Unused video outputs are currently attached to libvlc, so there is no
point in destroying them from the playlist (twice), then from libvlc.
This should also fix #1593 (multiple vout free).

(assuming rt-offset is zero)

POSIX warrants at least 32 priorities from
sched_get_priority_min(SCHED_RR) to sched_get_priority_max(SCHED_RR).
We were previously relying on 41 priorities, from 0 to 40.

Contrary to process priority increment ("niceness"), thread priority
increment increase in the priority of the thread. So multiplying
negative priorities with -1 was totally wrong here.

On Linux, there is only one non-realtime (SCHED_OTHER) priority, 0, so
the code was failing safe silently. Don't know on other systems.

This should avoid priority inversion.

So-called "RTSP" VoD remains extremely broken,
and I have no intention to fix it.

This reverts commit 05cb3470.

This reverts commit 0bcda15e.

This reverts commit e9fdaf06.

This reverts commit 77a7babe.

Simple demo:
vlc --sout="#duplicate{dst=display,dst='transcode{vcodec=mp4v}:std{access=http,dst=0.0.0.0:8080,mux=ts}'}" --ignore-config  --use-stream-immediate movie.avi &
vlc --use-stream-immediate http://127.0.0.1:8080
and
vlc --use-stream-immediate http://127.0.0.1:8080 --auto-adjust-pts-delay

You'll have to wait a bit until the pts delay is auto adjusted. Generally it takes 2-4 seconds, because access set a very high default pts delay value. One amelioration would be to lower the pts_delay when this option is set to allow a quicker convergence.

The general algorithm requires some tuning, but results are here.

Note, this only works if there is a video track. A similar function could be developped for the audio tracks.

Don't trust cherry-pick blindly.

./vlc movie.avi --sout="#transcode{aenc=dummy,venc=stats}:std{access=http,mux=dummy,dst=0.0.0.0:8081}"
./vlc -vvv http://127.0.0.1:8081 --demux=stats --vout=stats --codec=stats

stats: Initial import of a stat oriented module. Use with /vlc --ignore-config -I dummy -vvv movie.avi --sout="#transcode{aenc=dummy,venc=stats}:std{access=http,mux=dummy,dst=0.0.0.0:8080}"

It is much more efficient regarding latency as it doesn't bufferize more than needed, and let the module access take care of that eventually.

Enable with --use-stream-immediate. We may want to default it.

Note: --use-stream-immediate will be only effective on access that don't provide pf_block() for now. This is because I didn't benchmark against the Block method.

Here the gain that I did measure is about 200ms (less latency) when using the http access on a loopback.