="renderer\formatutils9.cpp">
+      <Filter>Source Files\Renderer9</Filter>
+    </ClCompile>
+    <ClCompile Include="renderer\formatutils11.cpp">
+      <Filter>Source Files\Renderer11</Filter>
+    </ClCompile>
+    <ClCompile Include="formatutils.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="BinaryStream.h">
@@ -493,6 +508,18 @@
     <ClInclude Include="renderer\shaders\compiled\clearsingle11ps.h">
       <Filter>Shaders\Compiled</Filter>
     </ClInclude>
+    <ClInclude Include="renderer\loadimage.h">
+      <Filter>Header Files\Renderer</Filter>
+    </ClInclude>
+    <ClInclude Include="renderer\formatutils11.h">
+      <Filter>Header Files\Renderer11</Filter>
+    </ClInclude>
+    <ClInclude Include="renderer\formatutils9.h">
+      <Filter>Header Files\Renderer9</Filter>
+    </ClInclude>
+    <ClInclude Include="formatutils.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="renderer\shaders\Blit.ps">

src/libGLESv2/renderer/formatutils11.h

+//
+// Copyright (c) 2013 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// formatutils11.h: Queries for GL image formats and their translations to D3D11
+// formats.
+
+#ifndef LIBGLESV2_RENDERER_FORMATUTILS11_H_
+#define LIBGLESV2_RENDERER_FORMATUTILS11_H_
+
+#include "libGLESv2/formatutils.h"
+
+namespace rx
+{
+
+namespace d3d11
+{
+
+MipGenerationFunction GetMipGenerationFunction(DXGI_FORMAT format);
+LoadImageFunction GetImageLoadFunction(GLint internalFormat, GLenum type, GLuint clientVersion);
+
+GLuint GetFormatPixelBytes(DXGI_FORMAT format);
+GLuint GetBlockWidth(DXGI_FORMAT format);
+GLuint GetBlockHeight(DXGI_FORMAT format);
+
+void MakeValidSize(bool isImage, DXGI_FORMAT format, GLsizei *requestWidth, GLsizei *requestHeight, int *levelOffset);
+
+}
+
+namespace gl_d3d11
+{
+
+DXGI_FORMAT GetTexFormat(GLint internalFormat, GLuint clientVersion);
+DXGI_FORMAT GetSRVFormat(GLint internalFormat, GLuint clientVersion);
+DXGI_FORMAT GetRTVFormat(GLint internalFormat, GLuint clientVersion);
+DXGI_FORMAT GetDSVFormat(GLint internalFormat, GLuint clientVersion);
+
+}
+
+namespace d3d11_gl
+{
+
+GLint GetInternalFormat(DXGI_FORMAT format);
+
+}
+
+}
+
+#endif // LIBGLESV2_RENDERER_FORMATUTILS11_H_

src/libGLESv2/renderer/formatutils9.h

+//
+// Copyright (c) 2013 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// formatutils9.h: Queries for GL image formats and their translations to D3D9
+// formats.
+
+#ifndef LIBGLESV2_RENDERER_FORMATUTILS9_H_
+#define LIBGLESV2_RENDERER_FORMATUTILS9_H_
+
+#include "libGLESv2/formatutils.h"
+
+namespace rx
+{
+
+class Renderer9;
+
+namespace d3d9
+{
+
+MipGenerationFunction GetMipGenerationFunction(D3DFORMAT format);
+LoadImageFunction GetImageLoadFunction(GLint internalFormat, const Renderer9 *renderer);
+
+GLuint GetFormatPixelBytes(D3DFORMAT format);
+GLuint GetBlockWidth(D3DFORMAT format);
+GLuint GetBlockHeight(D3DFORMAT format);
+GLuint GetBlockSize(D3DFORMAT format, GLuint width, GLuint height);
+
+void MakeValidSize(bool isImage, D3DFORMAT format, GLsizei *requestWidth, GLsizei *requestHeight, int *levelOffset);
+
+}
+
+namespace gl_d3d9
+{
+
+D3DFORMAT GetTexureFormat(GLint internalFormat, const Renderer9 *renderer);
+D3DFORMAT GetRenderFormat(GLint internalFormat, const Renderer9 *renderer);
+
+D3DMULTISAMPLE_TYPE GetMultisampleType(GLsizei samples);
+
+}
+
+namespace d3d9_gl
+{
+
+GLint GetInternalFormat(D3DFORMAT format);
+GLsizei GetSamplesCount(D3DMULTISAMPLE_TYPE type);
+bool IsFormatChannelEquivalent(D3DFORMAT d3dformat, GLenum format, GLuint clientVersion);
+
+}
+
+}
+
+#endif // LIBGLESV2_RENDERER_FORMATUTILS9_H_

src/libGLESv2/renderer/generatemip.h

@@ -54,6 +54,7 @@ struct A8R8G8B8
 };
 
 typedef A8R8G8B8 R8G8B8A8; // R8G8B8A8 type is functionally equivalent for mip purposes
+typedef A8R8G8B8 B8G8R8A8; // B8G8R8A8 type is functionally equivalent for mip purposes
 
 struct A16B16G16R16F
 {
@@ -93,6 +94,8 @@ struct R16G16F
     }
 };
 
