Re: "Discrete Hartley transform" vs " Discrete Fouuier transform"

On Oct 21, 8:46=A0am, "VelociChicken" <b...@[EMAIL PROTECTED]
> wrote:
> These graphs appear to show FHT in good favour:
>
> http://www.geocities.com/ResearchTriangle/8869/fft_summary.html

The lowest time on Mayer's graph is for FFTW, which does not use an
FHT.  (He doesn't say what version of FFTW he's benchmarking there,
but I'm guessing it's FFTW 2.x from the late 1990s.  See also graphs
from more recent machines and a wider variety of software on fftw.org/
speed)

Mayer's code is faster than the other code he benchmarked, but
arguably that is not because he uses an FHT, it is because he uses a
larger radix than all of the programs except FFTW; the other old
radix-2 and split-radix routines are no longer competitive on modern
general-purpose CPU because of poor cache and register/pipeline
utilization.

As I said before, Mayer's code is very good, and will probably beat a
lot of the real-input FFT code floating around on the web.  But
there's no evidence that this comes from any intrinsic property of
FHTs rather than from Mayer's skill as a programmer; certainly,
Mayer's code can be beaten by the fastest real-input FFT
implementations.  But if you want a nice, compact, clear program with
reasonable performance, I wouldn't hesitate to recommend Mayer's
routines.

Steven