[FFmpeg-devel] [DEVEL][PATCH] ffmpeg: fix channel_layout bug on non-default layout

[pkv.stream]

Le 10/11/2017 à 1:12 AM, Michael Niedermayer a écrit :
> On Thu, Nov 09, 2017 at 09:37:33PM +0100, pkv.stream wrote:
>> Hi Michael,
>>
>>>>   ffmpeg_opt.c |   11 ++++++++++-
>>>>   1 file changed, 10 insertions(+), 1 deletion(-)
>>>> 2af07f4366efdfaf1018bb2ea29be672befe0823  0001-ffmpeg-fix-channel_layout-bug-on-non-default-layout.patch
>>>>  From 4ec55dc88923108132307b41300a1abddf32e6f7 Mon Sep 17 00:00:00 2001
>>>> From: pkviet <pkv.stream at gmail.com>
>>>> Date: Mon, 2 Oct 2017 11:14:31 +0200
>>>> Subject: [PATCH] ffmpeg: fix channel_layout bug on non-default layout
>>>>
>>>> Fix for ticket 6706.
>>>> The -channel_layout option was not working when the channel layout was not
>>>> a default one (ex: for 4 channels, quad was interpreted as 4.0 which is
>>>> the default layout for 4 channels; or octagonal interpreted as 7.1).
>>>> This led to the spurious auto-insertion of an auto-resampler filter
>>>> remapping the channels even if input and output had identical channel
>>>> layouts.
>>>> The fix operates by directly calling the channel layout if defined in
>>>> options. If the layout is undefined, the default layout is selected as
>>>> before the fix.
>>>> ---
>>>>   ffmpeg_opt.c | 11 ++++++++++-
>>>>   1 file changed, 10 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/ffmpeg_opt.c b/ffmpeg_opt.c
>>>> index 100fa76..cf5a63c 100644
>>>> --- a/ffmpeg_opt.c
>>>> +++ b/ffmpeg_opt.c
>>>> @@ -1804,6 +1804,12 @@ static OutputStream *new_audio_stream(OptionsContext *o, AVFormatContext *oc, in
>>>>           MATCH_PER_STREAM_OPT(audio_channels, i, audio_enc->channels, oc, st);
>>>> +        AVDictionaryEntry *output_layout = av_dict_get(o->g->codec_opts,
>>>> +                                                       "channel_layout",
>>>> +                                                       NULL, AV_DICT_MATCH_CASE);
>>> This doesnt look right
>>>
>>> not an issue of the patch as such but
>>> why is the channel_layout option per file and not per stream?
>>
>> just my ignorance; do you mean this is not the right way to retrieve
>> the channel_layout option from an audio stream ?
> I think there is more buggy with how the channel_layout is handled
>
> try this:
> ./ffmpeg -i ~/videos/matrixbench_mpeg2.mpg -channel_layout 5 test.wav
> and this:
> ./ffmpeg -i ~/videos/matrixbench_mpeg2.mpg -channel_layout:a 5 test.wav
>
> while it may appear that the are both working this is deceiving.
> I think only the channel number gets actually used in the 2nd case
>
> Look at what your code with av_dict_get() reads.
> It does not obtain the 5 in the 2nd case

Hi Michael,

Thanks for the explanation;
actually I have the impression that it's worse than that.
In the second case indeed av_dict_get() doesn't get the 5 (was not aware 
of channel_layout:a syntax , is it documented somewhere ?)
In the first case, the 5 is recognized but the auto resampler spouts 
errors :

ffmpeg -loglevel debug -channel_layout octagonal -i test.mkv 
-channel_layout 5 test.wav
...
Successfully opened the file.
Parsing a group of options: output url test.wav.
Applying option channel_layout (set channel layout) with argument 5.
Successfully parsed a group of options.
Opening an output file: test.wav.
[file @ 0000018F66CF5C20] Setting default whitelist 'file,crypto'
Successfully opened the file.
Stream mapping:
   Stream #0:1 -> #0:0 (pcm_s24le (native) -> pcm_s16le (native))
Press [q] to stop, [?] for help
cur_dts is invalid (this is harmless if it occurs once at the start per 
stream)
     Last message repeated 3 times
detected 8 logical cores
[graph_0_in_0_1 @ 0000018F66D05760] Setting 'time_base' to value '1/48000'
[graph_0_in_0_1 @ 0000018F66D05760] Setting 'sample_rate' to value '48000'
[graph_0_in_0_1 @ 0000018F66D05760] Setting 'sample_fmt' to value 's32'
[graph_0_in_0_1 @ 0000018F66D05760] Setting 'channel_layout' to value 
'0x737'
[graph_0_in_0_1 @ 0000018F66D05760] tb:1/48000 samplefmt:s32 
samplerate:48000 chlayout:0x737
[format_out_0_0 @ 0000018F687E0C60] Setting 'sample_fmts' to value 's16'
[format_out_0_0 @ 0000018F687E0C60] Setting 'channel_layouts' to value '0x5'
[format_out_0_0 @ 0000018F687E0C60] auto-inserting filter 
'auto_resampler_0' between the filter 'Parsed_anull_0' and the filter 
'format_out_0_0'
[AVFilterGraph @ 0000018F66D48DE0] query_formats: 4 queried, 6 merged, 3 
already done, 0 delayed
[auto_resampler_0 @ 0000018F687E0EC0] [SWR @ 0000018F66D2CE80] Using 
fltp internally between filters
[auto_resampler_0 @ 0000018F687E0EC0] [SWR @ 0000018F66D2CE80] Output 
channel layout '2 channels (FL+FC)' is not supported
[auto_resampler_0 @ 0000018F687E0EC0] Failed to configure output pad on 
auto_resampler_0
Error reinitializing filters!
Failed to inject frame into filter network: Invalid argument
Error while processing the decoded data for stream #0:1

Not sure why swr_build_matrix is issuing an error. Is this a case of a 
bug hiding another bug ? namely (1) wrongly matrixing to default 
channel_layout instead of real one  , (2)  auto-inserted filter which 
seems only able to deal with non default layout ( that is to say,  
anything other than mono stereo 2.1 4.0 5.0 5.1 7.1 hexadecagonal).

If you have ideas to suggest for solving this mess, I'll try to work 
some more on this but this bug is growing too big for me, so without 
input, I'll leave it to senior devs.

Regards

>
> [...]
>
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel