synced hq2x with upstream changes; some performance tweaks to hq2x/hq3x (I think for bigger speed improvement we'd have to use ASM (with C++ fallback of course)...

svn-id: r10494

4 files changed, 768 insertions, 744 deletions

diff --git a/common/scaler.cpp b/common/scaler.cpp
index 4bb51cc84e..d8522440af 100644
--- a/common/scaler.cpp
+++ b/common/scaler.cpp
@@ -175,10 +175,9 @@ void AdvMame2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPi
 			A = B; B = C; C = *(p - nextlineSrc);
 			D = E; E = F; F = *(p);
 			G = H; H = I; I = *(p + nextlineSrc);
-
-			*(q) = D == B && B != F && D != H ? D : E;
+			*(q + 0) = D == B && B != F && D != H ? D : E;
 			*(q + 1) = B == F && B != D && F != H ? F : E;
-			*(q + nextlineDst) = D == H && D != B && H != F ? D : E;
+			*(q + nextlineDst + 0) = D == H && D != B && H != F ? D : E;
 			*(q + nextlineDst + 1) = H == F && D != H && B != F ? F : E;
 			q += 2;
 		}
diff --git a/common/scaler/hq2x.cpp b/common/scaler/hq2x.cpp
index de134373dd..4eef8679e1 100644
--- a/common/scaler/hq2x.cpp
+++ b/common/scaler/hq2x.cpp
@@ -22,49 +22,59 @@
 
 #include "common/scaler/intern.h"
 
-#define PIXEL00_0   *(q) = w[5];
-#define PIXEL00_10  *(q) = interpolate16_2<bitFormat,3,1>(w[5], w[1]);
-#define PIXEL00_11  *(q) = interpolate16_2<bitFormat,3,1>(w[5], w[4]);
-#define PIXEL00_12  *(q) = interpolate16_2<bitFormat,3,1>(w[5], w[2]);
-#define PIXEL00_20  *(q) = interpolate16_3<bitFormat,2,1,1>(w[5], w[4], w[2]);
-#define PIXEL00_30  *(q) = interpolate16_3<bitFormat,6,1,1>(w[5], w[4], w[2]);
-#define PIXEL00_40  *(q) = interpolate16_3<bitFormat,14,1,1>(w[5], w[4], w[2]);
-#define PIXEL00_50  *(q) = interpolate16_3<bitFormat,5,2,1>(w[5], w[2], w[4]);
-#define PIXEL00_51  *(q) = interpolate16_3<bitFormat,5,2,1>(w[5], w[4], w[2]);
-#define PIXEL00_60  *(q) = interpolate16_3<bitFormat,2,3,3>(w[5], w[4], w[2]);
+#define PIXEL00_0	*(q) = w[5];
+#define PIXEL00_10	*(q) = interpolate16_2<bitFormat,3,1>(w[5], w[1]);
+#define PIXEL00_11	*(q) = interpolate16_2<bitFormat,3,1>(w[5], w[4]);
+#define PIXEL00_12	*(q) = interpolate16_2<bitFormat,3,1>(w[5], w[2]);
+#define PIXEL00_20	*(q) = interpolate16_3<bitFormat,2,1,1>(w[5], w[4], w[2]);
+#define PIXEL00_21	*(q) = interpolate16_3<bitFormat,2,1,1>(w[5], w[1], w[2]);
+#define PIXEL00_22	*(q) = interpolate16_3<bitFormat,2,1,1>(w[5], w[1], w[4]);
+#define PIXEL00_60	*(q) = interpolate16_3<bitFormat,5,2,1>(w[5], w[2], w[4]);
+#define PIXEL00_61	*(q) = interpolate16_3<bitFormat,5,2,1>(w[5], w[4], w[2]);
+#define PIXEL00_70	*(q) = interpolate16_3<bitFormat,6,1,1>(w[5], w[4], w[2]);
+#define PIXEL00_90	*(q) = interpolate16_3<bitFormat,2,3,3>(w[5], w[4], w[2]);
+#define PIXEL00_100	*(q) = interpolate16_3<bitFormat,14,1,1>(w[5], w[4], w[2]);
 
-#define PIXEL01_0   *(q+1) = w[5];
-#define PIXEL01_10  *(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[3]);
-#define PIXEL01_11  *(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[2]);
-#define PIXEL01_12  *(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[6]);
-#define PIXEL01_20  *(q+1) = interpolate16_3<bitFormat,2,1,1>(w[5], w[2], w[6]);
-#define PIXEL01_30  *(q+1) = interpolate16_3<bitFormat,6,1,1>(w[5], w[2], w[6]);
-#define PIXEL01_40  *(q+1) = interpolate16_3<bitFormat,14,1,1>(w[5], w[2], w[6]);
-#define PIXEL01_50  *(q+1) = interpolate16_3<bitFormat,5,2,1>(w[5], w[6], w[2]);
-#define PIXEL01_51  *(q+1) = interpolate16_3<bitFormat,5,2,1>(w[5], w[2], w[6]);
-#define PIXEL01_60  *(q+1) = interpolate16_3<bitFormat,2,3,3>(w[5], w[2], w[6]);
+#define PIXEL01_0	*(q+1) = w[5];
+#define PIXEL01_10	*(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[3]);
+#define PIXEL01_11	*(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[2]);
+#define PIXEL01_12	*(q+1) = interpolate16_2<bitFormat,3,1>(w[5], w[6]);
+#define PIXEL01_20	*(q+1) = interpolate16_3<bitFormat,2,1,1>(w[5], w[2], w[6]);
+#define PIXEL01_21	*(q+1) = interpolate16_3<bitFormat,2,1,1>(w[5], w[3], w[6]);
+#define PIXEL01_22	*(q+1) = interpolate16_3<bitFormat,2,1,1>(w[5], w[3], w[2]);
+#define PIXEL01_60	*(q+1) = interpolate16_3<bitFormat,5,2,1>(w[5], w[6], w[2]);
+#define PIXEL01_61	*(q+1) = interpolate16_3<bitFormat,5,2,1>(w[5], w[2], w[6]);
+#define PIXEL01_70	*(q+1) = interpolate16_3<bitFormat,6,1,1>(w[5], w[2], w[6]);
+#define PIXEL01_90	*(q+1) = interpolate16_3<bitFormat,2,3,3>(w[5], w[2], w[6]);
+#define PIXEL01_100	*(q+1) = interpolate16_3<bitFormat,14,1,1>(w[5], w[2], w[6]);
 
