[petsc-users] MatGetSubmatrices from block matrix MPIBAIJ

[shu guo]

Hi Jed,

Thank you for the paper. It is very interesting and helpful.

Shu

On Tue, Oct 23, 2012 at 11:26 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> On Tue, Oct 23, 2012 at 10:19 AM, shu guo <greatgs2008 at gmail.com> wrote:
>>
>> I am solving a electromagnetic problem. Each node has 4 degree of
>> freedom, phi A1 A2 A3.
>>
>> The row in this blocked matrix is generated as
>>
>>                node1         |      node 2                 |         node
>> 3
>>                                  |                                |
>> phi1 A1_1 A1_2 A1_3 | phi2 A2_1 A2_2 A2_3| phi3 A3_1 A3_2 A3_3
>>
>> And I permute it as
>>
>> phi matrix      ||                A matrix
>> phi1 phi2 phi3 || A1_1 A1_2 A1_3 A2_1 A2_2 A2_3 A3_1 A3_2 A3_3
>>
>> I permuted it just want to sort each variables together.
>
>
> Why? This just destroys the locality that makes BAIJ efficient. See the
> effect of "field interlacing" here (still true today).
>
> http://www.mcs.anl.gov/~kaushik/Papers/parco01.pdf
>
>>
>>
>> > Just build an index set that has every part of the block. If you need
>> > arbitrary scalar subsets, use MPIAIJ (can use MatConvert).
>>
>> If I want the submatrix from original matrix, can you be more specific
>> that how the index set looks like? And is it possible for me to merely
>> get the submatrix of the ist DOF of each element? Thank you.
>
>
> Create an index set with [0,4,8,...], but use the AIJ format.
>
> You can request support for BAIJ here, but I'm afraid it's not going to be
> considered high priority because the potential benefits of BAIJ over AIJ are
> quite small if you are taking submatrices.
> https://bitbucket.org/petsc/petsc-dev/issues