1 | //===----------------------------------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #ifndef _LIBCPP___ALGORITHM_SIMD_UTILS_H |
10 | #define _LIBCPP___ALGORITHM_SIMD_UTILS_H |
11 | |
12 | #include <__algorithm/min.h> |
13 | #include <__bit/bit_cast.h> |
14 | #include <__bit/countl.h> |
15 | #include <__bit/countr.h> |
16 | #include <__config> |
17 | #include <__type_traits/is_arithmetic.h> |
18 | #include <__type_traits/is_same.h> |
19 | #include <__utility/integer_sequence.h> |
20 | #include <cstddef> |
21 | #include <cstdint> |
22 | |
23 | #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER) |
24 | # pragma GCC system_header |
25 | #endif |
26 | |
27 | _LIBCPP_PUSH_MACROS |
28 | #include <__undef_macros> |
29 | |
30 | // TODO: Find out how altivec changes things and allow vectorizations there too. |
31 | #if _LIBCPP_STD_VER >= 14 && defined(_LIBCPP_CLANG_VER) && !defined(__ALTIVEC__) |
32 | # define _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS 1 |
33 | #else |
34 | # define _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS 0 |
35 | #endif |
36 | |
37 | #if _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS && !defined(__OPTIMIZE_SIZE__) |
38 | # define _LIBCPP_VECTORIZE_ALGORITHMS 1 |
39 | #else |
40 | # define _LIBCPP_VECTORIZE_ALGORITHMS 0 |
41 | #endif |
42 | |
43 | #if _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS |
44 | |
45 | _LIBCPP_BEGIN_NAMESPACE_STD |
46 | |
47 | template <class _Tp> |
48 | inline constexpr bool __can_map_to_integer_v = |
49 | sizeof(_Tp) == alignof(_Tp) && (sizeof(_Tp) == 1 || sizeof(_Tp) == 2 || sizeof(_Tp) == 4 || sizeof(_Tp) == 8); |
50 | |
51 | template <size_t _TypeSize> |
52 | struct __get_as_integer_type_impl; |
53 | |
54 | template <> |
55 | struct __get_as_integer_type_impl<1> { |
56 | using type = uint8_t; |
57 | }; |
58 | |
59 | template <> |
60 | struct __get_as_integer_type_impl<2> { |
61 | using type = uint16_t; |
62 | }; |
63 | template <> |
64 | struct __get_as_integer_type_impl<4> { |
65 | using type = uint32_t; |
66 | }; |
67 | template <> |
68 | struct __get_as_integer_type_impl<8> { |
69 | using type = uint64_t; |
70 | }; |
71 | |
72 | template <class _Tp> |
73 | using __get_as_integer_type_t = typename __get_as_integer_type_impl<sizeof(_Tp)>::type; |
74 | |
75 | // This isn't specialized for 64 byte vectors on purpose. They have the potential to significantly reduce performance |
76 | // in mixed simd/non-simd workloads and don't provide any performance improvement for currently vectorized algorithms |
77 | // as far as benchmarks are concerned. |
78 | # if defined(__AVX__) || defined(__MVS__) |
79 | template <class _Tp> |
80 | inline constexpr size_t __native_vector_size = 32 / sizeof(_Tp); |
81 | # elif defined(__SSE__) || defined(__ARM_NEON__) |
82 | template <class _Tp> |
83 | inline constexpr size_t __native_vector_size = 16 / sizeof(_Tp); |
84 | # elif defined(__MMX__) |
85 | template <class _Tp> |
86 | inline constexpr size_t __native_vector_size = 8 / sizeof(_Tp); |
87 | # else |
88 | template <class _Tp> |
89 | inline constexpr size_t __native_vector_size = 1; |
90 | # endif |
91 | |
92 | template <class _ArithmeticT, size_t _Np> |
93 | using __simd_vector __attribute__((__ext_vector_type__(_Np))) = _ArithmeticT; |
