[FFmpeg-devel] H.264: illegal short term buffer state detected

[Reinhard Nissl]

Hi,

Jeff Downs schrieb:

>> The strange thing is, when I cat all those chunks together and play the
>> resulting file with ffplay or mplayer (without any further options), the
>> above mentioned error message doesn't appear at all. So I'm lost --
>> either I missed a switch which would make those players report this
>> message too, or the error is the result of how libxine feeds the data to
>> FFmpeg.
> 
> No, the messages should post with default command line (they are errors).

OK.

> My guess is it has to do with how libxine feeds the data, based on what 
> you are telling us.  Unfortunately, I can't go and look at libxine.

I understand that.

> Does libxine feed data with CODEC_FLAG2_CHUNKS set?  There is possibility 
> of a bug in field picture management there.

Running    grep CODEC_FLAG *   in xine-lib's FFmpeg interfacing
directory just yields CODEC_FLAG_EMU_EDGE.

>> Please tell me, what I shall do next (e. g. uploading the files).
> 
> Would be good if you can look at how libxine feeds data and let us know if 
> the CHUNKS flag is being set.  Barring that, you'll have to figure on what 
> libxine is doing different -- for all we know it could be a bug in the 
> handling of data on its way in to lavc.  The error you are reporting can 
> be caused by data loss, corrupt data, or feeding the same data twice.

I don't think that any of the last three reasons causes this issue
because -- as I wrote previously -- I had xine-lib write the data which
is given to and consumed by avcodec_decode_video() to separate files and
catting these files into a single file results in a properly playing stream.

I'm trying now to write a simple application which takes the above files
on command line and feeds them to avcodec_decode_video(). I hope that
this will show the same errors, giving you a chance to track this issue
down. But as I'm not used to libffmpeg coding, it will take quite a
while for creating this test scenario.

Bye.
-- 
Dipl.-Inform. (FH) Reinhard Nissl
mailto:rnissl at gmx.de