Exact final residual: -nan/ info= -202

[Ham]

Hello all,
I am trying to get X for the AX=B type of problem. I have a large banded sparse matrix A of double precision numbers. I chose GMRES as a solver. The code works with a small matrix size, but when I compile my code with a size of 4624 * 4624, i get the error bellow ,i also included the code.

% iter || residual-nrm2 || runtime || SpMV-count || info
%=================================================================================%
1 nan 0.256593 31 -202
%=================================================================================%

%=================================================================================%
% PGMRES(30) solver summary:
% initial residual: 0.000000e+00
% preconditioner setup: 0.0877 sec
% iterations: 1
% SpMV-count: 31
% exact final residual: -nan
% runtime: 0.2566 sec
% preconditioner runtime: 0.2334 sec
%=================================================================================%


Any help will be greatly appreciated.

Thanks, Ham

[Ham]

Hartwig,

Thank you for your answer.


I just displayed the vector B and in fact, it is equal to 0, but I can not modify it because this data belongs to the model on which I was working. On the other hand, these data give good results with the CPU gmres. I will try without the preconditioner.


Thanks again,

Ham

[Ham]

Hello all,

I would have two questions:
1- How to deactivate the preconditioner ?? We just put "dopts.precond_par.solver = Magma_NONE" (my code is already posted at the top)?. On the other hand, I still have the same error :(
2- My rhs b = 0, what you propose me as solver, or any solution?

Ps. The code I use to integrate magma is in this code http://docs.opentelemac.org/solve_8f_source.html

Thanks.

[hartwig anzt]

Dear Ham,

you try to solve a system Ax = b with b=0. Without using any HPC software, I can tell you the solution is x=0.

For your other question: Yes, "dopts.precond_par.solver = Magma_NONE" turns off the preconditioner.


I hope this helps, please feel free to ask other questions. Thanks, Hartwig