swscale: for >8bit scaling, read in native bit-depth.

For 9/10bit, it means we don't have to upscale to 16bit before
actual scaling or pixel format conversion, and thus a performance
gain.

6 files changed, 31 insertions, 83 deletions

diff --git a/libswscale/ppc/swscale_altivec.c b/libswscale/ppc/swscale_altivec.c
index 6fb3de0ee7..369e93b85a 100644
--- a/libswscale/ppc/swscale_altivec.c
+++ b/libswscale/ppc/swscale_altivec.c
@@ -220,7 +220,7 @@ yuv2yuvX_altivec_real(SwsContext *c,
     }
 }
 
-static void hScale_altivec_real(int16_t *dst, int dstW,
+static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
                                 const uint8_t *src, const int16_t *filter,
                                 const int16_t *filterPos, int filterSize)
 {
diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index ef8fc39c48..dacf40ed8a 100644
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -1766,59 +1766,6 @@ static void nv21ToUV_c(uint8_t *dstU, uint8_t *dstV,
 
 #define input_pixel(pos) (isBE(origin) ? AV_RB16(pos) : AV_RL16(pos))
 
-// FIXME Maybe dither instead.
-static av_always_inline void
-yuv9_OR_10ToUV_c_template(uint16_t *dstU, uint16_t *dstV,
-                          const uint16_t *srcU, const uint16_t *srcV,
-                          int width, enum PixelFormat origin, int depth)
-{
-    int i;
-
-    for (i = 0; i < width; i++) {
-        int upx = input_pixel(&srcU[i]);
-        int vpx = input_pixel(&srcV[i]);
-        dstU[i] =  (upx << (16 - depth)) | (upx >> (2 * depth - 16));
-        dstV[i] =  (vpx << (16 - depth)) | (vpx >> (2 * depth - 16));
-    }
-}
-
-static av_always_inline void
-yuv9_or_10ToY_c_template(uint16_t *dstY, const uint16_t *srcY,
-                         int width, enum PixelFormat origin, int depth)
-{
-    int i;
-
-    for (i = 0; i < width; i++) {
-        int px = input_pixel(&srcY[i]);
-        dstY[i] =  (px << (16 - depth)) | (px >> (2 * depth - 16));
-    }
-}
-
-#undef input_pixel
-
-#define YUV_NBPS(depth, BE_LE, origin) \
-static void BE_LE ## depth ## ToUV_c(uint8_t *_dstU, uint8_t *_dstV, \
-                                     const uint8_t *_srcU, const uint8_t *_srcV, \
-                                     int width, uint32_t *unused) \
-{ \
-    uint16_t *dstU = (uint16_t *) _dstU, *dstV = (uint16_t *) _dstV; \
-    const uint16_t *srcU = (const uint16_t *) _srcU, \
-                   *srcV = (const uint16_t *) _srcV; \
-    yuv9_OR_10ToUV_c_template(dstU, dstV, srcU, srcV, width, origin, depth); \
-} \
-static void BE_LE ## depth ## ToY_c(uint8_t *_dstY, const uint8_t *_srcY, \
-                                    int width, uint32_t *unused) \
-{ \
-    uint16_t *dstY = (uint16_t *) _dstY; \
-    const uint16_t *srcY = (const uint16_t *) _srcY; \
-    yuv9_or_10ToY_c_template(dstY, srcY, width, origin, depth); \
-}
-
-YUV_NBPS( 9, LE, PIX_FMT_YUV420P9LE);
-YUV_NBPS( 9, BE, PIX_FMT_YUV420P9BE);
-YUV_NBPS(10, LE, PIX_FMT_YUV420P10LE);
-YUV_NBPS(10, BE, PIX_FMT_YUV420P10BE);
-
 static void bgr24ToY_c(uint8_t *dst, const uint8_t *src,
                        int width, uint32_t *unused)
 {
@@ -1905,13 +1852,15 @@ static void rgb24ToUV_half_c(uint8_t *dstU, uint8_t *dstV, const uint8_t *src1,
     }
 }
 
-static void hScale16_c(int16_t *_dst, int dstW, const uint8_t *_src,
+static void hScale16_c(SwsContext *c, int16_t *_dst, int dstW, const uint8_t *_src,
                        const int16_t *filter,
                        const int16_t *filterPos, int filterSize)
 {
     int i;
     int32_t *dst = (int32_t *) _dst;
     const uint16_t *src = (const uint16_t *) _src;
+    int bits = av_pix_fmt_descriptors[c->srcFormat].comp[0].depth_minus1;
+    int sh = (bits <= 7) ? 11 : (bits - 4);
 
     for (i = 0; i < dstW; i++) {
         int j;
@@ -1922,12 +1871,12 @@ static void hScale16_c(int16_t *_dst, int dstW, const uint8_t *_src,
             val += src[srcPos + j] * filter[filterSize * i + j];
         }
         // filter=14 bit, input=16 bit, output=30 bit, >> 11 makes 19 bit
-        dst[i] = FFMIN(val >> 11, (1 << 19) - 1);
+        dst[i] = FFMIN(val >> sh, (1 << 19) - 1);
     }
 }
 
