vertical scaler with accurate rounding, some people on doom9 can see +-1 errors

the +-1 issue is limited to >2tap vertical filters, so bilinear upscale was unaffected
the new code is sometime faster sometimes slower but the difference is significant (~20%) so its optional and enabled with arnd=1

Originally committed as revision 19177 to svn://svn.mplayerhq.hu/mplayer/trunk/libswscale

diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index f45a4df..9fcb44c 100644 (file)
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -848,7 +848,7 @@ static double getSplineCoeff(double a, double b, double c, double d, double dist
                                                dist-1.0);
 }
 
-static inline void initFilter(int16_t **outFilter, int16_t **filterPos, int *outFilterSize, int xInc,
+static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outFilterSize, int xInc,
                              int srcW, int dstW, int filterAlign, int one, int flags,
                              SwsVector *srcFilter, SwsVector *dstFilter, double param[2])
 {
@@ -1127,10 +1127,18 @@ static inline void initFilter(int16_t **outFilter, int16_t **filterPos, int *out
             filterAlign = 1;
         }
 
+        if (flags & SWS_CPU_CAPS_MMX) {
+                // special case for unscaled vertical filtering
+                if(minFilterSize == 1 && filterAlign == 2)
+                        filterAlign= 1;
+        }
+
        ASSERT(minFilterSize > 0)
        filterSize= (minFilterSize +(filterAlign-1)) & (~(filterAlign-1));
        ASSERT(filterSize > 0)
        filter= av_malloc(filterSize*dstW*sizeof(double));
+        if(filterSize >= MAX_FILTER_SIZE)
+                return -1;
        *outFilterSize= filterSize;
 
        if(flags&SWS_PRINT_INFO)
@@ -1216,6 +1224,7 @@ static inline void initFilter(int16_t **outFilter, int16_t **filterPos, int *out
        }
 
        av_free(filter);
+        return 0;
 }
 
 #if defined(ARCH_X86) || defined(ARCH_X86_64)
@@ -2115,6 +2124,7 @@ SwsContext *sws_getContext(int srcW, int srcH, int origSrcFormat, int dstW, int
        /* precalculate vertical scaler filter coefficients */
        {
                const int filterAlign=
+                 (flags & SWS_CPU_CAPS_MMX) && (flags & SWS_ACCURATE_RND) ? 2 :
                  (flags & SWS_CPU_CAPS_ALTIVEC) ? 8 :
                  1;
 

diff --git a/libswscale/swscale.h b/libswscale/swscale.h
index 37c5ec7..0b35023 100644 (file)
--- a/libswscale/swscale.h
+++ b/libswscale/swscale.h
@@ -64,6 +64,7 @@ extern "C" {
 //input subsampling info
 #define SWS_FULL_CHR_H_INP     0x4000
 #define SWS_DIRECT_BGR         0x8000
+#define SWS_ACCURATE_RND       0x40000
 
 #define SWS_CPU_CAPS_MMX   0x80000000
 #define SWS_CPU_CAPS_MMX2  0x20000000

diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index bc8f08d..177074d 100644 (file)
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -126,7 +126,9 @@ typedef struct SwsContext{
 #define DSTW_OFFSET  "11*8+4*4*256*2" //do not change, its hardcoded in the asm
 #define ESP_OFFSET  "11*8+4*4*256*2+8"
 #define VROUNDER_OFFSET "11*8+4*4*256*2+16"
-                  
+#define U_TEMP       "11*8+4*4*256*2+24"
+#define V_TEMP       "11*8+4*4*256*2+32"
+
        uint64_t redDither   __attribute__((aligned(8)));
        uint64_t greenDither __attribute__((aligned(8)));
        uint64_t blueDither  __attribute__((aligned(8)));
@@ -144,6 +146,8 @@ typedef struct SwsContext{
        int dstW;
        uint64_t esp __attribute__((aligned(8)));
        uint64_t vRounder     __attribute__((aligned(8)));
+       uint64_t u_temp       __attribute__((aligned(8)));
+       uint64_t v_temp       __attribute__((aligned(8)));
 