-#define PIXEL10_0   *(q+nextlineDst) = w[5];
-#define PIXEL10_10  *(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[7]);
-#define PIXEL10_11  *(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[8]);
-#define PIXEL10_12  *(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[4]);
-#define PIXEL10_20  *(q+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[8], w[4]);
-#define PIXEL10_30  *(q+nextlineDst) = interpolate16_3<bitFormat,6,1,1>(w[5], w[8], w[4]);
-#define PIXEL10_40  *(q+nextlineDst) = interpolate16_3<bitFormat,14,1,1>(w[5], w[8], w[4]);
-#define PIXEL10_50  *(q+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[4], w[8]);
-#define PIXEL10_51  *(q+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[8], w[4]);
-#define PIXEL10_60  *(q+nextlineDst) = interpolate16_3<bitFormat,2,3,3>(w[5], w[8], w[4]);
+#define PIXEL10_0	*(q+nextlineDst) = w[5];
+#define PIXEL10_10	*(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[7]);
+#define PIXEL10_11	*(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[8]);
+#define PIXEL10_12	*(q+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[4]);
+#define PIXEL10_20	*(q+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[8], w[4]);
+#define PIXEL10_21	*(q+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[7], w[4]);
+#define PIXEL10_22	*(q+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[7], w[8]);
+#define PIXEL10_60	*(q+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[4], w[8]);
+#define PIXEL10_61	*(q+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[8], w[4]);
+#define PIXEL10_70	*(q+nextlineDst) = interpolate16_3<bitFormat,6,1,1>(w[5], w[8], w[4]);
+#define PIXEL10_90	*(q+nextlineDst) = interpolate16_3<bitFormat,2,3,3>(w[5], w[8], w[4]);
+#define PIXEL10_100	*(q+nextlineDst) = interpolate16_3<bitFormat,14,1,1>(w[5], w[8], w[4]);
 
-#define PIXEL11_0   *(q+1+nextlineDst) = w[5];
-#define PIXEL11_10  *(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[9]);
-#define PIXEL11_11  *(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[6]);
-#define PIXEL11_12  *(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[8]);
-#define PIXEL11_20  *(q+1+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[6], w[8]);
-#define PIXEL11_30  *(q+1+nextlineDst) = interpolate16_3<bitFormat,6,1,1>(w[5], w[6], w[8]);
-#define PIXEL11_40  *(q+1+nextlineDst) = interpolate16_3<bitFormat,14,1,1>(w[5], w[6], w[8]);
-#define PIXEL11_50  *(q+1+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[8], w[6]);
-#define PIXEL11_51  *(q+1+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[6], w[8]);
-#define PIXEL11_60  *(q+1+nextlineDst) = interpolate16_3<bitFormat,2,3,3>(w[5], w[6], w[8]);
+#define PIXEL11_0	*(q+1+nextlineDst) = w[5];
+#define PIXEL11_10	*(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[9]);
+#define PIXEL11_11	*(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[6]);
+#define PIXEL11_12	*(q+1+nextlineDst) = interpolate16_2<bitFormat,3,1>(w[5], w[8]);
+#define PIXEL11_20	*(q+1+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[6], w[8]);
+#define PIXEL11_21	*(q+1+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[9], w[8]);
+#define PIXEL11_22	*(q+1+nextlineDst) = interpolate16_3<bitFormat,2,1,1>(w[5], w[9], w[6]);
+#define PIXEL11_60	*(q+1+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[8], w[6]);
+#define PIXEL11_61	*(q+1+nextlineDst) = interpolate16_3<bitFormat,5,2,1>(w[5], w[6], w[8]);
+#define PIXEL11_70	*(q+1+nextlineDst) = interpolate16_3<bitFormat,6,1,1>(w[5], w[6], w[8]);
+#define PIXEL11_90	*(q+1+nextlineDst) = interpolate16_3<bitFormat,2,3,3>(w[5], w[6], w[8]);
+#define PIXEL11_100	*(q+1+nextlineDst) = interpolate16_3<bitFormat,14,1,1>(w[5], w[6], w[8]);
+
+#define YUV(x)	RGBtoYUV[w[x]]
 
 /**
  * The HQ2x high quality 2x graphics filter.
@@ -74,7 +84,6 @@
 template<int bitFormat>
 void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 	int  w[10];
-	int  yuv[10];
   
 	const uint32 nextlineSrc = srcPitch / sizeof(uint16);
 	const uint16 *p = (const uint16 *)srcPtr;
@@ -94,42 +103,41 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 	//	 +----+----+----+
 
 	while (height--) {
-		w[2] = *(p - 1 - nextlineSrc);  yuv[2] = RGBtoYUV[w[2]];
-		w[5] = *(p - 1);                yuv[5] = RGBtoYUV[w[5]];
-		w[8] = *(p - 1 + nextlineSrc);  yuv[8] = RGBtoYUV[w[8]];
+		w[2] = *(p - 1 - nextlineSrc);
+		w[5] = *(p - 1);
+		w[8] = *(p - 1 + nextlineSrc);
 
-		w[3] = *(p - nextlineSrc);      yuv[3] = RGBtoYUV[w[3]];
-		w[6] = *(p);                    yuv[6] = RGBtoYUV[w[6]];
-		w[9] = *(p + nextlineSrc);      yuv[9] = RGBtoYUV[w[9]];
+		w[3] = *(p - nextlineSrc);
+		w[6] = *(p);
+		w[9] = *(p + nextlineSrc);
 
 		int tmpWidth = width;
 		while (tmpWidth--) {
 			p++;
 