 // bilinear / bicubic scaling
-static void hScale_c(int16_t *dst, int dstW, const uint8_t *src,
+static void hScale_c(SwsContext *c, int16_t *dst, int dstW, const uint8_t *src,
                      const int16_t *filter, const int16_t *filterPos,
                      int filterSize)
 {
@@ -2063,7 +2012,7 @@ static av_always_inline void hyscale(SwsContext *c, int16_t *dst, int dstWidth,
     }
 
     if (!c->hyscale_fast) {
-        c->hScale(dst, dstWidth, src, hLumFilter, hLumFilterPos, hLumFilterSize);
+        c->hScale(c, dst, dstWidth, src, hLumFilter, hLumFilterPos, hLumFilterSize);
     } else { // fast bilinear upscale / crap downscale
         c->hyscale_fast(c, dst, dstWidth, src, srcW, xInc);
     }
@@ -2113,8 +2062,8 @@ static av_always_inline void hcscale(SwsContext *c, int16_t *dst1, int16_t *dst2
     }
 
     if (!c->hcscale_fast) {
-        c->hScale(dst1, dstWidth, src1, hChrFilter, hChrFilterPos, hChrFilterSize);
-        c->hScale(dst2, dstWidth, src2, hChrFilter, hChrFilterPos, hChrFilterSize);
+        c->hScale(c, dst1, dstWidth, src1, hChrFilter, hChrFilterPos, hChrFilterSize);
+        c->hScale(c, dst2, dstWidth, src2, hChrFilter, hChrFilterPos, hChrFilterSize);
     } else { // fast bilinear upscale / crap downscale
         c->hcscale_fast(c, dst1, dst2, dstWidth, src1, src2, srcW, xInc);
     }
@@ -2645,21 +2594,21 @@ static av_cold void sws_init_swScale_c(SwsContext *c)
         case PIX_FMT_PAL8     :
         case PIX_FMT_BGR4_BYTE:
         case PIX_FMT_RGB4_BYTE: c->chrToYV12 = palToUV_c; break;
+#if HAVE_BIGENDIAN
         case PIX_FMT_YUV444P9LE:
-        case PIX_FMT_YUV420P9LE:  c->chrToYV12 = LE9ToUV_c; break;
+        case PIX_FMT_YUV420P9LE:
         case PIX_FMT_YUV422P10LE:
         case PIX_FMT_YUV444P10LE:
-        case PIX_FMT_YUV420P10LE: c->chrToYV12 = LE10ToUV_c; break;
-        case PIX_FMT_YUV444P9BE:
-        case PIX_FMT_YUV420P9BE:  c->chrToYV12 = BE9ToUV_c; break;
-        case PIX_FMT_YUV444P10BE:
-        case PIX_FMT_YUV422P10BE:
-        case PIX_FMT_YUV420P10BE: c->chrToYV12 = BE10ToUV_c; break;
-#if HAVE_BIGENDIAN
+        case PIX_FMT_YUV420P10LE:
         case PIX_FMT_YUV420P16LE:
         case PIX_FMT_YUV422P16LE:
         case PIX_FMT_YUV444P16LE: c->chrToYV12 = bswap16UV_c; break;
 #else
+        case PIX_FMT_YUV444P9BE:
+        case PIX_FMT_YUV420P9BE:
+        case PIX_FMT_YUV444P10BE:
+        case PIX_FMT_YUV422P10BE:
+        case PIX_FMT_YUV420P10BE:
         case PIX_FMT_YUV420P16BE:
         case PIX_FMT_YUV422P16BE:
         case PIX_FMT_YUV444P16BE: c->chrToYV12 = bswap16UV_c; break;
@@ -2712,22 +2661,22 @@ static av_cold void sws_init_swScale_c(SwsContext *c)
     c->lumToYV12 = NULL;
     c->alpToYV12 = NULL;
     switch (srcFormat) {
+#if HAVE_BIGENDIAN
     case PIX_FMT_YUV444P9LE:
-    case PIX_FMT_YUV420P9LE:  c->lumToYV12 = LE9ToY_c; break;
+    case PIX_FMT_YUV420P9LE:
     case PIX_FMT_YUV444P10LE:
     case PIX_FMT_YUV422P10LE:
-    case PIX_FMT_YUV420P10LE: c->lumToYV12 = LE10ToY_c; break;
-    case PIX_FMT_YUV444P9BE:
-    case PIX_FMT_YUV420P9BE:  c->lumToYV12 = BE9ToY_c; break;
-    case PIX_FMT_YUV444P10BE:
-    case PIX_FMT_YUV422P10BE:
-    case PIX_FMT_YUV420P10BE: c->lumToYV12 = BE10ToY_c; break;
-#if HAVE_BIGENDIAN
+    case PIX_FMT_YUV420P10LE:
     case PIX_FMT_YUV420P16LE:
     case PIX_FMT_YUV422P16LE:
     case PIX_FMT_YUV444P16LE:
     case PIX_FMT_GRAY16LE: c->lumToYV12 = bswap16Y_c; break;
 #else
+    case PIX_FMT_YUV444P9BE:
+    case PIX_FMT_YUV420P9BE:
+    case PIX_FMT_YUV444P10BE:
+    case PIX_FMT_YUV422P10BE:
+    case PIX_FMT_YUV420P10BE:
     case PIX_FMT_YUV420P16BE:
     case PIX_FMT_YUV422P16BE:
     case PIX_FMT_YUV444P16BE:
diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index cfb19d9c13..e4b93c595b 100644
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -440,7 +440,7 @@ typedef struct SwsContext {
      *                   (and input coefficients thus padded with zeroes)
      *                   to simplify creating SIMD code.
      */
-    void (*hScale)(int16_t *dst, int dstW, const uint8_t *src,
+    void (*hScale)(struct SwsContext *c, int16_t *dst, int dstW, const uint8_t *src,
                    const int16_t *filter, const int16_t *filterPos,
                    int filterSize);
 
