[FFmpeg-devel] [PATCH v1 7/7] lavc/vaapi_dec: Add VVC decoder

[Nuo Mi]

>
> > > --- a/libavcodec/vaapi_decode.c
> > > +++ b/libavcodec/vaapi_decode.c
> > > @@ -455,6 +455,9 @@ static const struct {
> > >      MAP(AV1,         AV1_MAIN,        AV1Profile0),
> > >      MAP(AV1,         AV1_HIGH,        AV1Profile1),
> > >  #endif
> > > +#if VA_CHECK_VERSION(1, 22, 0)
> > > +    MAP(H266,        VVC_MAIN_10,     VVCMain10),
> > > +#endif
> > >
> > >  #undef MAP
> > >  };
> > > @@ -627,6 +630,10 @@ static int
> > > vaapi_decode_make_config(AVCodecContext *avctx,
> > >          case AV_CODEC_ID_VP8:
> > >              frames->initial_pool_size += 3;
> > >              break;
> > > +        case AV_CODEC_ID_H266:
> > > +            // Add additional 16 for maximum 16 frames delay in
> > > vvc native decode.
> > > +            frames->initial_pool_size += 32;
> >
> > One frame of 8k YUV444, 10 bits, is about 200MB. Thirty-two frames
> > amount to approximately 6GB.Can we dynamically allocate the buffer
> > pool?
>
> It's processing in other thread:
> https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=11316
>
> >
> > The software decoder requires a delay of 16 frames to ensure full
> > utilization of CPUs. In the future, we may consider increasing this
> > to 32 or even 64 frames.
> > However, for hardware decoding, given that all processing occurs on
> > the GPU, we do not require any delay.
>
> The delay can avoid sync hardware task immediately once it is
> submitted, which can avoid hardware switch tasks frequently and drop
> performance. If the number will increase, I'd prefer to set it as an
> option and diff the default value for hardware with software.

Why does VVC require such a large frame pool while other hardware codecs do
not?
What makes VVC so special?"