-			w[1] = w[2];                yuv[1] = yuv[2];
-			w[4] = w[5];                yuv[4] = yuv[5];
-			w[7] = w[8];                yuv[7] = yuv[8];
+			w[1] = w[2];
+			w[4] = w[5];
+			w[7] = w[8];
 
-			w[2] = w[3];                yuv[2] = yuv[3];
-			w[5] = w[6];                yuv[5] = yuv[6];
-			w[8] = w[9];                yuv[8] = yuv[9];
+			w[2] = w[3];
+			w[5] = w[6];
+			w[8] = w[9];
 
-			w[3] = *(p - nextlineSrc);	yuv[3] = RGBtoYUV[w[3]];
-			w[6] = *(p);				yuv[6] = RGBtoYUV[w[6]];
-			w[9] = *(p + nextlineSrc);	yuv[9] = RGBtoYUV[w[9]];
+			w[3] = *(p - nextlineSrc);
+			w[6] = *(p);
+			w[9] = *(p + nextlineSrc);
 
 			int pattern = 0;
-			int flag = 1;
-
-			for (int k = 1; k <= 9; k++) {
-				if (k == 5) continue;
+			const int yuv5 = YUV(5);
 
-				if (w[k] != w[5]) {
-					if (diffYUV(yuv[5], yuv[k]))
-						pattern |= flag;
-				}
-				flag <<= 1;
-			}
+			if (w[5] != w[1] && diffYUV(yuv5, YUV(1))) pattern |= 0x0001;
+			if (w[5] != w[2] && diffYUV(yuv5, YUV(2))) pattern |= 0x0002;
+			if (w[5] != w[3] && diffYUV(yuv5, YUV(3))) pattern |= 0x0004;
+			if (w[5] != w[4] && diffYUV(yuv5, YUV(4))) pattern |= 0x0008;
+			if (w[5] != w[6] && diffYUV(yuv5, YUV(6))) pattern |= 0x0010;
+			if (w[5] != w[7] && diffYUV(yuv5, YUV(7))) pattern |= 0x0020;
+			if (w[5] != w[8] && diffYUV(yuv5, YUV(8))) pattern |= 0x0040;
+			if (w[5] != w[9] && diffYUV(yuv5, YUV(9))) pattern |= 0x0080;
 