diff --git a/libswscale/utils.c b/libswscale/utils.c
index b7ccac94ba..eea32a130a 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -877,7 +877,6 @@ int sws_init_context(SwsContext *c, SwsFilter *srcFilter, SwsFilter *dstFilter)
         }
     }
 
-    // FIXME it's even nicer if bpp isn't 16, but max({src,dst}formatbpp)
     c->scalingBpp = FFMAX(av_pix_fmt_descriptors[srcFormat].comp[0].depth_minus1,
                           av_pix_fmt_descriptors[dstFormat].comp[0].depth_minus1) >= 8 ? 16 : 8;
     if (c->scalingBpp == 16)
diff --git a/libswscale/x86/swscale_template.c b/libswscale/x86/swscale_template.c
index 54c7edc25e..f58ac520e1 100644
--- a/libswscale/x86/swscale_template.c
+++ b/libswscale/x86/swscale_template.c
@@ -1859,7 +1859,7 @@ static void RENAME(rgb24ToUV)(uint8_t *dstU, uint8_t *dstV,
 
 #if !COMPILE_TEMPLATE_MMX2
 // bilinear / bicubic scaling
-static void RENAME(hScale)(int16_t *dst, int dstW,
+static void RENAME(hScale)(SwsContext *c, int16_t *dst, int dstW,
                            const uint8_t *src, const int16_t *filter,
                            const int16_t *filterPos, int filterSize)
 {
diff --git a/tests/ref/lavfi/pixfmts_scale b/tests/ref/lavfi/pixfmts_scale
index de41952a21..8901fe9132 100644
--- a/tests/ref/lavfi/pixfmts_scale
+++ b/tests/ref/lavfi/pixfmts_scale
@@ -31,12 +31,12 @@ uyvy422             314bd486277111a95d9369b944fa0400
 yuv410p             7df8f6d69b56a8dcb6c7ee908e5018b5
 yuv411p             1143e7c5cc28fe0922b051b17733bc4c
 yuv420p             fdad2d8df8985e3d17e73c71f713cb14
-yuv420p10be         8b5ad855229840a8fa87786cab83f856
-yuv420p10le         5264e87921d47b17411578d9c92672b1
+yuv420p10be         2343beaf83fccc2ab23a590b2049d38b
+yuv420p10le         94d511d783d175f573e7be5cce75ba4d
 yuv420p16be         f6ef3ba90f238b467c7e72ade927083d
 yuv420p16le         faf6aab3b1c16e8afbe160686dd360e0
-yuv420p9be          c14bf746d161face61e1f39b491bf7ef
-yuv420p9le          59457f9a51768bf1d4342238d50a9be3
+yuv420p9be          fdafb9ad473a559246c4cb0a1f416cd8
+yuv420p9le          fccfd3c3941da635b13739f579819b5a
 yuv422p             918e37701ee7377d16a8a6c119c56a40
 yuv422p16be         837945d3a771366a5a72a4ed095a4f53
 yuv422p16le         b8292ae9b52eb7afc3d8b93e8fd895b4