[FFmpeg-devel] Buffers Threads and refcounts

[Paul B Mahol]

On 1/8/13, Michael Niedermayer <michaelni at gmx.at> wrote:
> Hi all
>
> The fork plans to make some changes on how various buffers work, that
> is adding thread safe ref counting and doing all kinds of related and
> unrelated changes.
> They plan to stash this 5000+ lines diff changing 300+ files into
> few commits, to quote:
> "The changes are split into per-codec patches to simplify
>  review, but will be squashed on push."
>
> See:
> http://lists.libav.org/pipermail/libav-devel/2013-January/041437.html
> http://lists.libav.org/pipermail/libav-devel/2013-January/041447.html
> http://lists.libav.org/pipermail/libav-devel/2013-January/041576.html
>
> This quite large (and ineviable quite buggy) changeset will cause
> various problems
> * Moving AVFrame to libavutil causes ABI/API issues. Currently
>   if a new field is needed its just added and used, after the move
>   it will be needed to keep libavcodec, libavutil, libavfilter and
>   other users of AVFrame in sync or take extra care on accessing some
>   fields that might not have existed in other libs when they where
>   build.
> * The patchsets remove various buffer pools which will cause
>   significant speedloss.
>   It should be possible to reimplement these pools though
> * Interfacing with any application that doesnt have refcounted frames
>   will likely need a memcpy.
> * Various fields are removed from public API, these will break
>   usage of libavcodec with libpostproc, any reuse of motion vectors
>   or macroblock types on reencoding and probably all attempts to
>   extract statistics about the internals of a video codec, short of
>   parsing av_log() debug output.
>
> Above issues will hit libav as well as ffmpeg if we merge the patches.
> ffmpeg will be hit a bit harder because we have more optimizations
> that are incompatible with the refcounted buffers.
>
> The advantages of this patchset are
> 1. It simplifies the code, this is true, but this is
>    just because optimizations, features and not understood bugfixes
>    are removed. The rest is moving compexity around more than removing
>    it. And i doubt the code will be any simpler once all bugs are
>    fixed and once it runs at the same speed again.
>    (also its 5926 insertions(+), 5464 deletions(-)
>     that is 500 lines more than before the patches even with many
>     optimizations droped)
> 2. It simplifies the permissions in libavfilter at the expense of
>    speed. Currently libavfilter (of ffmpeg) supports almost any kind
>    of buffer, static, read only, .... After the patchset only the most
>    common kind of buffers are supported, that IS simpler and slower
>    because if your buffer doesnt fit in the API you need to copy it.
>    Which way is better, i dont know, but the permission system is
>    entirely independant of what these patchsets do. Such change should
>    have been done independantly
> 3. decoders dont return refcounted buffers, this patchset changes this
>    making users happier.
>    The same effect though can be achived with the current API/ABI by
>    just putting a av_refcounted_get_buffer() inside lavc and the
>    user app using that.
>
> in some cases it will need fewer memcpies in other cases, it will
> need more. And it will do the thread safe refcounting in all cases
> and that will slow all cases down, especially small frames like in
> some audio decoders/demuxers will suffer from this
>
> So in summary, these patchsets are bad, really bad in the short term
> and less so in the long term once all optimizations are put back
> and in the distant future it might even be an improvment and lead to
> slightly cleaner code overall.
>
> What to do about that ?
> If we do not merge them then this means an end to API/ABI
> compatibility with libav, this would be a big annoyance to
> applications using libav*, thus their oppinon is quite welcome here.
>
> If we do merge them this means alot of work, these are ~ 6000 changed
> lines done on a 2 year outdated fork. A fork missing many codecs,
> filter and (de)muxers that we do have. The code is buggy on its own
> and possible a lesser number of bugs will get introduced from the
> merge
>
> These changes should have been done on top of ffmpeg changing all
> codecs and filters And should have been reviewed by the authors who
> wrote the code that is being changed. But the authors are unwelcome in
> libav. The changes also should stay in a branch until all
> optimizations are back in place and bugs fixed instead of being
> commited to libav.
> But i doubt thats how it will be.
>
> If we do merge these patchsets, we must be aware that the code will
> likely be in not so good shape afterwards (nor will libav be in better)
> and will need time and lots of work.
> Also merging these patchsets that where worked on since 3+ months by
> libav developers will not be possible in 1 day, it will take longer
>
> Also consider that not merging them and the API/ABI incompatibility
> would affect distributions and applications. And that in this case
> There may be a need for volunteers to package ffmpeg for some
> distributions. (We either way still need volunteers for official
> debian & ubuntu packages of ffmpeg)
>
> My current plan is to attempt to integrate the patchsets once they hit
> libav, and fix as many of the bugs they contain as i can. But iam
> surely not opposed to the alternative if ffmpeg developers and
> application developers support going a different path.
>
> Also it would be great if the fork nonsense would end and the libav
> developers would join ffmpeg and we could all work together making
> these changes without bugs and without speedloss instead of the
> duplication of work this now will become ...

That is unthinkable.

> Also consider, if i wouldnt have had to work on daily merging of
> libav, we now probably would have a HEVC decoder and a useable AAC
> encoder but i rather had to spend my time on that. And now i will
> have to work on integrating and fixing this instead of more usefull
> things ...

I'm against auto-merging hell. If something is useful any interested
party will work on to put it into FFmpeg by doing work in separate branch
and post it on list for review once it is ready.

Thus if you really think that refcounted code as will be in fork is better
(I do not until someone shows me numbers) than current state it should be
merged into FFmpeg by using separate branch, and once quality reach
certain level it could get into master.