Team reduction for array type

[kyungjoo-kim]

Currently team reduction is designed for scalar type and I need team reduction for array and the array should be thread specific. The input array may have initial values.

template< typename iType, class Lambda, typename ValueType, class JoinType >
KOKKOS_INLINE_FUNCTION
void parallel_reduce(const Impl::TeamThreadRangeBoundariesStruct<iType,Impl::ThreadsExecTeamMember>& loop_boundaries,
const Lambda & lambda, const JoinType& join, ValueType& init_result) {

ValueType result = init_result;

for( iType i = loop_boundaries.start; i < loop_boundaries.end; i+=loop_boundaries.increment) {
ValueType tmp = ValueType();

/// this tmp initialization should be reconsider to use thread local array.

lambda(i,tmp);
join(result,tmp);

}

init_result = loop_boundaries.thread.team_reduce(result,Impl::JoinLambdaAdapter<ValueType,JoinType>(join));
}

[crtrott]

If this is still relevant please reissue in github kokkos/kokkos