 #ifdef HAVE_ALTIVEC
 

diff --git a/libswscale/swscale_template.c b/libswscale/swscale_template.c
index 9830e3d..4d3e5ea 100644 (file)
--- a/libswscale/swscale_template.c
+++ b/libswscale/swscale_template.c
@@ -67,7 +67,8 @@
 #include "swscale_altivec_template.c"
 #endif
 
-#define YSCALEYUV2YV12X(x, offset) \
+#define YSCALEYUV2YV12X(x, offset, dest, width) \
+               asm volatile(\
                        "xor %%"REG_a", %%"REG_a"       \n\t"\
                        "movq "VROUNDER_OFFSET"(%0), %%mm3\n\t"\
                        "movq %%mm3, %%mm4              \n\t"\
@@ -96,7 +97,73 @@
                        "movq %%mm3, %%mm4              \n\t"\
                        "lea " offset "(%0), %%"REG_d"  \n\t"\
                        "mov (%%"REG_d"), %%"REG_S"     \n\t"\
-                       "jb 1b                          \n\t"
+                       "jb 1b                          \n\t"\
+                        :: "r" (&c->redDither),\
+                        "r" (dest), "p" (width)\
+                        : "%"REG_a, "%"REG_d, "%"REG_S\
+                );
+
+#define YSCALEYUV2YV12X_ACCURATE(x, offset, dest, width) \
+               asm volatile(\
+                       "lea " offset "(%0), %%"REG_d"  \n\t"\
+                       "xor %%"REG_a", %%"REG_a"       \n\t"\
+                        "pxor %%mm4, %%mm4              \n\t"\
+                        "pxor %%mm5, %%mm5              \n\t"\
+                        "pxor %%mm6, %%mm6              \n\t"\
+                        "pxor %%mm7, %%mm7              \n\t"\
+                       "mov (%%"REG_d"), %%"REG_S"     \n\t"\
+                       ASMALIGN16 \
+                       "1:                             \n\t"\
+                       "movq " #x "(%%"REG_S", %%"REG_a", 2), %%mm0\n\t" /* srcData */\
+                       "movq 8+" #x "(%%"REG_S", %%"REG_a", 2), %%mm2\n\t" /* srcData */\
+                       "mov 4(%%"REG_d"), %%"REG_S"    \n\t"\
+                       "movq " #x "(%%"REG_S", %%"REG_a", 2), %%mm1\n\t" /* srcData */\
+                        "movq %%mm0, %%mm3              \n\t"\
+                        "punpcklwd %%mm1, %%mm0        \n\t"\
+                        "punpckhwd %%mm1, %%mm3        \n\t"\
+                       "movq 8(%%"REG_d"), %%mm1       \n\t" /* filterCoeff */\
+                        "pmaddwd %%mm1, %%mm0           \n\t"\
+                        "pmaddwd %%mm1, %%mm3           \n\t"\
+                        "paddd %%mm0, %%mm4             \n\t"\
+                        "paddd %%mm3, %%mm5             \n\t"\
+                       "movq 8+" #x "(%%"REG_S", %%"REG_a", 2), %%mm3\n\t" /* srcData */\
+                       "mov 16(%%"REG_d"), %%"REG_S"   \n\t"\
+                       "add $16, %%"REG_d"             \n\t"\
+                        "test %%"REG_S", %%"REG_S"      \n\t"\
+                        "movq %%mm2, %%mm0              \n\t"\
+                        "punpcklwd %%mm3, %%mm2        \n\t"\
+                        "punpckhwd %%mm3, %%mm0        \n\t"\
+                        "pmaddwd %%mm1, %%mm2           \n\t"\
+                        "pmaddwd %%mm1, %%mm0           \n\t"\
+                        "paddd %%mm2, %%mm6             \n\t"\
+                        "paddd %%mm0, %%mm7             \n\t"\
+                       " jnz 1b                        \n\t"\
+                       "psrad $16, %%mm4               \n\t"\
+                       "psrad $16, %%mm5               \n\t"\
+                       "psrad $16, %%mm6               \n\t"\
+                       "psrad $16, %%mm7               \n\t"\
+                       "movq "VROUNDER_OFFSET"(%0), %%mm0\n\t"\
+                       "packssdw %%mm5, %%mm4          \n\t"\
+                       "packssdw %%mm7, %%mm6          \n\t"\
+                        "paddw %%mm0, %%mm4             \n\t"\
+                        "paddw %%mm0, %%mm6             \n\t"\
+                       "psraw $3, %%mm4                \n\t"\
+                       "psraw $3, %%mm6                \n\t"\
+                       "packuswb %%mm6, %%mm4          \n\t"\
+                       MOVNTQ(%%mm4, (%1, %%REGa))\
+                       "add $8, %%"REG_a"              \n\t"\
+                       "cmp %2, %%"REG_a"              \n\t"\
+                       "lea " offset "(%0), %%"REG_d"  \n\t"\
+                        "pxor %%mm4, %%mm4              \n\t"\
+                        "pxor %%mm5, %%mm5              \n\t"\
+                        "pxor %%mm6, %%mm6              \n\t"\
+                        "pxor %%mm7, %%mm7              \n\t"\
+                       "mov (%%"REG_d"), %%"REG_S"     \n\t"\
+                       "jb 1b                          \n\t"\
+                        :: "r" (&c->redDither),\
+                        "r" (dest), "p" (width)\
+                        : "%"REG_a, "%"REG_d, "%"REG_S\
+                );
 
