> From: <[EMAIL PROTECTED]>
> Has anyone done any speed comparisons between
> - the FHT based code lame has now (which is damn fine)
> - "fastest fourier transform in the west" fftw (www.fftw.org)
According to the benchmarks on fftw.org (unbiased, I'm sure) FFTW is about
twice as fast as the RMayer routine used in LAME, although that lookup table
stuff I added closes the gap a bit (it increases FHT speed by ~20%). Since
FFTing accounts for roughly 10% of the execution time in LAME, using FFTW
should be a significant win.
> - pentium optimized fft - djbfft
http://pobox.com/~djb/djbfft.html
> (gives out-of-order output,that i'm not too sure how to grok)
A header is supplied to aid in grokking. We could always reorder it (more
lookup tables :). One other thing: djbfft overwrites the input data. Does
this matter for LAME?
> I like the fht-based one that's there at the moment because its
> general, self-contained and works fine on well on most architectures
(afaik).
But FFTW does seem to be a fairly easy library to build/use. Even without a
working ./configure, I persuaded it to work under MSVC quite quickly. :)
> The FFTW have announced a pentium optimized FFTW library. Only
does
> complex input though.
And produces out-of-order output, I think.
-- Mat.
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
