sws: add yuv2yuvX() back and updated to the just 1 plane API

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libswscale/swscale.c

@@ -2563,6 +2563,7 @@ static int swScale(SwsContext *c, const uint8_t* src[],
             dst[2] + dstStride[2] * chrDstY,
             (CONFIG_SWSCALE_ALPHA && alpPixBuf) ? dst[3] + dstStride[3] * dstY : NULL,
         };
+        int use_mmx_vfilter= c->use_mmx_vfilter;
 
         const int firstLumSrcY= vLumFilterPos[dstY]; //First line needed as input
         const int firstLumSrcY2= vLumFilterPos[FFMIN(dstY | ((1<<c->chrDstVSubSample) - 1), dstH-1)];
@@ -2653,6 +2654,7 @@ static int swScale(SwsContext *c, const uint8_t* src[],
             // hmm looks like we can't use MMX here without overwriting this array's tail
             find_c_packed_planar_out_funcs(c, &yuv2plane1, &yuv2planeX,  &yuv2nv12cX,
                                            &yuv2packed1, &yuv2packed2, &yuv2packedX);
+            use_mmx_vfilter= 0;
         }
 
         {
@@ -2669,6 +2671,19 @@ static int swScale(SwsContext *c, const uint8_t* src[],
                 vLumFilter +=    dstY * vLumFilterSize;
                 vChrFilter += chrDstY * vChrFilterSize;
 
+                av_assert0(use_mmx_vfilter != (
+                               yuv2planeX == yuv2planeX_10BE_c
+                            || yuv2planeX == yuv2planeX_10LE_c
+                            || yuv2planeX == yuv2planeX_9BE_c
+                            || yuv2planeX == yuv2planeX_9LE_c
+                            || yuv2planeX == yuv2planeX_16BE_c
+                            || yuv2planeX == yuv2planeX_16LE_c
+                            || yuv2planeX == yuv2planeX_8_c));
+                if(use_mmx_vfilter){
+                    vLumFilter= c->lumMmxFilter;
+                    vChrFilter= c->chrMmxFilter;
+                }
+
                 if (vLumFilterSize == 1) {
                     yuv2plane1(lumSrcPtr[0], dest[0], dstW, c->lumDither8, 0);
                 } else {
@@ -2686,11 +2701,14 @@ static int swScale(SwsContext *c, const uint8_t* src[],
                         yuv2planeX(vChrFilter, vChrFilterSize,
                                    chrUSrcPtr, dest[1], chrDstW, c->chrDither8, 0);
                         yuv2planeX(vChrFilter, vChrFilterSize,
-                                   chrVSrcPtr, dest[2], chrDstW, c->chrDither8, 3);
+                                   chrVSrcPtr, dest[2], chrDstW, c->chrDither8, use_mmx_vfilter ? (c->uv_offx2 >> 1) : 3);
                     }
                 }
 
                 if (CONFIG_SWSCALE_ALPHA && alpPixBuf){
+                    if(use_mmx_vfilter){
+                        vLumFilter= c->alpMmxFilter;
+                    }
                     if (vLumFilterSize == 1) {
                         yuv2plane1(alpSrcPtr[0], dest[3], dstW, c->lumDither8, 0);
                     } else {

libswscale/swscale_internal.h

@@ -412,6 +412,7 @@ typedef struct SwsContext {
 #if HAVE_VIS
     DECLARE_ALIGNED(8, uint64_t, sparc_coeffs)[10];
 #endif
+    int use_mmx_vfilter;
 
     /* function pointers for swScale() */
     yuv2planar1_fn yuv2plane1;

