[Ffmpeg-devel] ffmpeg hanging when transcoding mpeg4

[Aaron Williams]

Aaron Williams wrote:
> Aaron Williams wrote:
>   
>> Aaron Williams wrote:
>>   
>>     
>>> Michael Niedermayer wrote:
>>>   
>>>     
>>>       
>>>> Hi
>>>>
>>>> On Sat, Oct 21, 2006 at 02:32:11AM -0700, Aaron Williams wrote:
>>>>   
>>>>     
>>>>       
>>>>         
>>>>> I am trying to transcode a video for a friend into MPEG2 for use with a
>>>>> hardware MPEG decoder and am running into problems with FFMPEG stopping
>>>>> at the same place in the video and consuming all of the CPU.  I seem to
>>>>> be unable to do any source level debugging with gdb (even when compiled
>>>>> with -O2 and -g) so I cannot find the source of the hang.  It always
>>>>> occurs at about time 236.2 in the file.
>>>>>     
>>>>>       
>>>>>         
>>>>>           
>>>> you do use ffmpeg_g for debugging right?
>>>> and maybe try valgrind
>>>>   
>>>>     
>>>>       
>>>>         
>>> I do not see ffmpeg_g.  Do I need to do something special in configure
>>> to build that?  Also, valgrind is reporting problems, but again there
>>> appears to be no debugging information.
>>>
>>> I do see the following though:
>>> ==20851== Conditional jump or move depends on uninitialised value(s)
>>> ==20851==    at 0x4DEE61F: ff_mpeg4_decode_mb (in
>>> /usr/local/lib/libavcodec.so.51.20.0)
>>> ==20851==    by 0x4DFFB83: (within /usr/local/lib/libavcodec.so.51.20.0)
>>> ==20851==    by 0x4E00F40: ff_h263_decode_frame (in
>>> /usr/local/lib/libavcodec.so.51.20.0)
>>> ==20851==    by 0x4CCF222: avcodec_decode_video (in
>>> /usr/local/lib/libavcodec.so.51.20.0)
>>> ==20851==    by 0x40A5D3: (within /usr/local/bin/ffmpeg)
>>> ==20851==    by 0x40C2C4: main (in /usr/local/bin/ffmpeg)
>>>
>>>
>>> I can get the file to transcode if I scale it to certain resolutions. 
>>> The source is MPEG4 at 720x380, encoded with an aspect ratio of 16:9. 
>>> If I scale it to 720x480 it hangs.  If I scale it to 720x352 and add 64
>>> pixels of padding above and below it works.  It seems to be a bug
>>> somewhere in the scaling code.  If I convert to h264 instead of mpeg2
>>> then it seems to go through OK.
>>>
>>> I think the original video was improperly scaled for a 16:9 aspect ratio
>>> as well.
>>>
>>> I am trying to cut the size of the source file as well so I can post a
>>> link to a smaller file, but this is failing as well:
>>>
>>> /usr/local/bin/ffmpeg -i DCF-Brick-DivX-high.avi -vcodec copy -acodec
>>> copy -y -ss 220 -t 30 DCF-Brick3-hang.avi
>>> FFmpeg version SVN-r6280, Copyright (c) 2000-2006 Fabrice Bellard, et al.
>>>   configuration:  --enable-libogg --enable-vorbis --enable-faad
>>> --enable-faac --enable-xvid --enable-x264 --enable-a52 --enable-dts
>>> --enable-shared --enable-dc1394 --enable-gpl --enable-mp3lame
>>>   libavutil version: 49.0.1
>>>   libavcodec version: 51.20.0
>>>   libavformat version: 50.6.0
>>>   built on Oct 21 2006 02:51:51, gcc: 4.1.0 (SUSE Linux)
>>>
>>> Seems that stream 0 comes from film source: 29.98 (65534/2186) -> 29.97
>>> (2997/100)
>>> Input #0, avi, from 'DCF-Brick-DivX-high.avi':
>>>   Duration: 00:07:09.6, start: 0.000000, bitrate: 865 kb/s
>>>   Stream #0.0: Video: mpeg4, yuv420p, 720x380, 29.97 fps(r)
>>>   Stream #0.1: Audio: mp3, 48000 Hz, stereo, 128 kb/s
>>> Output #0, avi, to 'DCF-Brick3-hang.avi':
>>>   Stream #0.0: Video: mpeg4, yuv420p, 720x380, q=2-31, 29.98 fps(c)
>>>   Stream #0.1: Audio: mp3, 48000 Hz, stereo, 128 kb/s
>>> Stream mapping:
>>>   Stream #0.0 -> #0.0
>>>   Stream #0.1 -> #0.1
>>> Press [q] to stop encoding
>>> error, non monotone timestamps 9170 >= 9170
>>> error, non monotone timestamps 9173 >= 9173
>>> ...
>>> frame=  900 q=0.0 Lsize=    3589kB time=250.0 bitrate= 117.6kbits/s
>>> video:3039kB audio:476kB global headers:0kB muxing overhead 2.103378%
>>>
>>> -Aaron
>>>
>>>
>>> _______________________________________________
>>> ffmpeg-devel mailing list
>>> ffmpeg-devel at mplayerhq.hu
>>> http://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-devel
>>>
>>>
>>>
>>>
>>>
>>>   
>>>     
>>>       
>> I found the ffmpeg_g and that works a lot better for debugging.
>> It looks like it is stuck in the MPV_encode_picture where it
>> continuously goes back to vbv_retry.
>>
>>
>> (gdb) p s->pb
>> $3 = {bit_buf = 0, bit_left = 32, buf = 0x2b726eaa4010 "",
>>   buf_ptr = 0x2b726eab3dba, buf_end = 0x2b726ebf5810 }
>> (gdb) p s->rc_context.buffer_index
>> $4 = 1544315
>> (gdb)                         
>> (gdb) p s->next_lambda
>> $5 = 0
>> (gdb) p s->lambda
>> $6 = 0
>> (gdb)  p s->qscale
>> $7 = 2
>>
>> It looks like the problem is that s->lambda is zero and will never
>> scale, hence it gets stuck in an endless loop.
>>
>> -Aaron
>> _______________________________________________
>> ffmpeg-devel mailing list
>> ffmpeg-devel at mplayerhq.hu
>> http://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-devel
>>
>>
>>
>>   
>>     
> It looks a bit more fundamental than that.  In this case, s->qscale is
> also 2, so even if s->lambda is 1 it will also never increase.
>
> -Aaron
>   

--- mpegvideo.c~        2006-10-17 02:31:20.000000000 -0700
+++ mpegvideo.c 2006-10-21 13:48:57.000000000 -0700
@@ -2535,8 +2535,10 @@
             RateControlContext *rcc= &s->rc_context;
             int max_size= rcc->buffer_index/3;

-            if(put_bits_count(&s->pb) > max_size && s->qscale <
s->avctx->qmax){
+            if((put_bits_count(&s->pb) > max_size) && (s->qscale <
s->avctx->qmax)){
                 s->next_lambda= s->lambda*(s->qscale+1) / s->qscale;
+                if (s->lambda == s->next_lambda)
+                    s->next_lambda++;
                 s->mb_skipped = 0;        //done in MPV_frame_start()
                 if(s->pict_type==P_TYPE){ //done in encode_picture() so
we must undo it
                     if(s->flipflop_rounding || s->codec_id ==
CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)

I'm sure there's probably something cleaner, but this solves the problem
and it no longer hangs.
  Otherwise, it is possible to get stuck in an endless loop.

-Aaron