MMX(2) accelerated yuv420->rgb24 routine, trivial as essentially as it just

swaps R and B inputs for the existing yuv420->bgr24 routine.

Originally committed as revision 30254 to svn://svn.mplayerhq.hu/mplayer/trunk/libswscale

2 files changed, 118 insertions, 98 deletions

diff --git a/libswscale/x86/yuv2rgb_mmx.c b/libswscale/x86/yuv2rgb_mmx.c
index b03cce2ae7..1e69df1641 100644
--- a/libswscale/x86/yuv2rgb_mmx.c
+++ b/libswscale/x86/yuv2rgb_mmx.c
@@ -72,7 +72,8 @@ SwsFunc ff_yuv2rgb_init_mmx(SwsContext *c)
                 if (HAVE_7REGS) return yuva420_bgr32_MMX2;
                 break;
             } else return yuv420_bgr32_MMX2;
-        case PIX_FMT_BGR24:  return yuv420_rgb24_MMX2;
+        case PIX_FMT_RGB24:  return yuv420_rgb24_MMX2;
+        case PIX_FMT_BGR24:  return yuv420_bgr24_MMX2;
         case PIX_FMT_RGB565: return yuv420_rgb16_MMX2;
         case PIX_FMT_RGB555: return yuv420_rgb15_MMX2;
         }
@@ -89,7 +90,8 @@ SwsFunc ff_yuv2rgb_init_mmx(SwsContext *c)
                 if (HAVE_7REGS) return yuva420_bgr32_MMX;
                 break;
             } else return yuv420_bgr32_MMX;
-        case PIX_FMT_BGR24:  return yuv420_rgb24_MMX;
+        case PIX_FMT_RGB24:  return yuv420_rgb24_MMX;
+        case PIX_FMT_BGR24:  return yuv420_bgr24_MMX;
         case PIX_FMT_RGB565: return yuv420_rgb16_MMX;
         case PIX_FMT_RGB555: return yuv420_rgb15_MMX;
         }
diff --git a/libswscale/x86/yuv2rgb_template.c b/libswscale/x86/yuv2rgb_template.c
index aa17f196bd..b000dd2574 100644
--- a/libswscale/x86/yuv2rgb_template.c
+++ b/libswscale/x86/yuv2rgb_template.c
@@ -295,6 +295,106 @@ static inline int RENAME(yuv420_rgb15)(SwsContext *c, uint8_t* src[], int srcStr
     YUV2RGB_OPERANDS
 }
 
