On Wed, May 04, 2005 at 07:20:20PM +, Aaron Denney wrote:
> On 2005-05-03, David Roundy <[EMAIL PROTECTED]> wrote:
> > An interesting challenge would be to rewrite fftw in haskell, and see how
> > close one could come to the performance of the original... :)
>
> What precisely do you mean by that? FFTW is C code generated by OCaml?
> Do you want to retarget ti so that Haskell is generated, or rewrite the
> generator in Haskell? (Or both?)
Both. I'd be curious to see how well one could do with pure haskell code.
It may be possible to do reasonably well, and would certainly be
instructive. FFTW is not simpley C code generated by OCaml, the C code
also does runtime timing and optimization, generating new algorithms (by
combining existing primitives) on the fly. This sort of
code-generation-on-the-fly (or rather code-recombination) is what
functional languages ought to be good at, if only there isn't too much
overhead.
I'd be interested if someone *else* did this. FFTW is an amazing piece of
work, and I'd not want to bother duplicating their work. Another (also
crazy) idea would be to implement a runtime-optimized version of ATLAS.
That would be somewhat more practical, since I find it annoying that ATLAS
is compile-time optimized, so you need to recompile it on each computer
you use (or if you buy new RAM...). Of course, doing it all in haskell
would still be silliness, but it'd be cool... :) And I think the
algorithms involved in block matrix multiplies are much simpler than those
involved in ffts.
I don't know how it would work out in the real world, but it seems like one
*ought* to be able to write truly beautiful auto-tuned code in haskell!
Perhaps even some sort of a "time it as you go" mechanism could be used (if
one had a clock with enough precision), which would eliminate the annoying
feature of FFTW that one needs to create and keep track of plans.
(Mostly I'm just daydreaming here... if only I were an undergrad again and
had the time and energy for excessing hacking with very little hope of any
practical benefit.)
--
David Roundy
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
