如何用`std :: function`作为函数参数创建一个可变参数模板函数？

Question

How can I create a variadic template function with std::function as a function parameter that accepts a variadic number of arguments? I tried to reduce the problem to a MWE:

#include <functional>

template <class T> void run(std::function<void(T *)> fun, T *obj) { fun(obj); }

template <class T, class... Args>
void run_variadic(std::function<void(T *, Args...)> fun, T *obj, Args... args) {
  fun(obj, args...);
}

struct Foo {
  void bar() {}
};

int main() {
  Foo foo;
  std::function<void(Foo *)> fun = &Foo::bar;

  run(fun, &foo);                     // works
  run<Foo>(&Foo::bar, &foo);          // works
  run_variadic(fun, &foo);            // works
  run_variadic<Foo>(&Foo::bar, &foo); // does not compile
}

It seems like the mere presence of the variadic template parameter in run_variadic makes it impossible to directly call it with a member function pointer. clang 's error message is as follows:

main.cpp:21:3: error: no matching function for call to 'run_variadic'
  run_variadic<Foo>(&Foo::bar, &foo); // does not compile
  ^~~~~~~~~~~~~~~~~
main.cpp:6:6: note: candidate template ignored: could not match 'function<void (Foo *, type-parameter-0-1...)>' against 'void (Foo::*)()'
void run_variadic(std::function<void(T *, Args...)> fun, T *obj, Args&&... args) {
     ^
1 error generated.

Any suggestions on how I can fix run_variadic so that I do not have to go through the extra std::function object?

Background

I have a class hierarchy as

template <class T> class Abstract { ... };
class UnrelatedStuff { ... };
class Derived : public Abstract<UnrelatedStuff> { ... };

There are multiple Derived classes that all have to implement one or more methods to loop over a range of elements. The loop looks something like

#pragma omp parallel for
for (ZFSId i = begin; i != end; ++i) {
  callMemFun(i, and, other, args);
}

All loops should be OpenMP-accelerated. I want the accelerator stuff factored out and not repeated in each method of Derived that uses a loop, so that I only have to change one place if eg OpenMP would switch to OpenACC.

Thus I am looking for a way to put the loop (and its decoration) in its own function. Moving it to the Abstract base class is not an option either, since the loops are performance-critical and I cannot have an abstract function call in each loop iteration.

Answer 1

You are almost always certainly better off abstracting away the function object:

template <class Functor, class... Args>
void run(Functor&& f, Args&&... args) {
  f(std::forward<Args>(args)...);
}

This allows you to do the right thing at call site:

// function object is a lambda that binds to a member function:
run([&](auto... args) { foo.bar(args...); } /*, bar takes no args...*/);

I prefer a lambda to std::function or std::bind but you can also use those if they are already available:

run(std::function<void(Foo *)>{&Foo::bar}, &foo);
run(std::bind(&Foo::bar, &foo));
run(std::mem_fn(&Foo::bar), foo);

I provide a full example program below.

You have now edited the question with new information regarding what you are trying to do.

I'm pretty sure that you don't want to do this, since the OpenMP/OpenACC pragmas like parallel for usually require extra annotations for delivering reasonable performance, and they depend on what you are exactly trying to do at call site.

Still, if you really really want to go this route you can write your own for_each algorithm and dispatch according to an ExecutionAgent (see N3874 and N3731 ). If OpenMP, TBB, OpenACC parallel task are too slow, you can also easily provide overloads based on eg an ExecutionPolicy like this:

template<class RandomAccessRange, class Functor, 
         class ExecutionPolicy = execution::serial_t>
void for_each(RandomAccessRange&& r, Functor&& f, 
              ExecutionPolicy&& ex = ExecutionPolicy{}) {
  detail::for_each_(std::forward<RandomAccessRange>(r), 
                    std::forward<Functor>(f), 
                    std::forward<ExecutionPolicy>(ex));
}

And then you can implement overloads of for_each_ for each execution policy, eg:

namespace detail {

template<class RandomAccessRange, class Functor>
void for_each(RandomAccessRange&& r, Functor&& f, execution::serial_t) {
  boost::for_each(std::forward<RandomAccessRange>(r), std::forward<Functor>(f));
}

template<class RandomAccessRange, class Functor>
void for_each(RandomAccessRange&& r, Functor&& f, execution::openmp_t) {
  #pragma omp parallel for
  for (auto&& v : r) { f(v); } 
}

template<class RandomAccessRange, class Functor>
void for_each(RandomAccessRange&& r, Functor&& f, execution::openacc_t) {
  #pragma acc parallel for
  for (auto&& v : r) { f(v); } 
}

template<class RandomAccessRange, class Functor>
void for_each(RandomAccessRange&& r, Functor&& f, execution::tbb_t) {
  tbb::parallel_for_each(std::begin(std::forward<RandomAccessRange>(r)),
                         std::end(std::forward<RandomAccessRange>(r)),
                         std::forward<Functor>(f)); 
}

}  // namespace detail

Note that the ExecutionPolicy is just a tag, ie:

namespace execution {
  struct serial_t {}; static const constexpr serial_t serial{};
  struct openmp_t {}; static const constexpr openmp_t openmp{};
  struct openacc_t {}; static const constexpr openacc_t openacc{};
  struct tbb_t {}; static const constexpr tbb_t tbb{};
}  // namespace execution

This will at least give you an efficient TBB backend even tho the OpenMP/OpenACC performance will be mediocre at best. You can take a look at the parallel implementation of libstdc++ where they use OpenMP. Their for_each algorithm is over 1000 lines of code and uses work-stealing.

Full example program:

#include <functional>

template <class Functor, class... Args>
void run(Functor&& f, Args&&... args) {
  f(std::forward<Args>(args)...);
}

struct Foo { void bar() {} };

int main() {
  Foo foo;

  run([&](auto... args) { foo.bar(args...); } /*, bar takes no args*/);
  run(std::function<void(Foo *)>{ &Foo::bar}, &foo);
  run(std::bind(&Foo::bar, &foo));
  run(std::mem_fn(&Foo::bar), foo);
}

Answer 2

To answer your comment on the previous answer, that answer can be adapted to support pointers to member functions in the way that you've asked for. The previous answer already works for all callable objects, but not directly with a pointer to member function because those are not callable with the usual f(args) syntax. The following version uses tag dispatch to distinguish between pointers to member functions and traditional callable objects, applying the call syntax appropriate to each case.

template <class Functor, class... Args>
void run_helper(std::false_type, Functor f, Args&&... args)
{
    f(std::forward<Args>(args)...);
}

template <class Functor, class Arg0, class... Args>
void run_helper(std::true_type, Functor f, Arg0&& arg0, Args&&... args)
{
    (std::forward<Arg0>(arg0).*f)(std::forward<Args>(args)...);
}

template <class Functor, class... Args>
void run(Functor f, Args&&... args)
{
    run_helper(typename std::is_member_pointer<Functor>::type(),
               f, std::forward<Args>(args)...);
}

This can be used in all the same ways as the previous answer could, but also supports directly passing in a pointer to member function:

run(&Foo::bar, foo);

It even works with overloaded member functions and member functions which are templates, if you explicitly instantiate the run template to bind to a particular overloaded function or function template instantiation.

Live example: http://ideone.com/vsBS4H