The case for 16x16 blocks becomes 10 cpu cycles faster on pentium dual,
i could not find a speed difference in the case of subblocks though.

Originally committed as revision 16226 to svn://svn.ffmpeg.org/ffmpeg/trunk