MueLu: is reindexColumns needed in Trilinos-15

[jrobcary]

Question

@trilinos/muelu In Trilinos-14, we had to call reindexColumns before calling MueLu. Upon moving to Trilinos-15, the exact same code gives

3: unknown location:0: fatal error: in "poissonUpdater/poissonUpdater<std____1__pair<double_ std____1__integral_constant<unsigned long_ 2ul>>>": std::runtime_error: /Users/cary/projects/xsimall-dev1/trilinos/packages/tpetra/core/src/Tpetra_CrsGraph_def.hpp:3937:
3: 
3: Throw number = 1
3: 
3: Throw test that evaluated to true: (gblSuccess[0] == 0)
3: 
3: Tpetra::CrsGraph<int, long long, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<Kokkos::Serial, Kokkos::HostSpace>>::reindexColumns: It is not possible to continue.  The most likely reason is that the graph is locally indexed, but the column Map is missing (null) on some processes, due to a previous call to replaceColMap().

Upon commenting out the code that calls reindexColumns, the code runs and seems to give correct answers. Did something change in the code regarding the use of reindexColumns?

(I put this in the search bars at the trilinos github site and at https://trilinos.github.io/muelu.html and at https://trilinos.github.io/documentation.html, but nothing came up.)

[github-actions[bot]]

Automatic mention of the @trilinos/muelu team

[jrobcary]

A bit more complicated. If reindexColumns is taken out, the code fails at above 2000 degrees of freedom with

3: Throw number = 1
3: 
3: Throw test that evaluated to true: bIsZeroNSColumn == true
3: 
3: MueLu::TentativePFactory::MakeTentative: fine level NS part has a zero column in NS column 0

[cgcgcg]

Maybe try a debug build. My best guess in that some of the maps or matrices are set up incorrectly.

For the MueLu error: on what level does it fail? What this check says is the the near-nullspace has an outright zero column which is a bad thing.