 #define YSCALEYUV2YV121 \
                        "mov %2, %%"REG_a"              \n\t"\
@@ -159,8 +226,100 @@
                "test %%"REG_S", %%"REG_S"      \n\t"\
                " jnz 2b                        \n\t"\
 
-
-#define YSCALEYUV2RGBX \
+#define YSCALEYUV2PACKEDX_ACCURATE \
+               "xor %%"REG_a", %%"REG_a"       \n\t"\
+               ASMALIGN16\
+               "nop                            \n\t"\
+               "1:                             \n\t"\
+               "lea "CHR_MMX_FILTER_OFFSET"(%0), %%"REG_d"\n\t"\
+               "mov (%%"REG_d"), %%"REG_S"     \n\t"\
+                "pxor %%mm4, %%mm4              \n\t"\
+                "pxor %%mm5, %%mm5              \n\t"\
+                "pxor %%mm6, %%mm6              \n\t"\
+                "pxor %%mm7, %%mm7              \n\t"\
+               ASMALIGN16\
+               "2:                             \n\t"\
+               "movq (%%"REG_S", %%"REG_a"), %%mm0     \n\t" /* UsrcData */\
+               "movq 4096(%%"REG_S", %%"REG_a"), %%mm2 \n\t" /* VsrcData */\
+               "mov 4(%%"REG_d"), %%"REG_S"    \n\t"\
+               "movq (%%"REG_S", %%"REG_a"), %%mm1     \n\t" /* UsrcData */\
+                "movq %%mm0, %%mm3              \n\t"\
+                "punpcklwd %%mm1, %%mm0        \n\t"\
+                "punpckhwd %%mm1, %%mm3        \n\t"\
+                "movq 8(%%"REG_d"), %%mm1      \n\t" /* filterCoeff */\
+                "pmaddwd %%mm1, %%mm0           \n\t"\
+                "pmaddwd %%mm1, %%mm3           \n\t"\
+                "paddd %%mm0, %%mm4             \n\t"\
+                "paddd %%mm3, %%mm5             \n\t"\
+               "movq 4096(%%"REG_S", %%"REG_a"), %%mm3 \n\t" /* VsrcData */\
+                "mov 16(%%"REG_d"), %%"REG_S"  \n\t"\
+               "add $16, %%"REG_d"             \n\t"\
+                "test %%"REG_S", %%"REG_S"      \n\t"\
+                "movq %%mm2, %%mm0              \n\t"\
+                "punpcklwd %%mm3, %%mm2        \n\t"\
+                "punpckhwd %%mm3, %%mm0        \n\t"\
+                "pmaddwd %%mm1, %%mm2           \n\t"\
+                "pmaddwd %%mm1, %%mm0           \n\t"\
+                "paddd %%mm2, %%mm6             \n\t"\
+                "paddd %%mm0, %%mm7             \n\t"\
+               " jnz 2b                        \n\t"\
+                "psrad $16, %%mm4              \n\t"\
+                "psrad $16, %%mm5              \n\t"\
+                "psrad $16, %%mm6              \n\t"\
+                "psrad $16, %%mm7              \n\t"\
+                "movq "VROUNDER_OFFSET"(%0), %%mm0\n\t"\
+                "packssdw %%mm5, %%mm4         \n\t"\
+                "packssdw %%mm7, %%mm6         \n\t"\
+                "paddw %%mm0, %%mm4             \n\t"\
+                "paddw %%mm0, %%mm6             \n\t"\
+                "movq %%mm4, "U_TEMP"(%0)       \n\t"\
+                "movq %%mm6, "V_TEMP"(%0)       \n\t"\
+\
+               "lea "LUM_MMX_FILTER_OFFSET"(%0), %%"REG_d"\n\t"\
+               "mov (%%"REG_d"), %%"REG_S"     \n\t"\
+                "pxor %%mm1, %%mm1              \n\t"\
+                "pxor %%mm5, %%mm5              \n\t"\
+                "pxor %%mm7, %%mm7              \n\t"\
+                "pxor %%mm6, %%mm6              \n\t"\
+               ASMALIGN16\
+               "2:                             \n\t"\
+               "movq (%%"REG_S", %%"REG_a", 2), %%mm0  \n\t" /* Y1srcData */\
+               "movq 8(%%"REG_S", %%"REG_a", 2), %%mm2 \n\t" /* Y2srcData */\
+               "mov 4(%%"REG_d"), %%"REG_S"    \n\t"\
+               "movq (%%"REG_S", %%"REG_a", 2), %%mm4  \n\t" /* Y1srcData */\
+                "movq %%mm0, %%mm3              \n\t"\
+                "punpcklwd %%mm4, %%mm0        \n\t"\
+                "punpckhwd %%mm4, %%mm3        \n\t"\
+                "movq 8(%%"REG_d"), %%mm4      \n\t" /* filterCoeff */\
