[FFmpeg-devel,0/4,v2] ffmpeg: late A/V encoder init, AVFrame metadata usage

Message ID 20200913102622.168011-1-jeebjp@gmail.com
Headers show


Jan Ekström Sept. 13, 2020, 10:26 a.m. UTC
This patch set started with a very simple wish to not have to set color
related values manually each time when utilizing ffmpeg.c.

As of the second iteration, this patch set passes FATE on both of my machines.
Only the MXF muxer tests actually needed changing, as the MXF muxer can now
write the configured color range as it is no longer unspecified (in this case,
limited range as flagged by the AVFrame).

Unfortunately, audio still needs two locations where the encoder is
initialized, due to how avfilter_graph_request_oldest peeks and already puts
one AVFrame to be available from the filter graph (which is then utilized
as-is as an early return inside both av_buffersink_get_frame_flags and
av_buffersink_get_samples). If this would be improved in lavfi (or the call
to avfilter_graph_request_oldest removed), we could at least remove one of

Currently limited to using values for video and started with the basic values,
more can be added later if needed.

This probably fixes some trac issues, but with a quick look I couldn't find
anything that explicitly was due to lack of video color metadata passthrough.


Example 1:

I have an RGB 3-D render, which I would like to encode into BT.709 YCbCr.
The video filter I'm generally using for this (zscale) does flag the matrix in
the output AVFrame.

Yet to have the video encoder have the correct metadata set, I have to
set the value(s) manually.

With this patch set, the value(s) from the first AVFrame fed to do_video_out
will be utilized.

Example 2:

I have an input video that sets one or more of the following:
matrix/primaries/transfer function/range/chroma location.

I just want to re-encode it. All of this metadata gets stripped.

With this patch set, the value(s) from the first AVFrame fed to do_video_out
will be utilized.

Example 3:

I have a video which has incorrect metadata tagged. Before, I had to set
the correct data data manually.

With this patch set, since ffmpeg.c takes color related options as dictionary
keys, the AVFrame values will only be utilized if the user has not set the
option for a given stream. Thus, this use case still works.

Jan Ekström (4):
  ffmpeg: deduplicate init_output_stream usage logic
  ffmpeg: move AVFrame time base adjustment into a function
  ffmpeg: move A/V non-streamcopy initialization to a later point
  ffmpeg: pass decoded or filtered AVFrame to output stream

 fftools/ffmpeg.c            | 215 +++++++++++++++++++++++++-----------
 tests/ref/lavf/mxf_d10      |   2 +-
 tests/ref/lavf/mxf_dv25     |   2 +-
 tests/ref/lavf/mxf_dvcpro50 |   2 +-
 tests/ref/lavf/mxf_opatom   |   2 +-
 5 files changed, 155 insertions(+), 68 deletions(-)