[Trilinos-Users] [EXTERNAL] Trilinos-Users Digest, Vol 126, Issue 5

[Hoemmen, Mark]

> On Feb 11, 2016, at 8:30 AM, Cihan Altinay <c.altinay at uq.edu.au> wrote:
> 
> Hi Mark,
> 
> Thanks so much for the comprehensive overview of the state of block matrices! It's great news that there is active development on that front and I'll start using the github version to try things out.

Glad to help out!  Btw our fabulous postdoc Dr. Alicia Klinvex (who answered your questions before) is familiar with this capability and has contributed to it.

> I was quite surprised to read that Komplex is not required with the Tpetra/Belos stack - does that mean the solvers & preconditioners are specialized for complex problems, or are they using the approach taken in the Komplex module?

Most solvers and preconditioners that work with Tpetra objects are generic.  That is, they are written to work with either real or complex Scalar types.  As a result, they do not need to use the "equivalent real formulations" approach that the Komplex package uses.  The real and complex solvers and preconditioners use the same algorithm -- in fact, the same (templated) code, in most cases.

A few Tpetra-stack solvers and preconditioners only work with real Scalar types, but you should get informative error messages (either at compile time or at run time) if you try to use them with complex Scalar types.  For example, if Trilinos wraps a third-party sparse direct solver library, and you want to use that third-party library with Tpetra objects, then you may only use Scalar types that the library supports.

> Finally, your last comment suggests that there is currently no direct solver support for block matrices, correct?

That is correct.  We currently would only support this use case by conversion to an intermediate non-block sparse matrix.  For a complete factorization, you might not lose too much that way, since many of these factorizations already exploit dense substructure.

> Is this in the pipe at all? I'll file the request anyway...

We would still like to hear about your use case!  Incomplete factorizations and algebraic multigrid with block sparse matrices are higher priority, I think, but that may change depending on applications' needs.  The more detail you can give us in your feature request, the better.

Thanks!
mfh