[petsc-dev] MatISGetLocalMat: need to MatDestroy the returned local matrice ?

[Franck Houssen]

Hi Stefano,

I'am trying to use MatIS for testing (both results and performance with PETSc to be compared with other implementations) a domain decomposition method.

For now, I do:
MatCreateIS(...)
MatISSetPreallocation(...)
MatSetValues(...)
MatAssemblyBegin/End
And I tried to get domains back with MatISGetLocalMat.

This is not the order you mention !... :D

Franck

----- Mail original -----
> De: "Stefano Zampini" <stefano.zampini at gmail.com>
> À: "Franck Houssen" <franck.houssen at inria.fr>
> Cc: "Jed Brown" <jed at jedbrown.org>, petsc-dev at mcs.anl.gov
> Envoyé: Vendredi 19 Mai 2017 17:02:12
> Objet: Re: [petsc-dev] MatISGetLocalMat: need to MatDestroy the returned local	matrice ?
> 
> Franck,
> 
> if you get the local mat, you are free to do whatever you want with it. One
> possible case is that you create the MATIS object (i.e., basically a l2g map
> of dofs + local mat creation) and then do
> 
> MatISGetLocalMat(matis,&A);
> MatXAIJSetPreallocation(A,…..)
> MatSetValues(A,…)
> 
> What do you need it for?
> 
> Stefano
> 
> > On May 19, 2017, at 4:57 PM, Franck Houssen <franck.houssen at inria.fr>
> > wrote:
> > 
> > Thanks.
> > 
> > My understanding (?) is that modifying the local mat (with MatMult for
> > instance) changes also the global mat (as the local mat is a part of the
> > global one). Jed, I let you decide if this behavior (which is logical but
> > not always the desired one) is worth to be emphasized in the doc...
> > 
> > Franck
> > 
> > ----- Mail original -----
> >> De: "Jed Brown" <jed at jedbrown.org>
> >> À: "Franck Houssen" <franck.houssen at inria.fr>, petsc-dev at mcs.anl.gov
> >> Envoyé: Vendredi 19 Mai 2017 16:20:14
> >> Objet: Re: [petsc-dev] MatISGetLocalMat: need to MatDestroy the returned
> >> local	matrice ?
> >> 
> >> Thanks.  I just pushed this.
> >> 
> >> commit 96a6f12955a790d77f6eb7eb6b1dcf9dd5e1cf30
> >> Author: Jed Brown <jed at jedbrown.org>
> >> Date:   Fri May 19 08:17:45 2017 -0600
> >> 
> >>    MatISGetLocalMat: add note on reference counting
> >> 
> >>    Reported-by: Franck Houssen <franck.houssen at inria.fr>
> >> 
> >> diff --git a/src/mat/impls/is/matis.c b/src/mat/impls/is/matis.c
> >> index dc4a894aae..287c0aa554 100644
> >> --- a/src/mat/impls/is/matis.c
> >> +++ b/src/mat/impls/is/matis.c
> >> @@ -837,6 +837,9 @@ PetscErrorCode MatISGetLocalMat_IS(Mat mat,Mat *local)
> >>   matrix and want to provide it to the inner matrix object to improve the
> >>   performance
> >>   of the MatSetValues() operation.
> >> 
> >> +  This function does not increase the reference count for the local Mat.
> >> Do
> >> not destroy it and do not attempt to use
> >> +  your reference after destroying the parent mat.
> >> +
> >> .seealso: MATIS
> >> @*/
> >> PetscErrorCode MatISGetLocalMat(Mat mat,Mat *local)
> >> 
> >> Franck Houssen <franck.houssen at inria.fr> writes:
> >> 
> >>> Hello,
> >>> 
> >>> Do I need to MatDestroy the local matrix returned by MatISGetLocalMat ?
> >>> (the global matrix has been created with MatCreateIS)
> >>> 
> >>> Tried to look in the source tree... But I'am not sure to get the answer:
> >>> in
> >>> petsc-3.7.6/src/mat/impls/is/matis.c, MatISGetLocalMat seems to call
> >>> MatISGetLocalMat_C, but, I can't find MatISGetLocalMat_C in the source
> >>> tree !?...
> >>> 
> >>> Franck
> >> 
> 
>