[FFmpeg-devel] [PATCH 4/9] lavc: add generic-decode-layer private data

[Anton Khirnov]

Quoting Andreas Rheinhardt (2023-06-24 21:34:58)
> > diff --git a/libavcodec/internal.h b/libavcodec/internal.h
> > index dceae182c0..b672092ac4 100644
> > --- a/libavcodec/internal.h
> > +++ b/libavcodec/internal.h
> > @@ -49,7 +49,14 @@
> >  #   define STRIDE_ALIGN 8
> >  #endif
> >  
> > +typedef struct DecodeContext DecodeContext;
> > +
> >  typedef struct AVCodecInternal {
> > +    /**
> > +     * Generic decoding private data.
> > +     */
> > +    DecodeContext *d;
> 
> This approach has the downside of adding unnecessary indirecations; it
> furthermore adds pointers to DecodeContext and EncodeContext to
> AVCodecInternal, as if both could be valid at the same time.
> 
> The above could be overcome by using the typical approach to access
> these extra fields, as is used for FFStream etc.
> 
> Furthermore, I do not really think that is worth it: The number of
> fields affected by it are just so small. Has encoder code ever tried to
> set nb_draining_errors?

True, for decoding there are currently very few fields. But there are
many more that are encoding-only - I am not handling more of them in
this set because some of them are not trivial and I did not want to make
this patchset overly large, as its point is fixing the FATE test in 9/9.

And I belive there is a strong case for hiding generic-layer private
state from individual decoders/encoders, as it is prone to being
misunderstood and misused (e.g. AVCodecInternal.draining is currently
used in ways it most likely should not be).

Since the overhead of having a private decode context in addition to a
private encode context is very small, I prefer adding it now, at least
for consistency. Otherwise I'd have to add a shared encode-decode
context, which would be about the same amount of code.

-- 
Anton Khirnov