On Sat, Dec 18, 2021 at 02:36:12PM +0100, Michael Niedermayer wrote:
> On Fri, Dec 17, 2021 at 07:04:08PM +0100, Marton Balint wrote:
> > 
> > 
> > On Fri, 17 Dec 2021, Michael Niedermayer wrote:
> > 
> > > On Fri, Dec 17, 2021 at 01:04:19AM +0100, Marton Balint wrote:
> > > > 
> > > > 
> > > > On Thu, 16 Dec 2021, James Almer wrote:
> > > > 
> > > > > Resending the first two patches only, since this is meant to
> > > > > show the implementation of one of the several suggestions made
> > > > > in the previous set that need to be discussed and hopefully
> > > > > resolved in a call.
> > > > 
> > > > Can you push the full branch somewhere?
> > > > 
> > > > > 
> > > > > The proposals so far to extend the API to support either custom
> > > > > labels for channels are, or some form of extra user information.
> > > > > 
> > > > > - Fixed array of bytes to hold a label. Simple solution, but
> > > > >  the labels will have a hard limit that can only be extended
> > > > >  with a major bump. This is what i implemented in this version.
> > > > > - "char *name" per channel that the user may allocate and the
> > > > >  API will manage, duplicate and free. Simple solution, and the
> > > > >  name can be arbitrarily long, but inefficient (av_strdup() per
> > > > >  channel with a custom label on layout copy).
> > > > > - "const char *name" per channel for compile time constants, or
> > > > >  that the user may allocate and free. Very efficient, but for
> > > > >  non compile time strings ensuring they outlive the layout can
> > > > >  be tricky.
> > > > > - Refcounted AVChannelCustom with a dictionary. This can't be
> > > > >  done with AVBufferRef, so it would require some other form
> > > > >  of reference counting. And a dictionary may add quite a bit of
> > > > >  complexity to the API, as you can set anything on them.
> > > > 
> > > > Until we have proper refcounting API we can make the AVBufferRef in
> > > > AVChannelLayout a void *, and only allow channel_layout functions to
> > > > dereference it as an AVBufferRef. This would mean adding some extra helper
> > > > functions to channel layout, but overall it is not unsolvable.
> > > > 
> > > > The real question is that if you want to use refcounting and add helpers to
> > > > query / replace per-channel metadata, or you find the idea too heavy weight
> > > > and would like to stick to flat structs.
> > > 
> > > what is the advantage of refcounting for channel metadata ?
> > > is it about the used memory, about the reduced need to copy ?
> > 
> > Basicly it is the ability to store per-channel metadata in avdictionary,
> > because otherwise it would have to be copyed, and avdictionary is very
> > ineffective at copying because of many mallocs.
> > 
> > > 
> > > what kind of metadata and what size do you expect ?
> > > bytes, kilobytes, megabytes, gigabytes per channel ?
> > 
> > Usually, nothing, because most format don't have support for per-channel
> > metadata. In some cases it is going to be a couple of textual metadata
> > key-value pairs, such as language, label, group, speaker, positon, so 4-5
> > dynamically allocated string pairs, plus the AVDictionary itself, multiplied
> > by the number of channels in a layout.
> > 
> > > 
> > > what is the overhead for dynamic allocation and ref counting?
> > > that is at which point does it even make sense ?
> > 
> > I don't have exact measurements. It is generally felt that copying
> > AVDictionary per-channel is a huge overhead for something as lightweight as
> > an audio frame which is a 2-4 kB per channel at most and only a couple of
> > allocs usually not dependant on the number of channels. That's why
> > refcounting was proposed.
> 
> I was thinking more at a AVStream / AVCodecParameters level.

> How will a demuxer transport such metadata over a AVPacket into a decoder
> outputting metadata-filled AVFrames?

or is this never needed ?

thx

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Awnsering whenever a program halts or runs forever is
On a turing machine, in general impossible (turings halting problem).
On any real computer, always possible as a real computer has a finite number
of states N, and will either halt in less than N cycles or never halt.