+                "pmaddwd %%mm4, %%mm0           \n\t"\
+                "pmaddwd %%mm4, %%mm3           \n\t"\
+                "paddd %%mm0, %%mm1             \n\t"\
+                "paddd %%mm3, %%mm5             \n\t"\
+               "movq 8(%%"REG_S", %%"REG_a", 2), %%mm3 \n\t" /* Y2srcData */\
+                "mov 16(%%"REG_d"), %%"REG_S"  \n\t"\
+               "add $16, %%"REG_d"             \n\t"\
+                "test %%"REG_S", %%"REG_S"      \n\t"\
+                "movq %%mm2, %%mm0              \n\t"\
+                "punpcklwd %%mm3, %%mm2        \n\t"\
+                "punpckhwd %%mm3, %%mm0        \n\t"\
+                "pmaddwd %%mm4, %%mm2           \n\t"\
+                "pmaddwd %%mm4, %%mm0           \n\t"\
+                "paddd %%mm2, %%mm7             \n\t"\
+                "paddd %%mm0, %%mm6             \n\t"\
+               " jnz 2b                        \n\t"\
+                "psrad $16, %%mm1              \n\t"\
+                "psrad $16, %%mm5              \n\t"\
+                "psrad $16, %%mm7              \n\t"\
+                "psrad $16, %%mm6              \n\t"\
+                "movq "VROUNDER_OFFSET"(%0), %%mm0\n\t"\
+                "packssdw %%mm5, %%mm1         \n\t"\
+                "packssdw %%mm6, %%mm7         \n\t"\
+                "paddw %%mm0, %%mm1             \n\t"\
+                "paddw %%mm0, %%mm7             \n\t"\
+                "movq  "U_TEMP"(%0), %%mm3      \n\t"\
+                "movq  "V_TEMP"(%0), %%mm4      \n\t"\
+
+#define YSCALEYUV2RGBX(YSCALEYUV2PACKEDX) \
                YSCALEYUV2PACKEDX\
                "psubw "U_OFFSET"(%0), %%mm3    \n\t" /* (U-128)8*/\
                "psubw "V_OFFSET"(%0), %%mm4    \n\t" /* (V-128)8*/\
@@ -762,29 +921,21 @@ static inline void RENAME(yuv2yuvX)(SwsContext *c, int16_t *lumFilter, int16_t *
                                    uint8_t *dest, uint8_t *uDest, uint8_t *vDest, long dstW, long chrDstW)
 {
 #ifdef HAVE_MMX
-       if(uDest != NULL)
-       {
-               asm volatile(
-                               YSCALEYUV2YV12X(0, CHR_MMX_FILTER_OFFSET)
-                               :: "r" (&c->redDither),
-                               "r" (uDest), "p" (chrDstW)
-                               : "%"REG_a, "%"REG_d, "%"REG_S
-                       );
-
-               asm volatile(
-                               YSCALEYUV2YV12X(4096, CHR_MMX_FILTER_OFFSET)
-                               :: "r" (&c->redDither),
-                               "r" (vDest), "p" (chrDstW)
-                               : "%"REG_a, "%"REG_d, "%"REG_S
-                       );
-       }
-
-       asm volatile(
-                       YSCALEYUV2YV12X(0, LUM_MMX_FILTER_OFFSET)
-                       :: "r" (&c->redDither),
-                          "r" (dest), "p" (dstW)
-                       : "%"REG_a, "%"REG_d, "%"REG_S
-               );
+        if(c->flags & SWS_ACCURATE_RND){
+                if(uDest){
+                        YSCALEYUV2YV12X_ACCURATE(   0, CHR_MMX_FILTER_OFFSET, uDest, chrDstW)
+                        YSCALEYUV2YV12X_ACCURATE(4096, CHR_MMX_FILTER_OFFSET, vDest, chrDstW)
+                }
+
+                YSCALEYUV2YV12X_ACCURATE(0, LUM_MMX_FILTER_OFFSET, dest, dstW)
+        }else{
+                if(uDest){
+                        YSCALEYUV2YV12X(   0, CHR_MMX_FILTER_OFFSET, uDest, chrDstW)
+                        YSCALEYUV2YV12X(4096, CHR_MMX_FILTER_OFFSET, vDest, chrDstW)
+                }
+
+                YSCALEYUV2YV12X(0, LUM_MMX_FILTER_OFFSET, dest, dstW)
+        }
 #else
 #ifdef HAVE_ALTIVEC
 yuv2yuvX_altivec_real(lumFilter, lumSrc, lumFilterSize,
@@ -876,13 +1027,94 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                            uint8_t *dest, long dstW, long dstY)
 {
        long dummy=0;
+#ifdef HAVE_MMX
+    if(c->flags & SWS_ACCURATE_RND){
+                switch(c->dstFormat){
