[petsc-users] [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
Barry Smith
bsmith at mcs.anl.gov
Wed May 24 13:38:28 CDT 2017
> On May 24, 2017, at 4:45 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
>
> Coming from FEM, I believe the very confusing thing is that the local size of the user problem (math, physics point of view - DDM domain size) is not (can not be ?) the local size expected in MatCreateIS.
>
> My understanding is that the local size in MatIS is "just" related to backend implementation problems (it's logical that this local size is necessary, but, for another purpose: MPI machinery). Taking a few steps back, I can not see a case (I may be wrong) when a user does know how to compute or set "by hand" the local size that MatIS will expect: my understanding (once again, not sure) is that in most cases, the user will need local size to be PETSC_DECIDE in MatIS (because he doesn't want to "bother" with that or can not guess / compute it => unfortunatelly, as is, this jam the whole thing).
>
> I guess this kind of signature for MatIS would avoid/limit confusion in most cases and for most users :
> PetscErrorCode MatCreateIS(MPI_Comm comm,PetscInt bs,PetscInt M,PetscInt N,ISLocalToGlobalMapping rmap,ISLocalToGlobalMapping cmap,Mat *A,PetscInt m = PETSC_DECIDE,PetscInt n= PETSC_DECIDE)
> Or even
> PetscErrorCode MatCreateIS(MPI_Comm comm,PetscInt bs,PetscInt M,PetscInt N,ISLocalToGlobalMapping rmap,ISLocalToGlobalMapping cmap,Mat *A) // Always use PETSC_DECIDE backstage ?
>
You are correct that often m and n may be PETSC_DECIDE, however there are also valid reasons for them to be determined by the user and not just set automatically. With finite elements and PETSc one often partitions first the elements and then partitions the degrees of freedom on the elements subservient to the partitioning of the elements; by this I mean any degree of freedom that is on an element interior to a process in the element partitioning (degree of freedom in no way "shared" between processes) would be a assigned to that MPI process while "shared" elements are assigned by some rule to one of the processes that "share" the degree of freedom. In this case if the user computes the correct local m and n they will get exactly the partitioning of degrees of freedom they want (in the global vector) but if they let PETSc decide they won't get neccessarily the same partitioning.
The reason the m and n are the 3rd and 4th argument instead of the last arguments is to match the calls for, for example MatCreateAIJ() and MatCreateBAIJ() so that users understand the m and n have the same meaning as that case. Unfortunately this does not seem to have worked since the m and n arguments appear to have been confusing to you.
Barry
> Franck
>
> De: "Matthew Knepley" <knepley at gmail.com>
> À: "Franck Houssen" <franck.houssen at inria.fr>
> Cc: "Stefano Zampini" <stefano.zampini at gmail.com>, "PETSc" <petsc-users at mcs.anl.gov>, "PETSc" <petsc-dev at mcs.anl.gov>
> Envoyé: Mardi 23 Mai 2017 19:02:28
> Objet: Re: [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
>
> On Tue, May 23, 2017 at 11:51 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
> Not sure to know what question you're talking about ?!...
> I use MatIS to test some kind of domain decomposition methods. I define my own preconditioner for that: in the apply callback, I need to matmult my (matIS) matrix with the incoming vector.
>
> Okay. I will create an example using your suggestion.
>
> Thanks,
>
> Matt
>
> Franck
>
> De: "Matthew Knepley" <knepley at gmail.com>
> À: "Franck Houssen" <franck.houssen at inria.fr>
> Cc: "Stefano Zampini" <stefano.zampini at gmail.com>, "PETSc" <petsc-users at mcs.anl.gov>, "PETSc" <petsc-dev at mcs.anl.gov>
> Envoyé: Mardi 23 Mai 2017 18:46:34
> Objet: Re: [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
>
> On Tue, May 23, 2017 at 11:28 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
> OK, thanks. This is helpfull... But I really think the doc should be more verbose about that: this is really confusing and I didn't find any simple example to begin with which make all this even more confusing (personal opinion).
>
> Did you respond to my other question (how are you using them)? That would help me understand how to phrase it.
>
> Thanks,
>
> Matt
>
> Franck
>
>
> De: "Matthew Knepley" <knepley at gmail.com>
> À: "Franck Houssen" <franck.houssen at inria.fr>
> Cc: "Stefano Zampini" <stefano.zampini at gmail.com>, "PETSc" <petsc-users at mcs.anl.gov>, "PETSc" <petsc-dev at mcs.anl.gov>
> Envoyé: Mardi 23 Mai 2017 13:21:21
> Objet: Re: [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
>
> On Tue, May 23, 2017 at 4:53 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
> The first thing I did was to put 3, not 4 : I got an error thrown in MatCreateIS (see the git diff + stack below). As the error said I used globalSize = numberOfMPIProcessus * localSize : my understanding is that, when using MatIS, the global size needs to be the sum of all local sizes. Correct ?
>
> No. MatIS means that the matrix is not assembled. The easiest way (for me) to think of this is that processes do not have
> to hold full rows. One process can hold part of row i, and another processes can hold another part. However, there are still
> the same number of global rows.
>
> I have a 3x3 global matrix made of two overlapping 2x2 local matrix (= diagonal with 1.). Each local matrix correspond to one domain (each domain is delegated to one MPI proc, so, I have 2 MPI procs because I have 2 domains).
>
> So the global size is 3. The local size here is not the size of the local IS block, since that is a property only of MatIS. It is the
> size of the local piece of the vector you multiply. This allows PETSc to understand the parallel layout of the Vec, and how it
> matched the Mat.
>
> This is somewhat confusing because FEM people mean something different by "local" than we do here, and in fact we use this
> other definition of local when assembling operators.
