peclet::core::halo Namespace Reference

Namespaces
namespace	detail

Classes
struct	DomainMap
	Physical-space layout that maps a position to a global cell of the decomposition. More...
struct	GridFieldView
	A contiguous local array of T (one per cell of the extended block) viewed as a packable field. More...
class	GridHalo
	GPU ghost-layer exchange for a contiguous device field peclet::core::View<T> (one element per extended-block cell). More...
class	GridHaloTopology
class	NbxEngine
class	ParticleHalo
class	ParticleHaloTopology
class	ParticleMigrator
class	ParticleMigratorView
	Device counterpart of ParticleMigrator. More...

Functions
template<int Dim>
std::size_t	rebalanceByParticleCount (decomp::BlockDecomposer< Dim > &dec, ParticleMigrator< Dim > &mig, std::vector< typename detail::NonDeduced< Vec< Dim > >::type > &pos, std::vector< char > &payload, std::size_t stride, MPI_Comm comm=MPI_COMM_WORLD, std::vector< double > *weightOut=nullptr)
	Re-decompose dec so each block holds a near-equal particle count, then migrate the particles (positions + opaque fixed-stride payload) to their new owners via mig (which must reference dec).

Function Documentation

rebalanceByParticleCount()

template<int Dim>

std::size_t peclet::core::halo::rebalanceByParticleCount	(	decomp::BlockDecomposer< Dim > &	dec,
		ParticleMigrator< Dim > &	mig,
		std::vector< typename detail::NonDeduced< Vec< Dim > >::type > &	pos,
		std::vector< char > &	payload,
		std::size_t	stride,
		MPI_Comm	comm = MPI_COMM_WORLD,
		std::vector< double > *	weightOut = nullptr
	)

Re-decompose dec so each block holds a near-equal particle count, then migrate the particles (positions + opaque fixed-stride payload) to their new owners via mig (which must reference dec).

weightOut, when non-null, receives the agreed global per-cell count grid (x-fastest). Returns this rank's new local particle count. (pos is a non-deduced parameter — Dim is fixed by dec/mig — because int Dim cannot deduce from a std::array's size_t bound.)

Collective on comm: one MPI_Allreduce of the global per-cell count grid (size product(globalSize) doubles — mind it on a fine decomposition grid) and the migration's NBX round. payload holds stride bytes per particle. The new partition is the unaligned weighted ORB init(numBlocks, globalSize, weights); the coarse-first aligned form (init(…, weights, align) / chooseAlignedWeighted) is not used here.

Definition at line 47 of file particle_rebalance.hpp.

References peclet::core::decomp::BlockDecomposer< Dim >::globalSize(), peclet::core::decomp::BlockDecomposer< Dim >::init(), peclet::core::decomp::BlockDecomposer< Dim >::linearGlobal(), MPI_Allreduce(), MPI_Comm_size(), MPI_DOUBLE, and MPI_SUM.