From: walter_at_[hidden]
Date: 2001-03-13 15:15:58
Hello,
Dave Steffen wrote:
> I rather like the way Roldan Pozo's TNT library handles sparse
> matrices. There are a couple of different sparse vectors, which
have
> different performance characteristics. (IIRC one is based on
> std::vector, one on std::list, and one on std::map.) The sparse
> matrices are then templated on the type of sparse vector to use
> internally. Thus, you can choose the sparse storage scheme
depending
> on how you're going to use your sparse objects.
> This certainly doesn't allow for every possible way of handling
> sparse matrices, but it's not a bad start. ;-)
If I understand TNT correctly, it supports some sparse storage
schemes. But I think another goal to achieve is also some performance
enhancement using the right iterator concept. So if you require a
matrix library to deal with dense and sparse matrices supporting
expression templates too, things seem to get very complicated. I
don't know whether any existing library implements all these concepts
together.
>> Isn't BLAS, like, centuries old or something (spoken in valley-boy
>> ingenue voice)? See also http://math.nist.gov/spblas/
>
> Yep, but it works, and is standardized.
Sometimes I have the impression, that we have nothing comparable in
C++ yet.
(Or should we talk about std::valarray seriously?)
> Duplicating BLAS'
> functionality in C++ is what Roldan Pozo's TNT project was about
> (although progress on that project seems to have stopped).
>
> That sparse BLAS project is a C project that just screams C++ at
> me. The idea (IIRC) is that they support a jillion different sparse
> matrix storage schemes, which means that each computational routine
> has different versions for each scheme or combination of schemes.
> The number of actual routines (each with unique names) therefore
> grows combinatorically. They have some kind of fancy automatic code
> generator that (effectively) duplicates C++'s template facility, and
> actually pumps out all the different routines; I seem to recall that
> the whole shebang was at around 80K LOC and isn't done yet.
>
> Of course, we could just use templates and have the compiler do all
> that work. This seems pretty obvious to me; given that Roldan Pozo
> is invovled in the sparse BLAS project, and I know he knows C++, I'm
> at a loss to know why they didn't use C++ for the project to begin
with.
>
> OTOH, maybe that's a big, red, flashing neon sign telling Boost to
> get busy? ;-) ;-)
I have got another problem with sparse matrices also, which I would
like to mention here: IMHO dense matrices and classical linear
equation solvers together form a pair, for example. Sparse matrices
are normally used in the context of iterative solvers, which are
quite subtle to handle AFAIK.
Kind regards
J. Walter