[FFmpeg-devel] ffvorbis, inverted output

Benjamin Larsson banan
Sun Sep 28 10:55:52 CEST 2008

Siarhei Siamashka wrote:
> On Saturday 27 September 2008, Siarhei Siamashka wrote:
>> On Sunday 14 September 2008, Michael Niedermayer wrote:
>>> On Sun, Sep 14, 2008 at 04:50:59AM +0300, Siarhei Siamashka wrote:
>>>> Hi,
>>>> I tried to do PSNR comparison of libvorbis/ffvorbis/tremor and noticed
>>>> that output from ffvorbis is actually inverted (ex. output 0xFFFF in
>>>> libvorbis corresponds to 0x0001 in ffvorbis and so on) when compared to
>>>> the output from the other decoders.
>>>> Should this be fixed?
>>> yes, if all (/most) other vorbis decoders match and we differ from that
>> Can these two patches be used as a fix? The first one adds support for
>> scaled imdct output. The second one inverts output of the decoder (to match
>> libvorbis and tremor) using negative scale factor.
>> As additional bonus, 'copy_normalize' function from vorbis decoder is
>> simplified. Though I get some inconsistent benchmark results (performance
>> difference is negligible with one or another variant getting ahead
>> randomly) and would like someone to confirm that there is no performance
>> regression.
>> Getting scaled imdct output involves sqrt operation and scale factor uses
>> odd power of two, so there is some difference in PSNR compared to SVN
>> trunk (taking inversion into account):
>> stddev:    0.02 PSNR:127.97 bytes: 22057216/ 22057216
> Oops, an extra patch to ensure tables alignment is also needed. It is
> attached.
> Regarding performance, it really provides a visible improvement on some files
> (the ones which have lots of block size switching and consequently lots
> of 'copy_normalize' calls). After having done more tests, I observed vorbis
> decoding speedup up to 1%. In any case, output scaling support in IMDCT with
> zero overhead should be a good thing to have and might be useful in other
> decoders too.

I posted a patch a long time ago that calculated the window on the fly.
With the code it would be easy to scale the window in any way you want
and it would be as fast as now.

Benjamin Larsson

More information about the ffmpeg-devel mailing list