Parallel seam: minimal MPI abstraction (rank/size + collectives) with serial fallback. More...

This graph shows which files directly or indirectly include this file:

Namespaces
namespace	pops

Functions
bool	pops::comm_active ()

void	pops::comm_init (int =nullptr, char **=nullptr)

void	pops::comm_finalize ()

int	pops::my_rank ()

int	pops::n_ranks ()

void	pops::barrier ()

double	pops::all_reduce_sum (double x)

double	pops::all_reduce_max (double x)

double	pops::all_reduce_min (double x)

long	pops::all_reduce_sum (long x)

void	pops::all_reduce_sum_inplace (double *, int)

void	pops::all_reduce_or_inplace (char *, int)

Detailed Description

Parallel seam: minimal MPI abstraction (rank/size + collectives) with serial fallback.

Layer: include/pops/parallel. Role: expose my_rank()/n_ranks() and a fixed set of global reductions (sum/min/max on double and long, in-place sum on a double array, in-place OR on a marker array) behind a single facade. Without POPS_HAS_MPI everything compiles to a serial identity; the rest of the code never sees MPI_COMM_WORLD nor mpi.h directly. Contract: every collective operates on MPI_COMM_WORLD; each rank must call them in the same order (otherwise deadlock). The sum_inplace / or_inplace bricks feed respectively the multi-patch AMR reflux and the gathering of regrid tags before clustering; all_reduce_min guarantees a global dt identical on all ranks.

Invariants:

even compiled with MPI, if MPI is not initialized my_rank()=0 / n_ranks()=1 (comm_active() tests Initialized && !Finalized); call comm_init() at the start of main() for a distributed run;
the all_reduce_* functions are COLLECTIVE: all ranks participate or none;
in serial mode each function is the identity (no-op or returns the argument).

include/pops/parallel/comm.hpp File Reference¶

Namespaces

Functions

Detailed Description