+                case IMGFMT_BGR32:
+                       asm volatile(
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX_ACCURATE)
+                               WRITEBGR32(%4, %5, %%REGa)
+
+                       :: "r" (&c->redDither), 
+                          "m" (dummy), "m" (dummy), "m" (dummy),
+                          "r" (dest), "m" (dstW)
+                       : "%"REG_a, "%"REG_d, "%"REG_S
+                       );
+                        return;
+                case IMGFMT_BGR24:
+                       asm volatile(
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX_ACCURATE)
+                               "lea (%%"REG_a", %%"REG_a", 2), %%"REG_b"\n\t" //FIXME optimize
+                               "add %4, %%"REG_b"                      \n\t"
+                               WRITEBGR24(%%REGb, %5, %%REGa)
+
+                       :: "r" (&c->redDither), 
+                          "m" (dummy), "m" (dummy), "m" (dummy),
+                          "r" (dest), "m" (dstW)
+                       : "%"REG_a, "%"REG_b, "%"REG_d, "%"REG_S //FIXME ebx
+                       );
+                        return;
+                case IMGFMT_BGR15:
+                       asm volatile(
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX_ACCURATE)
+               /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
+#ifdef DITHER1XBPP
+                               "paddusb "MANGLE(b5Dither)", %%mm2\n\t"
+                               "paddusb "MANGLE(g5Dither)", %%mm4\n\t"
+                               "paddusb "MANGLE(r5Dither)", %%mm5\n\t"
+#endif
+
+                               WRITEBGR15(%4, %5, %%REGa)
+
+                       :: "r" (&c->redDither), 
+                          "m" (dummy), "m" (dummy), "m" (dummy),
+                          "r" (dest), "m" (dstW)
+                       : "%"REG_a, "%"REG_d, "%"REG_S
+                       );
+                        return;
+                case IMGFMT_BGR16:
+                       asm volatile(
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX_ACCURATE)
+               /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
+#ifdef DITHER1XBPP
+                               "paddusb "MANGLE(b5Dither)", %%mm2\n\t"
+                               "paddusb "MANGLE(g6Dither)", %%mm4\n\t"
+                               "paddusb "MANGLE(r5Dither)", %%mm5\n\t"
+#endif
+
+                               WRITEBGR16(%4, %5, %%REGa)
+
+                       :: "r" (&c->redDither), 
+                          "m" (dummy), "m" (dummy), "m" (dummy),
+                          "r" (dest), "m" (dstW)
+                       : "%"REG_a, "%"REG_d, "%"REG_S
+                       );
+                        return;
+                case IMGFMT_YUY2:
+                       asm volatile(
+                               YSCALEYUV2PACKEDX_ACCURATE
+               /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
+
+                               "psraw $3, %%mm3                \n\t"
+                               "psraw $3, %%mm4                \n\t"
+                               "psraw $3, %%mm1                \n\t"
+                               "psraw $3, %%mm7                \n\t"
+                               WRITEYUY2(%4, %5, %%REGa)
+
+                       :: "r" (&c->redDither), 
+                          "m" (dummy), "m" (dummy), "m" (dummy),
+                          "r" (dest), "m" (dstW)
+                       : "%"REG_a, "%"REG_d, "%"REG_S
+                       );
+                        return;
+                }
+    }else{
        switch(c->dstFormat)
        {
-#ifdef HAVE_MMX
        case IMGFMT_BGR32:
                {
                        asm volatile(
-                               YSCALEYUV2RGBX
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX)
                                WRITEBGR32(%4, %5, %%REGa)
 