94 | |
95 | template <class _VecT> |
96 | inline constexpr size_t __simd_vector_size_v = []<bool _False = false>() -> size_t { |
97 | static_assert(_False, "Not a vector!" ); |
98 | }(); |
99 | |
100 | template <class _Tp, size_t _Np> |
101 | inline constexpr size_t __simd_vector_size_v<__simd_vector<_Tp, _Np>> = _Np; |
102 | |
103 | template <class _Tp, size_t _Np> |
104 | _LIBCPP_HIDE_FROM_ABI _Tp __simd_vector_underlying_type_impl(__simd_vector<_Tp, _Np>) { |
105 | return _Tp{}; |
106 | } |
107 | |
108 | template <class _VecT> |
109 | using __simd_vector_underlying_type_t = decltype(std::__simd_vector_underlying_type_impl(_VecT{})); |
110 | |
111 | // This isn't inlined without always_inline when loading chars. |
112 | template <class _VecT, class _Iter> |
113 | _LIBCPP_NODISCARD _LIBCPP_ALWAYS_INLINE _LIBCPP_HIDE_FROM_ABI _VecT __load_vector(_Iter __iter) noexcept { |
114 | return [=]<size_t... _Indices>(index_sequence<_Indices...>) _LIBCPP_ALWAYS_INLINE noexcept { |
115 | return _VecT{__iter[_Indices]...}; |
116 | }(make_index_sequence<__simd_vector_size_v<_VecT>>{}); |
117 | } |
118 | |
119 | template <class _Tp, size_t _Np> |
120 | _LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI bool __all_of(__simd_vector<_Tp, _Np> __vec) noexcept { |
121 | return __builtin_reduce_and(__builtin_convertvector(__vec, __simd_vector<bool, _Np>)); |
122 | } |
123 | |
124 | template <class _Tp, size_t _Np> |
125 | _LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI size_t __find_first_set(__simd_vector<_Tp, _Np> __vec) noexcept { |
126 | using __mask_vec = __simd_vector<bool, _Np>; |
127 | |
128 | // This has MSan disabled du to https://github.com/llvm/llvm-project/issues/85876 |
129 | auto __impl = [&]<class _MaskT>(_MaskT) _LIBCPP_NO_SANITIZE("memory" ) noexcept { |
130 | # if defined(_LIBCPP_BIG_ENDIAN) |
131 | return std::min<size_t>( |
132 | _Np, std::__countl_zero(__builtin_bit_cast(_MaskT, __builtin_convertvector(__vec, __mask_vec)))); |
133 | # else |
134 | return std::min<size_t>( |
135 | _Np, std::__countr_zero(__builtin_bit_cast(_MaskT, __builtin_convertvector(__vec, __mask_vec)))); |
136 | # endif |
137 | }; |
138 | |
139 | if constexpr (sizeof(__mask_vec) == sizeof(uint8_t)) { |
140 | return __impl(uint8_t{}); |
141 | } else if constexpr (sizeof(__mask_vec) == sizeof(uint16_t)) { |
142 | return __impl(uint16_t{}); |
143 | } else if constexpr (sizeof(__mask_vec) == sizeof(uint32_t)) { |
144 | return __impl(uint32_t{}); |
145 | } else if constexpr (sizeof(__mask_vec) == sizeof(uint64_t)) { |
146 | return __impl(uint64_t{}); |
147 | } else { |
148 | static_assert(sizeof(__mask_vec) == 0, "unexpected required size for mask integer type" ); |
149 | return 0; |
150 | } |
151 | } |
152 | |
153 | template <class _Tp, size_t _Np> |
154 | _LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI size_t __find_first_not_set(__simd_vector<_Tp, _Np> __vec) noexcept { |
155 | return std::__find_first_set(~__vec); |
156 | } |
157 | |
158 | _LIBCPP_END_NAMESPACE_STD |
159 | |
160 | #endif // _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS |
161 | |
162 | _LIBCPP_POP_MACROS |
163 | |
164 | #endif // _LIBCPP___ALGORITHM_SIMD_UTILS_H |
165 | |