mirror of
https://github.com/livepeer/lpms
synced 2026-04-23 00:07:25 +08:00
Michal Adamczak
c7121378ee
This is quite deep and important change. Thing is, old code had "demuxer reuse" attempt. And unfortunately it neither made sense (next to nothing performance improvement, complex code, limited to MPEGTS streams), nor was working fine. During Low Latency work I tried to eliminate demuxer reusing because I wanted single demuxer logic to make input I/O plug-in easier. It turned out that so modified code failed certain tests such as Nvidia_AudioOnly. It took me a great deal of debugging, but I found out that there is bug in demuxer reuse procedure: namely only first muxer open attempt _really_ checks what is in the segment provided, the rest just assumes that same streams are there. So basically, in the series of "audio/video"/"audio without video" /"audio/video" segments all will be perceived as containing video! This "patched over" *AudioOnly tests, and let them run, but only because Transcoder "though" that it has video all the time. And this simply wasn't true. When I removed "muxer reusing" code, the Nvidia-dedicated "preserve video decoder" tactics was falling over the fact that pixel format changed (because it changed from AV_PIX_FMT_SOMETHING into AV_PIX_FMT_NONE) and tried to create hardware video decoder for AV_PIX_FMT_NONE, and failed. Fixing all that finally allowed me to get rid of recursive call to open_input when HW decoder needs to be reset. Which is good. During the work on this commit I also noticed that various pieces of code here and there assume that video stream is always there. Which is not true, and never was (for example, some tests were using segments without video in the middle of transcode sequence, and it worked only because Transcoder "failed to notice" there is no video anymore). So the whole code was carefully tested and examined against this dangerous assumption and checks were added. These changes, in turn, allowed to remove the limitations that allowed the Transcoder to start only with segments containing video. Transcoder will now happily process audio-only segments.