[petsc-users] Wrapper function to factorize a matrix

[Smith, Barry F.]

> On Oct 21, 2019, at 8:55 PM, Matthew Knepley via petsc-users <petsc-users at mcs.anl.gov> wrote:
> 
> On Mon, Oct 21, 2019 at 9:47 PM Zhang, Hong <hzhang at mcs.anl.gov> wrote:
> Matt:
> On Mon, Oct 21, 2019 at 9:30 PM Zhang, Hong via petsc-users <petsc-users at mcs.anl.gov> wrote:
> Shash,
> You may do it without using pc. See petsc/src/mat/examples/tests/ex125.c
> 
> Does this example let you invert the matrix and just apply to get the solve?
> No. It gets LU factors stored in F, then call MatSolve(F, b, x)  multiple times to get solution x for given b.
> 
> Do we have anything in PETSc now that gets the inverse? Storing and applying the inverse can be much faster
> than using the factors (it's done in supernodal methods, and it our patch smoothers). If not, we should talk about
> adding it.

   This is done in pbjaocobi. 

   There was some discussion in the most recent patch MR to try to provide a common API for "small" block factorization/solve that supported batching and other modern things. This would be a distinct construct from KSP and PC. I don't there is even a name yet. As it is done it should also be done for GPUs.

  Barry

> 
>   Thanks,
> 
>      Matt
>  
> Hong
> 
> On Mon, Oct 21, 2019 at 4:43 PM Shashwat Sharma via petsc-users <petsc-users at mcs.anl.gov> wrote:
> Hello,
> 
> For some small matrices, I'd like to use Petsc to perform direct LU factorization on a sequential dense or AIJ matrix, and then use the factored matrix later on via MatSolve. This occurs multiple times in my code, and each factored matrix is in turn used multiple times.
> 
> I tried to wrap this factorization process in a function, which should return the factored matrix, as follows:
> 
> void MatFactorize_Petsc(Mat &mat, Mat &mat_factored)
> {
> PC pc;
> PCCreate(MPI_COMM_SELF, &pc);
> PCSetOperators(pc, mat, mat);
> PCSetType(pc, PCLU);
> 
> PCFactorSetMatSolverType(pc, MATSOLVERPETSC); // Or SuperLU
> PCFactorSetUpMatSolverType(pc);
> 
> PCFactorGetMatrix(pc, &mat_factored);
> PCSetUp(pc);
> // PCDestroy(&pc);
> return;
> }
> 
> The command PCDestroy causes a segmentation fault, which I think happens because retrieving the factored matrix does not increase the reference count. Looking at the Petsc source, it basically returns a pointer to pc->data. So if I want to use mat_factored outside the function, I cannot destroy the PC object, which leads to memory leaks (as per Valgrind) even if I later call MatDestroy(&mat_factored).
> 
> I tried using a temp matrix to get the factors, and then doing MatDuplicate(temp, MAT_COPY_VALUES, &mat_factored), but MatDuplicate is not allowed on factored matrices; same for MatConvert.
> 
> Is there a way I can achieve the desired behaviour, where mat_factored is not "linked" to the pc or ksp object, keeping in mind that I'd like to be able to choose between SuperLU, Petsc and SuperLU_dist solvers?
> 
> Thanks,
> Shash
> 
> 
> 
> -- 
> 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
> 
> https://www.cse.buffalo.edu/~knepley/
> 
> 
> -- 
> 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
> 
> https://www.cse.buffalo.edu/~knepley/