dolfin team mailing list archive

[Joachim Berdal Haga <jobh@xxxxxxxxx>]

On 24 January 2011 18:33, Garth N. Wells <gnw20@xxxxxxxxx> wrote:

> On 24/01/11 17:17, Johan Hake wrote:
> > Isn't there a way to initialize a paralell vector with arbitrary
> distribution
> > pattern?
> >
> > There is a way to general a local range from a given entity size using
> > MPI.local_range. One should be able to generate a distributed Vector with
> this
> > information.
> >
>
> This isn't guaranteed to work since using MPI.local_range may produce a
> vector with a layout that is inconsistent with the matrix layout.
>
> A Vector could be created using the local range data from the Matrix.


It seems that this is only implemented for rows, not for columns. Any idea
how best to create a distributed vector suitable for the result of a
transpmult? Or is transpmult unsupported in parallel (it's probably not very
efficient)?

(Taking PETSc for example, it looks like using either
sparsity_pattern.local_range(1), MPI::local_range(N) or
PETSc::MatGetLocalSize.n for the local size might work. Or creating the
vector with specified global size and PETSC_DECIDE for the local size,
although this is a bit of a wart compared to the current *Vector interface.
But which to choose... and which are consistent with the matrix layout?)

-j.