@@ -1045,7 +1045,7 @@ static void fillPlane(uint8_t* plane, int stride, int width, int height, int y,

     }

 }

-static inline void rgb48ToY(uint8_t *dst, const uint8_t *src, long width,

+static inline void rgb48ToY(int16_t *dst, const uint8_t *src, long width,

                             uint32_t *unused)

 {

     int i;

@@ -1054,11 +1054,11 @@ static inline void rgb48ToY(uint8_t *dst, const uint8_t *src, long width,

         int g = src[i*6+2];

         int b = src[i*6+4];

-        dst[i] = (RY*r + GY*g + BY*b + (33<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

+        dst[i] = (RY*r + GY*g + BY*b + (32<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

     }

 }

-static inline void rgb48ToUV(uint8_t *dstU, uint8_t *dstV,

+static inline void rgb48ToUV(int16_t *dstU, int16_t *dstV,

                              const uint8_t *src1, const uint8_t *src2,

                              long width, uint32_t *unused)

 {

@@ -1069,12 +1069,12 @@ static inline void rgb48ToUV(uint8_t *dstU, uint8_t *dstV,

         int g = src1[6*i + 2];

         int b = src1[6*i + 4];

-        dstU[i] = (RU*r + GU*g + BU*b + (257<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

-        dstV[i] = (RV*r + GV*g + BV*b + (257<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

+        dstU[i] = (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

+        dstV[i] = (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

     }

 }

-static inline void rgb48ToUV_half(uint8_t *dstU, uint8_t *dstV,

+static inline void rgb48ToUV_half(int16_t *dstU, int16_t *dstV,

                                   const uint8_t *src1, const uint8_t *src2,

                                   long width, uint32_t *unused)

 {

@@ -1085,12 +1085,12 @@ static inline void rgb48ToUV_half(uint8_t *dstU, uint8_t *dstV,

         int g= src1[12*i + 2] + src1[12*i + 8];

         int b= src1[12*i + 4] + src1[12*i + 10];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<RGB2YUV_SHIFT)) >> (RGB2YUV_SHIFT+1);

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<RGB2YUV_SHIFT)) >> (RGB2YUV_SHIFT+1);

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT)) + (1<<(RGB2YUV_SHIFT-6))) >> (RGB2YUV_SHIFT-5);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT)) + (1<<(RGB2YUV_SHIFT-6))) >> (RGB2YUV_SHIFT-5);

     }

 }

-static inline void bgr48ToY(uint8_t *dst, const uint8_t *src, long width,

+static inline void bgr48ToY(int16_t *dst, const uint8_t *src, long width,

                             uint32_t *unused)

 {

     int i;

@@ -1099,11 +1099,11 @@ static inline void bgr48ToY(uint8_t *dst, const uint8_t *src, long width,

         int g = src[i*6+2];

         int r = src[i*6+4];

-        dst[i] = (RY*r + GY*g + BY*b + (33<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

+        dst[i] = (RY*r + GY*g + BY*b + (32<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

     }

 }

-static inline void bgr48ToUV(uint8_t *dstU, uint8_t *dstV,

+static inline void bgr48ToUV(int16_t *dstU, int16_t *dstV,

                              const uint8_t *src1, const uint8_t *src2,

                              long width, uint32_t *unused)

 {

@@ -1113,12 +1113,12 @@ static inline void bgr48ToUV(uint8_t *dstU, uint8_t *dstV,

         int g = src1[6*i + 2];

         int r = src1[6*i + 4];

-        dstU[i] = (RU*r + GU*g + BU*b + (257<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

-        dstV[i] = (RV*r + GV*g + BV*b + (257<<(RGB2YUV_SHIFT-1))) >> RGB2YUV_SHIFT;

+        dstU[i] = (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

+        dstV[i] = (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7))) >> (RGB2YUV_SHIFT-6);

     }

 }

-static inline void bgr48ToUV_half(uint8_t *dstU, uint8_t *dstV,

+static inline void bgr48ToUV_half(int16_t *dstU, int16_t *dstV,

                                   const uint8_t *src1, const uint8_t *src2,

                                   long width, uint32_t *unused)

 {

@@ -1128,13 +1128,13 @@ static inline void bgr48ToUV_half(uint8_t *dstU, uint8_t *dstV,

         int g= src1[12*i + 2] + src1[12*i + 8];

         int r= src1[12*i + 4] + src1[12*i + 10];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<RGB2YUV_SHIFT)) >> (RGB2YUV_SHIFT+1);

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<RGB2YUV_SHIFT)) >> (RGB2YUV_SHIFT+1);

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT)) + (1<<(RGB2YUV_SHIFT-6))) >> (RGB2YUV_SHIFT-5);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT)) + (1<<(RGB2YUV_SHIFT-6))) >> (RGB2YUV_SHIFT-5);

