[FFmpeg-devel] [PATCH] avcodec: Add AVClass to AVCodecParameters

[Hendrik Leppkes]

On Tue, May 17, 2016 at 6:26 PM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> Hi,
>
> On Tue, May 17, 2016 at 11:18 AM, Michael Niedermayer <
> michael at niedermayer.cc> wrote:
>
>> AVCodecContext contains all the fields that AVCodecParameters has.
>> Examples are width, height, bitrate, profile, sample_rate, channels
>>
>> In AVCodecContext these fields are accessible through AVOptions
>> in AVCodecParameters these fields are not accessible through AVOptions
>> and your code will crash mysteriously if you try to access them that
>> way
>>
>> This is a inconsistency that was not there before
>
>
> So, at the risk of going semi-offtopic, let me take a different angle at
> this problem.
>
> AVCodecParameter is supposed to be an intermediate layer that removes the
> need for lavf to access lavc structs like AVCodecContext, right?
>
> So, I worked (a long long time ago) on this software called GStreamer, and
> specifically, one of the plugins I worked on early on was gst-ffmpeg. I
> loved working on it because it practically added so many codecs to the
> player which were previously unsupported. Yay! Now, the API back then was
> horrible. You had an AVCodecContext (no AVOptions) and there were
> approximately 100 million fields that could be set. Most were undocumented,
> and it was totally unclear which needed to be set. For example, to play a
> WMA file, you might have to set block_align, but for other audio codecs,
> you do not [1].
>
> For video, sometimes you would have to set extradata (e.g. for SVQ3), but
> other times, you would not. Huffyuv required bits_per_sample, but h264 did
> not.
>
> So how do you write generic user-facing code where the demuxer and decoder
> actually are separate entities and merely interact over a serialized
> communication channel? You can't serialize AVCodecContext. I mean, you
> could, but that's a terrible idea. And nothing documented which fields were
> required to be set and which weren't.
>
> I was hoping that AVCodecParameters would solve this. Ideally, it would be
> a dictionary of fields that are to be shared between demuxer and decoder
> (or codec and muxer). But I think it has sort of strayed off of that intent
> and is just a flat minimized AVCodecContext as it existed in 2003.
>
> Can we fix this? I think AVOptions would be ideally suited to fix this.
> AVClass maybe helps also. But if introspection suggests I can set
> block_align on the h264 decoder, or width on an audio decoder, or who knows
> what. Then I don't think AVOptions or AVClass in AVCodecParameters helps at
> all. It just creates a minimized AVCodecContext from 2003 that will slowly
> become the current AVCodecContext, which is not all that useful for people
> that don't use ffmpeg.exe...
>
>

This is C, I don't want to have to run introspection on an object to
find out what I can set.
A clean and doumented struct is a far more effective way to get this
information across, which is also why I don't want AVOptions for such
a struct, if its not needed to set private options, it would just
double the need for docs (both in doxy and in the AVOption
definition), and to figure out which fields actually exist, you still
have to read the struct or a comment in the header file - so why not
use the struct in the first place?

Why use something "generic" with a dict or some weirdness, which loses
the strong typing and self-documentation of such a struct? Sounds like
you want to be working in another language, honestly. :)
AVOption will never "know" which fields a particular codec will need,
so that point is not going to get solved either way.

- Hendrik