Skip to content

Latest commit

 

History

History
126 lines (97 loc) · 3.39 KB

Custom-Reductions:-Custom-Reducers.md

File metadata and controls

126 lines (97 loc) · 3.39 KB

Custom arbitrary reductions are implemented using a reduction class and a "reduced" class. The "reduced" class is much like the custom scalar type used with BuiltIn Reducers and the reduction class implements the ReducerConcept

The following requirements must be fulfilled for the "reduced" class

  • Operators required for applied reduction class must be implemented.
  • The class / struct must either use the default copy constructor or have a specific copy constructor implemented.

The following requirements must be fulfilled for the reduction class

  • Typedefs reducer, value_type and result_view_type must be defined. see the Reducer Concept for details.
  • The reducer concept methods must be implemented
  • The exposed result_view_type must be defined in the memory space where the object is used

Example

This example performs a custom reduction on an array using a custom class and reducer.

#include <Kokkos_Core.hpp>

namespace sample {

template <class ScalarType, int N>
struct array_type {
  ScalarType myArray[N];

  KOKKOS_INLINE_FUNCTION
  array_type() { init(); }

  KOKKOS_INLINE_FUNCTION
  array_type(const array_type& rhs) {
    for (int i = 0; i < N; i++) {
      myArray[i] = rhs.myArray[i];
    }
  }

  KOKKOS_INLINE_FUNCTION  // initialize myArray to 0
      void
      init() {
    for (int i = 0; i < N; i++) {
      myArray[i] = 0;
    }
  }

  KOKKOS_INLINE_FUNCTION
  array_type& operator+=(const array_type& src) {
    for (int i = 0; i < N; i++) {
      myArray[i] += src.myArray[i];
    }
    return *this;
  }

  KOKKOS_INLINE_FUNCTION
  void operator+=(const volatile array_type& src) volatile {
    for (int i = 0; i < N; i++) {
      myArray[i] += src.myArray[i];
    }
  }
};

template <class T, class Space, int N>
struct SumMyArray {
 public:
  // Required
  typedef SumMyArray reducer;
  typedef array_type<T, N> value_type;
  typedef Kokkos::View<value_type*, Space, Kokkos::MemoryUnmanaged>
      result_view_type;

 private:
  value_type& value;

 public:
  KOKKOS_INLINE_FUNCTION
  SumMyArray(value_type& value_) : value(value_) {}

  // Required
  KOKKOS_INLINE_FUNCTION
  void join(value_type& dest, const value_type& src) const { dest += src; }

  KOKKOS_INLINE_FUNCTION
  void join(volatile value_type& dest, const volatile value_type& src) const {
    dest += src;
  }

  KOKKOS_INLINE_FUNCTION
  void init(value_type& val) const { val.init(); }

  KOKKOS_INLINE_FUNCTION
  value_type& reference() const { return value; }

  KOKKOS_INLINE_FUNCTION
  result_view_type view() const { return result_view_type(&value, 1); }

  KOKKOS_INLINE_FUNCTION
  bool references_scalar() const { return true; }
};
}  // namespace sample

int main(int argc, char* argv[]) {
  Kokkos::initialize(argc, argv);
  {
    int E = 1024;

    typedef sample::array_type<int, 4> ValueType;
    typedef sample::SumMyArray<int, Kokkos::HostSpace, 4> ArraySumResult;

    ValueType tr;

    Kokkos::parallel_reduce(
        E,
        KOKKOS_LAMBDA(const int i, ValueType& upd) {
          int ndx = i % 4;  // sum all of the i%4 entries (divide total by 4)
          upd.myArray[ndx] += 1;
        },
        ArraySumResult(tr));

    // Output result.
    printf("  Computed result %d, %d, %d, %d \n", tr.myArray[0], tr.myArray[1],
           tr.myArray[2], tr.myArray[3]);
  }
  Kokkos::finalize();
}