ncarray::ReductionTraits

ncarray::ReductionTraits

Definition

template<class Traits, typename T>
concept=  requires(ssize_t idx,
                                     T val,
                                     typename Traits::template AccumT<T> a,
                                     typename Traits::template AccumT<T> b,
                                     typename Traits::template AccumT<T> res,
                                     typename Traits::template AccumT<T>* dest,
                                     typename Traits::template AccumT<T> aval) {
    typename Traits::template AccumT<T>;
    { Traits::template identity<T>() };
    { Traits::template fill<T>() };
    { Traits::template transform<T>(idx, val) };
    { Traits::template reduce<T>(a, b) };
    { Traits::template store<T>(res, 0.0) };
    { Traits::template atomic<T>(dest, aval) };
  };

Detailed Description

Interface specifier for a valid reduction operation. Traits adhering to this specification can be used for performing axis-aware reductions. In the case that the reduction accumulates into a type that is different from the desired output type, an additional dual_atomic function (not included in this concept) would be required for operation on GPU. See argmax/argmin or variance traits for examples. The identity function returns the identity for the accumulator type. The fill function will, in general, return the identity; however, if the output type is not the same as the accumulator type this may be different. For example, the argmax reduction accumulates key/value pairs (max and index). However, the output is just the index, so fill returns an integer, while identity returns the key/value pair. The transform function converts an index and value into the acummulator type. The reduce function performs the necessary comparisons/reductions on a single pair of elements. The store function converts an accumulator type back to the output type. A dual_atomic function may be needed as well. The atomic function is used for a generalized atomic reduction on GPU.