Crash with explicit matrix multiplies in trilinos

[Joachim Haga]

Code such as

M = collapse(C*B)
prec = ML(M)

crashes with a memory access error.

It seems like EpetraExt.Multiply returns a malformed Epetra matrix when B and C are rectangular matrices, which then ML chokes on. This used to work with older Trilinos releases.

I will investigate further as time allows, but for now let's call it a known problem.

Example code: demo/navierstokes.py

[Joachim Haga (jobh)]

The problem was that multiplication of BC'd matrices leads to all-zero rows in the result. With newer Trilinos, these are not present in the sparsity pattern -- and ML doesn't like that at all. (A zero row with the diagonal present is handled fine though, it's just ignored.) I've worked around it by doing a preliminary matrix multiplication with dummy matrices, just to build a full sparsity pattern. Inefficient, but works.