MueLu: Use of ShiftedLaplacian with long long GO

[maartenarnst]

Currently, the ShiftedLaplacian preconditioner is enabled in MueLu only when Tpetra is instantiated with the int global ordinal.

• E.g.:

  Trilinos/packages/muelu/src/Utils/ForwardDeclaration/MueLu_ShiftedLaplacian_fwd.hpp

  Line 14 in fcf1459

  #if defined(HAVE_MUELU_IFPACK2) && defined(HAVE_MUELU_BELOS) && defined(HAVE_MUELU_TPETRA_INST_INT_INT)

It is also not clear whether tests using the ShiftedLaplacian preconditioner are still exercised by the Trilinos test pipelines. Because it seems that there are inconsistencies such as

• Trilinos/packages/muelu/test/helmholtz/Helmholtz3D.cpp

  Lines 91 to 92 in fcf1459

  	Galeri::Xpetra::Parameters<GO> matrixParameters_helmholtz(clp, nx, ny, nz, "Helmholtz3D", 0, stretchx, stretchy, stretchz,
  	h, delta, PMLXL, PMLXR, PMLYL, PMLYR, PMLZL, PMLZR, omega, 0.0, mx, my, mz, model);

• Trilinos/packages/galeri/src-xpetra/Galeri_XpetraParameters.hpp

  Lines 32 to 37 in fcf1459

  	Parameters(Teuchos::CommandLineProcessor& clp, GO nx = 16, GO ny = -1, GO nz = -1, const std::string& matrixType = "Laplace1D",
  	int keepBCs = 0, double stretchx = 1.0, double stretchy = 1.0, double stretchz = 1.0,
  	double Kxx = 1.0, double Kxy = 0.0, double Kyy = 1.0, double dt = 1.0, const std::string& meshType = "tri",
  	double h = 1.0, double delta = 0.0,
  	int PMLXL = 0, int PMLXR = 0, int PMLYL = 0, int PMLYR = 0, int PMLZL = 0, int PMLZR = 0,
  	double omega = 2.0*M_PI, double shift = 0.5, GO mx = -1, GO my = -1, GO mz = -1, int model = 0)

I.e. the kxx, kxy parameters and so on are not passed by the Helmholtz test, which leads to a compilation error in our build.

May be related to:

• MueLu: delete ShiftedLaplacian interface? #2608

We are wondering

• what the underlying cause is for not allowing the ShiftedLaplacian preconditioner for other global ordinals? Is it just a question of generalising the ifdef protections and enabling things in the eti?
• what is currently the recommended preconditioner to use for the Helmholtz equation?

Thanks in advance.

@cgcgcg @jhux2

@romintomasetti @GaleAxiom

[github-actions]

Automatic mention of the @trilinos/muelu team

[jhux2]

Adding @rstumin.

[rstumin]

I don't believe anyone is using the ShiftedLaplacian code ... so it is not clear to me whether it serves a useful purpose as a test?

[maartenarnst]

Just to clarify our question a bit more. We are looking for a preconditioner to solve Helmholtz problems in our Trilinos-based FE code. We saw the shifted Laplacian in Muelu. In our code, we use the default GO in Tpetra, which is typically long long (and not int). And so our question is not really about the status of the test in Muelu. Our question is whether/how the shifted Laplacian can work with GO other than int. And more generally, we wonder what is the currently recommended preconditioner for Helmholtz problems in Trilinos.

[cgcgcg]

I would be surprised if there is anything specific to GO=int. So it might be quite easy to enable long long. But this code was never fully integrated and by now probably has other issues that would need to be fixed. On the other hand, we added some capabilities for using Krylov solvers as level smoothers that could come in handy.