The code read/write code itself was 1 cycle faster, overall its
likely more due to cache effects

Originally committed as revision 22048 to svn://svn.ffmpeg.org/ffmpeg/trunk