| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | //====================================================================================================================== | ||
| 2 | /* | ||
| 3 | POLYFLOAT - Extended precision floating points | ||
| 4 | Copyright : POLYFLOAT Contributors & Maintainers | ||
| 5 | SPDX-License-Identifier: BSL-1.0 | ||
| 6 | */ | ||
| 7 | //====================================================================================================================== | ||
| 8 | #pragma once | ||
| 9 | |||
| 10 | #include <polyfloat/details/callable.hpp> | ||
| 11 | #include <polyfloat/types/concepts.hpp> | ||
| 12 | #include <polyfloat/types/traits.hpp> | ||
| 13 | #include <type_traits> | ||
| 14 | #include <polyfloat/module/core/two_fma_approx.hpp> | ||
| 15 | |||
| 16 | namespace plf | ||
| 17 | { | ||
| 18 | |||
| 19 | template<typename Options> | ||
| 20 | struct dot_t : eve::strict_tuple_callable<dot_t, Options, kahan_option, raw_option, pedantic_option> | ||
| 21 | { | ||
| 22 | template<typename... Ts> struct result : as_polyfloat_like<Ts...> | ||
| 23 | { | ||
| 24 | }; | ||
| 25 | |||
| 26 | template<concepts::polyfloat_like... Ts> | ||
| 27 | requires(eve::same_lanes_or_scalar<Ts...>) | ||
| 28 | 52 | POLYFLOAT_ABI typename result<Ts...>::type constexpr operator()(Ts... ts) const noexcept | |
| 29 | { | ||
| 30 | 52 | return POLYFLOAT_CALL(ts...); | |
| 31 | } | ||
| 32 | |||
| 33 | template<eve::non_empty_product_type Tup1, eve::non_empty_product_type Tup2> | ||
| 34 | 12 | POLYFLOAT_ABI constexpr kumi::apply_traits_t<result, kumi::result::cat_t<Tup1, Tup2>> operator()( | |
| 35 | Tup1 const& t1, Tup2 const& t2) const noexcept | ||
| 36 | requires(eve::same_lanes_or_scalar_tuple<Tup1> && eve::same_lanes_or_scalar_tuple<Tup2> && | ||
| 37 | (Tup1::size() == Tup2::size()) && !concepts::polyfloat_like<Tup1> && !concepts::polyfloat_like<Tup2>) | ||
| 38 | { | ||
| 39 | 12 | return POLYFLOAT_CALL(t1, t2); | |
| 40 | } | ||
| 41 | |||
| 42 | template<eve::non_empty_product_type Tup> | ||
| 43 | requires(eve::same_lanes_or_scalar_tuple<Tup> && !concepts::polyfloat_like<Tup>) | ||
| 44 | POLYFLOAT_ABI constexpr kumi::apply_traits_t<result, Tup> operator()(Tup const& t) const noexcept | ||
| 45 | requires(kumi::size_v<Tup> >= 1) | ||
| 46 | { | ||
| 47 | return POLYFLOAT_CALL(t); | ||
| 48 | } | ||
| 49 | |||
| 50 | 64 | POLYFLOAT_CALLABLE_OBJECT(dot_t, dot_); | |
| 51 | }; | ||
| 52 | //====================================================================================================================== | ||
| 53 | //! @addtogroup core | ||
| 54 | //! @{ | ||
| 55 | //! @var dot | ||
| 56 | //! @brief return the dot product of the vector of the first half of parameters by | ||
| 57 | //! the vector of the second half. | ||
| 58 | //! | ||
| 59 | //! @groupheader{Header file} | ||
| 60 | //! | ||
| 61 | //! @code | ||
| 62 | //! #include <polyfloat/module/core.hpp> | ||
| 63 | //! @endcode | ||
| 64 | //! | ||
| 65 | //! @groupheader{Callable Signatures} | ||
| 66 | //! | ||
| 67 | //! @code | ||
| 68 | //! namespace polyfloat | ||
| 69 | //! { | ||
| 70 | //! // Regular overloads | ||