     }

 }

 #define BGR2Y(type, name, shr, shg, shb, maskr, maskg, maskb, RY, GY, BY, S)\

-static inline void name(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)\

+static inline void name(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)\

 {\

     int i;\

     for (i=0; i<width; i++) {\

@@ -1142,7 +1142,7 @@ static inline void name(uint8_t *dst, const uint8_t *src, long width, uint32_t *

         int g= (((const type*)src)[i]>>shg)&maskg;\

         int r= (((const type*)src)[i]>>shr)&maskr;\

 \

-        dst[i]= (((RY)*r + (GY)*g + (BY)*b + (33<<((S)-1)))>>(S));\

+        dst[i]= (((RY)*r + (GY)*g + (BY)*b + (32<<((S)-1)) + (1<<(S-7)))>>((S)-6));\

     }\

 }

@@ -1155,16 +1155,16 @@ BGR2Y(uint16_t, bgr15ToY, 0, 0, 0, 0x001F, 0x03E0, 0x7C00, RY<<10, GY<<5, BY

 BGR2Y(uint16_t, rgb16ToY, 0, 0, 0, 0xF800, 0x07E0, 0x001F, RY    , GY<<5, BY<<11, RGB2YUV_SHIFT+8)

 BGR2Y(uint16_t, rgb15ToY, 0, 0, 0, 0x7C00, 0x03E0, 0x001F, RY    , GY<<5, BY<<10, RGB2YUV_SHIFT+7)

-static inline void abgrToA(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)

+static inline void abgrToA(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)

 {

     int i;

     for (i=0; i<width; i++) {

-        dst[i]= src[4*i];

+        dst[i]= src[4*i]<<6;

     }

 }

 #define BGR2UV(type, name, shr, shg, shb, shp, maskr, maskg, maskb, RU, GU, BU, RV, GV, BV, S) \

-static inline void name(uint8_t *dstU, uint8_t *dstV, const uint8_t *src, const uint8_t *dummy, long width, uint32_t *unused)\

+static inline void name(int16_t *dstU, int16_t *dstV, const uint8_t *src, const uint8_t *dummy, long width, uint32_t *unused)\

 {\

     int i;\

     for (i=0; i<width; i++) {\

@@ -1172,11 +1172,11 @@ static inline void name(uint8_t *dstU, uint8_t *dstV, const uint8_t *src, const

         int g= ((((const type*)src)[i]>>shp)&maskg)>>shg;\

         int r= ((((const type*)src)[i]>>shp)&maskr)>>shr;\

 \

-        dstU[i]= ((RU)*r + (GU)*g + (BU)*b + (257<<((S)-1)))>>(S);\

-        dstV[i]= ((RV)*r + (GV)*g + (BV)*b + (257<<((S)-1)))>>(S);\

+        dstU[i]= ((RU)*r + (GU)*g + (BU)*b + (256<<((S)-1)) + (1<<(S-7)))>>((S)-6);\

+        dstV[i]= ((RV)*r + (GV)*g + (BV)*b + (256<<((S)-1)) + (1<<(S-7)))>>((S)-6);\

     }\

 }\

-static inline void name ## _half(uint8_t *dstU, uint8_t *dstV, const uint8_t *src, const uint8_t *dummy, long width, uint32_t *unused)\

+static inline void name ## _half(int16_t *dstU, int16_t *dstV, const uint8_t *src, const uint8_t *dummy, long width, uint32_t *unused)\

 {\

     int i;\

     for (i=0; i<width; i++) {\

@@ -1189,8 +1189,8 @@ static inline void name ## _half(uint8_t *dstU, uint8_t *dstV, const uint8_t *sr

 \

         g>>=shg;\

 \

-        dstU[i]= ((RU)*r + (GU)*g + (BU)*b + (257<<(S)))>>((S)+1);\

-        dstV[i]= ((RV)*r + (GV)*g + (BV)*b + (257<<(S)))>>((S)+1);\

+        dstU[i]= ((RU)*r + (GU)*g + (BU)*b + (256U<<(S)) + (1<<(S-6)))>>((S)-6+1);\

+        dstV[i]= ((RV)*r + (GV)*g + (BV)*b + (256U<<(S)) + (1<<(S-6)))>>((S)-6+1);\

     }\

 }

@@ -1203,27 +1203,27 @@ BGR2UV(uint16_t, bgr15ToUV, 0, 0, 0, 0,   0x001F, 0x03E0,   0x7C00, RU<<10, GU<<

 BGR2UV(uint16_t, rgb16ToUV, 0, 0, 0, 0,   0xF800, 0x07E0,   0x001F, RU    , GU<<5, BU<<11, RV    , GV<<5, BV<<11, RGB2YUV_SHIFT+8)

 BGR2UV(uint16_t, rgb15ToUV, 0, 0, 0, 0,   0x7C00, 0x03E0,   0x001F, RU    , GU<<5, BU<<10, RV    , GV<<5, BV<<10, RGB2YUV_SHIFT+7)

