> On Dec 11, 2016, at 6:26 PM, Massoud Rezavand <[email protected]> wrote:
>
> Sorry, I forgot to say that my computational domain is decomposed with a
> parallel library using MPI, and the particles are defined with a distributed
> vector. So, the entries of the matrix are basically from that distributed
> vector.
Then likely you should use the same distribution for matrix rows as you do
for the vector.
Thus you can call VecGetLocalSize() and pass that local size in the
preallocation for the matrix, you can also call VecGetOwnershipRange() to get
the range of local rows in order to compute the preallocation for the matrix.
Barry
>
> Thanks,
> Massoud
>
> On Mon, Dec 12, 2016 at 1:21 AM, Massoud Rezavand <[email protected]>
> wrote:
> Thanks,
>
> as I already discussed with you, the matrix is coming from SPH
> discretization, which is not fixed on a grid and is changing over time.
>
> On Mon, Dec 12, 2016 at 1:10 AM, Barry Smith <[email protected]> wrote:
>
> > On Dec 11, 2016, at 6:04 PM, Massoud Rezavand <[email protected]> wrote:
> >
> > Thank you very much,
> >
> > So, if I am using PetscSplitOwnership() and then MatGetOwnershipRange() to
> > be prepared for preallocation, then MatSetSizes(A, local_size, local_size,
> > N, N) should be called with the calculated local_size from
> > PetscSplitOwnership() ?
>
> Confusion from the two responses. You cannot use MatGetOwnershipRange()
> for preallocation.
>
> Without preallocation:
>
> > > PetscInt local_size = PETSC_DECIDE;
> > >
> > > MatSetSizes(A, local_size, local_size, N, N);
>
> MatGetOwnershipRanges(...)
>
> With preallocation:
> > >
> > >
> > > 2)
> > >
> > > PetscInt local_size = PETSC_DECIDE;
> > >
> > > PetscSplitOwnership(PETSC_COMM_WORLD, &local_size, &N);
> > >
> > > MPI_Scan(&local_size, &end_row, 1, MPIU_INT, MPI_SUM, PETSC_COMM_WORLD);
> > > begin_row = end_row - local_size;
>
> MatMPIAIJSetPreallocation(.....).
>
>
> But note that normally if the matrix comes from a discretization on a grid
> you would not use either approach above. The parallel layout of the grid
> would determine the local sizes and you won't not obtain them with
> PetscSplitOwnership() or local_size = PETSC_DECIDE;
>
> Where is your matrix coming from?
>
> Barry
>
>
>
> > >
> > >
>
>
> >
> > Thanks,
> > Massoud
> >
> >
> > On Mon, Dec 12, 2016 at 12:35 AM, Jed Brown <[email protected]> wrote:
> > Massoud Rezavand <[email protected]> writes:
> >
> > > Dear PETSc team,
> > >
> > > What is the difference between the following two methods to get the local
> > > dimensions of a square matrix A? If they do the same, which one is
> > > recommended? Should I use MPI_Scan after both?
> >
> > I would typically use 1 because it's fewer calls and automatically uses
> > the correct communicator. You can use MatGetOwnershipRange() instead of
> > manually using MPI_Scan.
> >
> > > 1)
> > >
> > > PetscInt local_size = PETSC_DECIDE;
> > >
> > > MatSetSizes(A, local_size, local_size, N, N);
> > >
> > >
> > > 2)
> > >
> > > PetscInt local_size = PETSC_DECIDE;
> > >
> > > PetscSplitOwnership(PETSC_COMM_WORLD, &local_size, &N);
> > >
> > > MPI_Scan(&local_size, &end_row, 1, MPIU_INT, MPI_SUM, PETSC_COMM_WORLD);
> > > begin_row = end_row - local_size;
> > >
> > >
> > > Thanks in advance,
> > > Massoud
> >
>
>
>
