[Ffmpeg-devel] [RFC] faad/rtsp/aac bug...

[Ryan Martell]

Hey...

On Oct 30, 2006, at 7:03 PM, Michael Niedermayer wrote:

> Hi
>
> On Mon, Oct 30, 2006 at 06:38:44PM -0600, Ryan Martell wrote:
>> Hi--
>>
>> So there's a bug in rtsp.c/faad, that I could use a suggestion on
>> fixing.
>>
>> 1) in rtsp.c:
>>                     codec->sample_rate = i;
>>                     get_word_sep(buf, sizeof(buf), "/", &p);
>>                     i = atoi(buf);
>>                     if (i > 0)
>>                         codec->channels = i;
>>                     // TODO: there is a bug here; if it is a mono
>> stream, and less than 22000Hz, faad upconverts to stereo and twice  
>> the
>>                     //  frequency.  No problem, but the sample rate
>> is being set here by the sdp line.  Upcoming patch forthcoming. (rdm)
>>
>> 2) In faad.c:
>>     if (avctx->extradata){
>>         r = s->faacDecInit2(s->faac_handle, (uint8_t*) avctx-
>>> extradata,
>>                             avctx->extradata_size,
>>                             &samplerate, &channels);
>>         if (r < 0){
>>             av_log(avctx, AV_LOG_ERROR,
>>                    "faacDecInit2 failed r:%d   sr:%ld  ch:%ld  s:%d 
>> \n",
>>                    r, samplerate, (long)channels, avctx-
>>> extradata_size);
>>         } else {
>>             avctx->sample_rate = samplerate;
>>             avctx->channels = channels;
>>             s->init = 1;
>>         }
>>     }
>>
>> 3) The extradata specifies the channel count and the frequency.  In
>> my case, it specifies it as a 22050/1, which is also what the SDP
>> line says.
>>
>> 4) in the faad code, if the frequency is lower than a 22000Hz, it
>> doubles the frequency, and if it is a mono channel, it doubles it to
>> stereo.
>>
>> The hardware buffers are opened in ffplay, after the rtsp stuff is
>> parsed but before the aac code is called that tells it to
>> upconvert.   As a result, if it is a 22050 mono stream, it is really
>> slow sounding.
>>
>> I could include the logic from aac that talks about how to do this
>> (based on the extradata), but I believe that the upconversion/
>> doubling of channels might just be a faad thing, and put that in the
>> rtp code, but that seems potentially buggy.
>>
>> Any advice on where/how to fix this cleanly would be appreciated.
>
> the bug is clearly in faad, if the stream is 22khz mono faad should
> output that and nothing else
> id suggest to simply discard every second sample and the second chanel
> in libavcodec/faad.c if faad messes with the samplingrate (but first
> check that there is no parameter to faad which disables this silly
> behavior and second confirm that the stream really is 22kz and not  
> 44khz)

It really is 22050.  It's my stream, and if I replace the TODO above  
with

codec->sample_rate = 44100
codec->channels = 2;

which it upconverts it to, it works just fine.  I also checked the  
faad source to verify I wasn't nuts.  I'm not sure if this is  
something that it is supposed to do (ie from some specification) or  
not.  I'm like you- if it says 22050/1, it should output 22050/1.   
There may have been some reason relating to non-compatible sound  
hardware.

I don't think there's a flag, but I will definitely check.

Other than that though, you think I should hack on faad.c to fix it.   
I'll have to verify that that won't break the non-rtsp case.

Thanks,
-R