[FFmpeg-devel] [PATCH] remove out-dated ADPCM frame_size handling in libavformat

[Justin Ruggles]

Justin Ruggles wrote:

> Michael Niedermayer wrote:
> 
>> On Mon, Sep 06, 2010 at 08:11:38AM -0400, Justin Ruggles wrote:
>> [...]
>>> Index: tests/ref/acodec/g726
>>> ===================================================================
>>> --- tests/ref/acodec/g726	(revision 25042)
>>> +++ tests/ref/acodec/g726	(working copy)
>>> @@ -1,4 +1,4 @@
>>> -5d8cce28f83dd33c3c7eaf43a5db5294 *./tests/data/acodec/g726.wav
>>> -24082 ./tests/data/acodec/g726.wav
>>> -4f1ba1af75dee64625a1c852e6cd01d3 *./tests/data/g726.acodec.out.wav
>>> -stddev: 8504.69 PSNR: 17.74 MAXDIFF:31645 bytes:    96104/  1058400
>>> +fd090ddf05cc3401cc75c4a5ace1d05a *./tests/data/acodec/g726.wav
>>> +24052 ./tests/data/acodec/g726.wav
>>> +74abea06027375111eeac1b2f8c7d3af *./tests/data/g726.acodec.out.wav
>>> +stddev: 8554.55 PSNR: 17.69 MAXDIFF:29353 bytes:    95984/  1058400
>> the number of samples encoded seems to be changing and not equal to
>> the input
> 
> When the frame size in the encoder makes frames end on a byte boundary
> without any padding, the output is always identical.  Since codes are
> between 2 and 5 bits long, how would the decoder distinguish between
> padding to a byte boundary and another valid code?  I'll double-check,
> but it seems that the decoder currently treats padding as additional
> samples.

I've confirmed that this is the cause of the difference.  The parameters
used by the regression test give a 4-bit code size.  When the frame size
is odd, that leads to 1 extra sample being decoded by the decoder
because of padding.  In the current version, because of resampling from
44100 Hz to 8000 Hz, the frame size actually varies from frame-to-frame.

Current:
source samples             = 264600
resampled samples          =  47991
number of odd-sized frames =     61
decoded samples            =  48052
decoded data bytes         =  96104

Patched:
source samples             = 264600
resampled samples          =  47991
number of odd-sized frames =      1 (the last frame)
decoded samples            =  47992
decoded data bytes         =  95984

So choosing a frame size that forces the encoder to only use padding for
the last frame (which this patch does) seems to be the appropriate thing
to do.

-Justin