Clarification of resumeFill requirements for Tpetra::CrsMatrix

[keitat]

Tpetra::CrsMatrix requires resumeFill() when changing its sparsity pattern. However, I have observed that there are several set...() methods requiring resumeFill() even though no changes happen in the sparsity pattern. Could you clarify the following?

• setAllToScalar() (I got runtime error when resumeFill() is not called.)
• add( A ) (If the input matrix A had different sparsity pattern, do I have to call resumeFill() before the call? )
• Also, Tpetra does not support operations like A+ cI (c is a scalar and I is identity). What do you recommend for implementing this?

[mhoemmen]

You must call resumeFill whenever you change the CrsMatrix's values.

@keitat wrote:

Tpetra::CrsMatrix requires resumeFill() when changing its sparsity pattern.

Generally you're not allowed to change a CrsMatrix's sparsity pattern after first fillComplete. If fillComplete has never been called on the CrsMatrix, and if the CrsMatrix was not created with a const CrsGraph, then you may change the CrsMatrix's structure until first fillComplete.

[mhoemmen]

@keitat wrote:

Also, Tpetra does not support operations like A+ cI (c is a scalar and I is identity). What do you recommend for implementing this?

CrsMatrix has a method for getting the diagonal offsets. You could start with that.

MueLu has a specialized "Jacobi" kernel that does something like (A + omega D) * B, where omega is a constant and D is a diagonal matrix.

[mhoemmen]

I'm closing this issue for now. Please feel free to reopen or open a new issue if needed. Thanks!