void parfor_all_vertices(GraphType& graph,
FunctionType fn,
vertex_set vset = GraphType::complete_set(),
size_t nfibers = 10000,
size_t stacksize = 16384) {
distributed_control::get_instance()->barrier();
initialize_counters();
fiber_group group;
group.set_stacksize(stacksize);
warp_impl::parfor_all_vertices_impl<GraphType> parfor(graph, fn, vset);
for (size_t i = 0;i < nfibers; ++i) {
group.launch(boost::bind(&warp_impl::parfor_all_vertices_impl<GraphType>::run_fiber, &parfor));
}
group.join();
distributed_control::get_instance()->barrier();
graph.synchronize(vset);
}
auto pardot(Eigen::MatrixBase<V1> const& A, Eigen::MatrixBase<V2> const& B, int chunkSize = 1024) {
if(A.rows() < 10*chunkSize) {
return A.dot(B);
}
using T = std::decay_t<decltype(A(0))>;
struct r_struct{T val{0};};
SmallVector<ReductionVariable<r_struct>, 8> S(yafel::config::num_cores, r_struct{0});
parfor(0, A.rows(), [&](auto i) {
auto id = worker_global::worker_id;
S[id].val += A(i)*B(i);
},getGlobalScheduler(), chunkSize);
T total{0};
for(auto &s : S){
total += s.val;
}
return total;
}