-static inline void palToA(uint8_t *dst, const uint8_t *src, long width, uint32_t *pal)

+static inline void palToA(int16_t *dst, const uint8_t *src, long width, uint32_t *pal)

 {

     int i;

     for (i=0; i<width; i++) {

         int d= src[i];

-        dst[i]= pal[d] >> 24;

+        dst[i]= (pal[d] >> 24)<<6;

     }

 }

-static inline void palToY(uint8_t *dst, const uint8_t *src, long width, uint32_t *pal)

+static inline void palToY(int16_t *dst, const uint8_t *src, long width, uint32_t *pal)

 {

     int i;

     for (i=0; i<width; i++) {

         int d= src[i];

-        dst[i]= pal[d] & 0xFF;

+        dst[i]= (pal[d] & 0xFF)<<6;

     }

 }

-static inline void palToUV(uint8_t *dstU, uint8_t *dstV,

+static inline void palToUV(uint16_t *dstU, int16_t *dstV,

                            const uint8_t *src1, const uint8_t *src2,

                            long width, uint32_t *pal)

 {

@@ -1232,28 +1232,28 @@ static inline void palToUV(uint8_t *dstU, uint8_t *dstV,

     for (i=0; i<width; i++) {

         int p= pal[src1[i]];

-        dstU[i]= p>>8;

-        dstV[i]= p>>16;

+        dstU[i]= (uint8_t)(p>> 8)<<6;

+        dstV[i]= (uint8_t)(p>>16)<<6;

     }

 }

-static inline void monowhite2Y(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)

+static inline void monowhite2Y(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)

 {

     int i, j;

     for (i=0; i<width/8; i++) {

         int d= ~src[i];

         for(j=0; j<8; j++)

-            dst[8*i+j]= ((d>>(7-j))&1)*255;

+            dst[8*i+j]= ((d>>(7-j))&1)*16383;

     }

 }

-static inline void monoblack2Y(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)

+static inline void monoblack2Y(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)

 {

     int i, j;

     for (i=0; i<width/8; i++) {

         int d= src[i];

         for(j=0; j<8; j++)

-            dst[8*i+j]= ((d>>(7-j))&1)*255;

+            dst[8*i+j]= ((d>>(7-j))&1)*16383;

     }

 }

@@ -254,7 +254,7 @@ static inline void nv21ToUV_c(uint8_t *dstU, uint8_t *dstV,

     nvXXtoUV_c(dstV, dstU, src1, width);

 }

-static inline void bgr24ToY_c(uint8_t *dst, const uint8_t *src,

+static inline void bgr24ToY_c(int16_t *dst, const uint8_t *src,

                               long width, uint32_t *unused)

 {

     int i;

@@ -263,11 +263,11 @@ static inline void bgr24ToY_c(uint8_t *dst, const uint8_t *src,

         int g= src[i*3+1];

         int r= src[i*3+2];

-        dst[i]= ((RY*r + GY*g + BY*b + (33<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT);

+        dst[i]= ((RY*r + GY*g + BY*b + (32<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6));

     }

 }

-static inline void bgr24ToUV_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1,

+static inline void bgr24ToUV_c(int16_t *dstU, int16_t *dstV, const uint8_t *src1,

                                const uint8_t *src2, long width, uint32_t *unused)

 {

     int i;

@@ -276,13 +276,13 @@ static inline void bgr24ToUV_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1

         int g= src1[3*i + 1];

         int r= src1[3*i + 2];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT;

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT;

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6);

     }

     assert(src1 == src2);

 }

-static inline void bgr24ToUV_half_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1,

+static inline void bgr24ToUV_half_c(int16_t *dstU, int16_t *dstV, const uint8_t *src1,

                                     const uint8_t *src2, long width, uint32_t *unused)

 {

     int i;

@@ -291,13 +291,13 @@ static inline void bgr24ToUV_half_c(uint8_t *dstU, uint8_t *dstV, const uint8_t

         int g= src1[6*i + 1] + src1[6*i + 4];

         int r= src1[6*i + 2] + src1[6*i + 5];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<RGB2YUV_SHIFT))>>(RGB2YUV_SHIFT+1);

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<RGB2YUV_SHIFT))>>(RGB2YUV_SHIFT+1);

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<RGB2YUV_SHIFT) + (1<<(RGB2YUV_SHIFT-6)))>>(RGB2YUV_SHIFT-5);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<RGB2YUV_SHIFT) + (1<<(RGB2YUV_SHIFT-6)))>>(RGB2YUV_SHIFT-5);

     }

     assert(src1 == src2);

 }

-static inline void rgb24ToY_c(uint8_t *dst, const uint8_t *src, long width,

+static inline void rgb24ToY_c(int16_t *dst, const uint8_t *src, long width,

                               uint32_t *unused)

 {

     int i;

