[Ffmpeg-devel] Fixed vs. Floating Point AAC

[Oded Shimon]

On Fri, Mar 10, 2006 at 12:23:24AM +0100, Balatoni Denes wrote:
> Hi!
> 
> cs?t?rt?k 09 m?rcius 2006 23.07-kor Rich Felker ezeket a bolcs gondolatokat 
> fogalmazta meg:
> > How can 32*32>>32 possibly be slower than 32*32? It's just a matter of
> > whether you read the result from eax or edx...
> 
> You are right. I think everything should be using fixed point, because only 
> lamers use floating point, it's a horrible idea.

I'm bad at this, are you being sarcastic?

> Regarding the vorbis benchmark, it sounds interesting (I would be even more 
> interested in how ffmpeg would have compared). It is mostly the same code 
> (tremor and libvorbis), so I am somewhat puzzled at the result you are 
> claiming.

(mplayer is tremor)

oggdec 10.Entwined.ogg -o /dev/null   2.61s user 0.01s system 97% cpu 2.682 total
mplayer 10.Entwined.ogg -ao pcm       2.26s user 0.02s system 97% cpu 2.340 total
ffmpeg -i 10.Entwined.ogg bla.wav     1.82s user 0.19s system 64% cpu 3.136 total
./nutvorbis audiodump.ogg > /dev/null 1.58s user 0.03s system 91% cpu 1.775 total


oggdec audiodump.ogg -o /dev/null     27.92s user 0.21s system 96% cpu 29.140 total
mplayer audiodump.ogg -ao pcm         24.94s user 0.21s system 87% cpu 28.798 total
./ffmpeg -i audiodump.ogg bla.wav     26.64s user 3.24s system 77% cpu 38.421 total
./nutvorbis audiodump.ogg > /dev/null 17.79s user 0.24s system 95% cpu 18.942 total


Odd thing, the first statistic is a very old one, took it several months 
ago, i just retried ffmpeg on it and it was much slower (2.3s)... some 
ffmpeg vorbis (or imdct) has gotten quite slower (it used to be as fast or 
faster than nutvorbis). I'll look into it later.

Athlon XP 2500+ 1.83GHz

- ods15