std::vector<T> mpi_reduce(std::vector<T> const &a, communicator c, int root, bool all, MPI_Op op, std::true_type) {
std::vector<T> b(a.size());
if (!all)
MPI_Reduce((void *)a.data(), b.data(), a.size(), mpi_datatype<T>(), op, root, c.get());
else
MPI_Allreduce((void *)a.data(), b.data(), a.size(), mpi_datatype<T>(), op, c.get());
return b;
}
template <typename T> REQUIRES_IS_BASIC(T, T) mpi_reduce(T a, communicator c = {}, int root = 0, bool all = false, MPI_Op op = MPI_SUM) {
T b;
auto d = datatype<T>();
if (!all)
MPI_Reduce(&a, &b, 1, d, op, root, c.get());
else
MPI_Allreduce(&a, &b, 1, d, op, c.get());
return b;
}
//---------
static void broadcast(communicator c, A &a, int root) {
check_is_contiguous(a);
auto sh = a.shape();
MPI_Bcast(&sh[0], sh.size(), mpi_datatype<typename decltype(sh)::value_type>::invoke(), root, c.get());
if (c.rank() != root) a.resize(sh);
MPI_Bcast(a.data_start(), a.domain().number_of_elements(), D(), root, c.get());
}
template <typename T> std::vector<T> mpi_gather(std::vector<T> const &a, communicator c, int root, bool all, std::true_type) {
long size = mpi_reduce(a.size(), c, root, all);
std::vector<T> b((all || (c.rank() == root) ? size : 0));
auto recvcounts = std::vector<int>(c.size());
auto displs = std::vector<int>(c.size() + 1, 0);
int sendcount = a.size();
auto mpi_ty = mpi::mpi_datatype<int>();
if (!all)
MPI_Gather(&sendcount, 1, mpi_ty, &recvcounts[0], 1, mpi_ty, root, c.get());
else
MPI_Allgather(&sendcount, 1, mpi_ty, &recvcounts[0], 1, mpi_ty, c.get());
for (int r = 0; r < c.size(); ++r) displs[r + 1] = recvcounts[r] + displs[r];
if (!all)
MPI_Gatherv((void *)a.data(), sendcount, mpi_datatype<T>(), (void *)b.data(), &recvcounts[0], &displs[0], mpi_datatype<T>(),
root, c.get());
else
MPI_Allgatherv((void *)a.data(), sendcount, mpi_datatype<T>(), (void *)b.data(), &recvcounts[0], &displs[0], mpi_datatype<T>(),
c.get());
return b;
}
template <typename T> std::vector<T> mpi_scatter(std::vector<T> const &a, communicator c, int root, std::true_type) {
auto slow_size = a.size();
auto sendcounts = std::vector<int>(c.size());
auto displs = std::vector<int>(c.size() + 1, 0);
int recvcount = slice_length(slow_size - 1, c.size(), c.rank());
std::vector<T> b(recvcount);
for (int r = 0; r < c.size(); ++r) {
sendcounts[r] = slice_length(slow_size - 1, c.size(), r);
displs[r + 1] = sendcounts[r] + displs[r];
}
MPI_Scatterv((void *)a.data(), &sendcounts[0], &displs[0], mpi_datatype<T>(), (void *)b.data(), recvcount, mpi_datatype<T>(),
root, c.get());
return b;
}
// NOTE: We keep the naming mpi_XXX for the actual implementation functions
// so they can be defined in other namespaces and the mpi::reduce(T,...) function
// can find them via ADL
template <typename T> REQUIRES_IS_BASIC(T, void) mpi_broadcast(T &a, communicator c = {}, int root = 0) {
MPI_Bcast(&a, 1, datatype<T>(), root, c.get());
}
//---------
static void all_reduce_in_place(communicator c, A &a, int root) {
check_is_contiguous(a);
// assume arrays have the same size on all nodes...
MPI_Allreduce(MPI_IN_PLACE, a.data_start(), a.domain().number_of_elements(), D(), MPI_SUM, c.get());
}
void mpi_reduce_in_place(std::vector<T> &a, communicator c, int root, bool all, MPI_Op op, std::true_type) {
if (!all)
MPI_Reduce((c.rank() == root ? MPI_IN_PLACE : a.data()), a.data(), a.size(), mpi_datatype<T>(), op, root, c.get());
else
MPI_Allreduce(MPI_IN_PLACE, a.data(), a.size(), mpi_datatype<T>(), op, c.get());
}
template <typename T> void mpi_broadcast(std::vector<T> &a, communicator c, int root, std::true_type) {
size_t s = a.size();
mpi_broadcast(s, c, root);
if (c.rank() != root) a.resize(s);
MPI_Bcast(a.data(), a.size(), mpi_datatype<T>(), root, c.get());
}
