[FFmpeg-devel] [RFC]separation of multiple outputs' encoding

[Zhong Li]

Tomas Härdin <tjoppen at acc.umu.se> 于2020年5月19日周二 下午10:25写道：
>
> fre 2020-05-15 klockan 10:14 +0800 skrev Tao Zhang:
> > Marton Balint <cus at passwd.hu> 于2020年5月15日周五 上午2:33写道：
> > >
> > >
> > > On Thu, 14 May 2020, Tao Zhang wrote:
> > >
> > > > Hi,
> > > > FFmpeg supports multiple outputs created out of the same input in the
> > > > same process like
> > > > ffmpeg -i input -filter_complex '[0:v]yadif,split=3[out1][out2][out3]' \
> > > >        -map '[out1]' -s 1280x720 -acodec … -vcodec … output1 \
> > > >        -map '[out2]' -s 640x480  -acodec … -vcodec … output2 \
> > > >        -map '[out3]' -s 320x240  -acodec … -vcodec … output3
> > > > In ffmpeg.c, multiple outputs are processed sequentially like
> > > > for (i = 0; i < nb_output_streams; i++)
> > > >    encoding one frame;
> > > >
> > > > As below wiki noted, the slowest encoder will slow down the whole
> > > > process. Some encoders (like libx264) perform their encoding "threaded
> > > > and in the background", but x264_encoder_encode still cost some time.
> > > > And it is noticeable when multiple x264_encoder_encodes run in the same thread.
> > > > https://trac.ffmpeg.org/wiki/Creating%20multiple%20outputs
> > > >
> > > > For API users, they can use separate thread for multiple encoding in
> > > > their own code. But is there any way to rescue command line users?
> > >
> > > I am not sure I understand what you want. Processing will still be limited
> > > to the slowest encoder, because input processing will still be driven by
> > > the slowest encoder, even if the encoders run in separate threads.
> > >
> > > Buffering encoder input frames is not an option, because input frames are
> > > uncompressed, therefore huge. So if you want the faster x264 encoder to
> > > finish faster, you have to drive it from a different input, ultimately you
> > > should run 3 separate encode processes and accept that decoding and yadif
> > > processing happens in all 3 cases separately causing some overhead.
> > Currently FFmpeg works like below:
> > main thread:
> > encoding frame 1 for output 1; encoding frame 1 for output 2; encoding
> > frame 1 for output 3; encoding frame 2 for output 1; encoding frame 2
> > for output 2; encoding frame 2 for output 3;...
> >
> > What I want to do is
> > thread 1:
> > encoding frame 1 for output 1; encoding frame 2 for output 1;...
> > thread 2:
> > encoding frame 1 for output 2; encoding frame 2 for output 2;...
> > thread 3:
> > encoding frame 1 for output 3; encoding frame 2 for output 3;...
>
> Why not just run multiple copies of the ffmpeg program? Decoding should
> almost always be very minor work compared to encoding.
Not vey minor IMHO.  For same resolution, H264 decoding is about 10%
computing cost of encoding.
And if adding the filter computing cost (just as yadif filter listed
above), the redundant computing cost is huge.