* optimized remaing avg_pixels_xy2

Originally committed as revision 621 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/i386/dsputil_mmx_avg.h b/libavcodec/i386/dsputil_mmx_avg.h
index a958a0c..703cd0f 100644 (file)
--- a/libavcodec/i386/dsputil_mmx_avg.h
+++ b/libavcodec/i386/dsputil_mmx_avg.h
@@ -258,47 +258,39 @@ static void DEF(avg_pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size
 {
     __asm __volatile(
        MOVQ_BONE(%%mm7)
-       "xorl %%eax, %%eax              \n\t"
+       "lea (%3, %3), %%eax            \n\t"
        "movq (%1), %%mm0               \n\t"
-       "movq 1(%1), %%mm1              \n\t"
-       PAVGB" %%mm1, %%mm0             \n\t"
-       ".balign 16                     \n\t"
+       PAVGB" 1(%1), %%mm0             \n\t"
+       ".balign 8                      \n\t"
        "1:                             \n\t"
-       "movq (%2, %%eax), %%mm1        \n\t"
-       "movq (%3, %%eax), %%mm2        \n\t"
-       "movq 1(%2, %%eax), %%mm3       \n\t"
-       "movq 1(%3, %%eax), %%mm4       \n\t"
+       "movq (%1, %%eax), %%mm2        \n\t"
+       "movq (%1, %3), %%mm1           \n\t"
        "psubusb %%mm7, %%mm2           \n\t"
-       PAVGB" %%mm3, %%mm1             \n\t"
-       PAVGB" %%mm4, %%mm2             \n\t"
+       PAVGB" 1(%1, %3), %%mm1         \n\t"
+       PAVGB" 1(%1, %%eax), %%mm2      \n\t"
+       "addl %%eax, %1                 \n\t"
        PAVGB" %%mm1, %%mm0             \n\t"
        PAVGB" %%mm2, %%mm1             \n\t"
-       "movq (%4, %%eax), %%mm3        \n\t"
-       "movq (%5, %%eax), %%mm4        \n\t"
-       PAVGB" %%mm3, %%mm0             \n\t"
-       PAVGB" %%mm4, %%mm1             \n\t"
-       "movq %%mm0, (%4, %%eax)        \n\t"
-       "movq %%mm1, (%5, %%eax)        \n\t"
-       "addl %6, %%eax                 \n\t"
-       "movq (%2, %%eax), %%mm1        \n\t"
-       "movq (%3, %%eax), %%mm0        \n\t"
-       "movq 1(%2, %%eax), %%mm3       \n\t"
-       "movq 1(%3, %%eax), %%mm4       \n\t"
-       PAVGB" %%mm3, %%mm1             \n\t"
-       PAVGB" %%mm4, %%mm0             \n\t"
+       PAVGB" (%2), %%mm0              \n\t"
+       PAVGB" (%2, %3), %%mm1          \n\t"
+       "movq %%mm0, (%2)               \n\t"
+       "movq %%mm1, (%2, %3)           \n\t"
+       "movq (%1, %3), %%mm1           \n\t"
+       "movq (%1, %%eax), %%mm0        \n\t"
+       PAVGB" 1(%1, %3), %%mm1         \n\t"
+       PAVGB" 1(%1, %%eax), %%mm0      \n\t"
+       "addl %%eax, %2                 \n\t"
+       "addl %%eax, %1                 \n\t"
        PAVGB" %%mm1, %%mm2             \n\t"
        PAVGB" %%mm0, %%mm1             \n\t"
-       "movq (%4, %%eax), %%mm3        \n\t"
-       "movq (%5, %%eax), %%mm4        \n\t"
-       PAVGB" %%mm3, %%mm2             \n\t"
-       PAVGB" %%mm4, %%mm1             \n\t"
-       "movq %%mm2, (%4, %%eax)        \n\t"
-       "movq %%mm1, (%5, %%eax)        \n\t"
-       "addl %6, %%eax                 \n\t"
+       PAVGB" (%2), %%mm2              \n\t"
+       PAVGB" (%2, %3), %%mm1          \n\t"
+       "movq %%mm2, (%2)               \n\t"
+       "movq %%mm1, (%2, %3)           \n\t"
+       "addl %%eax, %2                 \n\t"
        "subl $4, %0                    \n\t"
-       " jnz 1b                        \n\t"
-       :"+g"(h)
-       :"D"(pixels), "S"(pixels+line_size), "r"(pixels+line_size*2), "r" (block),
-        "r" (block+line_size), "g"(line_size<<1)
+       "jnz 1b                         \n\t"
+       :"+g"(h), "+D"(pixels), "+S"(block)
+       :"r" (line_size)
        :"%eax",  "memory");
 }