Conversation
tlroy
added
ready for review
tupek2
left a comment
tupek2
left a comment
There was a problem hiding this comment.
Do the new preconditioners need to set height/width in their SetOperator implementations?
| void BlockTriangularPreconditioner::SetOperator(const mfem::Operator& jacobian) | ||
| { | ||
| // Cast the supplied jacobian to a block operator object | ||
| block_jacobian_ = dynamic_cast<const mfem::BlockOperator*>(&jacobian); |
There was a problem hiding this comment.
what will happen if someone passes a non-BlockOperator to the triangular preconditioner? ideally it should fall back to just a normal preconditioner somehow.
There was a problem hiding this comment.
We can work on that in a follow up, if you agree with that suggestion. There may be other ways to allow this flexibility.
There was a problem hiding this comment.
Just tried passing a SparseMatrix in one of my tests, and I get a seg fault. The code does work when passing a single block BlockOperator, so if we wanted we could just add something in SetOperator to do something like BlockOperator block_jacobian(offsets); block_jacobian.SetBlock(0, 0, jacobian);
| mfem_solvers_[0]->SetOperator(*A11); | ||
|
|
||
| // Extract the diagonal of A11 (no inversion!) | ||
| mfem::HypreParVector* Md = new mfem::HypreParVector(MPI_COMM_WORLD, A11->GetGlobalNumRows(), A11->GetRowStarts()); |
There was a problem hiding this comment.
can we get the communicator here from some other source? COMM_WORLD may or may not be appropriate.
| delete A21DinvA12; | ||
|
|
||
| // Set the Schur complement preconditioner for block (1,1) | ||
| mfem_solvers_[1]->SetOperator(*S_approx_); |
There was a problem hiding this comment.
do we need to call delete or something on S_approx_ to avoid a memory leak?
| BlockOperator J(block_offsets); | ||
|
|
||
| J.SetBlock(0, 0, A_hypre); | ||
| J.SetBlock(1, 1, new HypreParMatrix(*A_hypre)); // Deep copy |
|
|
||
| auto pv_writer = smith::createParaviewWriter(*mesh, sols, physics_name); | ||
| pv_writer.write(0, 0.0, sols); | ||
|
|
There was a problem hiding this comment.
Is there an assert anywhere in this test? for example to check that the solver converged?
2 participants
Add block preconditioners to be used in
LinearDifferentiableBlockSolver. This includes block diagonal (additive fieldsplit), block triangular (multiplicative fieldsplit) for arbitrary N x N block systems, as well as Schur complement factorizations for 2x2 systems, with a standard Schur complement approximation (selfp).