(commit by michael)

16-bit divide instead of 32-bit on x86 in msmpeg_pred_dc()

Originally committed as revision 264 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/msmpeg4.c b/libavcodec/msmpeg4.c
index c8df0c4..d737bbc 100644 (file)
--- a/libavcodec/msmpeg4.c
+++ b/libavcodec/msmpeg4.c
@@ -411,10 +411,34 @@ static int msmpeg4_pred_dc(MpegEncContext * s, int n,
        necessitate to modify mpegvideo.c. The problem comes from the
        fact they decided to store the quantized DC (which would lead
        to problems if Q could vary !) */
+#ifdef ARCH_X86
+    /* using 16bit divisions as they are large enough and 2x as fast */
+    asm volatile(
+        "movl %3, %%eax                \n\t"
+       "shrl $1, %%eax         \n\t"
+       "addl %%eax, %2         \n\t"
+       "addl %%eax, %1         \n\t"
+       "addl %0, %%eax         \n\t"
+       "xorl %%edx, %%edx      \n\t"
+        "divw %w3              \n\t"
+       "movzwl %%ax, %0        \n\t"
+       "movl %1, %%eax         \n\t"
+       "xorl %%edx, %%edx      \n\t"
+        "divw %w3              \n\t"
+       "movzwl %%ax, %1        \n\t"
+       "movl %2, %%eax         \n\t"
+       "xorl %%edx, %%edx      \n\t"
+        "divw %w3              \n\t"
+       "movzwl %%ax, %2        \n\t"
+        : "+r" (a), "+r" (b), "+r" (c)
+       : "r" (scale)
+       : "%eax", "%edx"
+    );
+#else    
     a = (a + (scale >> 1)) / scale;
     b = (b + (scale >> 1)) / scale;
     c = (c + (scale >> 1)) / scale;
-
+#endif
     /* XXX: WARNING: they did not choose the same test as MPEG4. This
        is very important ! */
     if (abs(a - b) <= abs(b - c)) {