[FFmpeg-devel] [PATCH] [libdav1d.c]: Add option to output all the spatial layers.

[James Almer]

On 11/14/2019 2:31 PM, Andrey Semashev wrote:
> On 2019-11-14 20:19, James Almer wrote:
>> On 11/14/2019 2:15 PM, Thierry Foucu wrote:
>>> Set the option to false by default, to match libaomdec wrapper.
>>> ---
>>>   libavcodec/libdav1d.c | 4 ++++
>>>   1 file changed, 4 insertions(+)
>>>
>>> diff --git a/libavcodec/libdav1d.c b/libavcodec/libdav1d.c
>>> index cf4b178f1d..5efa8eeb48 100644
>>> --- a/libavcodec/libdav1d.c
>>> +++ b/libavcodec/libdav1d.c
>>> @@ -40,6 +40,7 @@ typedef struct Libdav1dContext {
>>>       int tile_threads;
>>>       int frame_threads;
>>>       int apply_grain;
>>> +    int all_layers;
>>>   } Libdav1dContext;
>>>     static const enum AVPixelFormat pix_fmt[][3] = {
>>> @@ -134,6 +135,8 @@ static av_cold int libdav1d_init(AVCodecContext *c)
>>>       if (dav1d->apply_grain >= 0)
>>>           s.apply_grain = dav1d->apply_grain;
>>>   +    s.all_layers = dav1d->all_layers;
>>> +
>>>       s.n_tile_threads = dav1d->tile_threads
>>>                        ? dav1d->tile_threads
>>>                        : FFMIN(floor(sqrt(threads)),
>>> DAV1D_MAX_TILE_THREADS);
>>> @@ -378,6 +381,7 @@ static const AVOption libdav1d_options[] = {
>>>       { "tilethreads", "Tile threads", OFFSET(tile_threads),
>>> AV_OPT_TYPE_INT, { .i64 = 0 }, 0, DAV1D_MAX_TILE_THREADS, VD },
>>>       { "framethreads", "Frame threads", OFFSET(frame_threads),
>>> AV_OPT_TYPE_INT, { .i64 = 0 }, 0, DAV1D_MAX_FRAME_THREADS, VD },
>>>       { "filmgrain", "Apply Film Grain", OFFSET(apply_grain),
>>> AV_OPT_TYPE_BOOL, { .i64 = -1 }, -1, 1, VD },
>>> +    { "alllayers", "Output all spatial layers", OFFSET(all_layers),
>>> AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
>>>       { NULL }
>>>   };
>>
>> IMO, we don't want to output all layers under any circumstance. It'll
>> result in a mix of frames with duplicate pts and different dimensions.
> 
> IIRC, for VP9 with spatial SVC lavc decoder worked exactly like that -
> by returning a frame per spatial layer. Is this considered a bug?

I don't know if it's a bug, and never seen or tested vp9 svc samples
before. Ronald might know. But it sounds like having more than one video
"stream" per AVCodecContext, which i don't think was intended.
And if it was, then it would be up to the library user to figure what to
do with these frames.