phast.hpp<2>

      const unsigned trees_per_thread = num_trees / num_threads;
      const unsigned more_work_limit =
          num_trees - trees_per_thread * num_threads;
      std::vector<std::thread> workers;
      workers.reserve(num_threads);
      for (unsigned id = 0; id < num_threads; ++id) {
        workers.emplace_back([this, id, &ch, &starts, &distances,
                              trees_per_thread, more_work_limit]() {
          const unsigned begin =
              trees_per_thread * id + more_work_limit -
              (id < more_work_limit ? (more_work_limit - id) : 0);
          const unsigned end =
              trees_per_thread * (id + 1) + more_work_limit -
              ((id + 1) < more_work_limit ? (more_work_limit - (id + 1)) : 0);
          for (unsigned index = begin; index < end; ++index) {
            forward_search_init(ch, ((unsigned *)&starts)[index], distances,
                                index);
          }
        });
      }