SeqAn3 3.1.0
The Modern C++ library for sequence analysis.
simd_algorithm_avx512.hpp
Go to the documentation of this file.
1// -----------------------------------------------------------------------------------------------------
2// Copyright (c) 2006-2021, Knut Reinert & Freie Universität Berlin
3// Copyright (c) 2016-2021, Knut Reinert & MPI für molekulare Genetik
4// This file may be used, modified and/or redistributed under the terms of the 3-clause BSD-License
5// shipped with this file and also available at: https://github.com/seqan/seqan3/blob/master/LICENSE.md
6// -----------------------------------------------------------------------------------------------------
7
13#pragma once
14
15#include <array>
16
17#include <seqan3/utility/simd/concept.hpp>
18#include <seqan3/utility/simd/detail/builtin_simd_intrinsics.hpp>
19#include <seqan3/utility/simd/detail/builtin_simd.hpp>
20#include <seqan3/utility/simd/simd_traits.hpp>
21
22//-----------------------------------------------------------------------------
23// forward declare avx512 simd algorithms that use avx512 intrinsics
24//-----------------------------------------------------------------------------
25
26namespace seqan3::detail
27{
31template <simd::simd_concept simd_t>
32constexpr simd_t load_avx512(void const * mem_addr);
33
37template <simd::simd_concept simd_t>
38constexpr void store_avx512(void * mem_addr, simd_t const & simd_vec);
39
43template <simd::simd_concept simd_t>
44inline void transpose_matrix_avx512(std::array<simd_t, simd_traits<simd_t>::length> & matrix);
45
49template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
50constexpr target_simd_t upcast_signed_avx512(source_simd_t const & src);
51
55template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
56constexpr target_simd_t upcast_unsigned_avx512(source_simd_t const & src);
57
61template <uint8_t index, simd::simd_concept simd_t>
62constexpr simd_t extract_half_avx512(simd_t const & src);
63
67template <uint8_t index, simd::simd_concept simd_t>
68constexpr simd_t extract_quarter_avx512(simd_t const & src);
69
73template <uint8_t index, simd::simd_concept simd_t>
74constexpr simd_t extract_eighth_avx512(simd_t const & src);
75
76}
77
78//-----------------------------------------------------------------------------
79// implementation
80//-----------------------------------------------------------------------------
81
82#ifdef __AVX512F__
83
84namespace seqan3::detail
85{
86
87template <simd::simd_concept simd_t>
88constexpr simd_t load_avx512(void const * mem_addr)
89{
90 return reinterpret_cast<simd_t>(_mm512_loadu_si512(mem_addr));
91}
92
93template <simd::simd_concept simd_t>
94constexpr void store_avx512(void * mem_addr, simd_t const & simd_vec)
95{
96 _mm512_storeu_si512(mem_addr, reinterpret_cast<__m512i const &>(simd_vec));
97}
98
99// TODO: not implemented and used yet, if you implement it don't forget to add it to seqan3::simd::transpose
100template <simd::simd_concept simd_t>
101inline void transpose_matrix_avx512(std::array<simd_t, simd_traits<simd_t>::length> & matrix);
102
103template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
104constexpr target_simd_t upcast_signed_avx512(source_simd_t const & src)
105{
106 __m512i const & tmp = reinterpret_cast<__m512i const &>(src);
107 if constexpr (simd_traits<source_simd_t>::length == 64) // cast from epi8 ...
108 {
109 if constexpr (simd_traits<target_simd_t>::length == 32) // to epi16
110 return reinterpret_cast<target_simd_t>(_mm512_cvtepi8_epi16(_mm512_castsi512_si256(tmp)));
111 if constexpr (simd_traits<target_simd_t>::length == 16) // to epi32
112 return reinterpret_cast<target_simd_t>(_mm512_cvtepi8_epi32(_mm512_castsi512_si128(tmp)));
113 if constexpr (simd_traits<target_simd_t>::length == 8) // to epi64
114 return reinterpret_cast<target_simd_t>(_mm512_cvtepi8_epi64(_mm512_castsi512_si128(tmp)));
115 }
116 else if constexpr (simd_traits<source_simd_t>::length == 32) // cast from epi16 ...
