?iga Lenar?i? wrote:
>
>
> I have a working fft.lisp now (which uses Ray's algorithm). I don't
> know how to prepare a patch and probably this is not the final version
> of it. It
A patch is created with diff. Something like "diff -u old-version
new-version > foo.patch"
>
> I've attached fft.lisp - I hope the attachment gets through.
> I get
> (%i110) b: table( random(1.0), [i,1024*64])$
> Evaluation took 0.3000 seconds (0.3010 elapsed) using 18.562 MB.
> (%i111) fft(b)$
> Evaluation took 0.7180 seconds (0.7210 elapsed) using 164.494 MB.
> (%i112) myfft(b)$
> Evaluation took 0.3860 seconds (0.3870 elapsed) using 29.254 MB.
>
> Maxima's $fft is now only 2x slower than bordeaux-fft in SBCL. I'm
> still all for using bordeaux-fft since it's free GPLd code (and has
> been even
That might be expected if you're using a very recent version of SBCL
(and 19f or later CMUCL). I know in cmucl 19f, complex arithmetic uses
sse2 by default, and that can be significantly faster because the
operations are vectorized. The real/imag array version will suffer
quite a bit since the complex operations probably won't be vectorized.
>
> I might have switched the 'forward'/'backward' fft routines by accident..
Ooh, that would be very bad if no one noticed!
I will try to take a look at this soon.
Ray