Oh, I agree that it would be really nice to have a way to write high-performance code in pure Haskell --- it would be nice if I didn't have to drop to Fortran anymore just because it makes it easier to write high-performance numeric code! My only point was that it might not be worthwhile to spend too much time writing routines in Haskell from scratch to implement things like singular value decompositions when there are routines to do this that already exists in LAPACK and are probably superior to something to anything we could roll out ourselves, though on the other hand the act of implementing an SVD algorithm might be useful for giving guidance on the kind of high-level operations that are needed in the library. Cheers, Greg On May 17, 2010, at 8:01 PM, Roman Leshchinskiy wrote:
On 17/05/2010, at 05:17, Gregory Crosswhite wrote:
As an aside, while there are advantages to writing numerical analysis routines in Haskell, it might be better strategy to instead link in something like LAPACK and provide nice wrappers to it in Haskell, since this way you can harness the work of the experts who have spent a lot of time perfecting their code rather than re-inventing the wheel.
I don't see think this is an either/or question. A good array library ought to provide BLAS, Lapack, FFTW etc. bindings *and* allow writing high-performance code in pure Haskell. I haven't implemented any of these bindings for vector only because I'm still deciding what to do with multidimensional arrays.
Roman