[FFmpeg-devel] [PATCH] Implement PAFF in H.264

[Neil Brown]

On Wednesday September 19, tjordan at macrosystem.de wrote:
> Neil Brown schrieb:
> > On Tuesday September 18, heydowns at borg.com wrote:
> >> Attached is a patch that implements PAFF in H.264.
> >>
> >> I don't expect this to be accepted right away, due to size and
> >> complexity, but I want to get it out there for review and for people
> >> to try out and comment on.
> > 
> > I just tried it on my clips from my Sony HDR-SR1 Camcorder.
> > 
> > 1/ While somewhat faster than the reference decoder, it is still
> >    slower than real-time on my 2GHz Intel Core2 processor.  So it is
> >    still worth looking for performance improvements.
> maybe this is because the new SMP code can't be used? does the patch
> interfere with Andreas ?mans h264-smp code?

SMP???..... ahhh, there is a "-threads" option.  I hadn't tried that.

It doesn't seem to make a significant difference though.
For this 12 second 1080i clip, the reference decoded uses about
175 seconds of CPU time and ffmpeg uses about 30 (varies from 28 to
32) and changing the "-threads" number to 2, 3, or 4 does not affect
the wallclock time significantly (always around 30seconds).

When I use 2 or more threads, it says

[h264 @ 0xa64920]Cannot parallelize deblocking type 1, decoding such frames in sequential order

so maybe the file always uses deblocking type 1 (??).

So ffmpeg is about 6 times better than the reference decoder (which is
fantastic), but still 2.5 times worse than real-time.  So even if the
deblocking could be parallelised, I would need a 30% faster CPU, or
a third core, or a 30% improvement in the code.

I guess it will be a little while longer before I get real-time
decoding.  But I am still very happy with this improvement.

Thanks,
NeilBrown