1 files changed, 48 insertions, 23 deletions

diff --git a/graphics/yuv_to_rgb.cpp b/graphics/yuv_to_rgb.cpp
index e0af267106..78903d0cd8 100644
--- a/graphics/yuv_to_rgb.cpp
+++ b/graphics/yuv_to_rgb.cpp
@@ -300,19 +300,32 @@ void convertYUV420ToRGB(Graphics::Surface *dst, const byte *ySrc, const byte *uS
 		convertYUV420ToRGB<uint32>((byte *)dst->pixels, dst->pitch, lookup, ySrc, uSrc, vSrc, yWidth, yHeight, yPitch, uvPitch);
 }
 
-#define OUTPUT_4_PIXEL_COLUMN() \
+#define READ_QUAD(ptr, prefix) \
+	byte prefix##A = ptr[index]; \
+	byte prefix##B = ptr[index + 1]; \
+	byte prefix##C = ptr[index + uvPitch]; \
+	byte prefix##D = ptr[index + uvPitch + 1]
+
+#define DO_INTERPOLATION(out) \
+	out = (out##A * (4 - xDiff) * (4 - yDiff) + out##B * xDiff * (4 - yDiff) + \
+			out##C * yDiff * (4 - xDiff) + out##D * xDiff * yDiff) >> 4
+
+#define DO_YUV410_PIXEL() \
+	DO_INTERPOLATION(u); \
+	DO_INTERPOLATION(v); \
+	\
+	cr_r  = Cr_r_tab[v]; \
+	crb_g = Cr_g_tab[v] + Cb_g_tab[u]; \
+	cb_b  = Cb_b_tab[u]; \
+	\
 	PUT_PIXEL(*ySrc, dstPtr); \
-	PUT_PIXEL(*(ySrc + yPitch), dstPtr + dstPitch); \
-	PUT_PIXEL(*(ySrc + (yPitch << 1)), dstPtr + dstPitch * 2); \
-	PUT_PIXEL(*(ySrc + (yPitch * 3)), dstPtr + dstPitch * 3); \
+	dstPtr += sizeof(PixelInt); \
+	\
 	ySrc++; \
-	dstPtr += sizeof(PixelInt)
+	xDiff++
 
 template<typename PixelInt>
 void convertYUV410ToRGB(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 quarterHeight = yHeight >> 2;
-	int quarterWidth = yWidth >> 2;
-
 	// 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;
@@ -320,29 +333,41 @@ void convertYUV410ToRGB(byte *dstPtr, int dstPitch, const YUVToRGBLookup *lookup
 	const int16 *Cb_b_tab = Cb_g_tab + 256;
 	const uint32 *rgbToPix = lookup->_rgbToPix;
 
-	for (int h = 0; h < quarterHeight; h++) {
-		for (int w = 0; w < quarterWidth; w++) {
+	int quarterWidth = yWidth >> 2;
+
+	for (int y = 0; y < yHeight; y++) {
+		for (int x = 0; x < quarterWidth; x++) {
+			// Perform bilinear interpolation on the the chroma values
+			// Based on the algorithm found here: http://tech-algorithm.com/articles/bilinear-image-scaling/
+			// Feel free to optimize further
+			int targetY = y >> 2;
+			int xDiff = 0;
+			int yDiff = y & 3;
+			int index = targetY * uvPitch + x;
+
+			// Declare some variables for the following macros
+			byte u, v;
+			int16 cr_r, crb_g, cb_b;
 			register const uint32 *L;
 
-			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;
+			READ_QUAD(uSrc, u);
+			READ_QUAD(vSrc, v);
 
-			OUTPUT_4_PIXEL_COLUMN();
-			OUTPUT_4_PIXEL_COLUMN();
-			OUTPUT_4_PIXEL_COLUMN();
-			OUTPUT_4_PIXEL_COLUMN();
+			DO_YUV410_PIXEL();
+			DO_YUV410_PIXEL();
+			DO_YUV410_PIXEL();
+			DO_YUV410_PIXEL();
 		}
 
-		dstPtr += dstPitch * 3;
-		ySrc += (yPitch << 2) - yWidth;
-		uSrc += uvPitch - quarterWidth;
-		vSrc += uvPitch - quarterWidth;
+		dstPtr += dstPitch - yWidth * sizeof(PixelInt);
+		ySrc += yPitch - yWidth;
 	}
 }
 
+#undef READ_QUAD
+#undef DO_INTERPOLATION
+#undef DO_YUV410_PIXEL
+
 void convertYUV410ToRGB(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);