Triclinic 1xNx1 domain decomposition with Verlet scheme gives incorrect results
With the Verlet scheme and triclinic boxes, a 1D decomposition over y, i.e. 1xNx1, produces incorrect bounding boxes for the non-bonded grid, which results in missing non-bonded interactions. Unless box angles are very close to 90 degrees, many interactions will be missing and the system will crash after a few steps.
This was never noticed before, because the default 1D domain decomposition is Nx1x1, unless the box is longer in y than in x. Because the high number of missing interactions makes the system crash quickly, this will not lead to silent errors.
Fixed triclinic 1xNx1 domain decomposition
With the Verlet scheme, 1D triclinic domain decomposition along
dimension y produces incorrect bounding boxes for the non-bonded grid.
This led to a lot of missing non-bonded interactions, which quickly
crashes any simulation affected by this.
Add test cases to show bugs are fixed
- octahedron adds a fully trilinic test case
- dd121 adds a test case that will default to more
ranks in a direction other than X (here, Y)
The reference data is generated with a "last known good" patch from
release-5-0 from before the fix for #1631 was introduced, not that it
really matters. The reference build is unaffected by any of the bugs,
because it runs single-core.