Subject: Re: [linux-audio-dev] Resampling
From: Jussi Laako (jlaako_AT_pp.htv.fi)
Date: Sat Aug 18 2001 - 01:06:55 EEST
Alexander Ehlert wrote:
>
> other than exponents of 2. The problem is, that the so called
> split radix routines can be really slow. So I'm not actually sure
Split radix isn't slow in itself, in fact it's usually faster that
radix-2/4/8 transforms on architectures with large multilevel caches.
FFTW falls back to ordinary [O(n^2)] FFT algorithm for block sizes not 2, 3,
5 or 7 's exponent.
Biggest problem with performace is Intel CPU's that have _very_
unpredictable performance due to badly behaving cache when it comes to FPU,
upto PIII at least. FFTW is able to compensate this behaviour somewhat when
used with wisdom or measure mode. AMD CPU's are wery smooth.
- Jussi Laako
-- PGP key fingerprint: 161D 6FED 6A92 39E2 EB5B 39DD A4DE 63EB C216 1E4B Available at PGP keyservers
This archive was generated by hypermail 2b28 : Sat Aug 18 2001 - 01:06:28 EEST