+#if HAVE_MMX2
+#define RGB_PLANAR2PACKED24(red, blue)\
+        "movq "MANGLE(ff_M24A)", %%mm4     \n\t"\
+        "movq "MANGLE(ff_M24C)", %%mm7     \n\t"\
+        "pshufw $0x50, %%mm"blue", %%mm5   \n\t" /* B3 B2 B3 B2  B1 B0 B1 B0 */\
+        "pshufw $0x50, %%mm2, %%mm3     \n\t" /* G3 G2 G3 G2  G1 G0 G1 G0 */\
+        "pshufw $0x00, %%mm"red", %%mm6 \n\t" /* R1 R0 R1 R0  R1 R0 R1 R0 */\
+\
+        "pand   %%mm4, %%mm5            \n\t" /*    B2        B1       B0 */\
+        "pand   %%mm4, %%mm3            \n\t" /*    G2        G1       G0 */\
+        "pand   %%mm7, %%mm6            \n\t" /*       R1        R0       */\
+\
+        "psllq     $8, %%mm3            \n\t" /* G2        G1       G0    */\
+        "por    %%mm5, %%mm6            \n\t"\
+        "por    %%mm3, %%mm6            \n\t"\
+        MOVNTQ" %%mm6, (%1)             \n\t"\
+\
+        "psrlq     $8, %%mm2            \n\t" /* 00 G7 G6 G5  G4 G3 G2 G1 */\
+        "pshufw $0xA5, %%mm"blue", %%mm5\n\t" /* B5 B4 B5 B4  B3 B2 B3 B2 */\
+        "pshufw $0x55, %%mm2, %%mm3     \n\t" /* G4 G3 G4 G3  G4 G3 G4 G3 */\
+        "pshufw $0xA5, %%mm"red", %%mm6 \n\t" /* R5 R4 R5 R4  R3 R2 R3 R2 */\
+\
+        "pand "MANGLE(ff_M24B)", %%mm5  \n\t" /* B5       B4        B3    */\
+        "pand          %%mm7, %%mm3     \n\t" /*       G4        G3       */\
+        "pand          %%mm4, %%mm6     \n\t" /*    R4        R3       R2 */\
+\
+        "por    %%mm5, %%mm3            \n\t" /* B5    G4 B4     G3 B3    */\
+        "por    %%mm3, %%mm6            \n\t"\
+        MOVNTQ" %%mm6, 8(%1)            \n\t"\
+\
+        "pshufw $0xFF, %%mm"blue", %%mm5\n\t" /* B7 B6 B7 B6  B7 B6 B6 B7 */\
+        "pshufw $0xFA, %%mm2, %%mm3     \n\t" /* 00 G7 00 G7  G6 G5 G6 G5 */\
+        "pshufw $0xFA, %%mm"red", %%mm6 \n\t" /* R7 R6 R7 R6  R5 R4 R5 R4 */\
+        "movd 4 (%2, %0), %%mm0;" /* Load 4 Cb 00 00 00 00 u3 u2 u1 u0 */\
+\
+        "pand          %%mm7, %%mm5     \n\t" /*       B7        B6       */\
+        "pand          %%mm4, %%mm3     \n\t" /*    G7        G6       G5 */\
+        "pand "MANGLE(ff_M24B)", %%mm6  \n\t" /* R7       R6        R5    */\
+        "movd 4 (%3, %0), %%mm1;" /* Load 4 Cr 00 00 00 00 v3 v2 v1 v0 */\
+\
+        "por          %%mm5, %%mm3      \n\t"\
+        "por          %%mm3, %%mm6      \n\t"\
+        MOVNTQ"       %%mm6, 16(%1)     \n\t"\
+        "movq 8 (%5, %0, 2), %%mm6;" /* Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */\
+        "pxor         %%mm4, %%mm4      \n\t"
+#else
+#define RGB_PLANAR2PACKED24(red, blue)\
+        "pxor      %%mm4, %%mm4     \n\t"\
+        "movq      %%mm"blue", %%mm5\n\t" /* B */\