 			switch (pattern) {
 			case 0:
@@ -157,8 +165,8 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 34:
 			case 130:
 			case 162:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_22
+				PIXEL01_21
 				PIXEL10_20
 				PIXEL11_20
 				break;
@@ -167,9 +175,9 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 48:
 			case 49:
 				PIXEL00_20
-				PIXEL01_10
+				PIXEL01_22
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 64:
 			case 65:
@@ -177,16 +185,16 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 69:
 				PIXEL00_20
 				PIXEL01_20
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_21
+				PIXEL11_22
 				break;
 			case 8:
 			case 12:
 			case 136:
 			case 140:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 3:
@@ -194,7 +202,7 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 131:
 			case 163:
 				PIXEL00_11
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL10_20
 				PIXEL11_20
 				break;
@@ -202,7 +210,7 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 38:
 			case 134:
 			case 166:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_12
 				PIXEL10_20
 				PIXEL11_20
@@ -214,14 +222,14 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL00_20
 				PIXEL01_11
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 144:
 			case 145:
 			case 176:
 			case 177:
 				PIXEL00_20
-				PIXEL01_10
+				PIXEL01_22
 				PIXEL10_20
 				PIXEL11_12
 				break;
@@ -231,7 +239,7 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 197:
 				PIXEL00_20
 				PIXEL01_20
-				PIXEL10_10
+				PIXEL10_21
 				PIXEL11_11
 				break;
 			case 96:
@@ -241,13 +249,13 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL00_20
 				PIXEL01_20
 				PIXEL10_12
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 40:
 			case 44:
 			case 168:
 			case 172:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
 				PIXEL10_11
 				PIXEL11_20
@@ -258,26 +266,26 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 141:
 				PIXEL00_12
 				PIXEL01_20
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 18:
 			case 50:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 80:
 			case 81:
 				PIXEL00_20
-				PIXEL01_10
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL01_22
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
 					PIXEL11_20
@@ -285,37 +293,37 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 72:
 			case 76:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 10:
 			case 138:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL01_21
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 66:
-				PIXEL00_10
-				PIXEL01_10
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL00_22
+				PIXEL01_21
+				PIXEL10_21
+				PIXEL11_22
 				break;
 			case 24:
-				PIXEL00_10
-				PIXEL01_10
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL00_21
+				PIXEL01_22
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 7:
 			case 39:
@@ -351,21 +359,21 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 22:
 			case 54:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 208:
 			case 209:
 				PIXEL00_20
-				PIXEL01_10
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL01_22
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -373,195 +381,195 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 104:
 			case 108:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 11:
 			case 139:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL01_21
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 19:
 			case 51:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_11
 					PIXEL01_10
 				} else {
-					PIXEL00_50
-					PIXEL01_60
+					PIXEL00_60
+					PIXEL01_90
 				}
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 146:
 			case 178:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 					PIXEL11_12
 				} else {
-					PIXEL01_60
-					PIXEL11_51
+					PIXEL01_90
+					PIXEL11_61
 				}
 				PIXEL10_20
 				break;
 			case 84:
 			case 85:
 				PIXEL00_20
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL01_11
 					PIXEL11_10
 				} else {
-					PIXEL01_50
-					PIXEL11_60
+					PIXEL01_60
+					PIXEL11_90
 				}
-				PIXEL10_10
+				PIXEL10_21
 				break;
 			case 112:
 			case 113:
 				PIXEL00_20
-				PIXEL01_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL01_22
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL10_12
 					PIXEL11_10
 				} else {
-					PIXEL10_51
-					PIXEL11_60
+					PIXEL10_61
+					PIXEL11_90
 				}
 				break;
 			case 200:
 			case 204:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 					PIXEL11_11
 				} else {
-					PIXEL10_60
-					PIXEL11_50
+					PIXEL10_90
+					PIXEL11_60
 				}
 				break;
 			case 73:
 			case 77:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_12
 					PIXEL10_10
 				} else {
-					PIXEL00_51
-					PIXEL10_60
+					PIXEL00_61
+					PIXEL10_90
 				}
 				PIXEL01_20
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 42:
 			case 170:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 					PIXEL10_11
 				} else {
-					PIXEL00_60
-					PIXEL10_50
+					PIXEL00_90
+					PIXEL10_60
 				}
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL11_20
 				break;
 			case 14:
 			case 142:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 					PIXEL01_12
 				} else {
-					PIXEL00_60
-					PIXEL01_51
+					PIXEL00_90
+					PIXEL01_61
 				}
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 67:
 				PIXEL00_11
-				PIXEL01_10
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL01_21
+				PIXEL10_21
+				PIXEL11_22
 				break;
 			case 70:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_12
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_21
+				PIXEL11_22
 				break;
 			case 28:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 152:
-				PIXEL00_10
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL00_21
+				PIXEL01_22
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 194:
-				PIXEL00_10
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL00_22
+				PIXEL01_21
+				PIXEL10_21
 				PIXEL11_11
 				break;
 			case 98:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_22
+				PIXEL01_21
 				PIXEL10_12
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 56:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_21
+				PIXEL01_22
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 25:
 				PIXEL00_12
-				PIXEL01_10
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL01_22
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 26:
 			case 31:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 82:
 			case 214:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -569,14 +577,14 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 88:
 			case 248:
-				PIXEL00_10
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL00_21
+				PIXEL01_22
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -584,44 +592,44 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 74:
 			case 107:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_21
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 27:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_10
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 86:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
+				PIXEL10_21
 				PIXEL11_10
 				break;
 			case 216:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_21
+				PIXEL01_22
 				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -629,38 +637,38 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 106:
 				PIXEL00_10
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_21
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 30:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 210:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_10
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 120:
-				PIXEL00_10
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL00_21
+				PIXEL01_22
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -668,321 +676,321 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_10
 				break;
 			case 75:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL10_10
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 29:
 				PIXEL00_12
 				PIXEL01_11
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_22
+				PIXEL11_21
 				break;
 			case 198:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_12
-				PIXEL10_10
+				PIXEL10_21
 				PIXEL11_11
 				break;
 			case 184:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_21
+				PIXEL01_22
 				PIXEL10_11
 				PIXEL11_12
 				break;
 			case 99:
 				PIXEL00_11
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL10_12
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 57:
 				PIXEL00_12
-				PIXEL01_10
+				PIXEL01_22
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 71:
 				PIXEL00_11
 				PIXEL01_12
-				PIXEL10_10
-				PIXEL11_10
+				PIXEL10_21
+				PIXEL11_22
 				break;
 			case 156:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 226:
-				PIXEL00_10
-				PIXEL01_10
+				PIXEL00_22
+				PIXEL01_21
 				PIXEL10_12
 				PIXEL11_11
 				break;
 			case 60:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 195:
 				PIXEL00_11
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL01_21
+				PIXEL10_21
 				PIXEL11_11
 				break;
 			case 102:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_12
 				PIXEL10_12
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 153:
 				PIXEL00_12
-				PIXEL01_10
-				PIXEL10_10
+				PIXEL01_22
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 58:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 83:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 92:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 202:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_21
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
 				PIXEL11_11
 				break;
 			case 78:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 154:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 114:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 89:
 				PIXEL00_12
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_22
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 90:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 55:
 			case 23:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_11
 					PIXEL01_0
 				} else {
-					PIXEL00_50
-					PIXEL01_60
+					PIXEL00_60
+					PIXEL01_90
 				}
 				PIXEL10_20
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 182:
 			case 150:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 					PIXEL11_12
 				} else {
-					PIXEL01_60
-					PIXEL11_51
+					PIXEL01_90
+					PIXEL11_61
 				}
 				PIXEL10_20
 				break;
 			case 213:
 			case 212:
 				PIXEL00_20
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL01_11
 					PIXEL11_0
 				} else {
-					PIXEL01_50
-					PIXEL11_60
+					PIXEL01_60
+					PIXEL11_90
 				}
-				PIXEL10_10
+				PIXEL10_21
 				break;
 			case 241:
 			case 240:
 				PIXEL00_20
-				PIXEL01_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL01_22
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL10_12
 					PIXEL11_0
 				} else {
-					PIXEL10_51
-					PIXEL11_60
+					PIXEL10_61
+					PIXEL11_90
 				}
 				break;
 			case 236:
