DiFfRG::ParallelDoFs Namespace Reference

DoF-layout helpers shared by the CG, DG and FV discretizations. More...

Functions
template<int dim, int spacedim>
dealii::IndexSet	make_ghost_set (const dealii::DoFHandler< dim, spacedim > &dof_handler)
	The ghost (locally relevant) set used by the replicated-mesh policy: everything.
template<int dim, int spacedim>
void	build_support_points (const dealii::Mapping< dim, spacedim > &mapping, const dealii::DoFHandler< dim, spacedim > &dof_handler, std::vector< dealii::Point< spacedim > > &support_points)
	Fill a dense, mesh-wide support point array on every rank.

Detailed Description

DoF-layout helpers shared by the CG, DG and FV discretizations.

These exist because the answer to "which dofs does this rank hold?" differs between a serial and a partitioned triangulation, and every discretization needs the same answer.

Function Documentation

build_support_points()

template<int dim, int spacedim>

void DiFfRG::ParallelDoFs::build_support_points	(	const dealii::Mapping< dim, spacedim > &	mapping,
		const dealii::DoFHandler< dim, spacedim > &	dof_handler,
		std::vector< dealii::Point< spacedim > > &	support_points )

Fill a dense, mesh-wide support point array on every rank.

Why this is not just DoFTools::map_dofs_to_support_points: the std::vector overload of that function refuses to run on a DoFHandler built on a parallel::TriangulationBase – and parallel::shared::Triangulation is one – because its precondition is an array sized to the global dof count. Calling it on a partitioned mesh is an error, not a silent degradation.

The std::map overload does work in parallel, but returns only the locally relevant dofs. The union of those sets over all ranks is the whole mesh, so a max-reduction over a lowest()-initialized array reconstructs the dense global array exactly. Support points are pure geometry – every rank that holds a given dof computes bit-identical coordinates for it – so the reduction is a gather, not an approximation, and the result does not depend on the rank count.

Parameters

support_points

resized to dof_handler.n_dofs() and fully populated on every rank.

make_ghost_set()

template<int dim, int spacedim>

dealii::IndexSet DiFfRG::ParallelDoFs::make_ghost_set

(

const dealii::DoFHandler< dim, spacedim > &

dof_handler

)

The ghost (locally relevant) set used by the replicated-mesh policy: everything.

This is deliberately NOT DoFTools::extract_locally_relevant_dofs. The KT reconstruction stencil reaches three cells deep normal to a boundary and reads corner/diagonal cells that are not face neighbours at all (see FV/assembler/reconstruction_cache.hh), and the EoM search and the output path read arbitrary global indices. A one-ring relevant set is not wide enough for any of them, and being too narrow is a silent wrong answer rather than an error.

Ghosting everything costs exactly today's memory footprint, which is the whole reason the replicated-mesh rung is cheap. Narrowing this to extract_locally_relevant_dofs is the defining step of the fully-distributed rung, and every call site that reads a global index has to be revisited at the same time.