Additional operations for MATSBAIJ

[Hong Zhang]

Dave,

>    I would like to inquire whether there are plans to add some MatMatXXX
> support (in particular MatPtAP) for the MATSBAIJ types in the next version
> of petsc?

No, we do not have plan. It is more than few hours work to write
MatPtAP_SBAIJ(), parallel version requires much more time.

>
> The code used in our group combines the FEM with a PIC scheme. The addition
> of the particles increase the memory usage significantly, thus we could
> really benefit from the memory saved by only storing the upper triangular
> part of our global stiffness matrix. The FEM part of the code utilizes
> multigrid with Galerkin coarse grid operators constructed via MatPtAP.

Although SBAIJ matrix saves half memory of SPARSE matrices,
you pay price for data accessing and complecated data communication.
You may use petsc aij matrix with all the functional support, but
call symmetric solvers/preconditioners, e.g., cholesky or icc
which save half of memory for factored matrices (would be significant
in many cases).

Hong