[petsc-users] getting the near nullspace from PCSetCoordinates

[Mark Adams]

On Mon, Jan 2, 2017 at 8:29 PM, Jed Brown <jed at jedbrown.org> wrote:

> Barry Smith <bsmith at mcs.anl.gov> writes:
>
> >     How about MatSetCoordindates(Mat, Vec).
>
> This would assume an interpolatory basis for which each Mat (row?)bs
> corresponds to one Vec bs.  Should we consider mixed spaces?
>

I don't see how you do that without getting into discretizations, which is
not what we are addressing here and opens a big can of worms. Do we have
any users that write their own null space vectors because they are using a
non-interpolatory basis? (whatever the heck that is) The setCoordinates
thing was put in because this is convenient for a lot of users. And the new
matnullspace interface is great but it is just a minor tweak to that
interface IMO. Asking users to move over the the MatNullSpace interface is
only a few lines, which they can get from ksp/ex56, so not asking much if
we are happy with this interface (and it has been stable for years now).


> >     Then MatNullSpaceCreateRigidBody(Mat, MatNullSpace *);
>
> Perhaps something like this as a convenience, but I think it can be
> useful to call the current function without first creating a Mat to
> attach the coordinate Vec to.
>

Yea, maybe that makes more sense because it is a space and Vec is the
space(s).


>
> >     Then presumable GAMG can pass the appropriated coordinates down to
> the smaller matrices it creates internally and create the rigid body null
> spaces it wants as it moves to the smaller matrices?
> >
> >    Barry
> >
> >
> > You could have a MatGetCoordindates(Mat, Vec) and not change the calling
> sequence of MatNullSpaceCreateRigidBody() but I like the first alternative
> I suggested.
> >
> >
> >> On Jan 2, 2017, at 5:23 PM, Jed Brown <jed at jedbrown.org> wrote:
> >>
> >> Mark Adams <mfadams at lbl.gov> writes:
> >>
> >>> On Mon, Jan 2, 2017 at 11:47 AM, Jed Brown <jed at jedbrown.org> wrote:
> >>>
> >>>> Jeremy Theler <jeremy at seamplex.com> writes:
> >>>>
> >>>>> Hi all
> >>>>>
> >>>>> I want to check that the near nullspace I provide to GAMG gives
> "almost
> >>>>> null vectors" when multiplying each vector in the near nullspace
> against
> >>>>> the matrix problem.
> >>>>>
> >>>>> This way I can check that the unknown ordering I am using is
> consistent,
> >>>>> for example using by MatNullSpaceCreateRigidBody() or by computing
> the
> >>>>> nullspace by myself.
> >>>>
> >>>> Please use that and MatSetNearNullSpace().  It composes properly and
> you
> >>>> can check everything.
> >>>>
> >>>> PCSetCoordinates() happens to do double-duty for aggregation-based
> >>>> methods, but outside of semi-geometric methods, it is just ugly code
> >>>> duplication and makes assumptions that may be inappropriate (like
> >>>> elasticity with an interpolatory basis).
> >>>
> >>>
> >>> Yes, PCSetCoordinates is an old interface that is essentially
> deprecated.
> >>> Maybe we should officially deprecated this.
> >>
> >> I think we should officially deprecate it, but perhaps make something
> >> more general available as a Mat function (since some algorithms may use
> >> coordinates directly).  (Needing to dig up a PC to provide problem (as
> >> opposed to configuration) information is bad style.)
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20170103/f7c83df7/attachment.html>