libswscale/x86/swscale_template.c

@@ -66,38 +66,47 @@ dither_8to16(const uint8_t *srcDither, int rot)
 }
 #endif
 
-static void RENAME(yuv2yuv1)(SwsContext *c, const int16_t *lumSrc,
-                             const int16_t *chrUSrc, const int16_t *chrVSrc,
-                             const int16_t *alpSrc,
-                             uint8_t *dst[4], int dstW, int chrDstW)
+static void RENAME(yuv2yuvX)(const int16_t *filter, int filterSize,
+                           const int16_t **src, uint8_t *dest, int dstW,
+                           const uint8_t *dither, int offset)
 {
-    int p= 4;
-    const int16_t *src[4]= {
-        lumSrc + dstW,     chrUSrc + chrDstW,
-        chrVSrc + chrDstW, alpSrc + dstW
-    };
-    x86_reg counter[4]= { dstW, chrDstW, chrDstW, dstW };
-
-    while (p--) {
-        if (dst[p]) {
-            __asm__ volatile(
-                "mov %2, %%"REG_a"                    \n\t"
-                ".p2align               4             \n\t" /* FIXME Unroll? */
-                "1:                                   \n\t"
-                "movq  (%0, %%"REG_a", 2), %%mm0      \n\t"
-                "movq 8(%0, %%"REG_a", 2), %%mm1      \n\t"
-                "psraw                 $7, %%mm0      \n\t"
-                "psraw                 $7, %%mm1      \n\t"
-                "packuswb           %%mm1, %%mm0      \n\t"
-                MOVNTQ(%%mm0, (%1, %%REGa))
-                "add                   $8, %%"REG_a"  \n\t"
-                "jnc                   1b             \n\t"
-                :: "r" (src[p]), "r" (dst[p] + counter[p]),
-                   "g" (-counter[p])
-                : "%"REG_a
-            );
-        }
-    }
+    dither_8to16(dither, offset);
+    __asm__ volatile(\
+        "psraw        $4, %%mm3\n\t"
+        "psraw        $4, %%mm4\n\t"
+        "movq    %%mm3, %%mm6\n\t"
+        "movq    %%mm4, %%mm7\n\t"
+        "movslq %3, %%"REG_c"\n\t"
+        "mov                                 %0, %%"REG_d"  \n\t"\
+        "mov                        (%%"REG_d"), %%"REG_S"  \n\t"\
+        ".p2align                             4             \n\t" /* FIXME Unroll? */\
+        "1:                                                 \n\t"\
+        "movq                      8(%%"REG_d"), %%mm0      \n\t" /* filterCoeff */\
+        "movq                (%%"REG_S", %%"REG_c", 2), %%mm2      \n\t" /* srcData */\
+        "movq               8(%%"REG_S", %%"REG_c", 2), %%mm5      \n\t" /* srcData */\
+        "add                                $16, %%"REG_d"  \n\t"\
+        "mov                        (%%"REG_d"), %%"REG_S"  \n\t"\
+        "test                         %%"REG_S", %%"REG_S"  \n\t"\
+        "pmulhw                           %%mm0, %%mm2      \n\t"\
+        "pmulhw                           %%mm0, %%mm5      \n\t"\
+        "paddw                            %%mm2, %%mm3      \n\t"\
+        "paddw                            %%mm5, %%mm4      \n\t"\
+        " jnz                                1b             \n\t"\
+        "psraw                               $3, %%mm3      \n\t"\
+        "psraw                               $3, %%mm4      \n\t"\
+        "packuswb                         %%mm4, %%mm3      \n\t"
+        MOVNTQ2 "                         %%mm3, (%1, %%"REG_c")\n\t"
+        "add                          $8, %%"REG_c"         \n\t"\
+        "cmp                          %2, %%"REG_c"         \n\t"\
+        "movq    %%mm6, %%mm3\n\t"
+        "movq    %%mm7, %%mm4\n\t"
+        "mov                                 %0, %%"REG_d"  \n\t"\
+        "mov                        (%%"REG_d"), %%"REG_S"  \n\t"\
+        "jb                                  1b             \n\t"\
+        :: "g" (filter),
+           "r" (dest-offset), "g" ((x86_reg)(dstW+offset)), "m" (offset)
+        : "%"REG_d, "%"REG_S, "%"REG_c
+    );
 }
 
 static void RENAME(yuv2yuv1_ar)(const int16_t *src, uint8_t *dst, int dstW, const uint8_t *dither, int offset)
@@ -1869,7 +1878,7 @@ static av_cold void RENAME(sws_init_swScale)(SwsContext *c)
 {
     enum PixelFormat srcFormat = c->srcFormat,
                      dstFormat = c->dstFormat;
-
+    c->use_mmx_vfilter= 0;
     if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) && dstFormat != PIX_FMT_NV12
         && dstFormat != PIX_FMT_NV21 && !(c->flags & SWS_BITEXACT)) {
             c->yuv2plane1 = RENAME(yuv2yuv1_ar    );
@@ -1886,7 +1895,10 @@ static av_cold void RENAME(sws_init_swScale)(SwsContext *c)
                     }
                 }
             } else {
-                //c->yuv2yuv1 = RENAME(yuv2yuv1    );
+                int should_dither= isNBPS(c->srcFormat) || is16BPS(c->srcFormat);
+                //c->yuv2plane1 = should_dither ? RENAME(yuv2yuv1_ar    ) : RENAME(yuv2yuv1    );
+                c->use_mmx_vfilter= 1;
+                c->yuv2planeX = RENAME(yuv2yuvX    );
                 if (!(c->flags & SWS_FULL_CHR_H_INT)) {
                     switch (c->dstFormat) {
                     case PIX_FMT_RGB32:   c->yuv2packedX = RENAME(yuv2rgb32_X);   break;