 			case 232:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 					PIXEL11_11
 				} else {
-					PIXEL10_60
-					PIXEL11_50
+					PIXEL10_90
+					PIXEL11_60
 				}
 				break;
 			case 109:
 			case 105:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_12
 					PIXEL10_0
 				} else {
-					PIXEL00_51
-					PIXEL10_60
+					PIXEL00_61
+					PIXEL10_90
 				}
 				PIXEL01_20
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 171:
 			case 43:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 					PIXEL10_11
 				} else {
-					PIXEL00_60
-					PIXEL10_50
+					PIXEL00_90
+					PIXEL10_60
 				}
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL11_20
 				break;
 			case 143:
 			case 15:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 					PIXEL01_12
 				} else {
-					PIXEL00_60
-					PIXEL01_51
+					PIXEL00_90
+					PIXEL01_61
 				}
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_20
 				break;
 			case 124:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -990,38 +998,38 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_10
 				break;
 			case 203:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL10_10
 				PIXEL11_11
 				break;
 			case 62:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 211:
 				PIXEL00_11
 				PIXEL01_10
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 118:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
@@ -1031,9 +1039,9 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 217:
 				PIXEL00_12
-				PIXEL01_10
+				PIXEL01_22
 				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -1042,32 +1050,32 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 110:
 				PIXEL00_10
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 155:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_10
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 188:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
 				PIXEL10_11
 				PIXEL11_12
 				break;
 			case 185:
 				PIXEL00_12
-				PIXEL01_10
+				PIXEL01_22
 				PIXEL10_11
 				PIXEL11_12
 				break;
@@ -1075,28 +1083,28 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL00_12
 				PIXEL01_11
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 157:
 				PIXEL00_12
 				PIXEL01_11
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 103:
 				PIXEL00_11
 				PIXEL01_12
 				PIXEL10_12
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 227:
 				PIXEL00_11
-				PIXEL01_10
+				PIXEL01_21
 				PIXEL10_12
 				PIXEL11_11
 				break;
 			case 230:
-				PIXEL00_10
+				PIXEL00_22
 				PIXEL01_12
 				PIXEL10_12
 				PIXEL11_11
@@ -1104,45 +1112,45 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 199:
 				PIXEL00_11
 				PIXEL01_12
-				PIXEL10_10
+				PIXEL10_21
 				PIXEL11_11
 				break;
 			case 220:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 158:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 234:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_21
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -1150,161 +1158,161 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_11
 				break;
 			case 242:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 59:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 121:
 				PIXEL00_12
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_22
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 87:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 79:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 122:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 94:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 218:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 91:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 229:
@@ -1332,58 +1340,58 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_12
 				break;
 			case 186:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_11
 				PIXEL11_12
 				break;
 			case 115:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 93:
 				PIXEL00_12
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 206:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
 				PIXEL11_11
 				break;
@@ -1391,19 +1399,19 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 201:
 				PIXEL00_12
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_10
 				} else {
-					PIXEL10_30
+					PIXEL10_70
 				}
 				PIXEL11_11
 				break;
 			case 174:
 			case 46:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_10
 				} else {
-					PIXEL00_30
+					PIXEL00_70
 				}
 				PIXEL01_12
 				PIXEL10_11
@@ -1412,10 +1420,10 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 179:
 			case 147:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_10
 				} else {
-					PIXEL01_30
+					PIXEL01_70
 				}
 				PIXEL10_20
 				PIXEL11_12
@@ -1425,10 +1433,10 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL00_20
 				PIXEL01_11
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_10
 				} else {
-					PIXEL11_30
+					PIXEL11_70
 				}
 				break;
 			case 189:
@@ -1445,12 +1453,12 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 126:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -1458,48 +1466,48 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_10
 				break;
 			case 219:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_10
 				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 125:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_12
 					PIXEL10_0
 				} else {
-					PIXEL00_51
-					PIXEL10_60
+					PIXEL00_61
+					PIXEL10_90
 				}
 				PIXEL01_11
 				PIXEL11_10
 				break;
 			case 221:
 				PIXEL00_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL01_11
 					PIXEL11_0
 				} else {
-					PIXEL01_50
-					PIXEL11_60
+					PIXEL01_60
+					PIXEL11_90
 				}
 				PIXEL10_10
 				break;
 			case 207:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 					PIXEL01_12
 				} else {
-					PIXEL00_60
-					PIXEL01_51
+					PIXEL00_90
+					PIXEL01_61
 				}
 				PIXEL10_10
 				PIXEL11_11
@@ -1507,32 +1515,32 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 238:
 				PIXEL00_10
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 					PIXEL11_11
 				} else {
-					PIXEL10_60
-					PIXEL11_50
+					PIXEL10_90
+					PIXEL11_60
 				}
 				break;
 			case 190:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 					PIXEL11_12
 				} else {
-					PIXEL01_60
-					PIXEL11_51
+					PIXEL01_90
+					PIXEL11_61
 				}
 				PIXEL10_11
 				break;
 			case 187:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 					PIXEL10_11
 				} else {
-					PIXEL00_60
-					PIXEL10_50
+					PIXEL00_90
+					PIXEL10_60
 				}
 				PIXEL01_10
 				PIXEL11_12
@@ -1540,21 +1548,21 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 243:
 				PIXEL00_11
 				PIXEL01_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL10_12
 					PIXEL11_0
 				} else {
-					PIXEL10_51
-					PIXEL11_60
+					PIXEL10_61
+					PIXEL11_90
 				}
 				break;
 			case 119:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_11
 					PIXEL01_0
 				} else {
-					PIXEL00_50
-					PIXEL01_60
+					PIXEL00_60
+					PIXEL01_90
 				}
 				PIXEL10_12
 				PIXEL11_10
@@ -1563,19 +1571,19 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 233:
 				PIXEL00_12
 				PIXEL01_20
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
 				PIXEL11_11
 				break;
 			case 175:
 			case 47:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
 				PIXEL01_12
 				PIXEL10_11
@@ -1584,10 +1592,10 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 183:
 			case 151:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
 				PIXEL10_20
 				PIXEL11_12
@@ -1597,34 +1605,34 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL00_20
 				PIXEL01_11
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 250:
 				PIXEL00_10
 				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 123:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -1632,12 +1640,12 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_10
 				break;
 			case 95:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
@@ -1647,206 +1655,206 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 222:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 252:
-				PIXEL00_10
+				PIXEL00_21
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 249:
 				PIXEL00_12
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_22
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 235:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				PIXEL01_21
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
 				PIXEL11_11
 				break;
 			case 111:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				PIXEL11_10
+				PIXEL11_22
 				break;
 			case 63:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_11
-				PIXEL11_10
+				PIXEL11_21
 				break;
 			case 159:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
-				PIXEL10_10
+				PIXEL10_22
 				PIXEL11_12
 				break;
 			case 215:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
-				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				PIXEL10_21
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 246:
-				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				PIXEL00_22
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 254:
 				PIXEL00_10
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 253:
 				PIXEL00_12
 				PIXEL01_11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 251:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
 				PIXEL01_10
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
 				}
 				break;
 			case 239:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
 				PIXEL01_12
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
 				PIXEL11_11
 				break;
 			case 127:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
 					PIXEL01_20
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
 					PIXEL10_20
@@ -1854,32 +1862,32 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11_10
 				break;
 			case 191:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
 				PIXEL10_11
 				PIXEL11_12
 				break;
 			case 223:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
 					PIXEL00_20
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
 				PIXEL10_10
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
 					PIXEL11_20
@@ -1887,38 +1895,38 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 247:
 				PIXEL00_11
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
 				PIXEL10_12
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			case 255:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_0
 				} else {
-					PIXEL00_40
+					PIXEL00_100
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_0
 				} else {
-					PIXEL01_40
+					PIXEL01_100
 				}
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_0
 				} else {
-					PIXEL10_40
+					PIXEL10_100
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL11_0
 				} else {
-					PIXEL11_40
+					PIXEL11_100
 				}
 				break;
 			}
diff --git a/common/scaler/hq3x.cpp b/common/scaler/hq3x.cpp
index 6b40f676e2..864b8c31cf 100644
--- a/common/scaler/hq3x.cpp
+++ b/common/scaler/hq3x.cpp
@@ -77,6 +77,8 @@
 #define PIXEL22_5   *(q+2+nextlineDst2) = interpolate16_2<bitFormat,1,1>(w[6], w[8]);
 #define PIXEL22_C   *(q+2+nextlineDst2) = w[5];
 
