Applying a modified preconditioner matrix in PETSc

General Discussion
1

Hi all,

Is it possible to specific a modified preconditioner matrix when using PETSc in FreeFEM?

In PETSc a different Amat and Pmat can be specified using KSPSetOperators(KSP ksp,Mat Amat,Mat Pmat). Is this possible using FreeFEM?

Thanks in advance,
Greg

2

It is possible, but I had to make design choices that made the API “easy” for beginners, and slightly less “easy” for power users. I’d consider this a power user feature, thus the API is not very “easy”.

Here goes. KSPSetOperators() is always called by Amat in both slots. However, you can trick FreeFEM (or PETSc, whatever you prefer), by using a PCSHELL with your KSP. To do that, you have to define a func PetscScalar[int] Pmat(PetscScalar[int]& in) { PetscScalar[int] out; KSPSolve(P, in, out); return out; } and then call set(Amat, precon = Pmat, sparams = "-pc_type shell");.
Of course, you may want to adjust the Pmat function to your needs, see, e.g., FreeFem-sources/oseen-2d-PETSc.edp at develop · FreeFem/FreeFem-sources · GitHub.

If this does not fits your needs, please let me know.

1 Like
3

Thanks for the quick response. I’ll give that a go.

4

Is PetscScalar[int] definitely a permitted output type to a func? It looks like PetscScalar isn’t a valid identifier.

5

Is PetscScalar[int] definitely a permitted output type to a func ?

Yes.

$ git diff --unified=1 oseen-2d-PETSc.edp
diff --git a/examples/hpddm/oseen-2d-PETSc.edp b/examples/hpddm/oseen-2d-PETSc.edp
index db7774a3..d79f2595 100644
--- a/examples/hpddm/oseen-2d-PETSc.edp
+++ b/examples/hpddm/oseen-2d-PETSc.edp
@@ -109,3 +109,3 @@ set(dAp, sparams="-prefix_push stiffness_ -ksp_constant_null_space -pc_type gamg
 set(dMp, sparams="-prefix_push mass_ -pc_type jacobi -ksp_type cg -ksp_max_it 5 -prefix_pop", prefix="mass_");
-func real[int] PCD(real[int]& in) {
+func PetscScalar[int] PCD(PetscScalar[int]& in) {
     real[int] out(in.n);
$ ff-mpirun -n 4 oseen-2d-PETSc.edp -v 0
'mpiexec' -n 4 /Users/jolivet/repo/FreeFem-sources-opt/src/mpi/FreeFem++-mpi -nw 'oseen-2d-PETSc.edp' -v 0
  0 KSP Residual norm 1.126942766958e+01
  1 KSP Residual norm 1.211704042327e+00
  2 KSP Residual norm 1.196774444376e+00
  3 KSP Residual norm 9.349518221811e-01
  4 KSP Residual norm 6.165076499177e-01
  5 KSP Residual norm 5.160899777308e-01
  6 KSP Residual norm 4.258461730310e-01
  7 KSP Residual norm 3.233845833153e-01
  8 KSP Residual norm 2.549468207077e-01
  9 KSP Residual norm 2.082486066860e-01
 10 KSP Residual norm 1.658173156505e-01
 11 KSP Residual norm 1.265213502966e-01
 12 KSP Residual norm 9.290575051382e-02
 13 KSP Residual norm 6.324474572261e-02
 14 KSP Residual norm 3.979519396695e-02
 15 KSP Residual norm 2.400455372980e-02
 16 KSP Residual norm 1.467199804593e-02
 17 KSP Residual norm 9.934514655560e-03
 18 KSP Residual norm 7.109634777396e-03
 19 KSP Residual norm 5.471649973749e-03
 20 KSP Residual norm 4.547054280958e-03
 21 KSP Residual norm 4.178684030727e-03
 22 KSP Residual norm 3.924235944687e-03
 23 KSP Residual norm 3.406191499661e-03
 24 KSP Residual norm 2.343458788730e-03
 25 KSP Residual norm 1.199118692965e-03
 26 KSP Residual norm 6.228083971918e-04
 27 KSP Residual norm 3.377254147466e-04
 28 KSP Residual norm 1.996655242067e-04
 29 KSP Residual norm 1.108701321233e-04
Linear solve converged due to CONVERGED_RTOL iterations 29

By just saying “It looks like PetscScalar isn’t a valid identifier”, I can’t really help you further. My guess is that you have an old FreeFEM version which doesn’t have this commit https://github.com/FreeFem/FreeFem-sources/commit/723520e75ec36eb7f3802d30ec67319c249af51d#diff-6dd6dd51051ada00b3d8fa1b5e233cecd08ffe961add4f9fbdadf0805b73db1d.

1 Like