+        "movq      %%mm"red", %%mm6 \n\t" /* R */\
+        "punpcklbw %%mm2, %%mm"blue"\n\t" /* GBGBGBGB 0 */\
+        "punpcklbw %%mm4, %%mm"red" \n\t" /* 0R0R0R0R 0 */\
+        "punpckhbw %%mm2, %%mm5     \n\t" /* GBGBGBGB 2 */\
+        "punpckhbw %%mm4, %%mm6     \n\t" /* 0R0R0R0R 2 */\
+        "movq      %%mm"blue", %%mm7\n\t" /* GBGBGBGB 0 */\
+        "movq      %%mm5, %%mm3     \n\t" /* GBGBGBGB 2 */\
+        "punpcklwd %%mm"red", %%mm7 \n\t" /* 0RGB0RGB 0 */\
+        "punpckhwd %%mm"red", %%mm"blue"\n\t" /* 0RGB0RGB 1 */\
+        "punpcklwd %%mm6, %%mm5     \n\t" /* 0RGB0RGB 2 */\
+        "punpckhwd %%mm6, %%mm3     \n\t" /* 0RGB0RGB 3 */\
+\
+        "movq      %%mm7, %%mm2     \n\t" /* 0RGB0RGB 0 */\
+        "movq      %%mm"blue", %%mm6\n\t" /* 0RGB0RGB 1 */\
+        "movq      %%mm5, %%mm"red" \n\t" /* 0RGB0RGB 2 */\
+        "movq      %%mm3, %%mm4     \n\t" /* 0RGB0RGB 3 */\
+\
+        "psllq       $40, %%mm7     \n\t" /* RGB00000 0 */\
+        "psllq       $40, %%mm"blue"\n\t" /* RGB00000 1 */\
+        "psllq       $40, %%mm5     \n\t" /* RGB00000 2 */\
+        "psllq       $40, %%mm3     \n\t" /* RGB00000 3 */\
+\
+        "punpckhdq %%mm2, %%mm7     \n\t" /* 0RGBRGB0 0 */\
+        "punpckhdq %%mm6, %%mm"blue"\n\t" /* 0RGBRGB0 1 */\
+        "punpckhdq %%mm"red", %%mm5 \n\t" /* 0RGBRGB0 2 */\
+        "punpckhdq %%mm4, %%mm3     \n\t" /* 0RGBRGB0 3 */\
+\
+        "psrlq        $8, %%mm7     \n\t" /* 00RGBRGB 0 */\
+        "movq      %%mm"blue", %%mm6\n\t" /* 0RGBRGB0 1 */\
+        "psllq       $40, %%mm"blue"\n\t" /* GB000000 1 */\
+        "por       %%mm"blue", %%mm7\n\t" /* GBRGBRGB 0 */\
+        MOVNTQ"    %%mm7, (%1)      \n\t"\
+\
+        "psrlq       $24, %%mm6     \n\t" /* 0000RGBR 1 */\
+        "movq      %%mm5, %%mm"red" \n\t" /* 0RGBRGB0 2 */\
+        "psllq       $24, %%mm5     \n\t" /* BRGB0000 2 */\
+        "por       %%mm5, %%mm6     \n\t" /* BRGBRGBR 1 */\
+        MOVNTQ"    %%mm6, 8(%1)     \n\t"\
+\
+        "movq 8 (%5, %0, 2), %%mm6;" /* Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */\
+\
+        "psrlq       $40, %%mm"red" \n\t" /* 000000RG 2 */\
+        "psllq        $8, %%mm3     \n\t" /* RGBRGB00 3 */\
+        "por       %%mm3, %%mm"red" \n\t" /* RGBRGBRG 2 */\
+        MOVNTQ"    %%mm"red", 16(%1)\n\t"\
+\
+        "movd 4 (%3, %0), %%mm1;" /* Load 4 Cr 00 00 00 00 v3 v2 v1 v0 */\
+        "movd 4 (%2, %0), %%mm0;" /* Load 4 Cb 00 00 00 00 u3 u2 u1 u0 */\
+        "pxor      %%mm4, %%mm4     \n\t"
+#endif
+
 static inline int RENAME(yuv420_rgb24)(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
                                        int srcSliceH, uint8_t* dst[], int dstStride[])
