From 691c6c7f670e6fb49bee8ebac06fc3854543394b Mon Sep 17 00:00:00 2001
From: bkaradzic <branimirkaradzic@gmail.com>
Date: Wed, 14 Aug 2013 23:19:12 -0700
Subject: [PATCH] Added utility function imageSwizzleBGRA8 to convert RGBA8
 image to BGRA8.

---
 include/bgfx.h      |  3 +++
 src/image.cpp       | 63 +++++++++++++++++++++++++++++++++++++++++++++
 src/renderer_gl.cpp | 24 +++--------------
 3 files changed, 70 insertions(+), 20 deletions(-)
 create mode 100644 src/image.cpp

diff --git a/include/bgfx.h b/include/bgfx.h
index 3afff85e..ae5e87a9 100644
--- a/include/bgfx.h
+++ b/include/bgfx.h
@@ -511,6 +511,9 @@ namespace bgfx
 	///
 	void vertexConvert(const VertexDecl& _destDecl, void* _destData, const VertexDecl& _srcDecl, const void* _srcData, uint32_t _num = 1);
 
+	/// Swizzle RGBA8 image to BGRA8.
+	void imageSwizzleBGRA8(uint8_t* _rgbaData, uint32_t _width, uint32_t _height);
+
 	/// Returns renderer backend API type.
 	RendererType::Enum getRendererType();
 
diff --git a/src/image.cpp b/src/image.cpp
new file mode 100644
index 00000000..721364e2
--- /dev/null
+++ b/src/image.cpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright 2011-2013 Branimir Karadzic. All rights reserved.
+ * License: http://www.opensource.org/licenses/BSD-2-Clause
+ */
+
+#include <bx/float4_t.h>
+
+namespace bgfx
+{
+	static void imageSwizzleBGRA8Ref(uint8_t* _rgbaData, uint32_t _width, uint32_t _height) 
+	{
+		uint32_t dstpitch = _width*4;
+		for (uint32_t yy = 0; yy < _height; ++yy)
+		{
+			uint8_t* dst = &_rgbaData[yy*dstpitch];
+
+			for (uint32_t xx = 0; xx < _width; ++xx)
+			{
+				uint8_t tmp = dst[0];
+				dst[0] = dst[2];
+				dst[2] = tmp;
+				dst += 4;
+			}
+		}
+	}
+
+	void imageSwizzleBGRA8(uint8_t* _rgbaData, uint32_t _width, uint32_t _height)
+	{
+		if (0 != (_width&0xf)
+		||  _width < 16)
+		{
+			imageSwizzleBGRA8Ref(_rgbaData, _width, _height);
+			return;
+		}
+
+		uint32_t dstpitch = _width*4;
+		uint32_t num = dstpitch/16;
+
+		using namespace bx;
+
+		const float4_t mf0f0 = float4_isplat(0xff00ff00);
+		const float4_t m0f0f = float4_isplat(0x00ff00ff);
+
+		for (uint32_t yy = 0; yy < _height; ++yy)
+		{
+			uint8_t* ptr = &_rgbaData[yy*dstpitch];
+
+			for (uint32_t xx = 0; xx < num; ++xx)
+			{
+				const float4_t tabgr = float4_ld(ptr);
+				const float4_t t00ab = float4_srl(tabgr, 16);
+				const float4_t tgr00 = float4_sll(tabgr, 16);
+				const float4_t tgrab = float4_or(t00ab, tgr00);
+				const float4_t ta0g0 = float4_and(tabgr, mf0f0);
+				const float4_t t0g0b = float4_and(tgrab, m0f0f);
+				const float4_t targb = float4_or(ta0g0, t0g0b);
+				float4_st(ptr, targb);
+				ptr += 16;
+			}
+		}
+	}
+
+} // namespace bgfx
diff --git a/src/renderer_gl.cpp b/src/renderer_gl.cpp
index f27bdb96..87cbab6e 100644
--- a/src/renderer_gl.cpp
+++ b/src/renderer_gl.cpp
@@ -9,6 +9,7 @@
 #	include "renderer_gl.h"
 #	include <bx/timer.h>
 #	include <bx/uint32_t.h>
+#	include <bx/float4_t.h>
 
 namespace bgfx
 {
@@ -375,23 +376,6 @@ namespace bgfx
 
 	typedef void (*PostSwapBuffersFn)(uint32_t _width, uint32_t _height);
 
-	static void rgbaToBgra(uint8_t* _data, uint32_t _width, uint32_t _height) 
-	{
-		uint32_t dstpitch = _width*4;
-		for (uint32_t yy = 0; yy < _height; ++yy)
-		{
-			uint8_t* dst = &_data[yy*dstpitch];
-
-			for (uint32_t xx = 0; xx < _width; ++xx)
-			{
-				uint8_t tmp = dst[0];
-				dst[0] = dst[2];
-				dst[2] = tmp;
-				dst += 4;
-			}
-		}
-	}
-
 	static const char* getGLString(GLenum _name)
 	{
 		const char* str = (const char*)glGetString(_name);
@@ -775,7 +759,7 @@ namespace bgfx
 
 			if (GL_RGBA == fmt)
 			{
-				rgbaToBgra(data, width, height);
+				imageSwizzleBGRA8(data, width, height);
 			}
 
 			g_callback->screenShot(_filePath
@@ -1472,7 +1456,7 @@ namespace bgfx
 
 							if (swizzle)
 							{
-								rgbaToBgra(bits, width, height);
+								imageSwizzleBGRA8(bits, width, height);
 							}
 
 							texImage(target+side
@@ -1621,7 +1605,7 @@ namespace bgfx
 							if (NULL != data
 							&&  swizzle)
 							{
-								rgbaToBgra(data, width, height);
+								imageSwizzleBGRA8(data, width, height);
 							}
 
 							texImage(target+side