Re: [Boost-users] uBLAS bindings

From: Kresimir Fresl (fresl_at_[hidden])
Date: 2004-10-30 05:08:51

Daniel L Elliott wrote:

[...]
> I am working on adding identity matrices (and perhaps some others) to the 
> gesv method of the LAPACK bindings. 

As far as I know, LAPACK does not have a notion of identity matrix.

> I have a couple questions about the code.

> In the template <typename MatrA, typename MatrB, typename IVec> 
> inline int gesv (MatrA& a, IVec& ipiv, MatrB& b) method the static 
> asserts are checking for the matrix types allowed within the gesv type?

Yes. These are types which LAPACK's _gesv routines expect.
Bindings library are (almost) nothing more than a syntactic sugar
for Fortran or C numerical libraries (BLAS, ATLAS, LAPACK, ...).

> Instead of using gesv to invert a very large matrix (which would 
> require us to create a very large identity matrix), would it be 
> advisable to simply triangularize matrix A (maybe using LU) and 
> use the uBLAS solver and an identity_matrix?

The answer is partly outside of the scope of bindings, but the
idea is interesting. gesv.hpp contains bindings for LAPACK's _getrf
routines which compute LU factorisations. Regarding the second step,
LAPACK also has _getri routines which compute the inverse of a matrix,
using the factorisation obtained by _getrf. Adding _getri to bindings
is on the TO DO list. (There are getri in ATLAS bindings, in
clapack.hpp.)

On the other hand, I think that the use of the uBLAS solver with
ublas::identity_matrix is outside of the scope of bindings library.

> Also, any overview of how the bindings and traits work would be very 
> helpful!

Docs are in the sandbox:

    Intro: http://tinyurl.com/6acht

    Traits: http://tinyurl.com/3t4kz

    LAPACK bindings: http://tinyurl.com/4pagj

    ATLAS bindings: http://tinyurl.com/4ao22

Hope this helps,

fres