@@ -306,11 +306,11 @@ static inline void rgb24ToY_c(uint8_t *dst, const uint8_t *src, long width,

         int g= src[i*3+1];

         int b= src[i*3+2];

-        dst[i]= ((RY*r + GY*g + BY*b + (33<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT);

+        dst[i]= ((RY*r + GY*g + BY*b + (32<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6));

     }

 }

-static inline void rgb24ToUV_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1,

+static inline void rgb24ToUV_c(int16_t *dstU, int16_t *dstV, const uint8_t *src1,

                                const uint8_t *src2, long width, uint32_t *unused)

 {

     int i;

@@ -320,12 +320,12 @@ static inline void rgb24ToUV_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1

         int g= src1[3*i + 1];

         int b= src1[3*i + 2];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT;

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<(RGB2YUV_SHIFT-1)))>>RGB2YUV_SHIFT;

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<(RGB2YUV_SHIFT-1)) + (1<<(RGB2YUV_SHIFT-7)))>>(RGB2YUV_SHIFT-6);

     }

 }

-static inline void rgb24ToUV_half_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1,

+static inline void rgb24ToUV_half_c(int16_t *dstU, int16_t *dstV, const uint8_t *src1,

                                     const uint8_t *src2, long width, uint32_t *unused)

 {

     int i;

@@ -335,8 +335,8 @@ static inline void rgb24ToUV_half_c(uint8_t *dstU, uint8_t *dstV, const uint8_t

         int g= src1[6*i + 1] + src1[6*i + 4];

         int b= src1[6*i + 2] + src1[6*i + 5];

-        dstU[i]= (RU*r + GU*g + BU*b + (257<<RGB2YUV_SHIFT))>>(RGB2YUV_SHIFT+1);

-        dstV[i]= (RV*r + GV*g + BV*b + (257<<RGB2YUV_SHIFT))>>(RGB2YUV_SHIFT+1);

+        dstU[i]= (RU*r + GU*g + BU*b + (256<<RGB2YUV_SHIFT) + (1<<(RGB2YUV_SHIFT-6)))>>(RGB2YUV_SHIFT-5);

+        dstV[i]= (RV*r + GV*g + BV*b + (256<<RGB2YUV_SHIFT) + (1<<(RGB2YUV_SHIFT-6)))>>(RGB2YUV_SHIFT-5);

     }

 }

@@ -455,7 +455,8 @@ static inline void hyscale_c(SwsContext *c, uint16_t *dst, long dstWidth,

     }

     if (c->hScale16) {

-        c->hScale16(dst, dstWidth, (uint16_t*)src, srcW, xInc, hLumFilter, hLumFilterPos, hLumFilterSize, av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1);

+        int shift= isAnyRGB(c->srcFormat) || c->srcFormat==PIX_FMT_PAL8 ? 13 : av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1;

+        c->hScale16(dst, dstWidth, (uint16_t*)src, srcW, xInc, hLumFilter, hLumFilterPos, hLumFilterSize, shift);

     } else if (!c->hyscale_fast) {

         c->hScale(dst, dstWidth, src, srcW, xInc, hLumFilter, hLumFilterPos, hLumFilterSize);

     } else { // fast bilinear upscale / crap downscale

@@ -502,8 +503,9 @@ inline static void hcscale_c(SwsContext *c, uint16_t *dst, long dstWidth,

     }

     if (c->hScale16) {

-        c->hScale16(dst     , dstWidth, (uint16_t*)src1, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize, av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1);

-        c->hScale16(dst+VOFW, dstWidth, (uint16_t*)src2, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize, av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1);

+        int shift= isAnyRGB(c->srcFormat) || c->srcFormat==PIX_FMT_PAL8 ? 13 : av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1;

+        c->hScale16(dst     , dstWidth, (uint16_t*)src1, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize, shift);

+        c->hScale16(dst+VOFW, dstWidth, (uint16_t*)src2, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize, shift);

     } else if (!c->hcscale_fast) {

         c->hScale(dst     , dstWidth, src1, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize);

         c->hScale(dst+VOFW, dstWidth, src2, srcW, xInc, hChrFilter, hChrFilterPos, hChrFilterSize);

@@ -959,6 +961,9 @@ static void sws_init_swScale_c(SwsContext *c)

         }

     }

+    if(isAnyRGB(c->srcFormat) || c->srcFormat == PIX_FMT_PAL8)

+        c->hScale16= hScale16_c;

+

     switch (srcFormat) {

     case PIX_FMT_GRAY8A :

         c->alpSrcOffset = 1;

@@ -860,7 +860,7 @@ int sws_init_context(SwsContext *c, SwsFilter *srcFilter, SwsFilter *dstFilter)

             if (flags&SWS_PRINT_INFO)

                 av_log(c, AV_LOG_INFO, "output width is not a multiple of 32 -> no MMX2 scaler\n");

