[FFmpeg-devel] [PATCH] h264: don't touch H264Context->ref_count[] during MB decoding.

[Clément Bœsch]

On Thu, Oct 04, 2012 at 08:53:02AM -0700, Ronald S. Bultje wrote:
> Hi,
> 
> On Oct 4, 2012 8:46 AM, "Clément Bœsch" <ubitux at gmail.com> wrote:
> >
> > On Thu, Oct 04, 2012 at 05:37:27PM +0200, Clément Bœsch wrote:
> > > On Thu, Oct 04, 2012 at 07:51:18AM -0700, Ronald S. Bultje wrote:
> > > > Hi,
> > > >
> > > > On Oct 4, 2012 7:16 AM, "Clément Bœsch" <ubitux at gmail.com> wrote:
> > > > >
> > > > > On Thu, Oct 04, 2012 at 04:13:24PM +0200, Clément Bœsch wrote:
> > > > > [...]
> > > > > > > >      if(MB_MBAFF){
> > > > > > > > -        h->ref_count[0] <<= 1;
> > > > > > > > -        h->ref_count[1] <<= 1;
> > > > > > > > +        local_ref_count[0] <<= 1;
> > > > > > > > +        local_ref_count[1] <<= 1;
> > > > > > > >      }
> > > > > > >
> > > > > > > maybe
> > > > > > > local_ref_count[0] = h->ref_count[0] << MB_MBAFF
> > > > > > > is faster but otherwise whatever is fast & clean is fine with me
> > > > > > >
> > > > > >
> > > > > > Ah right, I'll push the attached version tomorrow unless there is
> an
> > > > > > objection then.
> > > > > >
> > > > >
> > > > > Derp, sent a bad draft version. Fixed.
> > > >
> > > > Has anyone been able to test if this decreases the amount of fate
> failures?
> > > >
> > >
> > > It's running since a while without any failure here. We'll be sure in a
> > > week or two with FATE. Helgrind still complains about the state issues I
> > > mentioned in the other mail though.
> > >
> >
> > Ah. It just failed with a higer cpu load. So it looks not enough.
> 
> OK. Do you have any feel if it fails less?
> 

It's hard to tell, it really is quite unpredictable: I'm just
experimenting with: while make -j20 fate-h264 THREADS=8; do done

(I'm not doing only the cama2_vtc_b test to allow the -j20 to work
properly and rape the CPU; it clearly helps triggering the problem)

It can run smoothly for long minutes or fail almost at the start. The
patch clearly is an improvement anyway IMO.

-- 
Clément B.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 490 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20121004/ed1295dc/attachment.asc>