+typedef A16B16G16R16F R16G16B16A16F;
+
 struct A32B32G32R32F
 {
     float R;
@@ -145,6 +148,11 @@ struct R32G32B32F
     }
 };
 
+typedef A32B32G32R32F R32G32B32A32F;
+
+namespace priv
+{
+
 template <typename T>
 static inline T *GetPixel(void *data, unsigned int x, unsigned int y, unsigned int z, unsigned int rowPitch, unsigned int depthPitch)
 {
@@ -346,6 +354,7 @@ static void GenerateMip_XYZ(unsigned int sourceWidth, unsigned int sourceHeight,
     }
 }
 
+
 typedef void (*MipGenerationFunction)(unsigned int sourceWidth, unsigned int sourceHeight, unsigned int sourceDepth,
                                       const unsigned char *sourceData, int sourceRowPitch, int sourceDepthPitch,
                                       unsigned int destWidth, unsigned int destHeight, unsigned int destDepth,
@@ -371,6 +380,8 @@ static MipGenerationFunction GetMipGenerationFunction(unsigned int sourceWidth,
     }
 }
 
+}
+
 template <typename T>
 static void GenerateMip(unsigned int sourceWidth, unsigned int sourceHeight, unsigned int sourceDepth,
                         const unsigned char *sourceData, int sourceRowPitch, int sourceDepthPitch,
@@ -380,7 +391,7 @@ static void GenerateMip(unsigned int sourceWidth, unsigned int sourceHeight, uns
     unsigned int mipHeight = std::max(1U, sourceHeight >> 1);
     unsigned int mipDepth = std::max(1U, sourceDepth >> 1);
 
-    MipGenerationFunction generationFunction = GetMipGenerationFunction<T>(sourceWidth, sourceHeight, sourceDepth);
+    priv::MipGenerationFunction generationFunction = priv::GetMipGenerationFunction<T>(sourceWidth, sourceHeight, sourceDepth);
     ASSERT(generationFunction != NULL);
 
     generationFunction(sourceWidth, sourceHeight, sourceDepth, sourceData, sourceRowPitch, sourceDepthPitch,

src/libGLESv2/renderer/loadimage.h

+//
+// Copyright (c) 2013 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// loadimage.h: Defines image loading functions
+
+#ifndef LIBGLESV2_RENDERER_LOADIMAGE_H_
+#define LIBGLESV2_RENDERER_LOADIMAGE_H_
+
+namespace rx
+{
+
+void loadAlphaDataToBGRA(int width, int height, int depth,
+                         const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                         void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadAlphaDataToNative(int width, int height, int depth,
+                           const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                           void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadAlphaDataToBGRASSE2(int width, int height, int depth,
+                             const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                             void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadAlphaFloatDataToRGBA(int width, int height, int depth,
+                              const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                              void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadAlphaHalfFloatDataToRGBA(int width, int height, int depth,
+                                  const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                  void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceDataToNative(int width, int height, int depth,
+                               const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                               void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceDataToBGRA(int width, int height, int depth,
+                             const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                             void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceFloatDataToRGBA(int width, int height, int depth,
+                                  const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                  void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceFloatDataToRGB(int width, int height, int depth,
+                                 const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                 void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceHalfFloatDataToRGBA(int width, int height, int depth,
+                                      const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                      void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceAlphaDataToNative(int width, int height, int depth,
+                                    const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                    void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceAlphaDataToBGRA(int width, int height, int depth,
+                                  const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                  void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceAlphaFloatDataToRGBA(int width, int height, int depth,
+                                       const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                       void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadLuminanceAlphaHalfFloatDataToRGBA(int width, int height, int depth,
+                                           const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                           void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBUByteDataToBGRX(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBUByteDataToRGBA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGB565DataToBGRA(int width, int height, int depth,
+                          const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                          void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGB565DataToRGBA(int width, int height, int depth,
+                          const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                          void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBFloatDataToRGBA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBFloatDataToNative(int width, int height, int depth,
+                              const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                              void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBHalfFloatDataToRGBA(int width, int height, int depth,
+                                const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBAUByteDataToBGRASSE2(int width, int height, int depth,
+                                 const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                 void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBAUByteDataToBGRA(int width, int height, int depth,
+                             const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                             void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBAUByteDataToNative(int width, int height, int depth,
+                               const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                               void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBA4444DataToBGRA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBA4444DataToRGBA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBA5551DataToBGRA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBA5551DataToRGBA(int width, int height, int depth,
+                            const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                            void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBAFloatDataToRGBA(int width, int height, int depth,
+                             const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                             void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadRGBAHalfFloatDataToRGBA(int width, int height, int depth,
+                                 const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                 void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+void loadBGRADataToBGRA(int width, int height, int depth,
+                        const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                        void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch);
+
+template <unsigned int blockWidth, unsigned int blockHeight, unsigned int blockSize>
+void loadCompressedBlockDataToNative(int width, int height, int depth,
+                                     const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+                                     void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch)
+{
+    int columns = (width + (blockWidth - 1)) / blockWidth;
+    int rows = (height + (blockHeight - 1)) / blockHeight;
+
+    for (int z = 0; z < depth; ++z)
+    {
+        for (int y = 0; y < rows; ++y)
+        {
+            void *source = (void*)((char*)input + y * inputRowPitch + z * inputDepthPitch);
+            void *dest = (void*)((char*)output + y * outputRowPitch + z * outputDepthPitch);
+
+            memcpy(dest, source, columns * blockSize);
+        }
+    }
+}
+
+}
+
+#endif // LIBGLESV2_RENDERER_LOADIMAGE_H_

