dsimcha wrote:
== Quote from Don (nos...@nospam.com)'s article
dsimcha wrote:
I'm going to need an FFT library to perform some convolutions at some point
soon. Two absolute, non-negotiable requirements are that it be written in
pure D and that it be Boost or compatibly (i.e. zlib or public domain)
licensed. I also prefer "simple and good enough"
What does "simple" mean?
If you're happy with lengths being restricted to powers of 2, it's
simple. Most of the complexity of something like FFTW comes from support
for arbitrary lengths.
Yeh, I only need powers of two. I realize this isn't very hard because I wrote
a
prototype of it a while back. However, this prototype would basically need to
be
rewritten b/c:
1. It only supports pure real inputs, meaning you can't use it to compute
inverse
FFTs.
2. I tried to write it using strides instead of rearranging the elements of the
arrays, mostly because I was curious what effect this would have on performance.
It turned out to be disastrous, presumably because it killed cache efficiency.
I believe that a moderately efficient cache-aware FFT could be made
using simple blocking with the info from core.cpuid.