+#define YUV(x)	RGBtoYUV[w[x]]
+
 /**
  * The HQ3x high quality 3x graphics filter.
  * Original author Maxim Stepin (see http://www.hiend3d.com/hq3x.html).
@@ -85,7 +87,6 @@
 template<int bitFormat>
 void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 	int   w[10];
-	int  yuv[10];
   
 	const uint32 nextlineSrc = srcPitch / sizeof(uint16);
 	const uint16 *p = (const uint16 *)srcPtr;
@@ -119,42 +120,41 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 	//	 +----+----+----+
 
 	while (height--) {
-		w[2] = *(p - 1 - nextlineSrc);  yuv[2] = RGBtoYUV[w[2]];
-		w[5] = *(p - 1);                yuv[5] = RGBtoYUV[w[5]];
-		w[8] = *(p - 1 + nextlineSrc);  yuv[8] = RGBtoYUV[w[8]];
+		w[2] = *(p - 1 - nextlineSrc);
+		w[5] = *(p - 1);
+		w[8] = *(p - 1 + nextlineSrc);
 
-		w[3] = *(p - nextlineSrc);      yuv[3] = RGBtoYUV[w[3]];
-		w[6] = *(p);                    yuv[6] = RGBtoYUV[w[6]];
-		w[9] = *(p + nextlineSrc);      yuv[9] = RGBtoYUV[w[9]];
+		w[3] = *(p - nextlineSrc);
+		w[6] = *(p);
+		w[9] = *(p + nextlineSrc);
 
 		int tmpWidth = width;
 		while (tmpWidth--) {
 			p++;
 
-			w[1] = w[2];                yuv[1] = yuv[2];
-			w[4] = w[5];                yuv[4] = yuv[5];
-			w[7] = w[8];                yuv[7] = yuv[8];
+			w[1] = w[2];
+			w[4] = w[5];
+			w[7] = w[8];
 
-			w[2] = w[3];                yuv[2] = yuv[3];
-			w[5] = w[6];                yuv[5] = yuv[6];
-			w[8] = w[9];                yuv[8] = yuv[9];
+			w[2] = w[3];
+			w[5] = w[6];
+			w[8] = w[9];
 
-			w[3] = *(p - nextlineSrc);	yuv[3] = RGBtoYUV[w[3]];
-			w[6] = *(p);				yuv[6] = RGBtoYUV[w[6]];
-			w[9] = *(p + nextlineSrc);	yuv[9] = RGBtoYUV[w[9]];
+			w[3] = *(p - nextlineSrc);
+			w[6] = *(p);
+			w[9] = *(p + nextlineSrc);
 
 			int pattern = 0;
-			int flag = 1;
-
-			for (int k = 1; k <= 9; k++) {
-				if (k == 5) continue;
+			const int yuv5 = YUV(5);
 
-				if (w[k] != w[5]) {
-					if (diffYUV(yuv[5], yuv[k]))
-						pattern |= flag;
-				}
-				flag <<= 1;
-			}
+			if (w[5] != w[1] && diffYUV(yuv5, YUV(1))) pattern |= 0x0001;
+			if (w[5] != w[2] && diffYUV(yuv5, YUV(2))) pattern |= 0x0002;
+			if (w[5] != w[3] && diffYUV(yuv5, YUV(3))) pattern |= 0x0004;
+			if (w[5] != w[4] && diffYUV(yuv5, YUV(4))) pattern |= 0x0008;
+			if (w[5] != w[6] && diffYUV(yuv5, YUV(6))) pattern |= 0x0010;
+			if (w[5] != w[7] && diffYUV(yuv5, YUV(7))) pattern |= 0x0020;
+			if (w[5] != w[8] && diffYUV(yuv5, YUV(8))) pattern |= 0x0040;
+			if (w[5] != w[9] && diffYUV(yuv5, YUV(9))) pattern |= 0x0080;
 
