Add TrilinosWrappers::SolverBelos #14352
Conversation
| for (unsigned int i = 0; i < n_cols; ++i) | ||
| (*this->vectors[i]) *= beta; | ||
|
|
||
| for (unsigned int i = 0; i < n_cols; ++i) | ||
| for (unsigned int j = 0; j < n_rows; ++j) | ||
| this->vectors[i]->add(alpha * B(j, i), *A.vectors[j]); |
There was a problem hiding this comment.
@kronbichler All the optimized vector operations are missing here. But I believe we do not have to implement those twice and can reuse them from #14349 once that one is generalized for block vectors.
There was a problem hiding this comment.
Correctly, I think we can simply create these general abstractions for blessed vectors (i.e., deal.II's own vectors Vector and LinearAlgebra::distributed::Vector) and the block vectors building on top of these vectors.
peterrum
changed the title
|
@kronbichler I have added some documentation and have worked on the solver control part. Belos only understands relative tolerance so that I need to translate the absolute one. |
|
|
||
| double relative_tolerance_to_be_achieved = | ||
| solver_control.tolerance() / norm_0; | ||
| const int max_steps = solver_control.max_steps(); |
There was a problem hiding this comment.
| const int max_steps = solver_control.max_steps(); | |
| const unsigned int max_steps = solver_control.max_steps(); |
There was a problem hiding this comment.
I have changed it. The reason for using int was that this is what Trilinos expects. I cast it now when giving the value to the parameter struct.
| Teuchos::RCP<Belos::SolverManager<value_type, MV, OP>> solver = | ||
| Teuchos::rcp( | ||
| new Belos::BlockGmresSolMgr<value_type, MV, OP>(problem, | ||
| belos_parameters_copy)); |
There was a problem hiding this comment.
Does this mean that right now, SolverBelos always uses GMRES? That seems unnecessarily restrictive.
There was a problem hiding this comment.
True. I have made gmres a parameter.
No description provided.