Re: [Lapackpp-devel] LaLinearSolve overloading Question

[Christian S. <sti...@tu...>]

Dear Terrance, 

(correspondence about the lapackpp package please always over the mailing list 
[archives, ideas from others], never individually.)

Am Samstag, 18. Dezember 2004 21:47 schrieb Terrrance O'Regan:
> I logged in annonomously to CVS but I mot sure what to sign out or how to
> install. If you could explain that would be great, but if you're busy I
> will wait until the new release.

As http://sourceforge.net/cvs/?group_id=99696 says: You issue these two 
commands:

 cvs -d:pserver:ano...@cv...:/cvsroot/lapackpp login 

which is the "login", and the password should be left empty, and then

 cvs -z3 -d:pserver:ano...@cv...:/cvsroot/lapackpp co 
lapackpp

which will give you a local copy of the current CVS status. In the file 
"ChangeLog", you will see the latest change to that CVS source code. If you 
already see a change by me yesterday, then you got these fixes. To compile 
from CVS, see the README.

> I am mainly using Lapackpp in the hopes that it will give me a super fast
> matrix inverter as I am going to be inverting a huge matrix thousands of
> times (and eventually use the complex functions).
>
> As of now my simply SOR algorithm is faster than LaLinearSolve for my very
> large sparse matrix.
>
> So I will test the non general matrix functions. Also, is Lapackpp the
> right thing for me, or should I consider using Lapack in the old fortran
> code (I believe I can use Lapack via C++)?

The objective of lapackpp is to be a C++ wrapper to the original Fortran Blas/
Lapack functions. So you need lapackpp if your application is in C++ and you 
want a C++ class interface to Blas/Lapack. If you don't need this, i.e. you 
can access Fortran directly, then by all means rather do this. I guess you 
will have much more flexibility and you have one source of errors less. Also, 
I think the Fortran interface of blas/lapack has extra data types for sparse 
matrices, but the Lapack++ does not have this. In that case you would surely 
be better off by directly using the fortran type for this.

As for inversion of sparse matrices: I don't know the "SOR algorithm". If you 
have an algorithm that performs faster than LaLinearSolve then simply use it. 
I tried to program a simple Gauss solver, and LaLinearSolve performed faster 
by a factor of 100, so I gave up on my own code and started Lapackpp. If in 
your application you can come up with faster algorithms, then lucky you. :-))

Christian