On 15 Feb 2008, lemming@henning-thielemann.de wrote: On Thu, 14 Feb 2008, Jed Brown wrote:
Hopefully these are mature enough to be generally useful. I would appreciate any comments regarding the interface. The FFTW API is not particularly nice for building a pure interface on.
because FFTW stores global data (called 'plan's)
Yes. In the Haskell interface, the existence of these plans is hidden From the user. In the advanced interface, you can influence the quality of the plan which may effect how long a transform takes to compute, but it will not effect the result. If FFTW added the ability to query the plan cache, or if we kept a mirror Haskell-side, then we could make better use of memory which would improve performance. Fortunately, we have a really fast malloc.
Some of the transforms could be made slightly faster at the expense of a much nastier interface, but unless someone needs the extra few percent, or is pushing the memory limits on their machine, I'm not inclined to expose the underlying nastiness.
I want to mention once again that there is already a Fast Fourier transform for arbitrary length that is coded entirely in Haskell. http://hackage.haskell.org/packages/archive/dsp/0.2/doc/html/Numeric-Transfo...
Indeed, this package does exist, but it is lacking compared to FFTW with regard to multi-dimensional transforms, real to real transforms, and performance (particularly for large sizes). The transforms in the dsp package are only a few times slower than FFTW for small sizes, but it quickly becomes orders of magnitude slower, especially when the size is not a power of two. It also uses several times more memory, due in part to always using boxed arrays. Modifying dsp to allow use of unboxed arrays would be nice. Jed