Intel xvec
Experimental C++ simd class library
Loading...
Searching...
No Matches
math.hpp
1//===----------------------------------------------------------------------===//
2//
3// Copyright (C) 2021 Intel Corporation
4//
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#pragma once
10
11#include <xvec/detail/config.hpp>
12
17
18#define _XVEC_SSE_UNARY_MATHS_OP(NAME) [](__m128 v) { return _mm_ ##NAME ##_ps (v); }, [](__m128d v) { return _mm_ ##NAME ##_pd (v); },
19#define _XVEC_SSE_BINARY_MATHS_OP(NAME) [](__m128 v0, __m128 v1) { return _mm_ ##NAME ##_ps (v0, v1); }, [](__m128d v0, __m128d v1) { return _mm_ ##NAME ##_pd (v0, v1); },
20
21#if defined (__AVX__)
22#define _XVEC_AVX_UNARY_MATHS_OP(NAME) [](__m256 v) { return _mm256_ ##NAME ##_ps (v); }, [](__m256d v) { return _mm256_ ##NAME ##_pd (v); },
23#define _XVEC_AVX_BINARY_MATHS_OP(NAME) [](__m256 v0, __m256 v1) { return _mm256_ ##NAME ##_ps (v0, v1); }, [](__m256d v0, __m256d v1) { return _mm256_ ##NAME ##_pd (v0, v1); },
24#else
25#define _XVEC_AVX_UNARY_MATHS_OP(NAME)
26#define _XVEC_AVX_BINARY_MATHS_OP(NAME)
27#endif
28
29#if defined (__AVX512F__)
30#define _XVEC_AVX512_UNARY_MATHS_OP(NAME) [](__m512 v) { return _mm512_ ##NAME ##_ps (v); }, [](__m512d v) { return _mm512_ ##NAME ##_pd (v); },
31#define _XVEC_AVX512_BINARY_MATHS_OP(NAME) [](__m512 v0, __m512 v1) { return _mm512_ ##NAME ##_ps (v0, v1); }, [](__m512d v0, __m512d v1) { return _mm512_ ##NAME ##_pd (v0, v1); },
32#else
33#define _XVEC_AVX512_UNARY_MATHS_OP(NAME)
34#define _XVEC_AVX512_BINARY_MATHS_OP(NAME)
35#endif
36
37#if defined (__AVX512FP16__)
38 #define _XVEC_AVX512FP16_UNARY_MATHS_OP(NAME) [](__m128h v) { return _mm_ ##NAME ##_ph (v); }, [](__m256h v) { return _mm256_ ##NAME ##_ph (v); }, [](__m512h v) { return _mm512_ ##NAME ##_ph (v); },
39 #define _XVEC_AVX512FP16_BINARY_MATHS_OP(NAME) [](__m128h v0, __m128h v1) { return _mm_ ##NAME ##_ph (v0, v1); }, [](__m256h v0, __m256h v1) { return _mm256_ ##NAME ##_ph (v0, v1); }, [](__m512h v0, __m512h v1) { return _mm512_ ##NAME ##_ph (v0, v1); },
40#else
41#define _XVEC_AVX512FP16_UNARY_MATHS_OP(NAME)
42#define _XVEC_AVX512FP16_BINARY_MATHS_OP(NAME)
43#endif
44
45// The following creates a lambda inside detail to provide overloads for
46// all the interesting intrinsics, and then sets up the actual xvec interface
47// function to invoke it.
48#define _XVEC_UNARY_MATHS_OP(NAME) \
49 template<_XVEC_NAMESPACE::simd::detail::math_floating_point _Vp> \
50 constexpr _XVEC_NAMESPACE::simd::detail::deduced_vec_t<_Vp> NAME (const _Vp& v) { \
51 _XVEC_NAMESPACE::simd::detail::target_overloads op { \
52 _XVEC_SSE_UNARY_MATHS_OP(NAME) \
53 _XVEC_AVX_UNARY_MATHS_OP(NAME) \
54 _XVEC_AVX512_UNARY_MATHS_OP(NAME) \
55 _XVEC_AVX512FP16_UNARY_MATHS_OP(NAME) \
56 [](auto unhandled) { static_assert(_XVEC_NAMESPACE::simd::detail::dependent_false<decltype(unhandled)>); return unhandled;} \
57 }; \
58 return _XVEC_NAMESPACE::simd::detail::maths_fn(v, op); \
59 }
60
61#define _XVEC_BINARY_MATHS_OP(NAME) \
62 template<_XVEC_NAMESPACE::simd::detail::math_floating_point _Vp> \
63 constexpr auto NAME (const _Vp& v0, const _Vp& v1) { \
64 _XVEC_NAMESPACE::simd::detail::target_overloads op { \
65 _XVEC_SSE_BINARY_MATHS_OP(NAME) \
66 _XVEC_AVX_BINARY_MATHS_OP(NAME) \
67 _XVEC_AVX512_BINARY_MATHS_OP(NAME) \
68 _XVEC_AVX512FP16_BINARY_MATHS_OP(NAME) \
69 [](auto unhandled, auto) { static_assert(_XVEC_NAMESPACE::simd::detail::dependent_false<decltype(unhandled)>); return unhandled;} \
70 }; \
71 return _XVEC_NAMESPACE::simd::detail::maths_fn(v0, v1, op); \
72 }
73
74namespace _XVEC_NAMESPACE::simd::detail
75{
85template<math_floating_point _Vp, typename OP_IMPL>
86constexpr deduced_vec_t<_Vp> maths_fn(const _Vp& v, OP_IMPL op_impl)
87{
88 using _Tp = typename _Vp::value_type;
89 if constexpr (detail::is_fp16_v<_Tp> && !detail::hasAvxFp16)
90 // Emulating FP16 on a non-native machine uses float to do the work.
91 return rebind_cast<_Tp>(maths_fn(rebind_cast<float>(v), op_impl));
92 else
93 {
94 // All other floating point types (including native FP16) get dispatched to the
95 // appropriate intrinsic.
96 auto wrapper = [=]<typename _Vec>(_Vec x) { return _Vec(op_impl(x.to_register())); };
97 return chunked_invoke(wrapper, v);
98 }
99}
100
101template<math_floating_point _Vp, typename OP_IMPL>
102constexpr deduced_vec_t<_Vp> maths_fn(const _Vp& v0, const _Vp& v1, OP_IMPL op_impl)
103{
104 using _Tp = typename _Vp::value_type;
105 if constexpr (_XVEC_NAMESPACE::simd::detail::is_fp16_v<_Tp> && !_XVEC_NAMESPACE::simd::detail::hasAvxFp16)
106 // Emulating FP16 on a non-native machine uses float to do the work.
107 return rebind_cast<_Tp>(maths_fn(rebind_cast<float>(v0), rebind_cast<float>(v1), op_impl));
108 else
109 {
110 // All other floating point types (including native FP16) get dispatched to the
111 // appropriate intrinsic.
112 auto wrapper = [=]<typename _Vec>(_Vec x, _Vec y) { return _Vec(op_impl(x.to_register(), y.to_register())); };
113 return chunked_invoke(wrapper, v0, v1);
114 }
115}
116
117} // namespace _XVEC_NAMESPACE::simd::detail
constexpr auto chunked_invoke(_Fn fn, const _Arg0 &arg0, const _Args &... args)
Invoke a callable on corresponding vector or mask chunks.
Definition invoke.hpp:61
constexpr detail::RebindCast< _Up > rebind_cast
Rebinding CPO.
Definition std_rebind.hpp:55
constexpr bool is_fp16_v
Definition utilities.hpp:90