[FFmpeg-devel] [RFC] [PATCH] huffyuvenc, multithread and context=1
michaelni at gmx.at
Wed Jun 11 19:43:10 CEST 2014
On Wed, Jun 11, 2014 at 07:26:13PM +0200, Christophe Gisquet wrote:
> 2014-06-10 22:28 GMT+02:00 Christophe Gisquet <christophe.gisquet at gmail.com>:
> > Case in point for 1000 frames of a 1920x1080 VC1 video, without
> > writing the output (hard drive speed being a limiting factor):
> > - context 0: 1286434kB, ~42fps conversion
> > - context 1: 653038kB, ~24fps
> > - context 1, patch 0001: 651001kB, ~38fps / 21fps(1 thread)
> > - context 1, patch 0002 (non-deterministic, should worsen with thread
> > count): ~657xxxkB, ~40fps
> More on this, considering context=0 uses a somewhat crude distribution
> (using probability 1/d). I tried using noisier sequences also as well
> as different distributions on a 6-cores PC:
> context=0,1/d: 851974 27226 1137281
> context=1: 471819 22604 972351
> context=1,ND: 472875 22673 972582
> context=0,1/d²: 619081 25069 1051500
> context=0,1/d³: 501983 30454 1290561
> context=0,lapl: 500650 31754 1304082
> The last result, for a Laplace distribution (0,0.8), shows how
> important adaption needs to be. On the other hand, using a probability
> of 1/d² seems to perform better than the default one.
maybe we should chnage the default ...
> And removing the determinism in context=1 is not really a show-stopper
> coding efficiency-wise. I'd personally prefer to go with that one, but
> that feels like a new option rather than -strict -3 (you may want
> non-experimental, standard stuff but with speed).
maybe some combination of:
CODEC_FLAG2_FAST and CODEC_FLAG_BITEXACT could be used
or an entirely new -context mode
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
It is what and why we do it that matters, not just one of them.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 181 bytes
Desc: Digital signature
More information about the ffmpeg-devel