Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | /*===---- fmaintrin.h - FMA intrinsics -------------------------------------=== |
| 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 | |||
| 10 | #ifndef __IMMINTRIN_H |
||
| 11 | #error "Never use <fmaintrin.h> directly; include <immintrin.h> instead." |
||
| 12 | #endif |
||
| 13 | |||
| 14 | #ifndef __FMAINTRIN_H |
||
| 15 | #define __FMAINTRIN_H |
||
| 16 | |||
| 17 | /* Define the default attributes for the functions in this file. */ |
||
| 18 | #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("fma"), __min_vector_width__(128))) |
||
| 19 | #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("fma"), __min_vector_width__(256))) |
||
| 20 | |||
| 21 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 22 | _mm_fmadd_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 23 | { |
||
| 24 | return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 25 | } |
||
| 26 | |||
| 27 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 28 | _mm_fmadd_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 29 | { |
||
| 30 | return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 31 | } |
||
| 32 | |||
| 33 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 34 | _mm_fmadd_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 35 | { |
||
| 36 | return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 37 | } |
||
| 38 | |||
| 39 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 40 | _mm_fmadd_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 41 | { |
||
| 42 | return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 43 | } |
||
| 44 | |||
| 45 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 46 | _mm_fmsub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 47 | { |
||
| 48 | return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 49 | } |
||
| 50 | |||
| 51 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 52 | _mm_fmsub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 53 | { |
||
| 54 | return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 55 | } |
||
| 56 | |||
| 57 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 58 | _mm_fmsub_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 59 | { |
||
| 60 | return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 61 | } |
||
| 62 | |||
| 63 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 64 | _mm_fmsub_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 65 | { |
||
| 66 | return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 67 | } |
||
| 68 | |||
| 69 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 70 | _mm_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 71 | { |
||
| 72 | return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 73 | } |
||
| 74 | |||
| 75 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 76 | _mm_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 77 | { |
||
| 78 | return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 79 | } |
||
| 80 | |||
| 81 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 82 | _mm_fnmadd_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 83 | { |
||
| 84 | return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, -(__v4sf)__B, (__v4sf)__C); |
||
| 85 | } |
||
| 86 | |||
| 87 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 88 | _mm_fnmadd_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 89 | { |
||
| 90 | return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, -(__v2df)__B, (__v2df)__C); |
||
| 91 | } |
||
| 92 | |||
| 93 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 94 | _mm_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 95 | { |
||
| 96 | return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 97 | } |
||
| 98 | |||
| 99 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 100 | _mm_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 101 | { |
||
| 102 | return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 103 | } |
||
| 104 | |||
| 105 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 106 | _mm_fnmsub_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 107 | { |
||
| 108 | return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, -(__v4sf)__B, -(__v4sf)__C); |
||
| 109 | } |
||
| 110 | |||
| 111 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 112 | _mm_fnmsub_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 113 | { |
||
| 114 | return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, -(__v2df)__B, -(__v2df)__C); |
||
| 115 | } |
||
| 116 | |||
| 117 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 118 | _mm_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 119 | { |
||
| 120 | return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 121 | } |
||
| 122 | |||
| 123 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 124 | _mm_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 125 | { |
||
| 126 | return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 127 | } |
||
| 128 | |||
| 129 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 130 | _mm_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 131 | { |
||
| 132 | return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 133 | } |
||
| 134 | |||
| 135 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 136 | _mm_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 137 | { |
||
| 138 | return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 139 | } |
||
| 140 | |||
| 141 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 142 | _mm256_fmadd_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 143 | { |
||
| 144 | return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 145 | } |
||
| 146 | |||
| 147 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 148 | _mm256_fmadd_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 149 | { |
||
| 150 | return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 151 | } |
||
| 152 | |||
| 153 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 154 | _mm256_fmsub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 155 | { |
||
| 156 | return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 157 | } |
||
| 158 | |||
| 159 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 160 | _mm256_fmsub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 161 | { |
||
| 162 | return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 163 | } |
||
| 164 | |||
| 165 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 166 | _mm256_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 167 | { |
||
| 168 | return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 169 | } |
||
| 170 | |||
| 171 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 172 | _mm256_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 173 | { |
||
| 174 | return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 175 | } |
||
| 176 | |||
| 177 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 178 | _mm256_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 179 | { |
||
| 180 | return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 181 | } |
||
| 182 | |||
| 183 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 184 | _mm256_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 185 | { |
||
| 186 | return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 187 | } |
||
| 188 | |||
| 189 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 190 | _mm256_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 191 | { |
||
| 192 | return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 193 | } |
||
| 194 | |||
| 195 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 196 | _mm256_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 197 | { |
||
| 198 | return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 199 | } |
||
| 200 | |||
| 201 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 202 | _mm256_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 203 | { |
||
| 204 | return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 205 | } |
||
| 206 | |||
| 207 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 208 | _mm256_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 209 | { |
||
| 210 | return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 211 | } |
||
| 212 | |||
| 213 | #undef __DEFAULT_FN_ATTRS128 |
||
| 214 | #undef __DEFAULT_FN_ATTRS256 |
||
| 215 | |||
| 216 | #endif /* __FMAINTRIN_H */ |