`parallel_reduce`¶

Header File: <Kokkos_Core.hpp>

Usage¶

Kokkos::parallel_reduce(name, policy, functor, reducer...);
Kokkos::parallel_reduce(name, policy, functor, result...);
Kokkos::parallel_reduce(name, policy, functor);
Kokkos::parallel_reduce(policy, functor, reducer...);
Kokkos::parallel_reduce(policy, functor, result...);
Kokkos::parallel_reduce(policy, functor);

Dispatches parallel work defined by functor according to the ExecutionPolicy and performs a reduction of the contributions provided by workers as defined by the execution policy. The optional label name is used by profiling and debugging tools. The reduction type is either a sum, is defined by the reducer or is deduced from an optional join operator on the functor. The reduction result is stored in result, or through the reducer handle. It is also provided to the functor.final() function if such a function exists. Multiple reducers can be used in a single parallel_reduce and thus, it is possible to compute the min and the max values in a single parallel_reduce.

Interface¶

template <class ExecPolicy, class FunctorType>
Kokkos::parallel_reduce(const std::string& name,
                        const ExecPolicy& policy,
                        const FunctorType& functor);

template <class ExecPolicy, class FunctorType>
Kokkos::parallel_reduce(const ExecPolicy& policy,
                        const FunctorType& functor);

template <class ExecPolicy, class FunctorType, class... ReducerArgument>
Kokkos::parallel_reduce(const std::string& name,
                        const ExecPolicy& policy,
                        const FunctorType& functor,
                        const ReducerArgument&... reducer);

template <class ExecPolicy, class FunctorType, class... ReducerArgument>
Kokkos::parallel_reduce(const ExecPolicy& policy,
                        const FunctorType& functor,
                        const ReducerArgument&... reducer);

template <class ExecPolicy, class FunctorType, class... ReducerArgumentNonConst>
Kokkos::parallel_reduce(const std::string& name,
                        const ExecPolicy& policy,
                        const FunctorType& functor,
                        ReducerArgumentNonConst&... reducer);

template <class ExecPolicy, class FunctorType, class... ReducerArgumentNonConst>
Kokkos::parallel_reduce(const ExecPolicy& policy,
                        const FunctorType& functor,
                        ReducerArgumentNonConst&... reducer);

Parameters:¶

name: A user provided string which is used in profiling and debugging tools via the Kokkos Profiling Hooks.
ExecPolicy: An ExecutionPolicy which defines iteration space and other execution properties. Valid policies are:
- IntegerType: defines a 1D iteration range, starting from 0 and going to a count.
- RangePolicy: defines a 1D iteration range.
- MDRangePolicy: defines a multi-dimensional iteration space.
- TeamPolicy: defines a 1D iteration range, each of which is assigned to a thread team.
- TeamVectorRange: defines a 1D iteration range to be executed by a thread-team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
- TeamVectorMDRange: defines a multi-dimensional iteration space to be executed by a thread-team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
- TeamThreadRange: defines a 1D iteration range to be executed by a thread-team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
- TeamThreadMDRange: defines a multi-dimensional iteration space to be executed by a thread-team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
- ThreadVectorRange: defines a 1D iteration range to be executed through vector parallelization dividing the threads within a team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
- ThreadVectorMDRange: defines a multi-dimensional iteration space to be executed through vector parallelization dividing the threads within a team. Only valid inside a parallel region executed through a TeamPolicy or a TaskTeam.
FunctorType: A valid functor with (at minimum) an operator() with a matching signature for the ExecPolicy combined with the reduced type.
ReducerArgument: Either a class fulfilling the “Reducer” concept or a Kokkos::View.
ReducerArgumentNonConst: A scalar type or an array type; see below for functor requirements.

