[Theora] Playback too slow?!

Alen Ladavac alenl-ml at croteam.com
Wed Mar 30 06:45:28 PST 2005


From: "Ivan Popov" <pin at medic.chalmers.se>
> It is the decoder-only mmx patches from
> http://lists.xiph.org/pipermail/theora-dev/2004-August/002348.html
> (http://kyoto.cool.ne.jp/vp3/developers/theora-a3-MMXd.zip)

Thanks! I believe we are using that, but will recheck.

> > I just thought you were using some magic options I don't know about. So,
> > guess we are back to those peephole optimizations on the mmx decoder. We
are
> > looking for the best mmx implementation to base our optimizations on,
and we
> > are probably going to multithread it and pre-buffer about a dozen frames
or
>
> I hope you will be able to split not only decoding from visualization,
> but also decoding itself into parallel threads, which would tremendously
> improve performance on multicore processors.

Interesting idea. How could that be done? I always thought that next frame
is dependent on the previous one. Or should the operations inside one frame
be paralellized?

> > so, to smooth out the peaks. Any additional ideas are welcome. This just
> > _has_ to work 30fps straight, no drops.
>
> I think the eye will not distinguish between 25 and 30 fps...
> you would spare quite a bit of cpu by decompressing 5 frames instead of 6.

I am still not sure why, but on consoles with TV output, frame-tearing
(swapping in the middle of refresh) just doesn't look good. It may be that
the audience is not used to frame-tearing, unlike PC users. So, since the
output is through an NTSC TV, it must go at 30 fps, because it must be
synched to vertical refresh. The PAL version has movies resampled to 25fps,
but we have to tune for the worst case scenario, which is NTSC. 20 fps would
work for NTSC as well (sync each 3 fields), but it is not smooth enough.

Thanks,
Alen



More information about the Theora mailing list