Fix a bug in mg ghost layer with periodicity and n_levels=1

[kronbichler]

This PR fixes a subtle bug in the setup of the level subdomain ids in the ghost layer of the multigrid levels: If we have a mesh with one single level (i.e., not refined, and a case where multigrid may not make much sense), we forgot to identify cells behind periodic boundaries as level ghost cells; this makes downstream multigrid data structures to break down.

The fix is to move the code that sets the periodic data structures before we call

dealii/source/distributed/tria.cc

Line 4727 in a0a79b2

copy_local_forest_to_triangulation();

Otherwise the code inside this call, namely the one that identifies the appropriate cells and vertices here:

dealii/source/distributed/tria.cc

Lines 3713 to 3739 in a0a79b2

	std::vector<std::vector<bool>> marked_vertices(this->n_levels());
	for (unsigned int lvl = 0; lvl < this->n_levels(); ++lvl)
	marked_vertices[lvl] = mark_locally_active_vertices_on_level(lvl);
	for (typename Triangulation<dim, spacedim>::cell_iterator cell =
	this->begin(0);
	cell != this->end(0);
	++cell)
	{
	typename dealii::internal::p4est::types<dim>::quadrant
	p4est_coarse_cell;
	const unsigned int tree_index =
	coarse_cell_to_p4est_tree_permutation[cell->index()];
	typename dealii::internal::p4est::types<dim>::tree *tree =
	init_tree(cell->index());
	dealii::internal::p4est::init_coarse_quadrant<dim>(
	p4est_coarse_cell);
	determine_level_subdomain_id_recursively<dim, spacedim>(
	*tree,
	tree_index,
	cell,
	p4est_coarse_cell,
	*parallel_forest,
	this->my_subdomain,
	marked_vertices);

does not include the periodic neighbors for a triangulation immediately after add_periodicity.

This was not a problem when we refined the mesh, as then the periodicity has already been set.

FYI @nfehn

[tjhei]

This look good to me. I hope the multigrid algorithm failed in an obvious way?

[kronbichler]

Yes, deadlock for DG. For continuous elements we got an error in MGConstrainedDoFs that we were trying to get dofs on artificial cells.

[tjhei]

Yes, deadlock for DG. For continuous elements we got an error in MGConstrainedDoFs that we were trying to get dofs on artificial cells.

I guess that is good enough. :-)