#pragma once

#include "launch.hpp"

namespace tf {

namespace detail {

// Function: scan_loop
template 
void scan_loop(
  tf::Runtime& rt,
  std::atomic& counter, 
  BufferT& buf, 
  B bop, 
  Iterator d_beg, 
  size_t W,
  size_t w, 
  size_t chunk_size
){
  // whoever finishes the last performs global scan
  if(counter.fetch_add(1, std::memory_order_acq_rel) == W-1) {
    for(size_t i=1; i;
  using value_type = typename std::iterator_traits::value_type;
  
  return [=] (Runtime& rt) mutable {

    // fetch the stateful values
    B_t s_beg = first;
    E_t s_end = last;
    D_t d_beg = d_first;

    if(s_beg == s_end) {
      return;
    }

    size_t W = rt.executor().num_workers();
    size_t N = std::distance(s_beg, s_end);

    // only myself - no need to spawn another graph
    if(W ;
  using value_type = typename std::iterator_traits::value_type;
  
  return [=] (Runtime& rt) mutable {

    // fetch the stateful values
    B_t s_beg = first;
    E_t s_end = last;
    D_t d_beg = d_first;

    if(s_beg == s_end) {
      return;
    }

    size_t W = rt.executor().num_workers();
    size_t N = std::distance(s_beg, s_end);

    // only myself - no need to spawn another graph
    if(W