GRAPHICS: Optimize the convertYUV420ToRGB function a bit more

A template is used to avoid a bytesPerPixel check on every pixel and less deferences are done

1 files changed, 37 insertions, 26 deletions

diff --git a/graphics/yuv_to_rgb.cpp b/graphics/yuv_to_rgb.cpp
index 831736cd75..037ea9a007 100644
--- a/graphics/yuv_to_rgb.cpp
+++ b/graphics/yuv_to_rgb.cpp
@@ -196,52 +196,63 @@ DECLARE_SINGLETON(Graphics::YUVToRGBManager);
 namespace Graphics {
 
 #define PUT_PIXEL(s, d) \
-	L = &lookup->_rgbToPix[(s)]; \
-	if (dst->format.bytesPerPixel == 2) \
-		*((uint16 *)(d)) = (L[cr_r] | L[crb_g] | L[cb_b]); \
-	else \
-		*((uint32 *)(d)) = (L[cr_r] | L[crb_g] | L[cb_b])
-
-void convertYUV420ToRGB(Graphics::Surface *dst, const byte *ySrc, const byte *uSrc, const byte *vSrc, int yWidth, int yHeight, int yPitch, int uvPitch) {
-	// Sanity checks
-	assert(dst && dst->pixels);
-	assert(dst->format.bytesPerPixel == 2 || dst->format.bytesPerPixel == 4);
-	assert(ySrc && uSrc && vSrc);
-	assert((yWidth & 1) == 0);
-	assert((yHeight & 1) == 0);
-
-	const YUVToRGBLookup *lookup = YUVToRGBMan.getLookup(dst->format);
-
-	byte *dstPtr = (byte *)dst->pixels;
+	L = &rgbToPix[(s)]; \
+	*((PixelInt *)(d)) = (L[cr_r] | L[crb_g] | L[cb_b])
 
+template<typename PixelInt>
+void convertYUV420ToRGB(byte *dstPtr, int dstPitch, const YUVToRGBLookup *lookup, const byte *ySrc, const byte *uSrc, const byte *vSrc, int yWidth, int yHeight, int yPitch, int uvPitch) {
 	int halfHeight = yHeight >> 1;
 	int halfWidth = yWidth >> 1;
 
+	// Keep the tables in pointers here to avoid a dereference on each pixel
+	const int16 *Cr_r_tab = lookup->_colorTab;
+	const int16 *Cr_g_tab = Cr_r_tab + 256;
+	const int16 *Cb_g_tab = Cr_g_tab + 256;
+	const int16 *Cb_b_tab = Cb_g_tab + 256;
+	const uint32 *rgbToPix = lookup->_rgbToPix;
+
 	for (int h = 0; h < halfHeight; h++) {
 		for (int w = 0; w < halfWidth; w++) {
-			register uint32 *L;
+			register const uint32 *L;
 
-			int16 cr_r  = lookup->_colorTab[*vSrc + 0 * 256];
-			int16 crb_g = lookup->_colorTab[*vSrc + 1 * 256] + lookup->_colorTab[*uSrc + 2 * 256];
-			int16 cb_b  = lookup->_colorTab[*uSrc + 3 * 256];
+			int16 cr_r  = Cr_r_tab[*vSrc];
+			int16 crb_g = Cr_g_tab[*vSrc] + Cb_g_tab[*uSrc];
+			int16 cb_b  = Cb_b_tab[*uSrc];
 			++uSrc;
 			++vSrc;
 
 			PUT_PIXEL(*ySrc, dstPtr);
-			PUT_PIXEL(*(ySrc + yPitch), dstPtr + dst->pitch);
+			PUT_PIXEL(*(ySrc + yPitch), dstPtr + dstPitch);
 			ySrc++;
-			dstPtr += dst->format.bytesPerPixel;
+			dstPtr += sizeof(PixelInt);
 			PUT_PIXEL(*ySrc, dstPtr);
-			PUT_PIXEL(*(ySrc + yPitch), dstPtr + dst->pitch);
+			PUT_PIXEL(*(ySrc + yPitch), dstPtr + dstPitch);
 			ySrc++;
-			dstPtr += dst->format.bytesPerPixel;
+			dstPtr += sizeof(PixelInt);
 		}
 
-		dstPtr += dst->pitch;
+		dstPtr += dstPitch;
 		ySrc += (yPitch << 1) - yWidth;
 		uSrc += uvPitch - halfWidth;
 		vSrc += uvPitch - halfWidth;
 	}
 }
 
+void convertYUV420ToRGB(Graphics::Surface *dst, const byte *ySrc, const byte *uSrc, const byte *vSrc, int yWidth, int yHeight, int yPitch, int uvPitch) {
+	// Sanity checks
+	assert(dst && dst->pixels);
+	assert(dst->format.bytesPerPixel == 2 || dst->format.bytesPerPixel == 4);
+	assert(ySrc && uSrc && vSrc);
+	assert((yWidth & 1) == 0);
+	assert((yHeight & 1) == 0);
+
+	const YUVToRGBLookup *lookup = YUVToRGBMan.getLookup(dst->format);
+
+	// Use a templated function to avoid an if check on every pixel
+	if (dst->format.bytesPerPixel == 2)
+		convertYUV420ToRGB<uint16>((byte *)dst->pixels, dst->pitch, lookup, ySrc, uSrc, vSrc, yWidth, yHeight, yPitch, uvPitch);
+	else
+		convertYUV420ToRGB<uint32>((byte *)dst->pixels, dst->pitch, lookup, ySrc, uSrc, vSrc, yWidth, yHeight, yPitch, uvPitch);
+}
+
 } // End of namespace Graphics