takes the most time, and it allows for more efficient unaligned access
and better control over memory latencies.

Originally committed as revision 711 to svn://svn.ffmpeg.org/ffmpeg/trunk