[FFmpeg-devel] Call for "make fate2" tests

[Vitor Sessak]

On 07/16/2010 04:36 PM, Ronald S. Bultje wrote:
> Hi,
>
> On Fri, Jul 16, 2010 at 10:24 AM, Vitor Sessak<vitor1001 at gmail.com>  wrote:
>> On 07/16/2010 04:03 PM, Ronald S. Bultje wrote:
>>>
>>> Hi,
>>>
>>> On Fri, Jul 16, 2010 at 4:57 AM, Vitor Sessak<vitor1001 at gmail.com>    wrote:
>>>>
>>>> On 07/16/2010 12:18 AM, Ronald S. Bultje wrote:
>>>>>
>>>>> On Thu, Jul 15, 2010 at 5:59 PM, Vitor Sessak<vitor1001 at gmail.com>
>>>>>   wrote:
>>>>>>
>>>>>> On 07/15/2010 11:28 PM, Ronald S. Bultje wrote:
>>>>>>>
>>>>>>> On Thu, Jul 15, 2010 at 5:19 PM, Vitor Sessak<vitor1001 at gmail.com>
>>>>>>>   wrote:
>>>>>>>>
>>>>>>>> There are a few formats that could use some tests for "make fate2".
>>>>>>>> Could
>>>>>>>> the maintainers send a patch or at least suggest one or more files to
>>>>>>>> test?
>>>>>>>
>>>>>>> WMAVoice!!!
>>>>>>
>>>>>> Ok course, how could I forgot it?! Can you send a sample/patch?
>>>>>
>>>>> http://samples.mplayerhq.hu/A-codecs/WMSP/
>>>>>
>>>>> I used the streaming_CBR-{7,11,19}K.wma, together they pretty much
>>>>> expose all features of the decoder.
>>>>
>>>> Unfortunately, wmavoice has the same problem of AMR: the floating-point
>>>> errors add up, making the output across arches vary considerably.
>>>> Comparing
>>>> my core duo with Mans' ppc box:
>>>>
>>>> stddev:    0.16 PSNR:111.87 MAXDIFF:   10 bytes:   264014/   264014
>>>> stddev:    1.12 PSNR: 95.34 MAXDIFF:   58 bytes:   264014/   264014
>>>> stddev:    2.50 PSNR: 88.36 MAXDIFF:   98 bytes:   527906/   527906
>>>
>>> I wonder if the vocoder experts can guess with me on whether we can
>>> create a "bitexact" version that does not have this problem.
>>
>> See the thread "Instability in AMR-NB" in ffmpeg-soc ML, it might be the
>> same problem here.
>
> That should be audible, I don't think any problem here was ever
> audibly noticeable.

Yes, but in that case it was about the difference between one 
implementation using int16_t and one using floats. With wmavoice we are 
comparing two float implementations with slightly different roundings, 
differences should be much smaller.

-Vitor