Numeric Traits

Numeric traits

Defined in header <Kokkos_NumericTraits.hpp> which is included from <Kokkos_Core.hpp>

Provides a replacement for numeric_limits from the standard library header <limits>. Implements a new facility that is being to the C++23 standard library and that breaks the monolithic numeric_limits class template apart into individual trait templates. For details, please refer to P1841.

Numeric traits are defined in the Kokkos::Experimental namespace since Kokkos 3.5

See below the list of available traits.

trait* denotes traits that were added in Kokkos 3.6

Numeric distinguished value traits infinity finite_min finite_max epsilon round_error norm_min denorm_min* reciprocal_overflow_threshold* quiet_NaN* signaling_NaN*

Numeric characteristics traits digits digits10 max_digits10 radix min_exponent min_exponent10 max_exponent max_exponent10

Standard library	Kokkos with C++17
`std::numeric_limits<Integral>::min()`	`finite_min_v<Integral>`
`std::numeric_limits<FloatingPoint>::min()`	`norm_min_v<FloatingPoint>`
`std::numeric_limits<Arithmetic>::lowest()`	`finite_min_v<Arithmetic>`
`std::numeric_limits<Arithmetic>::max()`	`finite_max_v<Arithmetic>`
`std::numeric_limits<FloatingPoint>::epsilon()`	`epsilon_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::round_error()`	`round_error_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::infinity()`	`infinity_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::quiet_NaN()`	`quiet_NaN_v<FloatingPoint>` (since 3.6)
`std::numeric_limits<FloatingPoint>::signaling_NaN()`	`signaling_NaN_v<FloatingPoint>` (since 3.6)
`std::numeric_limits<FloatingPoint>::denorm_min()`	`denorm_min_v<FloatingPoint>` (since 3.6)
`std::numeric_limits<Arithmetic>::digits`	`digits_v<Arithmetic>`
`std::numeric_limits<Arithmetic>::digits10`	`digits10_v<Arithmetic>`
`std::numeric_limits<FloatingPoint>::max_digits10`	`max_digits10_v<FloatingPoint>`
`std::numeric_limits<Arithmetic>::radix`	`radix_v<Arithmetic>`
`std::numeric_limits<FloatingPoint>::min_exponent`	`min_exponent_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::min_exponent10`	`min_exponent10_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::max_exponent`	`max_exponent_v<FloatingPoint>`
`std::numeric_limits<FloatingPoint>::max_exponent10`	`max_exponent10_v<FloatingPoint>`

Individual traits have value member constant that can be used with C++14 (e.g. epsilon<float>::value).

Individual traits are SFINAE-friendly, you can detect value presence/absence.

template <class T>
constexpr auto has_infinity(T)
    -> decltype(Kokkos::Experimental::infinity<T>::value, std::true_type{}) {
  return {};
}

constexpr std::false_type has_infinity(...) { return {}; }

template <class T>
KOKKOS_FUNCTION constexpr std::enable_if_t<has_infinity(T{}), T>
legacy_std_numeric_limits_infinity() {
  return Kokkos::Experimental::infinity<T>::value;
}

template <class T>
KOKKOS_FUNCTION constexpr std::enable_if_t<!has_infinity(T{}), T>
legacy_std_numeric_limits_infinity() {
  return T();
}