[vlc-devel] [PATCH] RFC: clock/aout: audio preroll discussion

[Thomas Guillem]

On Wed, Jun 20, 2018, at 16:23, Thomas Guillem wrote:
> 
> On Tue, Jun 19, 2018, at 19:37, Rémi Denis-Courmont wrote:
> > Le tiistaina 19. kesäkuuta 2018, 18.09.32 EEST Thomas Guillem a écrit :
> > > Currently, only the pulse audio module is able to start at a given time
> > > (because the API allows to start a stream corked). Every other modules will
> > > start right after the VLC Start callback.
> > 
> > Err no. Start() does not have any audio data. Most audio outputs start on the 
> > first Play(). And that first Play() actually conveys zero padding generated by 
> > the core to time the playback correctly.
> 
> I didn't know that. Good to know.
> 
> > 
> > The problem is that this does not generally work well. It requires that the 
> > latency is known correctly up-front. Not all audio drivers/hardware can 
> > support this.
> > 
> > > The problem is that the first play often need to be played in the future.
> > 
> > That's not a problem: you can always defer audio, if you know how long.
> > The opposite would be a problem.
> > 
> > > There is often a 30/50ms pts delay (tested on current master). With the
> > > future output clock, this delay should be arround the dejitter value of the
> > > vlc_clock_t (200 or 500ms). No problem for pulse audio, but every other
> > > modules will fail to start at this given time.
> > 
> > If the audio output makes up a poor latency estimate, then this happens. 
> > That's why PulseAudio does corking instead. If you remove the corking and 
> > timer from PulseAudio, PulseAudio will simply exhibit the same problem.
> > 
> > > We have some choices:
> > > 
> > >  1/ Simplify the date handling of aout modules: all modules behave the same
> > >  way, and the core handle the delay (by playing silence when needed). So we
> > >  move the delayed start functionally from PulseAudio to the core: this is
> > > what this patch is doing.
> > 
> > This is exactly what is already being done if TimeGet() succeeds before 
> > playback starts.
> 
> So if I understand correctly: for pulseaudio, TimeGet() will fail before 
> playback starts (because the uncork is delayed), but it's OK since this 
> module handle the deferred start itself. 
> 
> For other modules, they should return a valid TimeGet() before playback 
> is started in order to play some silence that will correspond to the 
> audio device hw delay. Right ?
> 
> But how do you handle, in the future clock branch, a first play that is 
> deferred by the clock dejitter time (200-500ms) ?
> The core need to play some silence somehow, or should every module 
> handle it themselves ?

Ah OK:

    drift += mdate () - dec_pts;
    /* Early audio output.
     * This is rare except at startup when the buffers are still empty. */

I see, this is already handled naturally. So, for the future clock branch, we just have to fix the first TimeGet of every aout modules (except PulseAudio).

> 
> > 
> > >  2/ Modify the aout API to handle 2 type of aout modules: one like
> > > pulseaudio (that handle a play date), and all the others one.
> > 
> > And this is the only way to get things working somewhat properly within the 
> > current system, where the playback time is determined ahead.
> > 
> > > I prefer the solution 1/ since it'll ease future debugging because all
> > > modules will behave the same way.
> > 
> > In other words, you want to keep the core as it is and rip out some code from 
> > PulseAudio. End result: PulseAudio will work worse and the other outputs will 
> > be exactly as broken as before.
> > 
> > Well no thanks.
> > 
> > > If you prefer the solution 2/, we need to think about a proper aout API.
> > 
> > I prefer solution 2 and the proper API is what we have right now.
> > 
> > -- 
> > 雷米‧德尼-库尔蒙
> > http://www.remlab.net/
> > 
> > 
> > 
> > _______________________________________________
> > vlc-devel mailing list
> > To unsubscribe or modify your subscription options:
> > https://mailman.videolan.org/listinfo/vlc-devel
> _______________________________________________
> vlc-devel mailing list
> To unsubscribe or modify your subscription options:
> https://mailman.videolan.org/listinfo/vlc-devel