bugfix: DIV_BY_ZERO in MatchingFunctionSpacePartitionerLonLatPolygon

[wdeconinck]

What happened

FE_DIV_BY_ZERO reported by @cducher with following reproducer:

/*
To run:
mpirun -np 2 --use-hwthread-cpus $exe_dir/atlas_bug.x
*/
#include <vector>

#include "atlas/library.h"
#include "atlas/parallel/mpi/mpi.h"
#include "atlas/meshgenerator.h"
#include "atlas/grid.h"
#include "atlas/grid/Grid.h"
#include "atlas/grid/StructuredGrid.h"
#include "atlas/grid/detail/grid/Gaussian.h"
#include "atlas/grid/Distribution.h"
#include "atlas/grid/Vertical.h"
#include "atlas/functionspace/StructuredColumns.h"
#include "atlas/functionspace/NodeColumns.h"
#include "atlas/field/Field.h"

int gridSizeX_{32};  // grid size X
int gridSizeY_{64};  // grid size Y
int nHalo_{3};       // N halo points
int nLevels_{19};

atlas::functionspace::StructuredColumns setupFunctionSpace() {

    long n_procs = atlas::mpi::comm().size();

    // ------------------ grid ------------------
    std::vector<long> gaussian_x(gridSizeX_,gridSizeY_);
    atlas::StructuredGrid grid = atlas::grid::detail::grid::reduced_gaussian( gaussian_x );
    atlas::grid::Distribution distribution(grid, 
                                           atlas::util::Config("type", "checkerboard") | 
                                           atlas::util::Config("bands", n_procs));

    // ----------- function space ---------------
    return atlas::functionspace::StructuredColumns(grid, 
                                                    distribution, 
                                                    atlas::util::Config("halo", nHalo_) | 
                                                    atlas::util::Config("levels", nLevels_) );

}

int main(int argc, char* argv[]) {

    atlas::initialize(argc,argv);

    atlas::functionspace::StructuredColumns fs_ = setupFunctionSpace();

    atlas::Field field_dummy = fs_.createField<double>(atlas::option::name("dummy") | 
                                                        atlas::option::variables(1) | 
                                                        atlas::option::levels(nLevels_));

    // target mesh and functional space
    atlas::RectangularLonLatDomain rd({140,50}, {-45,45});
    atlas::Grid areaGrid(fs_.grid(), rd); 

    atlas::MeshGenerator outputMeshGen_("structured");
    // Floating point exception
    atlas::Mesh outputMesh_ = outputMeshGen_.generate(
        areaGrid, atlas::grid::MatchingPartitioner(fs_)
    );

    atlas::finalize();
    return 0;
}

Environment

This was run on our HPC2020 system with relevant modules intel/2021.4, hpcx-openmpi/2.9.0
The reproducer is run on e.g. ecinteractive queue with mpirun -np 2 --use-hwthread-cpus.
This seems to be equivalent to MPI_SIZE=2 and OMP_NUM_THREADS=2!

Reason

When using the MatchingFunctionSpacePartitionerLonLatPolygon for small grids with OpenMP,
we encounter a DIV_BY_ZERO.
This is due to the integer chunk_size = grid.size() / (1000 * num_threads) is computed as zero.
In that case really it needs to be chunk_size = grid.size().