"Improved patch. Now supports subpel ME on all candidate MB types,
not just on the winner.

subpel_refine: (completely different scale from before)
0 => halfpel only
1 => 1 iteration of qpel on the winner (same as x264 r46)
2 => 2 iterations of qpel (about the same as my earlier patch, but faster
3 => halfpel on all MB types, qpel on the winner
4 => qpel on all
5 => more iterations

benchmarks:
mencoder dvd://1 -ovc x264 -x264encopts
qp_constant=19:fullinter:cabac:iframe=200:psnr

subpel_refine=1:  PSNR Global:46.82 kb/s:1048.1 fps:17.335
subpel_refine=2:  PSNR Global:46.83 kb/s:1034.4 fps:16.970
subpel_refine=3:  PSNR Global:46.84 kb/s:1023.3 fps:14.770
subpel_refine=4:  PSNR Global:46.87 kb/s:1010.8 fps:11.598
subpel_refine=5:  PSNR Global:46.88 kb/s:1006.9 fps:10.824"

 And

"The current code for calculating the cost of encoding which reference
frame a MB is predicted from, introduces a bias towards ref0 and
against P16x16.
Removing this bias produces an improvement of .4% - 2% bitrate,
depending on content and number of reference frames."



git-svn-id: svn://svn.videolan.org/x264/trunk@47 df754926-b1dd-0310-bc7b-ec298dee348c