I forgot one thing in my earlier email: What about making
MAT_NEW_NONZERO_LOCATION_ERR the default for DM matrices?
Barry
On May 1, 2011, at 6:45 AM, Jed Brown wrote:
> On Sun, May 1, 2011 at 05:07, Barry Smith <bsmith at mcs.anl.gov> wrote:
>
> Jed,
>
> You pose several issues here.
>
> 1) By default SBAIJ matrices generate an error if you set a value in the
> lower triangular portion.
>
> I think this is the wrong default. I've changed it to ignore lower
> triangular values unless you explicitly set the option.
>
> Excellent
>
>
> 2) MatSetType(A,MATSBAIJ); and friends deletes the matrix and creates a new
> one if the matrix is MATSEQSBAIJ or MATMPISBAIJ.
>
> This is silly. Given our model of setting types however there is no
> trivial way to fix this. We could try to handle setting a heirarchy of types
> properly (how to know if something is a subclass? etc?) or do something just
> for Mat to handle the
> rank = 1 and rank > 1 cases. I'm inclined to add a special Mat register
> for example MatRegisterBaseName("SBAIJ","SeqSBAIJ","MPISBAIJ"); and have
> MatSetType() first check if the type is in the BaseName list and if so grab
> the correct
> classname based on rank, then it will realize the matrix is not having a
> type change and will not destroy it. Note that this means routines like
> MatCreateSBAIJ() will disappear.
>
> I think you mean MatCreate_SBAIJ(). I also think this is better because there
> is a bit of duplicate code and you can never actually have an instance of
> MATSBAIJ.
>
> The lookup in the other direction would take more effort, but there is a lot
> of PetscTypeCompare that ends up needing to always check for both the SEQ and
> MPI variants. But I think managing a full type hierarchy is a very deep
> rabbit hole so I'm not convinced it's worth it.
>
>
> 3) The general wordyness of setting up the matrix fill from the DM.
>
> My planned model for the future was to change DMGetMatrix() to
> DMFillMatrix() and have code something like
>
> MatCreate(comm,&Ap);
> MatSetOptionsPrefix(Ap,"Ap_");
> MatSetFromOptions(Ap);
> MatSetOption(Ap,MAT_SYMMETRIC,PETSC_TRUE);
> DMFillMatrix(fsu,Ap);
>
> If is actually not any shorter but I think it is a better model.
>
> This is much less wordy because all the code is in one place (no separate
> part for PetscOptionsList) and there are no if statements. It would also be
> possible to write a helper like
>
> DMCreateMatWithOptions(fsu,"Ap_",MATSBAIJ,MAT_SPD,MAT_NEW_NONZERO_LOCATION_ERR,...,0,&Ap);
>
> (or without variadic handling of options).
>
