[petsc-users] MatPtAP

[Smith, Barry F.]

   How do you know it is exactly the MatPtAP() routine that is taking the large time? 

   Please send the output of a run with -log_view so we can see where the time is being spent.

   Barry


> On Jun 1, 2018, at 8:21 AM, Samuel Lanthaler <s.lanthaler at gmail.com> wrote:
> 
> Hi,
> 
> I was wondering what the most efficient way to use MatPtAP would be in the following situation: I am discretizing a PDE system. The discretization yields a matrix A that has a band structure (with k upper and lower bands, say). In order to implement the boundary conditions, I use a transformation matrix P which is essentially the unit matrix, except for the entries P_{ij} where i,j<k and n-i,n-j<k, so
> 
> P =  [ B, 0, 0, 0, ..., 0, 0 ]
>       [  0, 1, 0, 0, ..., 0, 0 ]
>       [                              ]
>       [                              ]
>       [                  ..., 1, 0 ]
>       [  0, 0, 0, 0, ..., 0, C ]
> 
> with B,C are (k-by-k) matrices.
> Right now, I'm simply constructing A, P and calling
> 
>    CALL MatPtAP(petsc_matA,petsc_matP,MAT_INITIAL_MATRIX,PETSC_DEFAULT_REAL,petsc_matPtAP,ierr)
> 
> where I haven't done anything to pestc_matPtAP, prior to this call. Is this the way to do it?
> 
> I'm asking because, currently, setting up the matrices A and P takes very little time, whereas the operation MatPtAP is taking quite long, which seems very odd... The matrices are of type MPIAIJ. In my problem, the total matrix dimension is around 10'000 and the matrix blocks (B,C) are of size ~100.
> 
> Thanks in advance for any ideas.
> 
> Cheers,
> Samuel