         }

-        if (usesHFilter || isNBPS(c->srcFormat) || is16BPS(c->srcFormat)) c->canMMX2BeUsed=0;

+        if (usesHFilter || isNBPS(c->srcFormat) || is16BPS(c->srcFormat) || isAnyRGB(c->srcFormat)) c->canMMX2BeUsed=0;

     }

     else

         c->canMMX2BeUsed=0;

@@ -1615,7 +1615,7 @@ static inline void RENAME(nv21ToUV)(uint8_t *dstU, uint8_t *dstV,

     RENAME(nvXXtoUV)(dstV, dstU, src1, width);

 }

-static inline void RENAME(bgr24ToY_mmx)(uint8_t *dst, const uint8_t *src, long width, enum PixelFormat srcFormat)

+static inline void RENAME(bgr24ToY_mmx)(int16_t *dst, const uint8_t *src, long width, enum PixelFormat srcFormat)

 {

     if(srcFormat == PIX_FMT_BGR24) {

@@ -1655,20 +1655,19 @@ static inline void RENAME(bgr24ToY_mmx)(uint8_t *dst, const uint8_t *src, long w

         "paddd                   %%mm3, %%mm2       \n\t"

         "paddd                   %%mm4, %%mm0       \n\t"

         "paddd                   %%mm4, %%mm2       \n\t"

-        "psrad                     $15, %%mm0       \n\t"

-        "psrad                     $15, %%mm2       \n\t"

+        "psrad                     $9, %%mm0       \n\t"

+        "psrad                     $9, %%mm2       \n\t"

         "packssdw                %%mm2, %%mm0       \n\t"

-        "packuswb                %%mm0, %%mm0       \n\t"

-        "movd                %%mm0, (%1, %%"REG_a") \n\t"

-        "add                        $4, %%"REG_a"   \n\t"

+        "movq                %%mm0, (%1, %%"REG_a") \n\t"

+        "add                        $8, %%"REG_a"   \n\t"

         " js                        1b              \n\t"

     : "+r" (src)

-    : "r" (dst+width), "g" ((x86_reg)-width)

+    : "r" (dst+width), "g" ((x86_reg)-2*width)

     : "%"REG_a

     );

 }

-static inline void RENAME(bgr24ToUV_mmx)(uint8_t *dstU, uint8_t *dstV, const uint8_t *src, long width, enum PixelFormat srcFormat)

+static inline void RENAME(bgr24ToUV_mmx)(int16_t *dstU, int16_t *dstV, const uint8_t *src, long width, enum PixelFormat srcFormat)

 {

     __asm__ volatile(

         "movq                    24(%4), %%mm6       \n\t"

@@ -1708,41 +1707,39 @@ static inline void RENAME(bgr24ToUV_mmx)(uint8_t *dstU, uint8_t *dstV, const uin

         "paddd                   %%mm3, %%mm2       \n\t"

         "paddd                   %%mm3, %%mm1       \n\t"

         "paddd                   %%mm3, %%mm4       \n\t"

-        "psrad                     $15, %%mm0       \n\t"

-        "psrad                     $15, %%mm2       \n\t"

-        "psrad                     $15, %%mm1       \n\t"

-        "psrad                     $15, %%mm4       \n\t"

+        "psrad                     $9, %%mm0       \n\t"

+        "psrad                     $9, %%mm2       \n\t"

+        "psrad                     $9, %%mm1       \n\t"

+        "psrad                     $9, %%mm4       \n\t"

         "packssdw                %%mm1, %%mm0       \n\t"

         "packssdw                %%mm4, %%mm2       \n\t"

-        "packuswb                %%mm0, %%mm0       \n\t"

-        "packuswb                %%mm2, %%mm2       \n\t"

-        "movd                %%mm0, (%1, %%"REG_a") \n\t"

-        "movd                %%mm2, (%2, %%"REG_a") \n\t"

-        "add                        $4, %%"REG_a"   \n\t"

+        "movq                %%mm0, (%1, %%"REG_a") \n\t"

+        "movq                %%mm2, (%2, %%"REG_a") \n\t"

+        "add                        $8, %%"REG_a"   \n\t"

         " js                        1b              \n\t"

     : "+r" (src)

-    : "r" (dstU+width), "r" (dstV+width), "g" ((x86_reg)-width), "r"(ff_bgr24toUV[srcFormat == PIX_FMT_RGB24])

+    : "r" (dstU+width), "r" (dstV+width), "g" ((x86_reg)-2*width), "r"(ff_bgr24toUV[srcFormat == PIX_FMT_RGB24])

     : "%"REG_a

     );

 }

-static inline void RENAME(bgr24ToY)(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)

+static inline void RENAME(bgr24ToY)(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)

 {

     RENAME(bgr24ToY_mmx)(dst, src, width, PIX_FMT_BGR24);

