thread_map.cpp

template<typename T>
std::vector<T> thread_map(int thread_num, std::vector<T> origin_arg, std::function<T(T)> func) {
    std::vector<std::thread> threads_;
    std::vector<T> output_(origin_arg.size());
    for(int i = 0; i < thread_num; ++i){
        auto runner = [i, func, origin_arg, thread_num, &output_]() {
            int array_size = origin_arg.size();
            int start = (array_size + thread_num - 1) / thread_num * i;
            int end = std::min((array_size + thread_num - 1) / thread_num * (i + 1), array_size);
            for(int j = start; j < end; ++j) {
                output_[j] = func(origin_arg[j]);
            }
        };
        threads_.emplace_back(std::thread(runner));
    }
    for(int i = 0; i < thread_num; ++i) {
        threads_[i].join();
    }
    return output_;
}