                        :: "r" (&c->redDither), 
@@ -891,11 +1123,11 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        : "%"REG_a, "%"REG_d, "%"REG_S
                        );
                }
-               break;
+               return;
        case IMGFMT_BGR24:
                {
                        asm volatile(
-                               YSCALEYUV2RGBX
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX)
                                "lea (%%"REG_a", %%"REG_a", 2), %%"REG_b"\n\t" //FIXME optimize
                                "add %4, %%"REG_b"                      \n\t"
                                WRITEBGR24(%%REGb, %5, %%REGa)
@@ -906,11 +1138,11 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        : "%"REG_a, "%"REG_b, "%"REG_d, "%"REG_S //FIXME ebx
                        );
                }
-               break;
+               return;
        case IMGFMT_BGR15:
                {
                        asm volatile(
-                               YSCALEYUV2RGBX
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX)
                /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
 #ifdef DITHER1XBPP
                                "paddusb "MANGLE(b5Dither)", %%mm2\n\t"
@@ -926,11 +1158,11 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        : "%"REG_a, "%"REG_d, "%"REG_S
                        );
                }
-               break;
+               return;
        case IMGFMT_BGR16:
                {
                        asm volatile(
-                               YSCALEYUV2RGBX
+                               YSCALEYUV2RGBX(YSCALEYUV2PACKEDX)
                /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
 #ifdef DITHER1XBPP
                                "paddusb "MANGLE(b5Dither)", %%mm2\n\t"
@@ -946,7 +1178,7 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        : "%"REG_a, "%"REG_d, "%"REG_S
                        );
                }
-               break;
+               return;
        case IMGFMT_YUY2:
                {
                        asm volatile(
@@ -965,9 +1197,10 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        : "%"REG_a, "%"REG_d, "%"REG_S
                        );
                }
-               break;
+               return;
+        }
+    }
 #endif
-       default:
 #ifdef HAVE_ALTIVEC
                /* The following list of supported dstFormat values should
                   match what's found in the body of altivec_yuv2packedX() */
@@ -982,8 +1215,6 @@ static inline void RENAME(yuv2packedX)(SwsContext *c, int16_t *lumFilter, int16_
                        yuv2packedXinC(c, lumFilter, lumSrc, lumFilterSize,
                                    chrFilter, chrSrc, chrFilterSize,
                                    dest, dstW, dstY);
-               break;
-       }
 }
 
