[FFmpeg-cvslog] r18959 - in trunk: libavcodec/mpeg12.c tests/seek.regression.ref

Mike Melanson mike
Mon Jun 15 04:46:52 CEST 2009


Reimar D?ffinger wrote:
> On Tue, May 26, 2009 at 09:40:46PM -0700, Mike Melanson wrote:
>> bcoudurier wrote:
>>> Author: bcoudurier
>>> Date: Wed May 27 02:14:32 2009
>>> New Revision: 18959
>>>
>>> Log:
>>> Fix off by one offset with fetch_timestamps, pts_parser_problem.mpg.
>>> Patch by Wolfram Gloger, wmglo at dentm dot med dot uni-muenchen dot de.
>>>
>>> Modified:
>>>    trunk/libavcodec/mpeg12.c
>>>    trunk/tests/seek.regression.ref
>> Do you have any idea why this might have broken the pva-demux test?
>>
>> http://fate.multimedia.cx/index.php?test_spec=298
>>
>> The file errors out with the following message:
>>
>> [NULL @ 0x120ab5730]error, non monotone timestamps 50151 >= 42951
>> av_interleaved_write_frame(): Error while opening file
> 
> I don't like it much, but since FATE staying broken or disabling the
> test seems even worse:
> How about changing the test to be based on the
> http://samples.mplayerhq.hu/PVA/_B5_Crusade_-_Kosmiczna_Krucjata_01pl_1276.pva
> or
> http://samples.mplayerhq.hu/PVA/_Babylon_5_5x01pl_1251.pva
> sample?
> That leaves 5/7 of our sample files broken but better than nothing...

I started testing the samples and noticed that all 7 appear to work just 
fine. Then I noticed that I omitted the "-acodec copy -vcodec copy" 
options. So everything is okay when decoding the audio and video.

Solution: Omit the "-vcodec copy" option (the MP2 is also decoding bit 
exact across platforms, but I don't trust that) and start decoding the 
video, though with the "-idct simple" option which is supposed to be bit 
exact across platforms and compilers.

New problem: Bit exactness is not maintained across platforms (PPC 
differs from x86_32/64).

Solution: Cheat and only decode some of the sample, just before problems 
are known to set in:

$BUILD_PATH/ffmpeg -idct simple -i 
$SAMPLES_PATH/pva/PVA_test-partial.pva -t 0.6 -acodec copy -f framecrc -

Hopefully, this solves this longstanding testing discrepancy satisfactorily.

-- 
     -Mike Melanson



More information about the ffmpeg-cvslog mailing list