[petsc-users] matrix reordering

[Jed Brown]

On Fri, Oct 19, 2012 at 6:03 PM, Mohammad Mirzadeh <mirzadeh at gmail.com>wrote:

> Cool. Jed, could you please point me to an example/tutorial for this? I'm
> going to give it a shot.
>

This works for SeqAIJ:

  ierr = MatGetOrdering(A,ordering,&rowperm,&colperm);CHKERRQ(ierr);
  ierr = MatPermute(A,rowperm,colperm,&Aperm);CHKERRQ(ierr);
  ierr = VecPermute(b,colperm,PETSC_FALSE);CHKERRQ(ierr);

  ierr =
KSPSetOperators(ksp,Aperm,Aperm,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
  ierr = KSPSolve(ksp,b,x);CHKERRQ(ierr);
  ierr = VecPermute(x,rowperm,PETSC_TRUE);CHKERRQ(ierr);

Now x is the solution back in the original ordering.


There is an example in petsc-dev now.

https://bitbucket.org/petsc/petsc-dev/src/tip/src/ksp/ksp/examples/tutorials/ex18.c

It should work with petsc-3.3 in serial, but you need petsc-dev for
parallel MatPermute.



>
> Thanks
>
>
> On Fri, Oct 19, 2012 at 3:57 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
>
>> _You_ can compute a MatOrdering, MatPermute and VecPermute, solve, and
>> permute back. Profile and if the solve is faster, go ahead and lift the
>> ordering code back into your mesh.
>>
>>
>> On Fri, Oct 19, 2012 at 5:48 PM, Mohammad Mirzadeh <mirzadeh at gmail.com>wrote:
>>
>>> Thanks Barry. I think I agree with you that right way to do it is
>>> through mesh setup. As a matter of fact I have done that for some of my
>>> problems. Yet, I can think of situations where that kind of support might
>>> be beneficial, especially if someone (a.k.a me in this case!) is looking
>>> for a 'dirty and fast' fix.
>>>
>>> Anyway, I still believe that in long term, one should do the
>>> partitioning in a pre-processing step anyways and so there may not be much
>>> incentive for adding such a support.
>>>
>>> Thanks again,
>>> Mohammad
>>>
>>>
>>> On Fri, Oct 19, 2012 at 3:00 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>>
>>>>
>>>>    Mohammad,
>>>>
>>>>      We've thought about adding this kind of support this over the
>>>> years but always came to the conclusion that "the right way to do it" is
>>>> during the mesh setup step.
>>>>
>>>>    Barry
>>>>
>>>>
>>>> On Oct 19, 2012, at 4:16 PM, Mohammad Mirzadeh <mirzadeh at gmail.com>
>>>> wrote:
>>>>
>>>> > Hi guys,
>>>> >
>>>> > Quick question. *After* the matrix is setup, is it possible to ask
>>>> PETSc to renumber the nodes internally in the KSPSolve phase to minimize
>>>> communications for MatMul inside the solver? Can I use the MatPartitioning
>>>> object for this or is that only intended to partition the mesh as a
>>>> pre-processing step?
>>>> >
>>>> > Thanks
>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20121020/872c9b46/attachment.html>