MueLu: Distance Laplacian dropping for multiple dof per node in CoalesceDropFactory_kokkos

[maxfirmbach]

Bug Report

@trilinos/muelu

Description

Using the distance Laplacian dropping from CoalesceDropFactory_kokkos for an elasticity problem with e.g. 2 or 3 dofs per node results in a segfault.

The main reason is the coordinate input and it's handling later on. It seems like the CoalesceDropFactory_kokkos expects an expanded dof coordinate vector build from the nodal one. In our application code we assemble the nodal coordinate vector and give it to MueLu. The multigrid hierarchy checks the coordinate vector map with ReplaceCoordinateMap and replaces it if e.g. A->getRowMap() != coordinates->getMap() is true. The replacement is not an extension to the matrix dof map though, but more a reordering based on the reconstructed matrix node map.
--> This will break in CoalesceDropFactory_kokkos as the maps of A and coordinates are not the same.

If I provide the dof based coordinate vector and A->getRowMap() == coordinates->getMap() is fullfilled, there's still a problem happening in CoalesceDropFactory_kokkos. As each dof at a point holds the same coordinates, looping over each matrix row results at some point in a zero distance breaking the 1/distance calculation for the strength measure with either segfaulting or random behavior.

So no matter what coordinate input I choose, it breaks either way. Currently I'm not sure what the correct input should look like and what coordinate vector we should expect in CoalesceDropFactory_kokkos.

@cgcgcg

Steps to Reproduce

I put a first WIP unit test here CoalesceDropFactory_kokkos-multiple-dof-distance-laplacian

[github-actions]

Automatic mention of the @trilinos/muelu team

[jhux2]

The multigrid hierarchy checks the coordinate vector map with ReplaceCoordinateMap and replaces it if e.g. A->getRowMap() != coordinates->getMap() is true. The replacement is not an extension to the matrix dof map though, but more a reordering based on the reconstructed matrix node map.
--> This will break in CoalesceDropFactory_kokkos as the maps of A and coordinates are not the same.

@maxfirmbach I don't understand this comment. If the blocksize > 1, then the map should be expanded.

https://github.com/trilinos/Trilinos/blob/2b67cd6a6fbfda5bb6c4c96159e53b181acc5485/packages/muelu/src/MueCentral/MueLu_Hierarchy_def.hpp#L1487-L1510C4

[cgcgcg]

I believe that for blockSize > 1 we want the map for the coordinates to be different from the rowmap of A. It is true that distance Laplacian for blockSize > 1 is buggy.

[maxfirmbach]

Problem also appears in #13893.