From: Gunter Winkler (guwi17_at_[hidden])
Date: 2006-11-09 07:47:24
On Thursday 09 November 2006 12:31, Maik Beckmann wrote:
> It seems that many people thought about and worked on good performing
> small fixed matrix implementations.
> i.e.: http://thread.gmane.org/gmane.comp.lib.boost.ublas/2837/focus=2843
Thanks for the pointer.
> Since it wasn't mention AFAIK on this list yet, I want to point out that
> lately a IMHO good looking approach has been posted to boost-devel:
I didn't check the code, but the replies state the problem: There are 3
proposals that slightly differ.
For my own programs I'd like to have a common interface for small, medium and
large matrices and vectors. Some parts mainly operate on objects of size 2 to
10, which are assembled into a large sparse matrix. Thus I frequently have to
cast from double arrays to usable matrix/vector types.
This leads to these conclusions:
1. A library for (small) fixed size vectors and matrices is useful (and should
be included into boost)
2. All boost libraries for linear algebra should have a common interface _and_
should contain types that are convertible or at least assignable to each
other. (This is IMHO the goal of GLAS)
3. Actually there should be 3 libraries
a) small (dense) objects -> tvmet, 2D-matrix, ...
b) large dense objects -> dense part ublas, blas bindings
c) (large) sparse objects -> sparse part of ublas, umfpack bindings