On Mar 1, 2018, at 10:56 AM, Edoardo alinovi <[email protected]> wrote:

Dear Barry,

thank you very much for the help. I will try to define the matrix as global and let you know if I have some problems. So just to be sure the following procedure is allowed:

- assemble matrix A with matAssemblyBegin/MatAssemblyEnd and rhs1 - solve the system for u= A\rhs1

- assemble rhs2 and do MatDiagonalSet() to obtain B by changing the main diagonal of A without redo matAssemblyBegin/MatAssemblyEnd - solve the system for v=B\rhs2

- assemble rhs3 and do like in the previous point to get C - solve the system for w=C\rhs3

Yes

Thanks,

Edoardo

2018-03-01 17:27 GMT+01:00 Smith, Barry F. <[email protected]>:

...

On Mar 1, 2018, at 10:24 AM, Edoardo alinovi <[email protected]> wrote:

Dear All,

thanks to your suggestions and Petsc, I have finished to program my finite volume code for CFD that runs in parallel and this is great.

I am now tryig to improve its performace and I have a question on matrix in MPIAij format.

Basically after the discetization, I have one matrix for each velocity component and they are equal except for the diagonal elements which changes because of the boundary conditions. What I am doing now is:

- fill each matrix using matSetValues row by row - finalize each matrix assembly with MatAssemblyBegin and MatAssemblyEnd

I am wondering if would be possible to build just one matrix doing MatAssemblyBegin/MatAssemblyEnd and for the other two change the diagonal values locally, thus avoiding the parallel communications.

You can use MatDiagonalSet() to add to or set the diagonal values of the matrix.

...

Also can I put a matrix object in a module and make it a global variable?

You can.

...

Can you provide/link me an example? I am using Fortran.

There may be examples in the master branch of PETSc. It is straightforward. Just do it and if it doesn't work send us email with full bug report.

...

Thank you very much for your help,

Edoardo