[FFmpeg-devel] [PATCH] VP8: correctly use optimal epel functions for splitmv mode

Ronald S. Bultje rsbultje
Sun Jun 27 19:57:30 CEST 2010


On Sat, Jun 26, 2010 at 8:19 PM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> currently, we apply MC/epel for splitmv coding as 4x4 subblocks (of
> 4x4px each) in the 16x16px MB. This is suboptimal, because the MVs are
> actually shared between multiple subblocks, so applying epel in
> 16x8/8x16/8x8 would be more optimal, particularly if we use SSE2/SSSE3
> optimizations.
> The attached patch tries to improve the situation.
> Once the SSE2/MMX patches are applied, this leads to about 10% speedup
> for splitmv MBs (5937 to 5486 cycles per whole splitmv-MB for sample
> 15 in the vector testsuite). Of course this depends on the coding of
> the MB and thus on the sample. With SSSE3 it probably leads to even
> better speedups, but I can't test that because my CPU is old.

New patch against SVN after David's bilinear filter addition.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: vp8_subblocks.patch
Type: application/octet-stream
Size: 6377 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20100627/be0bab30/attachment.obj>

More information about the ffmpeg-devel mailing list