[FFmpeg-devel] [PATCH 2] hevc: use different thread wait/report for tmvp

Ronald S. Bultje rsbultje at gmail.com
Sun Jul 20 18:58:27 CEST 2014


Hi,


On Sun, Jul 20, 2014 at 12:31 PM, Christophe Gisquet <
christophe.gisquet at gmail.com> wrote:

> Hi,
>
> 2014-07-20 13:35 GMT+02:00 Ronald S. Bultje <rsbultje at gmail.com>:
> > Do you get a speedup? I tinkered with some related problem in vp9 once
> and
> > I never got a significant speedup out of it...
>
> Whatever your definition of significant, I don't think the figure I have
> is.
>
> It's around 1% on a dual core system on average, but that's near the
> measurement error on the other hand. I'll probably be able to provide
> figures for systems with more cores within a few days. Another
> possible comparison point is the impact in openhevc. I've heard it is
> more than 20% faster per core over ffhevc, therefore I'd expect such
> things to cause a bigger impact.
>
> Technically, I wager hevc and vp9 are different here. hevc has a
> second level of filtering called sao, that causes a delay of one CTB
> line/64 pixel lines. So, effectively, parsing and pixel output are
> probably farther apart than in vp9.


That's a fair explanation - and yes, 1% is totally worth it (imo).

Ronald


More information about the ffmpeg-devel mailing list