Shaka packager doesn't handle non-interleaved multi track fragmented MP4 files correctly.

[duggaraju]

System info

Operating System: Any
Shaka Packager Version:

Issue and steps to reproduce the problem

v2.6.1-634af65-release
Packager Command:
ffmpeg -i test.mp4 -c copy test.ismv // convert mp4 to fragmented mp4
packager "in=test.ismv,stream=audio,out=audio.mp4" "in=test.ismv,stream=video,out=video.mp4" --mpd_output a.mpd

Extra steps to reproduce the problem?
(1)
convert MP4 to fragmented MP4 (alternate audio/video tracks)
ffmpeg -i test.mp4 -c copy test.ismv // convert mp4 to fragmented mp4

(2)
packager "in=test.ismv,stream=audio,out=audio.mp4" "in=test.ismv,stream=video,out=video.mp4" --mpd_output a.mpd

What is the expected result?
the packager succeeds.

What happens instead?
A ton of warnings about timestamps gaps/overlaps and the playback doesn't work.

The issue is that shaka mixes the timestamps of audio and video together due to bug in the TrackRunIterator.

[duggaraju]

Shaka assumes that every moof contains all tracks. If the tracks were non interleaved then each moof would only contain a single track and the code would mix up the tiemstamps of the tracks together (treat last audio dts as next video dts and vice versa)

[duggaraju]

ffmpeg -i test.mp4 -c copy -movflags cmaf+separate_moof+delay_moov+skip_trailer dashmp4
packager in=dash.mp4,stream=audio,out=audio.mp4 in=dash.mp4,stream=video,out=video.mp4 --mpd_output dash.mpd

ffmpeg -i test.mp4 -c copy -f ismv smooth.mp4
would also created something similar.

[duggaraju]

@joeyparrish Found the issue and have the fix but wanted to confirm something.
The Flush() test is crashing. It Appends the same fragmented MP4 data twice.
The first time Init() is called the moov has 2 tracks.
But when the second time it is appended, the move is cleared and hence the crash. Not sure if this is a test bug or expected.
The simplest fix is do std::max(moov_->tracks.size(), moof.tracks.size()) but wanted to confirm the same.

[duggaraju]

Looks like I can't always expect Init() to be called before Init(moof). There is code explicitly to ignore the second moov if it is already parsed and Init() wont be called causing the Flush() test to crash. So to be safe do not assume Init() would be called but instead handle this in Init(moof)