 			switch (pattern) {
 			case 0:
@@ -354,7 +354,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 18:
 			case 50:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_1M
 					PIXEL12_C
@@ -377,7 +377,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_1
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_1M
@@ -394,7 +394,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_2
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_1M
 					PIXEL21_C
@@ -407,7 +407,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 10:
 			case 138:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 					PIXEL01_C
 					PIXEL10_C
@@ -500,7 +500,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 22:
 			case 54:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -523,7 +523,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_1
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -540,7 +540,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_2
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -553,7 +553,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 11:
 			case 139:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -571,7 +571,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 19:
 			case 51:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_1L
 					PIXEL01_C
 					PIXEL02_1M
@@ -590,7 +590,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 146:
 			case 178:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_1M
 					PIXEL12_C
@@ -609,7 +609,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 84:
 			case 85:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL02_1U
 					PIXEL12_C
 					PIXEL21_C
@@ -628,7 +628,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 112:
 			case 113:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL20_1L
 					PIXEL21_C
@@ -647,7 +647,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 200:
 			case 204:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_1M
 					PIXEL21_C
@@ -666,7 +666,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 73:
 			case 77:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_1U
 					PIXEL10_C
 					PIXEL20_1M
@@ -685,7 +685,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 42:
 			case 170:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 					PIXEL01_C
 					PIXEL10_C
@@ -704,7 +704,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 14:
 			case 142:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 					PIXEL01_C
 					PIXEL02_1R
@@ -811,7 +811,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 26:
 			case 31:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL10_C
 				} else {
@@ -819,7 +819,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL10_3
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 					PIXEL12_C
 				} else {
@@ -834,7 +834,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 82:
 			case 214:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 				} else {
@@ -845,7 +845,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11
 				PIXEL12_C
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL21_C
 					PIXEL22_C
 				} else {
@@ -859,7 +859,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL01_1
 				PIXEL02_1M
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 				} else {
@@ -867,7 +867,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_4
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL22_C
 				} else {
@@ -877,7 +877,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 74:
 			case 107:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 				} else {
@@ -888,7 +888,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 					PIXEL21_C
 				} else {
@@ -898,7 +898,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 27:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -916,7 +916,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 86:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -938,7 +938,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -954,7 +954,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -967,7 +967,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 30:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -989,7 +989,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_1
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1005,7 +1005,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1017,7 +1017,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 75:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1166,13 +1166,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1D
 				break;
 			case 58:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1187,7 +1187,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 83:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1197,7 +1197,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1M
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -1210,20 +1210,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 202:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
@@ -1233,7 +1233,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
@@ -1242,7 +1242,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1R
 				break;
 			case 78:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
@@ -1252,7 +1252,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
@@ -1261,13 +1261,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 154:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1282,7 +1282,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 114:
 				PIXEL00_1M
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1292,7 +1292,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -1305,26 +1305,26 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 90:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1332,13 +1332,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -1346,7 +1346,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 55:
 			case 23:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_1L
 					PIXEL01_C
 					PIXEL02_C
@@ -1365,7 +1365,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 182:
 			case 150:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1384,7 +1384,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 213:
 			case 212:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL02_1U
 					PIXEL12_C
 					PIXEL21_C
@@ -1403,7 +1403,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 241:
 			case 240:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL20_1L
 					PIXEL21_C
@@ -1422,7 +1422,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 236:
 			case 232:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1441,7 +1441,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 109:
 			case 105:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_1U
 					PIXEL10_C
 					PIXEL20_C
@@ -1460,7 +1460,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 171:
 			case 43:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1479,7 +1479,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 143:
 			case 15:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL02_1R
@@ -1502,7 +1502,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1U
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1514,7 +1514,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 203:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1532,7 +1532,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 62:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1554,7 +1554,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_1
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1566,7 +1566,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 118:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1588,7 +1588,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1604,7 +1604,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1R
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1616,7 +1616,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 155:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1726,12 +1726,12 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1U
 				PIXEL10_C
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1742,12 +1742,12 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 158:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1763,7 +1763,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1D
 				break;
 			case 234:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
@@ -1772,7 +1772,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1786,7 +1786,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 242:
 				PIXEL00_1M
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1794,7 +1794,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_1
 				PIXEL11
 				PIXEL20_1L
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1805,7 +1805,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 59:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1814,7 +1814,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL01_3
 					PIXEL10_3
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -1831,7 +1831,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1840,7 +1840,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_4
 					PIXEL21_3
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -1848,7 +1848,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 87:
 				PIXEL00_1L
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1861,14 +1861,14 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11
 				PIXEL20_1M
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 79:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1880,7 +1880,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1R
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
@@ -1889,20 +1889,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 122:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
 				}
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -1911,19 +1911,19 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_4
 					PIXEL21_3
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 94:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -1934,38 +1934,38 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				PIXEL10_C
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 218:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
 				}
 				PIXEL10_C
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -1976,7 +1976,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 91:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -1985,20 +1985,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL01_3
 					PIXEL10_3
