a little less than 200 cpu cycles faster with the cathedral sample.

Originally committed as revision 21785 to svn://svn.ffmpeg.org/ffmpeg/trunk