inline std::vector<int> _autorange_workers(const alps::mpi::communicator &comm, bool include_boss) {
std::vector<int> out;
size_t Nprocs(comm.size() - !include_boss);
if (!Nprocs) throw (std::logic_error("No workers to evaluate"));
for (size_t p = 0; p < comm.size(); p++) {
if (include_boss || comm.rank() != p) {
out.push_back(p);
};
};
return out;
}
template<typename T, typename Op, typename C> void reduce_impl(const alps::mpi::communicator & comm, T const & in_values, Op op, int root, boost::true_type, C) {
// using alps::mpi::reduce;
// reduce(comm, in_values, op, root);
using alps::mpi::get_mpi_datatype;
if (comm.rank()==root) {
throw std::runtime_error("reduce_impl(): 4-arg overload is called by root rank."+ALPS_STACKTRACE);
}
MPI_Reduce((void*)&in_values, NULL, 1, get_mpi_datatype(T()),
alps::mpi::is_mpi_op<Op, T>::op(), root, comm);
}
template<typename T, typename Op, typename C> void reduce_impl(const alps::mpi::communicator & comm, T const & in_values, T & out_values, Op op, int root, boost::true_type, C) {
using alps::mpi::reduce;
// using boost::mpi::reduce;
// reduce(comm, (T)in_values, out_values, op, root); // TODO: WTF? - why does boost not define unsigned long long as native datatype
using alps::mpi::get_mpi_datatype;
if (comm.rank()!=root) {
// // usleep((comm.rank()+1)*1000000); // DEBUG!
// std::cerr << "DEBUG:WARNING: rank=" << comm.rank() << " is not root=" << root
// << " but called 5-argument reduce_impl()." + ALPS_STACKTRACE << std::endl;
}
void* sendbuf=const_cast<T*>(&in_values);
if (sendbuf == &out_values) {
sendbuf=MPI_IN_PLACE;
}
MPI_Reduce(sendbuf, &out_values, 1, get_mpi_datatype(T()),
alps::mpi::is_mpi_op<Op, T>::op(), root, comm);
}
void broadcast_tail(const alps::mpi::communicator& comm,
int& min_order, int& max_order,
std::vector<TAILT>& tails,
const TAILT& tail_init,
int root)
{
using alps::mpi::broadcast;
broadcast(comm, min_order, root);
broadcast(comm, max_order, root);
if (min_order==TAIL_NOT_SET) return;
if (comm.rank()!=root) {
tails.resize(max_order+1, tail_init);
}
for (int i=min_order; i<=max_order; ++i) {
tails[i].broadcast(comm,root);
}
}