Requirements:¶

If ExecPolicy is not MDRangePolicy, the functor has a member function of the form operator() (const HandleType& handle, ReducerValueType& value) const or operator() (const WorkTag, const HandleType& handle, ReducerValueType& value) const.
- If ExecPolicy::work_tag is void or if ExecPolicy is an IntegerType, the overload without a WorkTag argument is used.
- HandleType is an IntegerType if ExecPolicy is an IntegerType else it is ExecPolicy::member_type.
If ExecPolicy is MDRangePolicy the functor has a member function of the form operator() (const IntegerType& i0, ... , const IntegerType& iN, ReducerValueType& value) const or operator() (const WorkTag, const IntegerType& i0, ... , const IntegerType& iN, ReducerValueType& value) const.
- If ExecPolicy::work_tag is void, the overload without a WorkTag argument is used.
- N must match ExecPolicy::rank.
If the functor is a lambda, ReducerArgument must satisfy the Reducer concept or ReducerArgumentNonConst must be a POD type with operator += and operator = or a Kokkos::View. In the latter case, a sum reduction is applied where the identity is assumed to be given by the default constructor of the value type (and not by reduction_identity`). If provided, the init/ join/ final member functions must not take a WorkTag argument even for tagged reductions.
If ExecPolicy is TeamThreadRange a “reducing” functor is not allowed and the ReducerArgument must satisfy the Reducer concept or ReducerArgumentNonConst must be a POD type with operator += and operator = or a Kokkos::View. In the latter case, a sum reduction is applied where the identity is assumed to be given by the default constructor of the value type (and not by reduction_identity`).
If ExecPolicty is TeamVectorMDRange, TeamThreadMDRange, or ThreadVectorMDRange, only a ReducerArgumentNonConst is allowed, and it must be a POD type with operator += and operator =.
The reduction argument type ReducerValueType of the functor operator must be compatible with the ReducerArgument (or ReducerArgumentNonConst) and must match the arguments of the init, join and final functions of the functor if those exist and no reducer is specified (ReducerArgument doesn’t satisfy the Reducer concept but is a scalar, array or Kokkos::View). In case of tagged reductions, i.e., when specifying a tag in the policy, the functor’s potential init/ join/ final member functions must also be tagged.
If ReducerArgument (or ReducerArgumentNonConst)
- is a scalar type then ReducerValueType must be of the same type.
- is a Kokkos::View then ReducerArgument::rank must be 0 and ReducerArgument::non_const_value_type must match ReducerValueType.
- satisfies the Reducer concept then ReducerArgument::value_type must match ReducerValueType.
- is an array
  - ReducerValueType must match the array signature.
  - the functor must define FunctorType::value_type the same as ReducerValueType.
  - the functor must declare a public member variable int value_count which is the length of the array.
  - the functor must implement the function void init( ReducerValueType dst[] ) const.
  - the functor must implement the function void join( ReducerValueType dst[], ReducerValueType src[] ) const.
  - If the functor implements the final function, the argument must also match those of init and join.

Semantics¶

For every element of the iteration space defined in policy the functors call operator is invoked exactly once, except for TeamPolicy and TeamThreadRange where the call operator is invoked by each vector lane of the team and thread respectively.
Neither concurrency nor order of execution are guaranteed.
The call is potentially asynchronous if the ReducerArgument is not a scalar type.
The ReducerArgument content will be overwritten, i.e. the value does not need to be initialized to the reduction-neutral element.
The input value to the operator may contain a partial reduction result, Kokkos may only combine the thread local contributions in the end. The operator must modify the input reduction value according to the requested reduction type.

Examples¶

Further examples are provided in the Custom Reductions and ExecutionPolicy documentation.

#include <Kokkos_Core.hpp>
#include <cstdio>

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

    int N = atoi(argv[1]);
    double result;
    Kokkos::parallel_reduce("Loop1", N, KOKKOS_LAMBDA (const int& i, double& lsum) {
        lsum += 1.0*i;
    }, result);

    printf("Result: %i %lf\n", N, result);
    Kokkos::finalize();
}

#include <Kokkos_Core.hpp>
#include <cstdio>

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

    int N = atoi(argv[1]);
    double sum, min;
    Kokkos::parallel_reduce("Loop1", N, KOKKOS_LAMBDA (const int& i, double& lsum, double& lmin) {
        lsum += 1.0*i;
        lmin = lmin < 1.0*i ? lmin : 1.0*i;
    }, sum, Kokkos::Min<double>(min));

    printf("Result: %i %lf %lf\n", N, sum, min);
    Kokkos::finalize();
}

#include <Kokkos_Core.hpp>
#include <cstdio>

struct TagMax {};
struct TagMin {};

struct Foo {
    KOKKOS_INLINE_FUNCTION
    void operator() (const TagMax, const Kokkos::TeamPolicy<>::member_type& team, double& lmax) const {
        if (team.league_rank() % 17 + team.team_rank() % 13 > lmax)
            lmax = team.league_rank() % 17 + team.team_rank() % 13;
    }
    KOKKOS_INLINE_FUNCTION
    void operator() (const TagMin, const Kokkos::TeamPolicy<>::member_type& team, double& lmin) const {
        if (team.league_rank() % 17 + team.team_rank() % 13 < lmin)
            lmin = team.league_rank() % 17 + team.team_rank() % 13;
    }
};

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

    int N = atoi(argv[1]);

    Foo foo;
    double max, min;
    Kokkos::parallel_reduce(Kokkos::TeamPolicy<TagMax>(N,Kokkos::AUTO), foo, Kokkos::Max<double>(max));
    Kokkos::parallel_reduce("Loop2", Kokkos::TeamPolicy<TagMin>(N,Kokkos::AUTO), foo, Kokkos::Min<double>(min));
    Kokkos::fence();

    printf("Result: %lf %lf\n", min, max);

    Kokkos::finalize();
}

parallel_reduce¶