 /**
@@ -2813,6 +3044,22 @@ i--;
                int16_t **chrSrcPtr= chrPixBuf + chrBufIndex + firstChrSrcY - lastInChrBuf + vChrBufSize;
 #ifdef HAVE_MMX
                int i;
+            if(flags & SWS_ACCURATE_RND){
+                        for(i=0; i<vLumFilterSize; i+=2){
+                                lumMmxFilter[2*i+0]= lumSrcPtr[i  ];
+                                lumMmxFilter[2*i+1]= lumSrcPtr[i+(vLumFilterSize>1)];
+                                lumMmxFilter[2*i+2]=
+                                lumMmxFilter[2*i+3]= vLumFilter[dstY*vLumFilterSize + i    ]
+                                                + (vLumFilterSize>1 ? vLumFilter[dstY*vLumFilterSize + i + 1]<<16 : 0);
+                        }
+                        for(i=0; i<vChrFilterSize; i+=2){
+                                chrMmxFilter[2*i+0]= chrSrcPtr[i  ];
+                                chrMmxFilter[2*i+1]= chrSrcPtr[i+(vChrFilterSize>1)];
+                                chrMmxFilter[2*i+2]=
+                                chrMmxFilter[2*i+3]= vChrFilter[chrDstY*vChrFilterSize + i    ]
+                                                + (vChrFilterSize>1 ? vChrFilter[chrDstY*vChrFilterSize + i + 1]<<16 : 0);
+                        }
+            }else{
                for(i=0; i<vLumFilterSize; i++)
                {
                        lumMmxFilter[4*i+0]= (int32_t)lumSrcPtr[i];
@@ -2827,6 +3074,7 @@ i--;
                        chrMmxFilter[4*i+3]= 
                                ((uint16_t)vChrFilter[chrDstY*vChrFilterSize + i])*0x10001;
                }
+            }
 #endif
                if(dstFormat == IMGFMT_NV12 || dstFormat == IMGFMT_NV21){
                        const int chrSkipMask= (1<<c->chrDstVSubSample)-1;
@@ -2868,6 +3116,10 @@ i--;
                        {
                                int lumAlpha= vLumFilter[2*dstY+1];
                                int chrAlpha= vChrFilter[2*dstY+1];
+                                lumMmxFilter[2]=
+                                lumMmxFilter[3]= vLumFilter[2*dstY   ]*0x10001;
+                                chrMmxFilter[2]=
+                                chrMmxFilter[3]= vChrFilter[2*chrDstY]*0x10001;
                                RENAME(yuv2packed2)(c, *lumSrcPtr, *(lumSrcPtr+1), *chrSrcPtr, *(chrSrcPtr+1),
                                                 dest, dstW, lumAlpha, chrAlpha, dstY);
                        }