[FFmpeg-devel] Fixing av_seek

[Don Moir]

For the most part, I have seeking in my app working really well. I have 
flags for exact time seeking or key frame seeking. When doing exact time 
seeking, I have a seek slop factor that currently ranges from 0 to 1/2 
second. I compute this depending on the file and it just governs how close 
you want it to the timestamp of interest. In the odd case where the actual 
seeked to time is far off from the requested timestamp, it could go back to 
keyframe seeking and this could be controlled with flags at a users 
degression. This is a pretty rare event though considering that large files 
generally have an adequate supply of keyframes and with low res or short 
files, it's quick enough to do it.

Most of the work has been in over coming a few problems with 
avformat_seek_file and the related decoders.

The problem is you cannot have avformat_seek_file failing when it should not 
and even if it succeeds it must be doing the right thing.

Some cases when avformat_seek_file fails or fails to do the right thing:

o - No frames are marked as keyframes. avformat_seek_file returns -1. I 
think I only have one file like this so we could say this is a rare case. I 
just make sure I at least seek to the first frame. In order to keep 
avformat_seek_file from failing, you have to mark the first frame as a key 
frame and I do this. The one file I have like this is filecopy.avi and it 
can be found here: https://ffmpeg.org/trac/ffmpeg/ticket/504

o - The ogg decoder read_seek function is weak. It does not make an attempt 
to create it's index_entries and without these seeking is going to be bad. 
This is outlined here: https://ffmpeg.org/trac/ffmpeg/ticket/941

o - The matroska decoder read_seek function has a couple bugs that effects 
at least one file I have and that is outlined here: 
https://ffmpeg.org/trac/ffmpeg/ticket/958

With the above changes, things work very well but there are a few odd cases 
still. These cases don't effect things too much. Mostly just a lag after 
seeking is done and before the next frame can be displayed. This happens 
only on a few ogg files and matroska files that I know of. It could be 
because of improper timestamps, improper index_entries, or even just a bad 
file. Not sure.

I have a case with WMV2 where after seeking is done, audio is far behind. 
This was easy for me to deal with and I have not investigated it further at 
present.

I tested about 200 files of varing sizes, formats, etc and it works pretty 
good when the above fixes are applied. For most file formats, I don't have 
to do much and they just work.