experimental/__simd/vec_ext.h

0001 // -*- C++ -*-
0002 //===----------------------------------------------------------------------===//
0003 //
0004 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
0005 // See https://llvm.org/LICENSE.txt for license information.
0006 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
0007 //
0008 //===----------------------------------------------------------------------===//
0009
0010 #ifndef _LIBCPP_EXPERIMENTAL___SIMD_VEC_EXT_H
0011 #define _LIBCPP_EXPERIMENTAL___SIMD_VEC_EXT_H
0012
0013 #include <__assert>
0014 #include <__bit/bit_ceil.h>
0015 #include <__config>
0016 #include <__cstddef/size_t.h>
0017 #include <__type_traits/integral_constant.h>
0018 #include <__utility/forward.h>
0019 #include <__utility/integer_sequence.h>
0020 #include <experimental/__simd/declaration.h>
0021 #include <experimental/__simd/traits.h>
0022 #include <experimental/__simd/utility.h>
0023
0024 #if _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
0025
0026 _LIBCPP_BEGIN_NAMESPACE_EXPERIMENTAL
0027 inline namespace parallelism_v2 {
0028 namespace simd_abi {
0029 template <int _Np>
0030 struct __vec_ext {
0031   static constexpr size_t __simd_size = _Np;
0032 };
0033 } // namespace simd_abi
0034
0035 template <int _Np>
0036 inline constexpr bool is_abi_tag_v<simd_abi::__vec_ext<_Np>> = _Np > 0 && _Np <= 32;
0037
0038 template <class _Tp, int _Np>
0039 struct __simd_storage<_Tp, simd_abi::__vec_ext<_Np>> {
0040   _Tp __data __attribute__((__vector_size__(std::__bit_ceil((sizeof(_Tp) * _Np)))));
0041
0042   _LIBCPP_HIDE_FROM_ABI _Tp __get(size_t __idx) const noexcept {
0043     _LIBCPP_ASSERT_VALID_ELEMENT_ACCESS(__idx < _Np, "Index is out of bounds");
0044     return __data[__idx];
0045   }
0046   _LIBCPP_HIDE_FROM_ABI void __set(size_t __idx, _Tp __v) noexcept {
0047     _LIBCPP_ASSERT_VALID_ELEMENT_ACCESS(__idx < _Np, "Index is out of bounds");
0048     __data[__idx] = __v;
0049   }
0050 };
0051
0052 template <class _Tp, int _Np>
0053 struct __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>
0054     : __simd_storage<decltype(experimental::__choose_mask_type<_Tp>()), simd_abi::__vec_ext<_Np>> {};
0055
0056 template <class _Tp, int _Np>
0057 struct __simd_operations<_Tp, simd_abi::__vec_ext<_Np>> {
0058   using _SimdStorage _LIBCPP_NODEBUG = __simd_storage<_Tp, simd_abi::__vec_ext<_Np>>;
0059   using _MaskStorage _LIBCPP_NODEBUG = __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>;
0060
0061   static _LIBCPP_HIDE_FROM_ABI _SimdStorage __broadcast(_Tp __v) noexcept {
0062     _SimdStorage __result;
0063     for (int __i = 0; __i < _Np; ++__i) {
0064       __result.__set(__i, __v);
0065     }
0066     return __result;
0067   }
0068
0069   template <class _Generator, size_t... _Is>
0070   static _LIBCPP_HIDE_FROM_ABI _SimdStorage __generate_init(_Generator&& __g, std::index_sequence<_Is...>) {
0071     return _SimdStorage{{__g(std::integral_constant<size_t, _Is>())...}};
0072   }
0073
0074   template <class _Generator>
0075   static _LIBCPP_HIDE_FROM_ABI _SimdStorage __generate(_Generator&& __g) noexcept {
0076     return __generate_init(std::forward<_Generator>(__g), std::make_index_sequence<_Np>());
0077   }
0078
0079   template <class _Up>
0080   static _LIBCPP_HIDE_FROM_ABI void __load(_SimdStorage& __s, const _Up* __mem) noexcept {
0081     for (size_t __i = 0; __i < _Np; __i++)
0082       __s.__data[__i] = static_cast<_Tp>(__mem[__i]);
0083   }
0084
0085   template <class _Up>
0086   static _LIBCPP_HIDE_FROM_ABI void __store(_SimdStorage __s, _Up* __mem) noexcept {
0087     for (size_t __i = 0; __i < _Np; __i++)
0088       __mem[__i] = static_cast<_Up>(__s.__data[__i]);
0089   }
0090
0091   static _LIBCPP_HIDE_FROM_ABI void __increment(_SimdStorage& __s) noexcept { __s.__data = __s.__data + 1; }
0092
0093   static _LIBCPP_HIDE_FROM_ABI void __decrement(_SimdStorage& __s) noexcept { __s.__data = __s.__data - 1; }
0094
0095   static _LIBCPP_HIDE_FROM_ABI _MaskStorage __negate(_SimdStorage __s) noexcept { return {!__s.__data}; }
0096
0097   static _LIBCPP_HIDE_FROM_ABI _SimdStorage __bitwise_not(_SimdStorage __s) noexcept { return {~__s.__data}; }
0098
0099   static _LIBCPP_HIDE_FROM_ABI _SimdStorage __unary_minus(_SimdStorage __s) noexcept { return {-__s.__data}; }
0100 };
0101
0102 template <class _Tp, int _Np>
0103 struct __mask_operations<_Tp, simd_abi::__vec_ext<_Np>> {
0104   using _MaskStorage _LIBCPP_NODEBUG = __mask_storage<_Tp, simd_abi::__vec_ext<_Np>>;
0105
0106   static _LIBCPP_HIDE_FROM_ABI _MaskStorage __broadcast(bool __v) noexcept {
0107     _MaskStorage __result;
0108     auto __all_bits_v = experimental::__set_all_bits<_Tp>(__v);
0109     for (int __i = 0; __i < _Np; ++__i) {
0110       __result.__set(__i, __all_bits_v);
0111     }
0112     return __result;
0113   }
0114
0115   static _LIBCPP_HIDE_FROM_ABI void __load(_MaskStorage& __s, const bool* __mem) noexcept {
0116     for (size_t __i = 0; __i < _Np; __i++)
0117       __s.__data[__i] = experimental::__set_all_bits<_Tp>(__mem[__i]);
0118   }
0119
0120   static _LIBCPP_HIDE_FROM_ABI void __store(_MaskStorage __s, bool* __mem) noexcept {
0121     for (size_t __i = 0; __i < _Np; __i++)
0122       __mem[__i] = static_cast<bool>(__s.__data[__i]);
0123   }
0124 };
0125
0126 } // namespace parallelism_v2
0127 _LIBCPP_END_NAMESPACE_EXPERIMENTAL
0128
0129 #endif // _LIBCPP_STD_VER >= 17 && defined(_LIBCPP_ENABLE_EXPERIMENTAL)
0130 #endif // _LIBCPP_EXPERIMENTAL___SIMD_VEC_EXT_H