[FFmpeg-devel] [PATCH] Use ff_thread_once() to initialize sin/cos static tables.

[Ronald S. Bultje]

Hi,

On Sun, Oct 25, 2015 at 9:39 AM, wm4 <nfxjfg at googlemail.com> wrote:

> On Sun, 25 Oct 2015 07:56:49 -0400
> "Ronald S. Bultje" <rsbultje at gmail.com> wrote:
>
> > Hi,
> >
> > On Sun, Oct 25, 2015 at 7:48 AM, Derek Buitenhuis <
> > derek.buitenhuis at gmail.com> wrote:
> >
> > > On 10/25/2015 11:44 AM, Michael Niedermayer wrote:
> > > >> One could argue this is not a representative sample. It's a single
> > > small file,
> > > >> which must always init (as opposed to a longer running process such
> as
> > > Chrome.
> > > >> Whereas if you have a longer sample, it wouldn't even be within the
> > > margin of
> > > >> error, I bet.
> > > >
> > > > If the usecase is to probe many files, then file duration would not
> > > > help hideing the init time. (to fill a GUI list or whatever)
> > >
> > > I reject this notion. You are completely discounting that people use
> > > libavcodec *as a library*, and not via CLI.
> > >
> > > > Also startup time for normal playback itself does matter, as its
> > > > part of user vissible responsiveness, the user clicks play in his
> > > > players playlist and if that isnt instantaneaously starting playback
> > > > the user notices. So IMO startup time is not irrelevant
> > >
> > > I agree, but neither is thread-safety. Perhaps there is a middle
> ground.
> >
> >
> > So this is likely because we init all tables instead of just these that
> we
> > need, right? So how about having one ff_once per table? That should be
> > trivial to implement.
> >
> > Obviously anyone using shell scripts and calls to CLI ffmpeg versions to
> do
> > probing of production-level numbers of files on production systems
> deserves
> > the results (s)he'll be getting.
>
> Having more AVOnce to reduce the load might be ok.
>
> But am I the only one who thinks it's ridiculous to reject a patch
> fixing thread-safety issues, because someone MIGHT probe 100s of small
> files on a weak ARM? I'd say correctness first, then performance.


The result is not incorrect, not on any real platform. Show me any real
platform where an actual bug occurs.

Ronald