[petsc-users] question about arbitrary eigenvector selection in SLEPC

[Matthew Knepley]

On Thu, May 29, 2014 at 10:03 AM, Giacomo Mulas <gmulas at oa-cagliari.inaf.it>
wrote:

> Hi Jose, and list.
>
> I am in the process of writing the function to use with
> EPSSetArbitrarySelection().
>
> Inside it, I will need to take some given component (which one is included
> in the info passed via ctx) of the eigenvector and square it. To do this,
> since the eigenvector is not necessarily local, I will need to first do a
> scatter to a local 1-component vector. So this would be like:
>
> ... some omitted machinery to cast the info from *ctx to more easily
> accessible form...
>

There might be an easier way to do this:
  PetscScalar val = 0.0, gval;

  VecGetOwnershipRange(xr, &low, &high);
  if ((myindex >= low) && (myindex < high)) {
    VecGetArray(localx1,&a);
    val = a[myindex-low];
    VecRestoreArray(localx1, &a);
  }
  MPI_Allreduce(&val, &gval, 1, MPIU_SCALAR, MPI_SUM, PETSC_COMM_WORLD);

Now everyone has the value at myindex.

   Matt


>   ierr = ISCreateStride(PETXC_COMM_WORLD,1,myindex,1,&is1_from);
> CHKERRQ(ierr);
>   ierr = ISCreateStride(PETSC_COMM_WORLD,1,0,1,&is1_to);CHKERRQ(ierr);
>   ierr = VecCreateSeq(PETSC_COMM_SELF, 1, &localx1);CHKERRQ(ierr);
>   ierr = VecScatterCreate(xr,is1_from,localx1,is1_to,&scatter1);
> CHKERRQ(ierr);
>   ierr = VecScatterBegin(scatter1,xr,localx1,INSERT_VALUES,
>                          SCATTER_FORWARD);
>   ierr = VecScatterEnd(scatter1,xr,localx1,INSERT_VALUES,
>                        SCATTER_FORWARD);
>   ierr = VecGetArray(localx1,&comp);
>   *rr = comp*comp;
>   ierr = VecRestoreArray(localx1, &comp);
>   ierr = VecDestroy(localx1);
>   ierr = VecScatterDestroy(&scatter1);
>   ierr = ISDestroy(&is1_from);
>   ierr = ISDestroy(&is1_to);
>   *ri = 0;
>
> ... some internal housekeeping omitted
>
>   return 0;
>
> The questions are:
>
> 1) when the arbitrary function is called, is it called on all nodes
> simultaneously, so that collective functions can be expected to work
> properly, being called on all involved nodes at the same time? Should all
> processes compute the *rr and *ri to be returned, and return the same
> value?
> would it be more efficient to create a unit vector uv containing only one
> nonzero component, and then use VecDot(xr, uv, &comp), instead of pulling
> the component I need and squaring it as I did above?
>
>
> 2) since the stride, the 1-component vector, the scatter are presumably
> the same through all calls within one EPSSolve, can I take them out of the
> arbitrary function, and make them available to it through *ctx? For this
> to work, the structure of xr, the eigenvector passed to the arbitrary
> function, must be known outside of EPSSolve.
>
>
> Thanks, bye
> Giacomo
>
> --
> _________________________________________________________________
>
> Giacomo Mulas <gmulas at oa-cagliari.inaf.it>
> _________________________________________________________________
>
> INAF - Osservatorio Astronomico di Cagliari
> via della scienza 5 - 09047 Selargius (CA)
>
> tel.   +39 070 71180244
> mob. : +39 329 6603810
> _________________________________________________________________
>
> "When the storms are raging around you, stay right where you are"
>                          (Freddy Mercury)
> _________________________________________________________________
>



-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20140529/2625a811/attachment.html>