 }

-static inline void RENAME(bgr24ToUV)(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1, const uint8_t *src2, long width, uint32_t *unused)

+static inline void RENAME(bgr24ToUV)(int16_t *dstU, int16_t *dstV, const uint8_t *src1, const uint8_t *src2, long width, uint32_t *unused)

 {

     RENAME(bgr24ToUV_mmx)(dstU, dstV, src1, width, PIX_FMT_BGR24);

     assert(src1 == src2);

 }

-static inline void RENAME(rgb24ToY)(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)

+static inline void RENAME(rgb24ToY)(int16_t *dst, const uint8_t *src, long width, uint32_t *unused)

 {

     RENAME(bgr24ToY_mmx)(dst, src, width, PIX_FMT_RGB24);

 }

-static inline void RENAME(rgb24ToUV)(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1, const uint8_t *src2, long width, uint32_t *unused)

+static inline void RENAME(rgb24ToUV)(int16_t *dstU, int16_t *dstV, const uint8_t *src1, const uint8_t *src2, long width, uint32_t *unused)

 {

     assert(src1==src2);

     RENAME(bgr24ToUV_mmx)(dstU, dstV, src1, width, PIX_FMT_RGB24);

@@ -2323,7 +2320,7 @@ static void RENAME(sws_init_swScale)(SwsContext *c)

         case PIX_FMT_YUV420P16LE:

         case PIX_FMT_YUV422P16LE:

         case PIX_FMT_YUV444P16LE: c->hScale16= RENAME(hScale16); break;

-    }   

+    }

     if (!c->chrSrcHSubSample) {

         switch(srcFormat) {

         case PIX_FMT_BGR24  : c->chrToYV12 = RENAME(bgr24ToUV); break;

@@ -2348,4 +2345,7 @@ static void RENAME(sws_init_swScale)(SwsContext *c)

         default: break;

         }

     }

+

+    if(isAnyRGB(c->srcFormat))

+        c->hScale16= RENAME(hScale16);

 }

@@ -67,13 +67,13 @@ DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toY1Coeff) = 0x0C88000040870C88ULL;

 DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toY2Coeff) = 0x20DE4087000020DEULL;

 DECLARE_ASM_CONST(8, uint64_t, ff_rgb24toY1Coeff) = 0x20DE0000408720DEULL;

 DECLARE_ASM_CONST(8, uint64_t, ff_rgb24toY2Coeff) = 0x0C88408700000C88ULL;

-DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toYOffset) = 0x0008400000084000ULL;

+DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toYOffset) = 0x0008010000080100ULL;

 DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toUV)[2][4] = {

     {0x38380000DAC83838ULL, 0xECFFDAC80000ECFFULL, 0xF6E40000D0E3F6E4ULL, 0x3838D0E300003838ULL},

     {0xECFF0000DAC8ECFFULL, 0x3838DAC800003838ULL, 0x38380000D0E33838ULL, 0xF6E4D0E30000F6E4ULL},

 };

-DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toUVOffset)= 0x0040400000404000ULL;

+DECLARE_ASM_CONST(8, uint64_t, ff_bgr24toUVOffset)= 0x0040010000400100ULL;

 #endif /* SWSCALE_X86_SWSCALE_TEMPLATE_H */

@@ -16,15 +16,15 @@ e176bd14185788110e055f945de7f95f *./tests/data/pixfmt/yuvj420p.yuv

