[MPlayer-dev-eng] Re: Vorbis decoding performance

[Diego Biurrun]

On Wed, Aug 02, 2006 at 10:38:39PM +0200, Balatoni Denes wrote:
> 
> Thanks for the benchmarks, looks good!
> IMHO the reason why ffvorbis was so slow on the k6-III is the
> non-SSE floating point mdct (k6-III has no SSE, right?).

Yes, K6-III has 3dnow and 3dnowext and mmx, but no sse.

> Otherwise, ffvorbis was 20-30% faster than the floating-point official 
> libvorbis when I was testing with ffmpeg a year ago (Uoti's benchmark shows 
> something similar).
> 
> BTW if there was a fixed-point mdct in ffmpeg (xiph's can be stolen, but 
> that's not too creative, is it) a blazing fast ffvorbis could be made, 
> assuming that 32 bit is enough for mdct coefficients - the vorbis spec 
> explicitly warns against this, and "mandates" 48 bits (and in fact tremor 
> uses some tricky floating-point for this reason), but 32 bit is easily enough 
> for all current vorbis streams/encoders.

Any volunteers?  Outperforming Xiph's decoder in all situations would be
way cool :-)

Diego