-
Notifications
You must be signed in to change notification settings - Fork 0
Kokkos::parallel_reduce
Header File: Kokkos_Core.hpp
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.
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...);template <class ExecPolicy, class FunctorType, class ReducerArgument1, class... ReducerArguments>
Kokkos::parallel_reduce(const std::string& name,
const ExecPolicy& policy,
const FunctorType& functor,
const ReducerArgument& reducer...);template <class ExecPolicy, class FunctorType, class ReducerArgument1, class... ReducerArguments>
Kokkos::parallel_reduce(const ExecPolicy& policy,
const FunctorType& functor,
const ReducerArgument& reducer...);template <class ExecPolicy, class FunctorType, class ReducerArgumentNonConst1, class... ReducerArgumentNonConsts>
Kokkos::parallel_reduce(const std::string& name,
const ExecPolicy& policy,
const FunctorType& functor,
ReducerArgumentNonConst& reducer...);template <class ExecPolicy, class FunctorType, class ReducerArgumentNonConst1, class ReducerArgumentNonConsts>
Kokkos::parallel_reduce(const ExecPolicy& policy,
const FunctorType& functor,
ReducerArgumentNonConst& reducer...);template <class ExecPolicy, class FunctorType>
Kokkos::parallel_reduce(const std::string& name,
const ExecPolicy& policy,
const FunctorType& functor,
const ResultType& result...);template <class ExecPolicy, class FunctorType>
Kokkos::parallel_reduce(const ExecPolicy& policy,
const FunctorType& functor,
const ResultType& result...);-
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.
-
TeamThreadRange: defines a 1D iteration range to be executed by a thread-team. Only valid inside a parallel region executed through a
TeamPolicyor aTaskTeam. -
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
TeamPolicyor aTaskTeam.
-
- FunctorType: A valid functor with (at minimum) an
operator()with a matching signature for theExecPolicycombined with the reduced type. - ReducerArgument: Either a class fullfilling the "Reducer" concept or a
Kokkos::View - ReducerArgumentNonConst: a class fullfilling the "Reducer" concept, a POD type with
operator +=andoperator =, or aKokkos::View. The ReducerArgumentNonConst can also be an array or a pointer; see below for functor requirements.
- If
ExecPolicyis notMDRangePolicy, thefunctorhas a member function of the formoperator() (const HandleType& handle, ReducerValueType& value) constoroperator() (const WorkTag, const HandleType& handle, ReducerValueType& value) const.- The
WorkTagfree form of the operator is used ifExecPolicyis anIntegerTypeor ifExecPolicy::work_tagisvoid. -
HandleTypeis anIntegerTypeifExecPolicyis anIntegerTypeelse it isExecPolicy::member_type.
- The
- If
ExecPolicyisMDRangePolicythefunctorhas a member function of the formoperator() (const IntegerType& i0, ... , const IntegerType& iN, ReducerValueType& value) constoroperator() (const WorkTag, const IntegerType& i0, ... , const IntegerType& iN, ReducerValueType& value) const.- The
WorkTagfree form of the operator is used ifExecPolicy::work_tagis notvoid. -
Nmust matchExecPolicy::rank.
- The
- If the
functoris a lambda,ReducerArgumentmust satisfy theReducerconcept orReducerArgumentNonConstmust be a POD type withoperator +=andoperator =or aKokkos::View. In the latter case, the defaultSumreduction is applied. - If
ExecPolicyisTeamThreadRangea "reducing"functoris not allowed and theReducerArgumentmust satisfy theReducerconcept orReducerArgumentNonConstmust be a POD type withoperator +=andoperator =or aKokkos::View. In the latter case, the defaultSumreduction is applied. - The reduction argument type
ReducerValueTypeof thefunctoroperator must be compatible with theReducerArgument(orReducerArgumentNonConst) and must match the arguments of theinit,joinandfinalfunctions of the functor if those exist. - If
ReducerArgument(orReducerArgumentNonConst)- is a scalar type then
ReducerValueTypemust be of the same type. - is a
Kokkos::ViewthenReducerArgument::rankmust be 0 andReducerArgument::non_const_value_typemust matchReducerValueType. - satisfies the
Reducerconcept thenReducerArgument::value_typemust matchReducerValueType. - is an array or a pointer
- ReducerValueType must match the array or the pointer signature.
- the functor must define FunctorType::value_type the same as ReducerValueType.
- the functor must declare a public member variable
int value_countwhich is the length of the array. - the functor must implement the function
void init( ReducerValueType dst [] ) constorvoid init( ReducerValueType * dst) constdepending on whether ReducerArgumentNonConst is an array or pointer respectively. - the functor must implement the function
void join( ReducerValueType dst[], ReducerValueType src[] ) constorvoid join( ReducerValueType * dst, ReducerValueType * src ) constdepending on whether.ReducerArgumentNonConst is an array or pointer respectively. - If the functor implements the
finalfunction, the argument must also match those of init and join.
- is a scalar type then
- Neither concurrency nor order of execution are guaranteed.
- The call is potentially asynchronous if the
ReducerArgumentis not a scalar type. - The
ReducerArgumentcontent 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.
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,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();
}Home:
- Introduction
- Machine Model
- Programming Model
- Compiling
- Initialization
- View
- Parallel Dispatch
- Hierarchical Parallelism
- Custom Reductions
- Atomic Operations
- Subviews
- Interoperability
- Kokkos and Virtual Functions
- Initialization and Finalization
- View
- Data Parallelism
- Execution Policies
- Spaces
- Task Parallelism
- Utilities
- STL Compatibility
- Numerics
- Detection Idiom