SolutionView< VectorType > Class Template Reference#

DiFfRG: DiFfRG::SolutionView< VectorType > Class Template Reference

DiFfRG

Discretization Framework for functional Renormalization Group flows

A read-only, fully-replicated view of the solution. More...

#include <solution_view.hh>

Public Types
using	NumberType = get_type::NumberType<VectorType>

using	size_type = dealii::types::global_dof_index

Public Member Functions
	SolutionView ()=default

void	reinit (const dealii::IndexSet &, const dealii::IndexSet &, MPI_Comm)
	Establish the layout. No-op for the serial policy.

void	refresh (const VectorType &owned)
	Make the view reflect `owned`.

const VectorType &	get () const

	operator const VectorType & () const

NumberType	operator[] (const size_type i) const

size_type	size () const

Private Attributes
const VectorType *	source = nullptr

Detailed Description

template<typename VectorType>
class DiFfRG::SolutionView< VectorType >

A read-only, fully-replicated view of the solution.

A large amount of code under discretization/ reads the solution at arbitrary global dof indices: the EoM search (common/eom.hh), the KT reconstruction caches (FV/assembler/reconstruction_cache.hh), the output path (data/fe_output.cc) and the support point lookups. None of that is expressible against a vector that only holds this rank's rows.

Rather than rewrite all of it, this class maintains one replica holding every index, refreshed once per residual/Jacobian evaluation. Those call sites then read the replica and need no logic change at all. At replicated-mesh sizes the replica is exactly today's memory footprint.

The serial specialization is a pure passthrough – it stores a pointer, copies nothing, and compiles to the same code as passing the vector directly – so a serial build pays nothing for the abstraction.

Warning: The two policies differ in one respect that matters if you hold a view across a write: the serial view aliases the source vector, so it observes later mutations, while the distributed view is a snapshot taken at refresh(). Always refresh() immediately before reading and never hold a view across an assembly step, and the two behave identically.

Narrowing the replica from "everything" to the locally relevant dofs is the defining step of the fully-distributed rung. It is deliberately confined to this one class so that step has a single place to happen, and so the call sites that will need real work are exactly the ones that name this type.