[FFmpeg-devel] [PATCH] H.264: faster skip decoding

Michael Niedermayer michaelni
Sun Jan 17 02:45:32 CET 2010


On Sat, Jan 16, 2010 at 07:43:03PM -0500, Jason Garrett-Glaser wrote:
> On Sat, Jan 16, 2010 at 7:17 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> > On Sat, Jan 16, 2010 at 02:37:14PM -0500, Jason Garrett-Glaser wrote:
> >> On Sat, Jan 16, 2010 at 2:34 PM, Jason Garrett-Glaser
> >> <darkshikari at gmail.com> wrote:
> >> > On Sat, Jan 16, 2010 at 2:25 PM, Jason Garrett-Glaser
> >> > <darkshikari at gmail.com> wrote:
> >> >> On a very skip-heavy source:
> >> >>
> >> >> After:
> >> >> real ? ?12m52.048s
> >> >> user ? ?10m6.360s
> >> >> sys ? ? 1m53.584s
> >> >>
> >> >> Before:
> >> >> real ? ?13m21.203s
> >> >> user ? ?10m28.450s
> >> >> sys ? ? 1m53.646s
> >> >>
> >> >> Dark Shikari
> >> >>
> >> >
> >> > Ugh, it seems the hashes don't match, despite the fact that this patch
> >> > really shouldn't break anything. ?Ideas are welcome as to what's wrong
> >> > here, because the idea seems to be sound.
> >> >
> >> > Dark Shikari
> >> >
> >>
> >> And the problem is deblocking. ?Ignore this patch, it doesn't work
> >> with ffmpeg's deblocking model. ?Maybe if we did if(IS_SKIP(mbtype) &&
> >> deblocking is off)...
> >
> > Iam currently trying to move the deblocing into a seperate pass
> 
> Are you sure this is a good idea? 

i dont know its an experiment
if its slower it goes to /dev/null


> It might be much better on L1I, but
> it might be a lot worse on everything else, since everything will be
> pushed out of cache and then back into cache... on the other hand, it
> will allow threading with disable_deblock_idc = 1.

ff_h264_filter_mb() is around 16kb alone not counting the code it calls
having to load this into L1 code cache for every MB is a big speed killer.
Its also totally deadlocked as any optimization that makes it larger
makes it slower

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Opposition brings concord. Out of discord comes the fairest harmony.
-- Heraclitus
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20100117/e3f8ebb6/attachment.pgp>



More information about the ffmpeg-devel mailing list