[petsc-users] Preallocation: Allocate more than the local column

[Matthew Knepley]

On Mon, Apr 18, 2016 at 7:02 AM, Florian Lindner <mailinglists at xgm.de>
wrote:

> Hello,
>
> I create a MATAIJ on 4 processors using local sizes
>
> (0) Set matrix A to local size 2 x 5
> (1) Set matrix A to local size 2 x 2
> (2) Set matrix A to local size 2 x 2
> (3) Set matrix A to local size 2 x 2
>
> which results in a global size of 8 x 11.
>
> Afaik the local size given is not relevant for the parallel layout, so
> each processors holds 2 rows x 11 cols. Accordingly I can set these using
> MatSetValue and alike
>
> But it seems to be important when I do preallocation, i.e. work on the
> local submatrices:
>
> (0) Local Submatrix Rows = 0 / 2, Local Submatrix Cols = 0 / 5
> (1) Local Submatrix Rows = 2 / 4, Local Submatrix Cols = 5 / 7
> (2) Local Submatrix Rows = 4 / 6, Local Submatrix Cols = 7 / 9
> (3) Local Submatrix Rows = 6 / 8, Local Submatrix Cols = 9 / 11
>
> (retrieved with MatGetOwnershipRange / MatGetOwnershipRangeColumn)
>
> therefore, if I do a preallocation using MatMPIAIJSetPreallocation I get
> errors like
>
> [1]PETSC ERROR: MatSeqAIJSetPreallocation_SeqAIJ() line 3567 in
> /home/florian/software/petsc/src/mat/impls/aij/seq/aij.c
>     nnz cannot be greater than row length: local row 0 value 4 rowlength 2
>
> Easiest way to work around that would be to set the number of local cols
> (here always 5/2/2/2) to 11. Setting the number of global columns to 11
> would probably have the same effect.
>

So you want to preallocate a rectangular matrix?

The error above is correct because the diagonal blocks on proc 1 is 2x2.
You have to put the other nonzeros in the off-diagonal portion.

    Matt


> Problem is that each processor only knows the number of local columns, so
> I would need some communication for that.
>
> Another possibility that comes to my mind is to declare the Mat only after
> the preallocation has been determined and during preallocation the longest
> line has been computed.
>
> Are my assumptions correct so far?
> How would you cope with that?
>
> Best Regards,
> Florian
>
>
>


-- 
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/20160418/904cf3fc/attachment.html>