[FFmpeg-devel] [PATCH 1/4] fftools/ffmpeg_enc: strip DOVI config record for AV1

[Anton Khirnov]

Quoting Niklas Haas (2024-03-21 13:11:32)
> On Thu, 21 Mar 2024 11:16:57 +0100 Anton Khirnov <anton at khirnov.net> wrote:
> > Quoting Niklas Haas (2024-03-19 20:16:39)
> > > From: Niklas Haas <git at haasn.dev>
> > > 
> > > AV1 streams don't use configuration records, so delete them when
> > > encoding to AV1. Ideally this would be, as the comment suggests, handled
> > > at the frame-level (and stripped by the av1 encoder), but given the
> > > status quo of copying the packet-level data here directly, we should
> > > definitely make an effort to strip it.
> > > ---
> > >  fftools/ffmpeg_enc.c | 25 ++++++++++++++-----------
> > >  1 file changed, 14 insertions(+), 11 deletions(-)
> > 
> > I'm very much not a fan of having codec-specific code in ffmpeg CLI. It
> > implies that every single caller must now be aware of this
> > (undocumented?) interaction of this specific side data with this
> > specific codec ID.
> 
> Note: This is an existing bug, not introduced by this series. This
> series just makes it obvious. The status quo is that, beacuse of this
> logic in ffmpeg_enc.c, we incorrectly forward dolby vision configuration
> records when transcoding to AV1.

I know pretty much nothing about dolby vision, so could you please
explain why precisely is this incorrect? And at what point in the
transcoding chain does the side data become invalid?

> Or, indeed, when transcoding to *any* format - since current FFmpeg also
> does not propagate dolby vision RPUs, we generate broken files pretty
> much always when transcoding dolby vision. So we definitely need to
> strip the metadata from the stream muxer *somewhere*. Where else comes
> to mind?
> 
> This also gets into another topic I wanted to touch on, which is that
> the presence of dynamic dolby vision metadata currently hinders the
> ability of libavfilter to treat the video primaries/gamma as
> a negotiable colorspace property (the way it is done currently for YUV
> matrix/range). This is because when interpreted as such, DV metadata
> fundamentally changes the colorspace of the incoming video stream.
> Ideally we would like some way to negotiate DV metadata on the
> query_formats() level.
> 
> Ideally, we'd want something like AVCOL_SPC_DOLBYVISION, but we can't
> easily introduce that without breaking ISO/IEC 23091 compatibility..

In principle it could be yet another negotiated field, could it not? You
just added a bunch of those recently, what's another one?

-- 
Anton Khirnov