 304128 ./tests/data/pixfmt/yuvj422p.yuv

 c10442da177c9f1d12be3c53be6fa12c *./tests/data/pixfmt/yuvj444p.yuv

 304128 ./tests/data/pixfmt/yuvj444p.yuv

-c6e0f9b5817f484b175c1ec4ffb4e9c9 *./tests/data/pixfmt/rgb24.yuv

+6bb61113e7b70eb09dbcec356122a0e2 *./tests/data/pixfmt/rgb24.yuv

 304128 ./tests/data/pixfmt/rgb24.yuv

-c6e0f9b5817f484b175c1ec4ffb4e9c9 *./tests/data/pixfmt/bgr24.yuv

+6bb61113e7b70eb09dbcec356122a0e2 *./tests/data/pixfmt/bgr24.yuv

 304128 ./tests/data/pixfmt/bgr24.yuv

-c6e0f9b5817f484b175c1ec4ffb4e9c9 *./tests/data/pixfmt/rgb32.yuv

+6bb61113e7b70eb09dbcec356122a0e2 *./tests/data/pixfmt/rgb32.yuv

 304128 ./tests/data/pixfmt/rgb32.yuv

-66d39d464bd89ded2a124897f0a75ade *./tests/data/pixfmt/rgb565.yuv

+efa7c0337cc00c796c6df615223716f1 *./tests/data/pixfmt/rgb565.yuv

 304128 ./tests/data/pixfmt/rgb565.yuv

-c894c3bd8d2631ed1964500b90a0c350 *./tests/data/pixfmt/rgb555.yuv

+0df2a477af1415a1b8fbf2a3e552bc39 *./tests/data/pixfmt/rgb555.yuv

 304128 ./tests/data/pixfmt/rgb555.yuv

 6be306b0cce5f8e6c271ea17fef9745b *./tests/data/pixfmt/gray.yuv

 304128 ./tests/data/pixfmt/gray.yuv

@@ -1,28 +1,28 @@

-abgr                d894cb97f6c80eb21bdbe8a4eea62d86

-argb                54346f2b2eef10919e0f247241df3b24

-bgr24               570f8d6b51a838aed022ef67535f6bdc

-bgr48be             fcc0f2dbf45d325f84f816c74cbeeebe

-bgr48le             3f9c2b23eed3b8d196d1c14b38ce50f5

+abgr                cff82561a074874027ac1cc896fd2730

+argb                756dd1eaa5baca2238ce23dbdc452684

+bgr24               e44192347a45586c6c157e3059610cd1

+bgr48be             62e6043fbe9734e63ad679999ca8011c

+bgr48le             61237dad4fa5f3e9109db85f53cd25d9

 bgr4_byte           ee1d35a7baf8e9016891929a2f565c0b

-bgr555le            36b745067197f9ca8c1731cac51329c9

-bgr565le            3a514a298c6161a071ddf9963c06509d

+bgr555le            41e3e0961478dc634bf68a7bbd670cc9

+bgr565le            614897eaeb422bd9a972f8ee51909be5

 bgr8                7f007fa6c153a16e808a9c51605a4016

-bgra                a5e7040f9a80cccd65e5acf2ca09ace5

+bgra                01cfdda1f72fcabb6c46424e27f8c519

 gray                d7786a7d9d99ac74230cc045cab5632c

 gray16be            5ba22d4802b40ec27e62abb22ad1d1cc

 gray16le            2d5e83aa875a4c3baa6fecf55e3223bf

-monob               88c4c050758e64d120f50c7eff694381

-monow               d31772ebaa877fc2a78565937f7f9673

+monob               cb62f31b701c6e987b574974d1b31e32

+monow               fd5d417ab7728acddffc06870661df61

 nv12                4676d59db43d657dc12841f6bc3ab452

 nv21                69c699510ff1fb777b118ebee1002f14

-rgb24               514692e28e8ff6860e415ce4fcf6eb8c

-rgb48be             1894cd30dabcd3180518e4d5f09f25e7

-rgb48le             1354e6e27ce3c1d4d4989ee56030c94b

+rgb24               13ff53ebeab74dc05492836f1cfbd2c1

+rgb48be             7f6b1f8139c6a64eadf9dfa867ac20e8

+rgb48le             2756d8710c152cbc367656de4d0f1b76

 rgb4_byte           d81ffd3add95842a618eec81024f0b5c

-rgb555le            a350ef1dc2c9688ed49e7ba018843795

-rgb565le            6f4bb711238baa762d73305213f8d035

+rgb555le            bd698d86c03170c4a16607c0fd1f750f

+rgb565le            bfa0c639d80c3c03fd0c9e5f34296a5e

 rgb8                091d0170b354ef0e97312b95feb5483f

-rgba                a3d362f222098a00e63867f612018659

+rgba                16873e3ac914e76116629a5ff8940ac4

 uyvy422             314bd486277111a95d9369b944fa0400

 yuv410p             7df8f6d69b56a8dcb6c7ee908e5018b5

 yuv411p             1143e7c5cc28fe0922b051b17733bc4c

@@ -1,4 +1,4 @@

 97894502b4cb57aca1105b6333f72dae *./tests/data/vsynth1/flashsv.flv

 14681925 ./tests/data/vsynth1/flashsv.flv

-947cb24ec45a453348ae6fe3fa278071 *./tests/data/flashsv.vsynth1.out.yuv

-stddev:    2.85 PSNR: 39.03 MAXDIFF:   49 bytes:  7603200/  7603200

+791e1fb999deb2e4156e2286d48c4ed1 *./tests/data/flashsv.vsynth1.out.yuv

+stddev:    2.84 PSNR: 39.04 MAXDIFF:   49 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 519e26bb1ac0f3db8f90b36537f2f760 *./tests/data/vsynth1/jpegls.avi

 9089812 ./tests/data/vsynth1/jpegls.avi

-947cb24ec45a453348ae6fe3fa278071 *./tests/data/jpegls.vsynth1.out.yuv

-stddev:    2.85 PSNR: 39.03 MAXDIFF:   49 bytes:  7603200/  7603200

+791e1fb999deb2e4156e2286d48c4ed1 *./tests/data/jpegls.vsynth1.out.yuv

+stddev:    2.84 PSNR: 39.04 MAXDIFF:   49 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 267a152a73cbc5ac4694a6e3b254be34 *./tests/data/vsynth1/msvideo1.avi