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
 				}
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -2049,13 +2049,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1D
 				break;
 			case 186:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -2070,7 +2070,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 115:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -2080,7 +2080,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -2093,20 +2093,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 206:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
@@ -2116,7 +2116,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
@@ -2132,7 +2132,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_1M
 				} else {
 					PIXEL20_2
@@ -2142,7 +2142,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 174:
 			case 46:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_1M
 				} else {
 					PIXEL00_2
@@ -2160,7 +2160,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 147:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_1M
 				} else {
 					PIXEL02_2
@@ -2182,7 +2182,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_1M
 				} else {
 					PIXEL22_2
@@ -2212,7 +2212,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 126:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -2222,7 +2222,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL12_3
 				}
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -2234,7 +2234,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 219:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -2246,7 +2246,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -2257,7 +2257,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 125:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL00_1U
 					PIXEL10_C
 					PIXEL20_C
@@ -2275,7 +2275,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 221:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL02_1U
 					PIXEL12_C
 					PIXEL21_C
@@ -2293,7 +2293,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL20_1M
 				break;
 			case 207:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL02_1R
@@ -2311,7 +2311,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1R
 				break;
 			case 238:
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -2329,7 +2329,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_1
 				break;
 			case 190:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -2347,7 +2347,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL21_1
 				break;
 			case 187:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -2365,7 +2365,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1D
 				break;
 			case 243:
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL20_1L
 					PIXEL21_C
@@ -2383,7 +2383,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11
 				break;
 			case 119:
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL00_1L
 					PIXEL01_C
 					PIXEL02_C
@@ -2408,7 +2408,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
@@ -2418,7 +2418,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 175:
 			case 47:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
@@ -2436,7 +2436,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 151:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2458,7 +2458,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
@@ -2469,7 +2469,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL01_C
 				PIXEL02_1M
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 				} else {
@@ -2477,7 +2477,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_4
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL22_C
 				} else {
@@ -2486,7 +2486,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 123:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 				} else {
@@ -2497,7 +2497,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 					PIXEL21_C
 				} else {
@@ -2507,7 +2507,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 95:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL10_C
 				} else {
@@ -2515,7 +2515,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL10_3
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 					PIXEL12_C
 				} else {
@@ -2529,7 +2529,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 222:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 				} else {
@@ -2540,7 +2540,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11
 				PIXEL12_C
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL21_C
 					PIXEL22_C
 				} else {
@@ -2554,7 +2554,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1U
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 				} else {
@@ -2562,7 +2562,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_4
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
@@ -2574,13 +2574,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL02_1M
 				PIXEL10_C
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL22_C
 				} else {
@@ -2589,7 +2589,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 235:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 				} else {
@@ -2600,7 +2600,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
@@ -2609,7 +2609,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1R
 				break;
 			case 111:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
@@ -2619,7 +2619,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 					PIXEL21_C
 				} else {
@@ -2629,13 +2629,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 63:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 					PIXEL12_C
 				} else {
@@ -2649,7 +2649,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 159:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL10_C
 				} else {
@@ -2657,7 +2657,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL10_3
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2671,7 +2671,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 215:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2680,7 +2680,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL11
 				PIXEL12_C
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL21_C
 					PIXEL22_C
 				} else {
@@ -2690,7 +2690,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 246:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 				} else {
@@ -2702,7 +2702,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
@@ -2710,7 +2710,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				break;
 			case 254:
 				PIXEL00_1M
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 				} else {
@@ -2718,14 +2718,14 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL02_4
 				}
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 				} else {
 					PIXEL10_3
 					PIXEL20_4
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL21_C
 					PIXEL22_C
@@ -2742,20 +2742,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 251:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 				} else {
@@ -2764,7 +2764,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				PIXEL02_1M
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL10_C
 					PIXEL20_C
 					PIXEL21_C
@@ -2773,7 +2773,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL20_2
 					PIXEL21_3
 				}
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL12_C
 					PIXEL22_C
 				} else {
@@ -2782,7 +2782,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				break;
 			case 239:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
@@ -2792,7 +2792,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_1
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
@@ -2801,7 +2801,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1R
 				break;
 			case 127:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL01_C
 					PIXEL10_C
@@ -2810,7 +2810,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL01_3
 					PIXEL10_3
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 					PIXEL12_C
 				} else {
@@ -2818,7 +2818,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 					PIXEL12_3
 				}
 				PIXEL11
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 					PIXEL21_C
 				} else {
@@ -2828,13 +2828,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1M
 				break;
 			case 191:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2847,14 +2847,14 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL22_1D
 				break;
 			case 223:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 					PIXEL10_C
 				} else {
 					PIXEL00_4
 					PIXEL10_3
 				}
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL01_C
 					PIXEL02_C
 					PIXEL12_C
@@ -2865,7 +2865,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				}
 				PIXEL11
 				PIXEL20_1M
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL21_C
 					PIXEL22_C
 				} else {
@@ -2876,7 +2876,7 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 			case 247:
 				PIXEL00_1L
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2886,20 +2886,20 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL12_C
 				PIXEL20_1L
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
 				}
 				break;
 			case 255:
-				if (diffYUV(yuv[4], yuv[2])) {
+				if (diffYUV(YUV(4), YUV(2))) {
 					PIXEL00_C
 				} else {
 					PIXEL00_2
 				}
 				PIXEL01_C
-				if (diffYUV(yuv[2], yuv[6])) {
+				if (diffYUV(YUV(2), YUV(6))) {
 					PIXEL02_C
 				} else {
 					PIXEL02_2
@@ -2907,13 +2907,13 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 				PIXEL10_C
 				PIXEL11
 				PIXEL12_C
-				if (diffYUV(yuv[8], yuv[4])) {
+				if (diffYUV(YUV(8), YUV(4))) {
 					PIXEL20_C
 				} else {
 					PIXEL20_2
 				}
 				PIXEL21_C
-				if (diffYUV(yuv[6], yuv[8])) {
+				if (diffYUV(YUV(6), YUV(8))) {
 					PIXEL22_C
 				} else {
 					PIXEL22_2
diff --git a/common/scaler/intern.h b/common/scaler/intern.h
index e65e6f311a..02773bca93 100644
--- a/common/scaler/intern.h
+++ b/common/scaler/intern.h
@@ -95,11 +95,28 @@ static inline bool diffYUV(int yuv1, int yuv2) {
 	static const int trY   = 0x00300000;
 	static const int trU   = 0x00000700;
 	static const int trV   = 0x00000006;
-
+	
+	int diff;
+	
+	diff = ((yuv1 & Ymask) - (yuv2 & Ymask));
+	if (diff < 0) diff = - diff;
+	if (diff > trY) return true;
+
+	diff = ((yuv1 & Umask) - (yuv2 & Umask));
+	if (diff < 0) diff = - diff;
+	if (diff > trU) return true;
+
+	diff = ((yuv1 & Vmask) - (yuv2 & Vmask));
+	if (diff < 0) diff = - diff;
+	if (diff > trV) return true;
+
+	return false;
+/*
 	return
 	  ( ( ABS((yuv1 & Ymask) - (yuv2 & Ymask)) > trY ) ||
 	    ( ABS((yuv1 & Umask) - (yuv2 & Umask)) > trU ) ||
 	    ( ABS((yuv1 & Vmask) - (yuv2 & Vmask)) > trV ) );
+*/
 }
 
 /**