src/libGLESv2/renderer/loadimageSSE2.cpp

+#include "precompiled.h"
+//
+// Copyright (c) 2002-2012 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// loadimage.cpp: Defines image loading functions. It's
+// in a separated file for GCC, which can enable SSE usage only per-file,
+// not for code blocks that use SSE2 explicitly.
+
+#include "libGLESv2/renderer/loadimage.h"
+
+namespace rx
+{
+
+    void loadAlphaDataToBGRASSE2(int width, int height, int depth,
+        const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+        void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch)
+    {
+        const unsigned char *source = NULL;
+        unsigned int *dest = NULL;
+        __m128i zeroWide = _mm_setzero_si128();
+
+        for (int z = 0; z < depth; z++)
+        {
+            for (int y = 0; y < height; y++)
+            {
+                source = static_cast<const unsigned char*>(input) + y * inputRowPitch + z * inputDepthPitch;
+                dest = reinterpret_cast<unsigned int*>(static_cast<unsigned char*>(output) + y * outputRowPitch + z * outputDepthPitch);
+
+                int x;
+                // Make output writes aligned
+                for (x = 0; ((reinterpret_cast<intptr_t>(&dest[x]) & 0xF) != 0 && x < width); x++)
+                {
+                    dest[x] = static_cast<unsigned int>(source[x]) << 24;
+                }
+
+                for (; x + 7 < width; x += 8)
+                {
+                    __m128i sourceData = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&source[x]));
+                    // Interleave each byte to 16bit, make the lower byte to zero
+                    sourceData = _mm_unpacklo_epi8(zeroWide, sourceData);
+                    // Interleave each 16bit to 32bit, make the lower 16bit to zero
+                    __m128i lo = _mm_unpacklo_epi16(zeroWide, sourceData);
+                    __m128i hi = _mm_unpackhi_epi16(zeroWide, sourceData);
+
+                    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), lo);
+                    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x + 4]), hi);
+                }
+
+                // Handle the remainder
+                for (; x < width; x++)
+                {
+                    dest[x] = static_cast<unsigned int>(source[x]) << 24;
+                }
+            }
+        }
+    }
+
+    void loadRGBAUByteDataToBGRASSE2(int width, int height, int depth,
+        const void *input, unsigned int inputRowPitch, unsigned int inputDepthPitch,
+        void *output, unsigned int outputRowPitch, unsigned int outputDepthPitch)
+    {
+        const unsigned int *source = NULL;
+        unsigned int *dest = NULL;
+        __m128i brMask = _mm_set1_epi32(0x00ff00ff);
+
+        for (int z = 0; z < depth; z++)
+        {
+            for (int y = 0; y < height; y++)
+            {
+                source = reinterpret_cast<const unsigned int*>(static_cast<const unsigned char*>(input) + y * inputRowPitch + z * inputDepthPitch);
+                dest = reinterpret_cast<unsigned int*>(static_cast<unsigned char*>(output) + y * outputRowPitch + z * outputDepthPitch);
+                int x = 0;
+
+                // Make output writes aligned
+                for (x = 0; ((reinterpret_cast<intptr_t>(&dest[x]) & 15) != 0) && x < width; x++)
+                {
+                    unsigned int rgba = source[x];
+                    dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
+                }
+
+                for (; x + 3 < width; x += 4)
+                {
+                    __m128i sourceData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&source[x]));
+                    // Mask out g and a, which don't change
+                    __m128i gaComponents = _mm_andnot_si128(brMask, sourceData);
+                    // Mask out b and r
+                    __m128i brComponents = _mm_and_si128(sourceData, brMask);
+                    // Swap b and r
+                    __m128i brSwapped = _mm_shufflehi_epi16(_mm_shufflelo_epi16(brComponents, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
+                    __m128i result = _mm_or_si128(gaComponents, brSwapped);
+                    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), result);
+                }
+
+                // Perform leftover writes
+                for (; x < width; x++)
+                {
+                    unsigned int rgba = source[x];
+                    dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
+                }
+            }
+        }
+    }
+
+}