[FFmpeg-devel] [FFmpeg-cvslog] vf_scale: use colorspace negotiation API

[Niklas Haas]

On Mon, 29 Jan 2024 04:19:43 +0100 Michael Niedermayer <michael at niedermayer.cc> wrote:
> On Sun, Dec 31, 2023 at 09:49:47PM +0000, Niklas Haas wrote:
> > ffmpeg | branch: master | Niklas Haas <git at haasn.dev> | Tue Oct 31 13:52:53 2023 +0100| [45e09a30419cc2a7251e72689142e021ecdfe6d9] | committer: Niklas Haas
> > 
> > vf_scale: use colorspace negotiation API
> > 
> > This filter will always accept any input format, even if the user sets
> > a specific in_range/in_color_matrix. This is to preserve status quo with
> > current behavior, where passing a specific in_color_matrix merely
> > overrides the incoming frames' attributes. (Use `vf_format` to force
> > a specific input range)
> > 
> > Because changing colorspace and color_range now requires reconfiguring
> > the link, we can lift sws_setColorspaceDetails out of scale_frame and
> > into config_props. (This will also get re-called if the input frame
> > properties change)
> > 
> > > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=45e09a30419cc2a7251e72689142e021ecdfe6d9
> > ---
> > 
> >  libavfilter/vf_scale.c | 188 +++++++++++++++++++++++++++----------------------
> >  1 file changed, 103 insertions(+), 85 deletions(-)
> 
> This seems to remove metadata
> 
> ./ffmpeg -f lavfi -i color=black -pix_fmt yuv422p10le -bitexact -vframes 1 -f framecrc -
> 
>  Stream #0:0: Video: rawvideo (Y3[10][10] / 0xA0A3359), yuv422p10le(tv, progressive), 320x240 [SAR 1:1 DAR 4:3], q=2-31, 38400 kb/s, 25 fps, 25 tbn
> vs
>  Stream #0:0: Video: rawvideo (Y3[10][10] / 0xA0A3359), yuv422p10le(progressive), 320x240 [SAR 1:1 DAR 4:3], q=2-31, 38400 kb/s, 25 fps, 25 tbn
> 
> 
>  when its stored in a file, it results in a different output, example:
> ./ffmpeg -f lavfi -i color=black -pix_fmt yuv422p10le -bitexact -vframes 1 -c:v ffv1 -level 3 -y ffv1-lavf-black-4636-new.mkv
> 
> 
> I have not investigated what is correct, but i would think "no range information"
> would seem worse than whatever the correct is

Looking at this quickly, what is happening here is not that vf_scale is
stripping metadata, it's that vf_scale used to *add* default metadata after
a pixel format conversion - even though it did not touch the pixel
range, and the incoming frames were untagged. This was arguably a bug.
The new behavior is actually less surprising here - the vsrc generates
untagged frames, so vf_scale preserves them being untagged.

(Observe -vf showinfo,format=yuv422p10le,showinfo before vs after patch)

In any case, the *proper* fix here would be to make vsrc_testsrc aware
of the colorspace negotiation API. Ideally, it should switch from
ff_draw_init to ff_draw_init2, and use the negotiated YUV colorspace and
range. (pal(75|100)bars are also bugged)