| 71 | //! constexpr auto dot(auto value... xs, auto value... ys) noexcept; // 1 | ||
| 72 | //! constexpr auto dot(kumi::tuple xs, kumi::tuple ys) noexcept; // 2 | ||
| 73 | //! } | ||
| 74 | //! @endcode | ||
| 75 | //! | ||
| 76 | //! **Parameters** | ||
| 77 | //! | ||
| 78 | //! * `z1`, `z2`: Values to process. | ||
| 79 | //! | ||
| 80 | //! **Return value** | ||
| 81 | //! | ||
| 82 | //! 1. dot product. \f$\sum_s x_s*y_s\f$. | ||
| 83 | //! 2. use the content of the tuples | ||
| 84 | //! | ||
| 85 | //! @groupheader{Example} | ||
| 86 | //! | ||
| 87 | //! @godbolt{doc/core/dot.cpp} | ||
| 88 | //====================================================================================================================== | ||
| 89 | |||
| 90 | inline constexpr auto dot = eve::functor<dot_t>; | ||
| 91 | //====================================================================================================================== | ||
| 92 | //! @} | ||
| 93 | //====================================================================================================================== | ||
| 94 | } | ||
| 95 | |||
| 96 | namespace plf::_ | ||
| 97 | { | ||
| 98 | |||
| 99 | template<eve::callable_options O, concepts::polyfloat_like... Ts> | ||
| 100 | 52 | constexpr auto dot_(POLYFLOAT_DELAY(), O const& o, Ts const&... args) noexcept | |
| 101 | requires(sizeof...(Ts) > 1 && sizeof...(Ts) % 2 == 0) | ||
| 102 | { | ||
| 103 | if constexpr (sizeof...(Ts) == 2) | ||
| 104 | { | ||
| 105 | 34 | return plf::mul[o](args...); | |
| 106 | } | ||
| 107 | else | ||
| 108 | { | ||
| 109 | using r_t = as_polyfloat_like_t<Ts...>; | ||
| 110 | using u_t = eve::element_type_t<r_t>; | ||
| 111 | 38 | auto cvt = [](auto a) { return plf::convert(a, as<u_t>()); }; | |
| 112 | 18 | auto coeffs = kumi::tuple{cvt(args)...}; | |
| 113 | 54 | auto [f, s] = kumi::split(coeffs, kumi::index<sizeof...(Ts) / 2>); | |
| 114 | 55 | return add[o](kumi::map([](auto a, auto b) { return a * b; }, f, s)); | |
| 115 | } | ||
| 116 | } | ||
| 117 | |||
| 118 | template<eve::non_empty_product_type PT1, eve::non_empty_product_type PT2, eve::callable_options O> | ||
| 119 | 12 | constexpr auto dot_(POLYFLOAT_DELAY(), O const& o, PT1 f, PT2 s) noexcept | |
| 120 | requires(kumi::as_tuple_t<PT1>::size() == kumi::as_tuple_t<PT2>::size() && !concepts::polyfloat_like<PT1> && | ||
| 121 | !concepts::polyfloat_like<PT2>) | ||
| 122 | { | ||
| 123 | using Tup1 = kumi::as_tuple_t<PT1>; | ||
| 124 | using Tup2 = kumi::as_tuple_t<PT2>; | ||
| 125 | 12 | constexpr auto siz = Tup1::size(); | |
| 126 | if constexpr (siz == 1) | ||
| 127 | { | ||
| 128 | return get<0>(f) * get<0>(s); | ||
| 129 | } | ||
| 130 | else | ||
| 131 | { | ||
| 132 | using r1_t = kumi::apply_traits_t<plf::as_polyfloat_like, Tup1>; | ||
| 133 | using r2_t = kumi::apply_traits_t<plf::as_polyfloat_like, Tup2>; | ||
| 134 | using r_t = plf::as_polyfloat_like_t<r1_t, r2_t>; | ||
| 135 | using u_t = eve::element_type_t<r_t>; | ||
| 136 | 12 | auto cvt = [](auto a) { return plf::convert(a, as<u_t>()); }; | |
| 137 | 36 | return add[o](kumi::map([cvt](auto a, auto b) { return cvt(a * b); }, f, s)); | |
| 138 | } | ||
| 139 | } | ||
| 140 | |||
| 141 | } | ||
| 142 |