Different execution policies at runtime

user_2413857 0
Viewed 778

In C++17, a number of functions in the algorithm header now can take an execution policy. I can for example define and call a function like this:

template <class ExecutionPolicy>
void f1(const std::vector<std::string>& vec, const std::string& elem, ExecutionPolicy&& policy) {
    const auto it = std::find(
        std::forward<ExecutionPolicy>(policy),
        vec.cbegin(), vec.cend(), elem
    );
}

std::vector<std::string> vec;
f1(vec, "test", std::execution::seq);

However I haven't found a good way to use different policies at runtime. For example when I want to use a different policy depending on some input file.

I toyed around with variants, but in the end the problem was always the different types of std::execution::seq, std::execution::par and std::execution::par_unseq.

A working but cumbersome solution would look like this:

void f2(const std::vector<std::string>& vec, const std::string& elem, const int policy) {
    const auto it = [&]() {
        if (policy == 0) {
            return std::find(
                std::execution::seq,
                vec.cbegin(), vec.cend(), elem
            );
        }
        else if (policy == 1) {
            return std::find(
                std::execution::par,
                vec.cbegin(), vec.cend(), elem
            );
        }
        else{
            return std::find(
                std::execution::par_unseq,
                vec.cbegin(), vec.cend(), elem
            );
        }
    };
}

f2(vec, "test", 0);

Is there any more elegant solution I'm overlooking?

edit: maybe I should be more precise. Let's say the goal is to save the policy in a variable that can have either of the three policies. That variable should be a parameter to the function.

parallel-processing c++
2 Answers

The standard approach here is to separate the selection of a type from the use of the type: the latter takes the form of a function template instantiated several times by the former non-template function (or function template with fewer template parameters).

To avoid duplicating the normal parameters between these two layers, use a generic lambda as the template. To avoid duplicating the selection logic, make a function template that calls whatever lambda with the appropriate policy:

enum Policy {seq,par,par_unseq};

template<class F>
auto maybe_parallel(F f,Policy p) {
  switch(p) {
  case seq: return f(std::execution::seq);
  case par: return f(std::execution::par);
  default: return f(std::execution::par_unseq);
  }
}

auto f2(const std::vector<std::string>& vec,
        const std::string& elem,Policy p) {
  return maybe_parallel
    ([&](auto &pol) {return std::find(pol,vec.begin(),vec.end(),elem);},p);
}
user_85862270

An alternative approach is to use std::variant and std::visit for this purpose.

using parallel_policy = std::variant
                        <
                         std::execution::sequenced_policy,
                         std::execution::parallel_policy,
                         std::execution::parallel_unsequenced_policy,
                         std::execution::unsequenced_policy
                        >;

void f2(const std::vector<std::string>& vec,
        const std::string&              elem,
        const parallel_policy           policy
       )
{
    std::visit
    (
        [&](auto policy_real)
        {
            f1(vec, elem, policy_real);
        },
        policy
    );
}
user_2543060
Related