std::tuple<ptrdiff_t, ptrdiff_t>
partition(const distributed_matrix<B> &A, idx_t npart, std::vector<ptrdiff_t> &perm) const {
communicator comm = A.comm();
idx_t n = A.loc_rows();
int active = (n > 0);
std::vector<idx_t> ptr;
std::vector<idx_t> col;
symm_graph(A, ptr, col);
idx_t wgtflag = 0;
idx_t numflag = 0;
idx_t options = 0;
idx_t edgecut = 0;
idx_t ncon = 1;
std::vector<real_t> tpwgts(npart, 1.0 / npart);
std::vector<real_t> ubvec(ncon, 1.05);
std::vector<idx_t> part(n);
if (!n) part.reserve(1); // So that part.data() is not NULL
MPI_Comm scomm;
MPI_Comm_split(comm, active ? 0 : MPI_UNDEFINED, comm.rank, &scomm);
if (active) {
communicator sc(scomm);
std::vector<idx_t> vtxdist = sc.exclusive_sum(n);
sc.check(
METIS_OK == ParMETIS_V3_PartKway( &vtxdist[0], &ptr[0],
&col[0], NULL, NULL, &wgtflag, &numflag, &ncon,
&npart, &tpwgts[0], &ubvec[0], &options, &edgecut,
&part[0], &scomm),
"Error in ParMETIS"
);
MPI_Comm_free(&scomm);
}
return graph_perm_index(comm, npart, part, perm);
}
void
ParMETISGraphPartitionerImpl::p_partition(void)
{
int me(this->processor_rank());
std::vector<idx_t> vtxdist;
std::vector<idx_t> xadj;
std::vector<idx_t> adjncy;
ParMETISGraphWrapper wrap(p_adjacency_list);
wrap.get_csr_local(vtxdist, xadj, adjncy);
int nnodes(vtxdist[me+1] - vtxdist[me]);
#if 0
for (int p = 0; p < this->processor_size(); ++p) {
if (this->processor_rank() == p) {
std::cout << "Processor " << p << ": nodes: ";
for (Index n = 0; n < nnodes; ++n) {
std::cout << p_adjacency_list.node_index(n) << ",";
}
std::cout << std::endl;
std::cout << "Processor " << p << ": vtxdist: ";
std::copy(vtxdist.begin(), vtxdist.end(),
std::ostream_iterator<idx_t>(std::cout, ","));
std::cout << std::endl;
std::cout << "Processor " << p << ": xadj: ";
std::copy(xadj.begin(), xadj.end(),
std::ostream_iterator<idx_t>(std::cout, ","));
std::cout << std::endl;
std::cout << "Processor " << p << ": adjncy: ";
std::copy(adjncy.begin(), adjncy.end(),
std::ostream_iterator<idx_t>(std::cout, ","));
std::cout << std::endl;
}
this->communicator().barrier();
}
#endif
// Call the partitioner (try to use variable names that match the documentation)
int status;
idx_t ncon(1);
idx_t wgtflag(3), numflag(0);
idx_t nparts(this->processor_size());
std::vector<idx_t> vwgt(nnodes, 1);
std::vector<idx_t> adjwgt(adjncy.size(), 2);
std::vector<real_t> tpwgts(nparts*ncon, 1.0/static_cast<real_t>(nparts));
real_t ubvec(1.05);
std::vector<idx_t> options(3);
options[0] = 1;
options[1] = 127;
options[2] = 14;
MPI_Comm comm(this->communicator());
idx_t edgecut;
std::vector<idx_t> part(nnodes);
status = ParMETIS_V3_PartKway(&vtxdist[0],
&xadj[0],
&adjncy[0],
&vwgt[0],
&adjwgt[0],
&wgtflag,
&numflag,
&ncon,
&nparts,
&tpwgts[0],
&ubvec,
&options[0],
&edgecut, &part[0],
&comm);
if (status != 0) {
// FIXME: throw an exception
}
// "part" contains the destination processors; transfer this to the
// local array
wrap.set_partition(vtxdist, part);
wrap.get_partition(p_node_destinations);
}
