Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | /*===---- fma4intrin.h - FMA4 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 __X86INTRIN_H |
||
| 11 | #error "Never use <fma4intrin.h> directly; include <x86intrin.h> instead." |
||
| 12 | #endif |
||
| 13 | |||
| 14 | #ifndef __FMA4INTRIN_H |
||
| 15 | #define __FMA4INTRIN_H |
||
| 16 | |||
| 17 | #include <pmmintrin.h> |
||
| 18 | |||
| 19 | /* Define the default attributes for the functions in this file. */ |
||
| 20 | #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("fma4"), __min_vector_width__(128))) |
||
| 21 | #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("fma4"), __min_vector_width__(256))) |
||
| 22 | |||
| 23 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 24 | _mm_macc_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 25 | { |
||
| 26 | return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 27 | } |
||
| 28 | |||
| 29 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 30 | _mm_macc_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 31 | { |
||
| 32 | return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 33 | } |
||
| 34 | |||
| 35 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 36 | _mm_macc_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 37 | { |
||
| 38 | return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 39 | } |
||
| 40 | |||
| 41 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 42 | _mm_macc_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 43 | { |
||
| 44 | return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 45 | } |
||
| 46 | |||
| 47 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 48 | _mm_msub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 49 | { |
||
| 50 | return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 51 | } |
||
| 52 | |||
| 53 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 54 | _mm_msub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 55 | { |
||
| 56 | return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 57 | } |
||
| 58 | |||
| 59 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 60 | _mm_msub_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 61 | { |
||
| 62 | return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 63 | } |
||
| 64 | |||
| 65 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 66 | _mm_msub_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 67 | { |
||
| 68 | return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 69 | } |
||
| 70 | |||
| 71 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 72 | _mm_nmacc_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 73 | { |
||
| 74 | return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 75 | } |
||
| 76 | |||
| 77 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 78 | _mm_nmacc_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 79 | { |
||
| 80 | return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 81 | } |
||
| 82 | |||
| 83 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 84 | _mm_nmacc_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 85 | { |
||
| 86 | return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 87 | } |
||
| 88 | |||
| 89 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 90 | _mm_nmacc_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 91 | { |
||
| 92 | return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 93 | } |
||
| 94 | |||
| 95 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 96 | _mm_nmsub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 97 | { |
||
| 98 | return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 99 | } |
||
| 100 | |||
| 101 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 102 | _mm_nmsub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 103 | { |
||
| 104 | return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 105 | } |
||
| 106 | |||
| 107 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 108 | _mm_nmsub_ss(__m128 __A, __m128 __B, __m128 __C) |
||
| 109 | { |
||
| 110 | return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 111 | } |
||
| 112 | |||
| 113 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 114 | _mm_nmsub_sd(__m128d __A, __m128d __B, __m128d __C) |
||
| 115 | { |
||
| 116 | return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 117 | } |
||
| 118 | |||
| 119 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 120 | _mm_maddsub_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 121 | { |
||
| 122 | return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); |
||
| 123 | } |
||
| 124 | |||
| 125 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 126 | _mm_maddsub_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 127 | { |
||
| 128 | return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, (__v2df)__C); |
||
| 129 | } |
||
| 130 | |||
| 131 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
||
| 132 | _mm_msubadd_ps(__m128 __A, __m128 __B, __m128 __C) |
||
| 133 | { |
||
| 134 | return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); |
||
| 135 | } |
||
| 136 | |||
| 137 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
||
| 138 | _mm_msubadd_pd(__m128d __A, __m128d __B, __m128d __C) |
||
| 139 | { |
||
| 140 | return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, -(__v2df)__C); |
||
| 141 | } |
||
| 142 | |||
| 143 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 144 | _mm256_macc_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 145 | { |
||
| 146 | return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 147 | } |
||
| 148 | |||
| 149 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 150 | _mm256_macc_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 151 | { |
||
| 152 | return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 153 | } |
||
| 154 | |||
| 155 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 156 | _mm256_msub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 157 | { |
||
| 158 | return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 159 | } |
||
| 160 | |||
| 161 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 162 | _mm256_msub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 163 | { |
||
| 164 | return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 165 | } |
||
| 166 | |||
| 167 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 168 | _mm256_nmacc_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 169 | { |
||
| 170 | return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 171 | } |
||
| 172 | |||
| 173 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 174 | _mm256_nmacc_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 175 | { |
||
| 176 | return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 177 | } |
||
| 178 | |||
| 179 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 180 | _mm256_nmsub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 181 | { |
||
| 182 | return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 183 | } |
||
| 184 | |||
| 185 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 186 | _mm256_nmsub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 187 | { |
||
| 188 | return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 189 | } |
||
| 190 | |||
| 191 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 192 | _mm256_maddsub_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 193 | { |
||
| 194 | return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C); |
||
| 195 | } |
||
| 196 | |||
| 197 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 198 | _mm256_maddsub_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 199 | { |
||
| 200 | return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, (__v4df)__C); |
||
| 201 | } |
||
| 202 | |||
| 203 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 204 | _mm256_msubadd_ps(__m256 __A, __m256 __B, __m256 __C) |
||
| 205 | { |
||
| 206 | return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C); |
||
| 207 | } |
||
| 208 | |||
| 209 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 210 | _mm256_msubadd_pd(__m256d __A, __m256d __B, __m256d __C) |
||
| 211 | { |
||
| 212 | return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C); |
||
| 213 | } |
||
| 214 | |||
| 215 | #undef __DEFAULT_FN_ATTRS128 |
||
| 216 | #undef __DEFAULT_FN_ATTRS256 |
||
| 217 | |||
| 218 | #endif /* __FMA4INTRIN_H */ |