 2162264 ./tests/data/vsynth1/msvideo1.avi

-cf15dd12b8347567ae350383bf4ef4bb *./tests/data/msvideo1.vsynth1.out.yuv

-stddev:   11.81 PSNR: 26.68 MAXDIFF:  151 bytes:  7603200/  7603200

+c0665fac1bd896b6fe7fe0eead805bd5 *./tests/data/msvideo1.vsynth1.out.yuv

+stddev:   11.80 PSNR: 26.69 MAXDIFF:  151 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 d14041925ce5ec5001dc519276b1a1ab *./tests/data/vsynth1/qtrle.mov

 15263232 ./tests/data/vsynth1/qtrle.mov

-243325fb2cae1a9245efd49aff936327 *./tests/data/qtrle.vsynth1.out.yuv

-stddev:    3.42 PSNR: 37.43 MAXDIFF:   48 bytes:  7603200/  7603200

+93695a27c24a61105076ca7b1f010bbd *./tests/data/qtrle.vsynth1.out.yuv

+stddev:    3.42 PSNR: 37.44 MAXDIFF:   48 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 05f0719cb52486d9a4beb9cfae3f2571 *./tests/data/vsynth1/rgb.avi

 15213260 ./tests/data/vsynth1/rgb.avi

-243325fb2cae1a9245efd49aff936327 *./tests/data/rgb.vsynth1.out.yuv

-stddev:    3.42 PSNR: 37.43 MAXDIFF:   48 bytes:  7603200/  7603200

+93695a27c24a61105076ca7b1f010bbd *./tests/data/rgb.vsynth1.out.yuv

+stddev:    3.42 PSNR: 37.44 MAXDIFF:   48 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 0667077971e0cb63b5f49c580006e90e *./tests/data/vsynth2/flashsv.flv

 12368953 ./tests/data/vsynth2/flashsv.flv

-592b3321994e26a990deb3a0a1415de9 *./tests/data/flashsv.vsynth2.out.yuv

-stddev:    0.65 PSNR: 51.84 MAXDIFF:   14 bytes:  7603200/  7603200

+3a984506f1ebfc9fb73b6814cab201cc *./tests/data/flashsv.vsynth2.out.yuv

+stddev:    0.66 PSNR: 51.73 MAXDIFF:   14 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 4fc53937f048c900ae6d50fda9dba206 *./tests/data/vsynth2/jpegls.avi

 8334630 ./tests/data/vsynth2/jpegls.avi

-592b3321994e26a990deb3a0a1415de9 *./tests/data/jpegls.vsynth2.out.yuv

-stddev:    0.65 PSNR: 51.84 MAXDIFF:   14 bytes:  7603200/  7603200

+3a984506f1ebfc9fb73b6814cab201cc *./tests/data/jpegls.vsynth2.out.yuv

+stddev:    0.66 PSNR: 51.73 MAXDIFF:   14 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 5dddbbd6616d9be4bc0fd0c9650bd9e3 *./tests/data/vsynth2/msvideo1.avi

 1259308 ./tests/data/vsynth2/msvideo1.avi

-e2e7a952135f6307a74f2e178dc0df20 *./tests/data/msvideo1.vsynth2.out.yuv

-stddev:    7.42 PSNR: 30.71 MAXDIFF:  123 bytes:  7603200/  7603200

+cd83ffcbc73573044e3aead3094229e5 *./tests/data/msvideo1.vsynth2.out.yuv

+stddev:    7.42 PSNR: 30.72 MAXDIFF:  123 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 d8c1604dc46d9aa4ec0385e6722c6989 *./tests/data/vsynth2/qtrle.mov

 14798419 ./tests/data/vsynth2/qtrle.mov

-b2418e0e3a9a8619b31219cbcf24dc82 *./tests/data/qtrle.vsynth2.out.yuv

-stddev:    1.26 PSNR: 46.06 MAXDIFF:   13 bytes:  7603200/  7603200

+98d0e2854731472c5bf13d8638502d0a *./tests/data/qtrle.vsynth2.out.yuv

+stddev:    1.26 PSNR: 46.10 MAXDIFF:   13 bytes:  7603200/  7603200

@@ -1,4 +1,4 @@

 f2e9c419023c743bf99aa5b2e55ad233 *./tests/data/vsynth2/rgb.avi

 15213260 ./tests/data/vsynth2/rgb.avi

-b2418e0e3a9a8619b31219cbcf24dc82 *./tests/data/rgb.vsynth2.out.yuv

-stddev:    1.26 PSNR: 46.06 MAXDIFF:   13 bytes:  7603200/  7603200

+98d0e2854731472c5bf13d8638502d0a *./tests/data/rgb.vsynth2.out.yuv

+stddev:    1.26 PSNR: 46.10 MAXDIFF:   13 bytes:  7603200/  7603200