std::shared_ptr<matrix> operator()(const matrix &A, unsigned block_size = 1) const {
communicator comm = A.comm();
idx_t n = A.loc_rows();
ptrdiff_t row_beg = A.loc_col_shift();
// Partition the graph.
int active = (n > 0);
int active_ranks = comm.reduce(MPI_SUM, active);
idx_t npart = std::max(1, active_ranks / prm.shrink_ratio);
if (comm.rank == 0)
std::cout << "Partitioning[ParMETIS] " << active_ranks << " -> " << npart << std::endl;
std::vector<ptrdiff_t> perm(n);
ptrdiff_t col_beg, col_end;
if (npart == 1) {
col_beg = (comm.rank == 0) ? 0 : A.glob_rows();
col_end = A.glob_rows();
for(ptrdiff_t i = 0; i < n; ++i) {
perm[i] = row_beg + i;
}
} else {
if (block_size == 1) {
std::tie(col_beg, col_end) = partition(A, npart, perm);
} else {
typedef typename math::scalar_of<value_type>::type scalar;
typedef backend::builtin<scalar> sbackend;
ptrdiff_t np = n / block_size;
distributed_matrix<sbackend> A_pw(A.comm(),
pointwise_matrix(*A.local(), block_size),
pointwise_matrix(*A.remote(), block_size)
);
std::vector<ptrdiff_t> perm_pw(np);
std::tie(col_beg, col_end) = partition(A_pw, npart, perm_pw);
col_beg *= block_size;
col_end *= block_size;
for(ptrdiff_t ip = 0; ip < np; ++ip) {
ptrdiff_t i = ip * block_size;
ptrdiff_t j = perm_pw[ip] * block_size;
for(unsigned k = 0; k < block_size; ++k)
perm[i + k] = j + k;
}
}
}
return graph_perm_matrix<Backend>(comm, col_beg, col_end, perm);
}
pointwise_aggregates(const Matrix &A, const params &prm) : count(0)
{
if (prm.block_size == 1) {
plain_aggregates aggr(A, prm);
count = aggr.count;
strong_connection.swap(aggr.strong_connection);
id.swap(aggr.id);
} else {
strong_connection.resize( nonzeros(A) );
id.resize( rows(A) );
Matrix Ap = pointwise_matrix(A, prm.block_size);
plain_aggregates pw_aggr(Ap, prm);
count = pw_aggr.count * prm.block_size;
#pragma omp parallel
{
std::vector<ptrdiff_t> marker(Ap.nrows, -1);
#ifdef _OPENMP
int nt = omp_get_num_threads();
int tid = omp_get_thread_num();
size_t chunk_size = (Ap.nrows + nt - 1) / nt;
size_t chunk_start = tid * chunk_size;
size_t chunk_end = std::min(Ap.nrows, chunk_start + chunk_size);
#else
size_t chunk_start = 0;
size_t chunk_end = Ap.nrows;
#endif
for(size_t ip = chunk_start, ia = ip * prm.block_size; ip < chunk_end; ++ip) {
ptrdiff_t row_beg = Ap.ptr[ip];
ptrdiff_t row_end = row_beg;
for(unsigned k = 0; k < prm.block_size; ++k, ++ia) {
id[ia] = prm.block_size * pw_aggr.id[ip] + k;
for(ptrdiff_t ja = A.ptr[ia], ea = A.ptr[ia+1]; ja < ea; ++ja) {
ptrdiff_t cp = A.col[ja] / prm.block_size;
if (marker[cp] < row_beg) {
marker[cp] = row_end;
strong_connection[ja] = pw_aggr.strong_connection[row_end];
++row_end;
} else {
strong_connection[ja] = pw_aggr.strong_connection[ marker[cp] ];
}
}
}
}
}
}
}