117 {
118 if constexpr (simd_traits<target_simd_t>::length == 16) // to epi32
119 return reinterpret_cast<target_simd_t>(_mm512_cvtepi16_epi32(_mm512_castsi512_si256(tmp)));
120 if constexpr (simd_traits<target_simd_t>::length == 8) // to epi64
121 return reinterpret_cast<target_simd_t>(_mm512_cvtepi16_epi64(_mm512_castsi512_si128(tmp)));
122 }
123 else // cast from epi32 to epi64
124 {
125 static_assert(simd_traits<source_simd_t>::length == 16, "Expected 32 bit scalar type.");
126 return reinterpret_cast<target_simd_t>(_mm512_cvtepi32_epi64(_mm512_castsi512_si256(tmp)));
127 }
128}
129
130template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
131constexpr target_simd_t upcast_unsigned_avx512(source_simd_t const & src)
132{
133 __m512i const & tmp = reinterpret_cast<__m512i const &>(src);
134 if constexpr (simd_traits<source_simd_t>::length == 64) // cast from epi8 ...
135 {
136 if constexpr (simd_traits<target_simd_t>::length == 32) // to epi16
137 return reinterpret_cast<target_simd_t>(_mm512_cvtepu8_epi16(_mm512_castsi512_si256(tmp)));
138 if constexpr (simd_traits<target_simd_t>::length == 16) // to epi32
139 return reinterpret_cast<target_simd_t>(_mm512_cvtepu8_epi32(_mm512_castsi512_si128(tmp)));
140 if constexpr (simd_traits<target_simd_t>::length == 8) // to epi64
141 return reinterpret_cast<target_simd_t>(_mm512_cvtepu8_epi64(_mm512_castsi512_si128(tmp)));
142 }
143 else if constexpr (simd_traits<source_simd_t>::length == 32) // cast from epi16 ...
144 {
145 if constexpr (simd_traits<target_simd_t>::length == 16) // to epi32
146 return reinterpret_cast<target_simd_t>(_mm512_cvtepu16_epi32(_mm512_castsi512_si256(tmp)));
147 if constexpr (simd_traits<target_simd_t>::length == 8) // to epi64
148 return reinterpret_cast<target_simd_t>(_mm512_cvtepu16_epi64(_mm512_castsi512_si128(tmp)));
149 }
150 else // cast from epi32 to epi64
151 {
152 static_assert(simd_traits<source_simd_t>::length == 16, "Expected 32 bit scalar type.");
153 return reinterpret_cast<target_simd_t>(_mm512_cvtepu32_epi64(_mm512_castsi512_si256(tmp)));
154 }
155}
156
157// TODO: not implemented and used yet, if you implement it don't forget to add it to seqan3::detail::extract_half
158template <uint8_t index, simd::simd_concept simd_t>
159constexpr simd_t extract_half_avx512(simd_t const & src);
160
161// TODO: not implemented and used yet, if you implement it don't forget to add it to seqan3::detail::extract_quarter
162template <uint8_t index, simd::simd_concept simd_t>
163constexpr simd_t extract_quarter_avx512(simd_t const & src);
164
165// TODO: not implemented and used yet, if you implement it don't forget to add it to seqan3::detail::extract_eighth
166template <uint8_t index, simd::simd_concept simd_t>
167constexpr simd_t extract_eighth_avx512(simd_t const & src);
168
169} // namespace seqan3::detail
170
171#endif // __AVX512F__
builtin_simd.hpp
Provides seqan3::detail::builtin_simd, seqan3::detail::is_builtin_simd and seqan3::simd::simd_traits<...
builtin_simd_intrinsics.hpp
Provides intrinsics include for builtin simd.
simd_traits.hpp
Provides seqan3::simd::simd_traits.
concept.hpp
Provides seqan3::simd::simd_concept.