 {
@@ -306,106 +406,24 @@ static inline int RENAME(yuv420_rgb24)(SwsContext *c, uint8_t* src[], int srcStr
         YUV2RGB_INIT
         YUV2RGB
         /* mm0=B, %%mm2=G, %%mm1=R */
-#if HAVE_MMX2
-        "movq "MANGLE(ff_M24A)", %%mm4     \n\t"
-        "movq "MANGLE(ff_M24C)", %%mm7     \n\t"
-        "pshufw $0x50, %%mm0, %%mm5     \n\t" /* B3 B2 B3 B2  B1 B0 B1 B0 */
-        "pshufw $0x50, %%mm2, %%mm3     \n\t" /* G3 G2 G3 G2  G1 G0 G1 G0 */
-        "pshufw $0x00, %%mm1, %%mm6     \n\t" /* R1 R0 R1 R0  R1 R0 R1 R0 */
-
-        "pand   %%mm4, %%mm5            \n\t" /*    B2        B1       B0 */
-        "pand   %%mm4, %%mm3            \n\t" /*    G2        G1       G0 */
-        "pand   %%mm7, %%mm6            \n\t" /*       R1        R0       */
-
-        "psllq     $8, %%mm3            \n\t" /* G2        G1       G0    */
-        "por    %%mm5, %%mm6            \n\t"
-        "por    %%mm3, %%mm6            \n\t"
-        MOVNTQ" %%mm6, (%1)             \n\t"
-
-        "psrlq     $8, %%mm2            \n\t" /* 00 G7 G6 G5  G4 G3 G2 G1 */
-        "pshufw $0xA5, %%mm0, %%mm5     \n\t" /* B5 B4 B5 B4  B3 B2 B3 B2 */
-        "pshufw $0x55, %%mm2, %%mm3     \n\t" /* G4 G3 G4 G3  G4 G3 G4 G3 */
-        "pshufw $0xA5, %%mm1, %%mm6     \n\t" /* R5 R4 R5 R4  R3 R2 R3 R2 */
-
-        "pand "MANGLE(ff_M24B)", %%mm5     \n\t" /* B5       B4        B3    */
-        "pand          %%mm7, %%mm3     \n\t" /*       G4        G3       */
-        "pand          %%mm4, %%mm6     \n\t" /*    R4        R3       R2 */
-
-        "por    %%mm5, %%mm3            \n\t" /* B5    G4 B4     G3 B3    */
-        "por    %%mm3, %%mm6            \n\t"
-        MOVNTQ" %%mm6, 8(%1)            \n\t"
-
-        "pshufw $0xFF, %%mm0, %%mm5     \n\t" /* B7 B6 B7 B6  B7 B6 B6 B7 */
-        "pshufw $0xFA, %%mm2, %%mm3     \n\t" /* 00 G7 00 G7  G6 G5 G6 G5 */
-        "pshufw $0xFA, %%mm1, %%mm6     \n\t" /* R7 R6 R7 R6  R5 R4 R5 R4 */
-        "movd 4 (%2, %0), %%mm0;" /* Load 4 Cb 00 00 00 00 u3 u2 u1 u0 */
-
-        "pand          %%mm7, %%mm5     \n\t" /*       B7        B6       */
-        "pand          %%mm4, %%mm3     \n\t" /*    G7        G6       G5 */
-        "pand "MANGLE(ff_M24B)", %%mm6     \n\t" /* R7       R6        R5    */
-        "movd 4 (%3, %0), %%mm1;" /* Load 4 Cr 00 00 00 00 v3 v2 v1 v0 */
-\
-        "por          %%mm5, %%mm3      \n\t"
-        "por          %%mm3, %%mm6      \n\t"
-        MOVNTQ"       %%mm6, 16(%1)     \n\t"
-        "movq 8 (%5, %0, 2), %%mm6;" /* Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
-        "pxor         %%mm4, %%mm4      \n\t"
+        RGB_PLANAR2PACKED24("0", "1")
 
