slightly change horizontal lift3 so it needs 1 bit less headroom

Originally committed as revision 10159 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/i386/snowdsp_mmx.c

@@ -333,18 +333,18 @@ void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){
                 "movq   12(%1), %%mm6        \n\t"
                 "paddd    (%1), %%mm2        \n\t"
                 "paddd   8(%1), %%mm6        \n\t"
-                "movq    %%mm2, %%mm0        \n\t"
-                "movq    %%mm6, %%mm4        \n\t"
-                "pslld      $2, %%mm2        \n\t"
-                "pslld      $2, %%mm6        \n\t"
+                "pxor    %%mm0, %%mm0        \n\t" //note: the 2 xor could be avoided if we would flip the rounding direction
+                "pxor    %%mm4, %%mm4        \n\t"
                 "psubd   %%mm2, %%mm0        \n\t"
                 "psubd   %%mm6, %%mm4        \n\t"
                 "psrad      $1, %%mm0        \n\t"
                 "psrad      $1, %%mm4        \n\t"
-                "movq     (%0), %%mm2        \n\t"
-                "movq    8(%0), %%mm6        \n\t"
                 "psubd   %%mm0, %%mm2        \n\t"
                 "psubd   %%mm4, %%mm6        \n\t"
+                "movq     (%0), %%mm0        \n\t"
+                "movq    8(%0), %%mm4        \n\t"
+                "paddd   %%mm0, %%mm2        \n\t"
+                "paddd   %%mm4, %%mm6        \n\t"
                 "movq    %%mm2, (%2)         \n\t"
                 "movq    %%mm6, 8(%2)        \n\t"
                 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])