Declare user-defined reduction of OpenMP for class template

Question

When using OpenMP, I would like to declare an user-defined reduction for a class template.

#include <omp.h>
#include <iostream>

template<typename T>
class Foo
{
  public:
  T Data_;

  template<typename U> friend Foo<U> operator+( const Foo<U>& lhs, const Foo<U>& rhs );
};

template<typename U>
Foo<U> operator+( const Foo<U>& lhs, const Foo<U>& rhs )
{
  Foo<U> Addition;

  Addition.Data_ = lhs.Data_ + rhs.Data_;

  return Addition;
}

#pragma omp declare reduction( + : template<typename U> Foo<U> : omp_out = omp_out + omp_in ) initializer (omp_priv=omp_orig)

int main( int argc, char* argv[] )
{
  Foo<int> Array[100];
  
  for ( int i = 0 ; i < 100 ; ++i )
  {
    Array[i].Data_ = i;
  }
  
  Foo<int> Sum {0};
  
  #pragma omp parallel for num_threads(4) reduction( + : Sum )
  for ( int i = 0 ; i < 100 ; ++i )
  {
    Sum.Data_ += Array[i].Data_;
  }
  
  std::cout << Sum.Data_ << std::endl;
  
  return 0;
}

But I got the following errors:

error: expected type-specifier before 'template'
#pragma omp declare reduction( + : template<typename U> Foo<U> : omp_out = omp_out + omp_in ) initializer (omp_priv=omp_orig)
                                   ^~~~~~~~

I can fix the error by replacing template<typename U> Foo<U> with Foo<int>.

But I would like to know is there any solution by keeping using template.

Thanks.

Victor Eijkhout · Accepted Answer · 2022-01-13 13:43:15Z

3

Supposing you have a templated function

template<typename T>
T f(T x,T y) {};

you can templatize the reduction:

#pragma omp declare reduction                                   \
  (rwzt:T:omp_out=f<T>(omp_out,omp_in))

And use that as:

template<typename T>
T generic_reduction( const vector<T>& tdata ) {
  #pragma omp declare reduction                                   \
    (rwzt:T:omp_out=f<T>(omp_out,omp_in))

  T tmin;
  #pragma omp parallel for reduction(rwzt:tmin)
   for ( stuff ) {}
  return tmin;
}

auto tm = generic_reduction<float>( /* some vector<float> */ );

I'm somewhat bothered by the fact that this needs to define a named function to contain them the reduction, rather than having all code inlined. I can not figure out a way to do this with a lambda in the calling environment.

edited Jan 13, 2022 at 13:43

answered Jan 12, 2022 at 16:41

Victor Eijkhout

5,9902 gold badges29 silver badges31 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

Jérôme Richard Over a year ago

Good. This seems to be well supported by mainstream compilers. Just a minor remark: I think it is better to use const vector<T>& instead of vector<T> to avoid unneeded copies.

Victor Eijkhout Over a year ago

@JérômeRichard With vectors of a size worth parallelizing you're absolutely right. Code edited.

Collectives™ on Stack Overflow

Declare user-defined reduction of OpenMP for class template

1 Answer 1

2 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related