-#else
-
-        "pxor      %%mm4, %%mm4     \n\t"
-        "movq      %%mm0, %%mm5     \n\t" /* B */
-        "movq      %%mm1, %%mm6     \n\t" /* R */
-        "punpcklbw %%mm2, %%mm0     \n\t" /* GBGBGBGB 0 */
-        "punpcklbw %%mm4, %%mm1     \n\t" /* 0R0R0R0R 0 */
-        "punpckhbw %%mm2, %%mm5     \n\t" /* GBGBGBGB 2 */
-        "punpckhbw %%mm4, %%mm6     \n\t" /* 0R0R0R0R 2 */
-        "movq      %%mm0, %%mm7     \n\t" /* GBGBGBGB 0 */
-        "movq      %%mm5, %%mm3     \n\t" /* GBGBGBGB 2 */
-        "punpcklwd %%mm1, %%mm7     \n\t" /* 0RGB0RGB 0 */
-        "punpckhwd %%mm1, %%mm0     \n\t" /* 0RGB0RGB 1 */
-        "punpcklwd %%mm6, %%mm5     \n\t" /* 0RGB0RGB 2 */
-        "punpckhwd %%mm6, %%mm3     \n\t" /* 0RGB0RGB 3 */
-
-        "movq      %%mm7, %%mm2     \n\t" /* 0RGB0RGB 0 */
-        "movq      %%mm0, %%mm6     \n\t" /* 0RGB0RGB 1 */
-        "movq      %%mm5, %%mm1     \n\t" /* 0RGB0RGB 2 */
-        "movq      %%mm3, %%mm4     \n\t" /* 0RGB0RGB 3 */
-
-        "psllq       $40, %%mm7     \n\t" /* RGB00000 0 */
-        "psllq       $40, %%mm0     \n\t" /* RGB00000 1 */
-        "psllq       $40, %%mm5     \n\t" /* RGB00000 2 */
-        "psllq       $40, %%mm3     \n\t" /* RGB00000 3 */
-
-        "punpckhdq %%mm2, %%mm7     \n\t" /* 0RGBRGB0 0 */
-        "punpckhdq %%mm6, %%mm0     \n\t" /* 0RGBRGB0 1 */
-        "punpckhdq %%mm1, %%mm5     \n\t" /* 0RGBRGB0 2 */
-        "punpckhdq %%mm4, %%mm3     \n\t" /* 0RGBRGB0 3 */
-
-        "psrlq        $8, %%mm7     \n\t" /* 00RGBRGB 0 */
-        "movq      %%mm0, %%mm6     \n\t" /* 0RGBRGB0 1 */
-        "psllq       $40, %%mm0     \n\t" /* GB000000 1 */
-        "por       %%mm0, %%mm7     \n\t" /* GBRGBRGB 0 */
-        MOVNTQ"    %%mm7, (%1)      \n\t"
-
-        "movd 4 (%2, %0), %%mm0;" /* Load 4 Cb 00 00 00 00 u3 u2 u1 u0 */
-
-        "psrlq       $24, %%mm6     \n\t" /* 0000RGBR 1 */
-        "movq      %%mm5, %%mm1     \n\t" /* 0RGBRGB0 2 */
-        "psllq       $24, %%mm5     \n\t" /* BRGB0000 2 */
-        "por       %%mm5, %%mm6     \n\t" /* BRGBRGBR 1 */
-        MOVNTQ"    %%mm6, 8(%1)     \n\t"
+    YUV2RGB_ENDLOOP(3)
+    YUV2RGB_OPERANDS
+}
 
-        "movq 8 (%5, %0, 2), %%mm6;" /* Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
+static inline int RENAME(yuv420_bgr24)(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
+                                       int srcSliceH, uint8_t* dst[], int dstStride[])
+{
+    int y, h_size;
 
-        "psrlq       $40, %%mm1     \n\t" /* 000000RG 2 */
-        "psllq        $8, %%mm3     \n\t" /* RGBRGB00 3 */
-        "por       %%mm3, %%mm1     \n\t" /* RGBRGBRG 2 */
-        MOVNTQ"    %%mm1, 16(%1)    \n\t"
+    YUV422_UNSHIFT
+    YUV2RGB_LOOP(3)
 
-        "movd 4 (%3, %0), %%mm1;" /* Load 4 Cr 00 00 00 00 v3 v2 v1 v0 */
-        "pxor      %%mm4, %%mm4     \n\t"
-#endif
+        YUV2RGB_INIT
+        YUV2RGB
+        /* mm0=B, %%mm2=G, %%mm1=R */
+        RGB_PLANAR2PACKED24("1", "0")
 
     YUV2RGB_ENDLOOP(3)
     YUV2RGB_OPERANDS