>
> Matt
>
> This is the simplest possible example: I have two 2x2 (local) diag matrix that overlap so that the global matrix built from them is 1, 2, 1 on the diagonal (local contributions add up in the middle).
> I need to MatMult this global matrix with a global vector filled with 1.
>
> Franck
>
> Git diff :
>
> --- a/matISLocalMat.cpp
> +++ b/matISLocalMat.cpp
> @@ -16,7 +16,7 @@ int main(int argc,char **argv) {
> int size = 0; MPI_Comm_size(MPI_COMM_WORLD, &size); if (size != 2) return 1;
> int rank = 0; MPI_Comm_rank(MPI_COMM_WORLD, &rank);
>
> - PetscInt localSize = 2, globalSize = localSize*2 /*2 MPI*/;
> + PetscInt localSize = 2, globalSize = 3;
> PetscInt localIdx[2] = {0, 0};
> if (rank == 0) {localIdx[0] = 0; localIdx[1] = 1;}
> else {localIdx[0] = 1; localIdx[1] = 2;}
>
>
>
> Stack error:
>
> [0]PETSC ERROR: Nonconforming object sizes
> [0]PETSC ERROR: Sum of local lengths 4 does not equal global length 3, my local length 2
> [0]PETSC ERROR: [0] ISG2LMapApply line 17 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/vec/is/utils/isltog.c
> [0]PETSC ERROR: [0] MatSetValues_IS line 692 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/impls/is/matis.c
> [0]PETSC ERROR: [0] MatSetValues line 1157 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/interface/matrix.c
> [0]PETSC ERROR: [0] MatISSetPreallocation_IS line 95 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/impls/is/matis.c
> [0]PETSC ERROR: [0] MatISSetPreallocation line 80 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/impls/is/matis.c
> [0]PETSC ERROR: [0] PetscSplitOwnership line 80 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/sys/utils/psplit.c
> [0]PETSC ERROR: [0] PetscLayoutSetUp line 129 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/vec/is/utils/pmap.c
> [0]PETSC ERROR: [0] MatSetLocalToGlobalMapping_IS line 628 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/impls/is/matis.c
> [0]PETSC ERROR: [0] MatSetLocalToGlobalMapping line 1899 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/interface/matrix.c
> [0]PETSC ERROR: [0] MatCreateIS line 986 /home/fghoussen/Documents/INRIA/petsc-3.7.6/src/mat/impls/is/matis.c
>
>
>
> De: "Stefano Zampini" <stefano.zampini at gmail.com>
> À: "Matthew Knepley" <knepley at gmail.com>
> Cc: "Franck Houssen" <franck.houssen at inria.fr>, "PETSc" <petsc-users at mcs.anl.gov>, "PETSc" <petsc-dev at mcs.anl.gov>
> Envoyé: Dimanche 21 Mai 2017 23:02:37
> Objet: Re: [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
>
> Franck,
>
> PETSc takes care of doing the matrix-vector multiplication properly using MatIS. As Matt said, the layout of the vectors is the usual parallel layout.
> The local sizes of the MatIS matrix (i.e. the local size of the left and right vectors used in MatMult) are not the sizes of the local subdomain matrices in MatIS.
>
>
> On May 21, 2017, at 6:47 PM, Matthew Knepley <knepley at gmail.com> wrote:
>
> On Sun, May 21, 2017 at 11:26 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
> Using PETSc MatIS, how to matmult a global IS matrix and a global vector ? Example is attached : I don't get what I expect that is a vector such that proc0 = [1, 2] and proc1 = [2, 1]
>
> 1) I think the global size of your matrix is wrong. You seem to want 3, not 4
>
> 2) Global vectors have a non-overlapping row partition. You might be thinking of local vectors
>
> Thanks,
>
> Matt
>
>
>
>
> --
> 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
>
> http://www.caam.rice.edu/~mk51/
>
> De: "Stefano Zampini" <stefano.zampini at gmail.com>
> À: "Matthew Knepley" <knepley at gmail.com>
> Cc: "Franck Houssen" <franck.houssen at inria.fr>, "PETSc" <petsc-users at mcs.anl.gov>, "PETSc" <petsc-dev at mcs.anl.gov>
> Envoyé: Dimanche 21 Mai 2017 23:02:37
> Objet: Re: [petsc-dev] Using PETSc MatIS, how to matmult a global IS matrix and a global vector ?
>
> Franck,
>
> PETSc takes care of doing the matrix-vector multiplication properly using MatIS. As Matt said, the layout of the vectors is the usual parallel layout.
> The local sizes of the MatIS matrix (i.e. the local size of the left and right vectors used in MatMult) are not the sizes of the local subdomain matrices in MatIS.
>
>
> On May 21, 2017, at 6:47 PM, Matthew Knepley <knepley at gmail.com> wrote:
>
> On Sun, May 21, 2017 at 11:26 AM, Franck Houssen <franck.houssen at inria.fr> wrote:
> Using PETSc MatIS, how to matmult a global IS matrix and a global vector ? Example is attached : I don't get what I expect that is a vector such that proc0 = [1, 2] and proc1 = [2, 1]
>
> 1) I think the global size of your matrix is wrong. You seem to want 3, not 4
>
> 2) Global vectors have a non-overlapping row partition. You might be thinking of local vectors
>
> Thanks,
>
> Matt
>
> Franck
>
>
>
> --
> 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
>
> http://www.caam.rice.edu/~mk51/
>
>
>
>
>
> --
> 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
>
> http://www.caam.rice.edu/~mk51/
>
>
>
>
> --
> 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
>
> http://www.caam.rice.edu/~mk51/
>
>
>
>
> --
> 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
>
> http://www.caam.rice.edu/~mk51/
>
More information about the petsc-users
mailing list