Re: [Lapackpp-devel] LaBandFactDouble Question

[Christophe P. <pi...@tl...>]

Dear Christian,

Thank you for the answer.
Christian Stimming wrote:

> LaBandMatDouble A(m,n);
> // fill A ..., then:
> LaBandFactDouble AF(A);
> LaBandMatFactorize(A, AF);
> // A and AF both refer to the factorization, which is fine:
> LaLinearSolve(AF, X, B); // ... and so on.

It is first what I tried, but it seems that the factorize matrix and the 
pivot are not correct ( I compare the results with lapack77 code that is 
proved to be correct). When I print the content of each matrix AF 
doesn't have the correct form, and the pivot vector is completely 
incoherent.
So I tried what you suggested last time, and redesigned a class based on 
the philosophy you used in gfqrc.h. The thing is  I can get the correct 
matrix this time (I mean the correct factorize matrix), but the pivot is 
obviously erroneous (out of bounds references to the matrix). I don't 
think that my design was correct, it was a quick prototype for my 
problem, but the integer pointer to the memory have a strange behavior .
I will keep looking at my implementation of this class, and if I managed 
to make it work, I will be glad to send it to you.
Best regards,
Christophe Picard

> This does not mean that the current lapackpp code is correct in any 
> case, but at least it should work for you. The point here is that the 
> original matrix A is no longer being used after constructing the 
> factorization.
>
> The only real solution would be a full redesign of the factorization 
> classes, somehow along the lines of the gfqrc.h. But for now, the 
> above code should work for you.
>
> Best regards,
>
> Christian Stimming