OpenGL2: Replace R_MipMapsRGB() with faster version.

2 files changed, 36 insertions, 31 deletions

diff --git a/src/renderergl2/tr_extramath.h b/src/renderergl2/tr_extramath.h
index f1ea5767..131d9f6b 100644
--- a/src/renderergl2/tr_extramath.h
+++ b/src/renderergl2/tr_extramath.h
@@ -56,9 +56,6 @@ void Mat4SimpleInverse( const mat4_t in, mat4_t out);
 #define ByteToFloat(a)          ((float)(a) * 1.0f/255.0f)
 #define FloatToByte(a)          (byte)((a) * 255.0f)
 
-#define RGBtosRGB(a)            (((a) < 0.0031308f) ? (12.92f * (a)) : (1.055f * pow((a), 0.41666f) - 0.055f))
-#define sRGBtoRGB(a)            (((a) <= 0.04045f)  ? ((a) / 12.92f) : (pow((((a) + 0.055f) / 1.055f), 2.4)) )
-
 static ID_INLINE int VectorCompare4(const vec4_t v1, const vec4_t v2)
 {
 	if(v1[0] != v2[0] || v1[1] != v2[1] || v1[2] != v2[2] || v1[3] != v2[3])
diff --git a/src/renderergl2/tr_image.c b/src/renderergl2/tr_image.c
index 6f005247..b7237f53 100644
--- a/src/renderergl2/tr_image.c
+++ b/src/renderergl2/tr_image.c
@@ -1333,43 +1333,51 @@ static void R_MipMap2( byte *in, int inWidth, int inHeight ) {
 
 static void R_MipMapsRGB( byte *in, int inWidth, int inHeight)
 {
-	int			i, j, k;
-	int			outWidth, outHeight;
-	byte		*temp;
+	int x, y, c, stride;
+	const byte *in2;
+	float total;
+	static float downmipSrgbLookup[256];
+	static int downmipSrgbLookupSet = 0;
+	byte *out = in;
 
-	outWidth = inWidth >> 1;
-	outHeight = inHeight >> 1;
-	temp = ri.Hunk_AllocateTempMemory( outWidth * outHeight * 4 );
+	if (!downmipSrgbLookupSet) {
+		for (x = 0; x < 256; x++)
+			downmipSrgbLookup[x] = ((x < 10) ? x / 3294.6f : powf((x / 255.0f + 0.055f) / 1.055f, 2.4f)) * 0.25f;
+		downmipSrgbLookupSet = 1;
+	}
 
-	for ( i = 0 ; i < outHeight ; i++ ) {
-		byte *outbyte = temp + (  i          * outWidth ) * 4;
-		byte *inbyte1 = in   + (  i * 2      * inWidth  ) * 4;
-		byte *inbyte2 = in   + ( (i * 2 + 1) * inWidth  ) * 4;
-		for ( j = 0 ; j < outWidth ; j++ ) {
-			for ( k = 0 ; k < 3 ; k++ ) {
-				float total, current;
+	if (inWidth == 1 && inHeight == 1)
+		return;
 
-				current = ByteToFloat(inbyte1[0]); total  = sRGBtoRGB(current);
-				current = ByteToFloat(inbyte1[4]); total += sRGBtoRGB(current);
-				current = ByteToFloat(inbyte2[0]); total += sRGBtoRGB(current);
-				current = ByteToFloat(inbyte2[4]); total += sRGBtoRGB(current);
+	if (inWidth == 1 || inHeight == 1) {
+		for (x = (inWidth * inHeight) >> 1; x; x--) {
+			for (c = 3; c; c--, in++) {
+				total  = (downmipSrgbLookup[*(in)] + downmipSrgbLookup[*(in + 4)]) * 2.0f;
 
-				total *= 0.25f;
+				*out++ = (byte)(powf(total, 1.0f / 2.2f) * 255.0f);
+			}
+			*out++ = (*(in) + *(in + 4)) >> 1; in += 5;
+		}
+		
+		return;
+	}
+
+	stride = inWidth * 4;
+	inWidth >>= 1; inHeight >>= 1;
 
-				inbyte1++;
-				inbyte2++;
+	in2 = in + stride;
+	for (y = inHeight; y; y--, in += stride, in2 += stride) {
+		for (x = inWidth; x; x--) {
+			for (c = 3; c; c--, in++, in2++) {
+				total = downmipSrgbLookup[*(in)]  + downmipSrgbLookup[*(in + 4)]
+				      + downmipSrgbLookup[*(in2)] + downmipSrgbLookup[*(in2 + 4)];
 
-				current = RGBtosRGB(total);
-				*outbyte++ = FloatToByte(current);
+				*out++ = (byte)(powf(total, 1.0f / 2.2f) * 255.0f);
 			}
-			*outbyte++ = (inbyte1[0] + inbyte1[4] + inbyte2[0] + inbyte2[4]) >> 2;
-			inbyte1 += 5;
-			inbyte2 += 5;
+
+			*out++ = (*(in) + *(in + 4) + *(in2) + *(in2 + 4)) >> 2; in += 5, in2 += 5;
 		}
 	}
-
-	Com_Memcpy( in, temp, outWidth * outHeight * 4 );
-	ri.Hunk_FreeTempMemory( temp );
 }
 
 /*