1 /* Copyright (C) 2019-2022 Free Software Foundation, Inc. 2 3 This file is part of GCC. 4 5 GCC is free software; you can redistribute it and/or modify 6 it under the terms of the GNU General Public License as published by 7 the Free Software Foundation; either version 3, or (at your option) 8 any later version. 9 10 GCC is distributed in the hope that it will be useful, 11 but WITHOUT ANY WARRANTY; without even the implied warranty of 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 GNU General Public License for more details. 14 15 Under Section 7 of GPL version 3, you are granted additional 16 permissions described in the GCC Runtime Library Exception, version 17 3.1, as published by the Free Software Foundation. 18 19 You should have received a copy of the GNU General Public License and 20 a copy of the GCC Runtime Library Exception along with this program; 21 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 22 <http://www.gnu.org/licenses/>. */ 23 24 #ifndef _IMMINTRIN_H_INCLUDED 25 #error "Never use <avx512bf16vlintrin.h> directly; include <immintrin.h> instead." 26 #endif 27 28 #ifndef _AVX512BF16VLINTRIN_H_INCLUDED 29 #define _AVX512BF16VLINTRIN_H_INCLUDED 30 31 #if !defined(__AVX512VL__) || !defined(__AVX512BF16__) 32 #pragma GCC push_options 33 #pragma GCC target("avx512bf16,avx512vl") 34 #define __DISABLE_AVX512BF16VL__ 35 #endif /* __AVX512BF16__ */ 36 37 /* Internal data types for implementing the intrinsics. */ 38 typedef short __v16bh __attribute__ ((__vector_size__ (32))); 39 typedef short __v8bh __attribute__ ((__vector_size__ (16))); 40 41 /* The Intel API is flexible enough that we must allow aliasing with other 42 vector types, and their scalar components. */ 43 typedef short __m256bh __attribute__ ((__vector_size__ (32), __may_alias__)); 44 typedef short __m128bh __attribute__ ((__vector_size__ (16), __may_alias__)); 45 46 typedef unsigned short __bfloat16; 47 /* vcvtne2ps2bf16 */ 48 49 extern __inline __m256bh 50 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 51 _mm256_cvtne2ps_pbh (__m256 __A, __m256 __B) 52 { 53 return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi(__A, __B); 54 } 55 56 extern __inline __m256bh 57 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 58 _mm256_mask_cvtne2ps_pbh (__m256bh __A, __mmask16 __B, __m256 __C, __m256 __D) 59 { 60 return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi_mask(__C, __D, __A, __B); 61 } 62 63 extern __inline __m256bh 64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 65 _mm256_maskz_cvtne2ps_pbh (__mmask16 __A, __m256 __B, __m256 __C) 66 { 67 return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi_maskz(__B, __C, __A); 68 } 69 70 extern __inline __m128bh 71 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 72 _mm_cvtne2ps_pbh (__m128 __A, __m128 __B) 73 { 74 return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi(__A, __B); 75 } 76 77 extern __inline __m128bh 78 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 79 _mm_mask_cvtne2ps_pbh (__m128bh __A, __mmask8 __B, __m128 __C, __m128 __D) 80 { 81 return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi_mask(__C, __D, __A, __B); 82 } 83 84 extern __inline __m128bh 85 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 86 _mm_maskz_cvtne2ps_pbh (__mmask8 __A, __m128 __B, __m128 __C) 87 { 88 return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi_maskz(__B, __C, __A); 89 } 90 91 /* vcvtneps2bf16 */ 92 93 extern __inline __m128bh 94 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 95 _mm256_cvtneps_pbh (__m256 __A) 96 { 97 return (__m128bh)__builtin_ia32_cvtneps2bf16_v8sf(__A); 98 } 99 100 extern __inline __m128bh 101 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 102 _mm256_mask_cvtneps_pbh (__m128bh __A, __mmask8 __B, __m256 __C) 103 { 104 return (__m128bh)__builtin_ia32_cvtneps2bf16_v8sf_mask(__C, __A, __B); 105 } 106 107 extern __inline __m128bh 108 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 109 _mm256_maskz_cvtneps_pbh (__mmask8 __A, __m256 __B) 110 { 111 return (__m128bh)__builtin_ia32_cvtneps2bf16_v8sf_maskz(__B, __A); 112 } 113 114 extern __inline __m128bh 115 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 116 _mm_cvtneps_pbh (__m128 __A) 117 { 118 return (__m128bh)__builtin_ia32_cvtneps2bf16_v4sf(__A); 119 } 120 121 extern __inline __m128bh 122 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 123 _mm_mask_cvtneps_pbh (__m128bh __A, __mmask8 __B, __m128 __C) 124 { 125 return (__m128bh)__builtin_ia32_cvtneps2bf16_v4sf_mask(__C, __A, __B); 126 } 127 128 extern __inline __m128bh 129 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 130 _mm_maskz_cvtneps_pbh (__mmask8 __A, __m128 __B) 131 { 132 return (__m128bh)__builtin_ia32_cvtneps2bf16_v4sf_maskz(__B, __A); 133 } 134 135 /* vdpbf16ps */ 136 137 extern __inline __m256 138 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 139 _mm256_dpbf16_ps (__m256 __A, __m256bh __B, __m256bh __C) 140 { 141 return (__m256)__builtin_ia32_dpbf16ps_v8sf(__A, __B, __C); 142 } 143 144 extern __inline __m256 145 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 146 _mm256_mask_dpbf16_ps (__m256 __A, __mmask8 __B, __m256bh __C, __m256bh __D) 147 { 148 return (__m256)__builtin_ia32_dpbf16ps_v8sf_mask(__A, __C, __D, __B); 149 } 150 151 extern __inline __m256 152 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 153 _mm256_maskz_dpbf16_ps (__mmask8 __A, __m256 __B, __m256bh __C, __m256bh __D) 154 { 155 return (__m256)__builtin_ia32_dpbf16ps_v8sf_maskz(__B, __C, __D, __A); 156 } 157 158 extern __inline __m128 159 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 160 _mm_dpbf16_ps (__m128 __A, __m128bh __B, __m128bh __C) 161 { 162 return (__m128)__builtin_ia32_dpbf16ps_v4sf(__A, __B, __C); 163 } 164 165 extern __inline __m128 166 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 167 _mm_mask_dpbf16_ps (__m128 __A, __mmask8 __B, __m128bh __C, __m128bh __D) 168 { 169 return (__m128)__builtin_ia32_dpbf16ps_v4sf_mask(__A, __C, __D, __B); 170 } 171 172 extern __inline __m128 173 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 174 _mm_maskz_dpbf16_ps (__mmask8 __A, __m128 __B, __m128bh __C, __m128bh __D) 175 { 176 return (__m128)__builtin_ia32_dpbf16ps_v4sf_maskz(__B, __C, __D, __A); 177 } 178 179 extern __inline __bfloat16 180 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 181 _mm_cvtness_sbh (float __A) 182 { 183 __v4sf __V = {__A, 0, 0, 0}; 184 __v8hi __R = __builtin_ia32_cvtneps2bf16_v4sf_mask ((__v4sf)__V, 185 (__v8hi)_mm_undefined_si128 (), (__mmask8)-1); 186 return __R[0]; 187 } 188 189 extern __inline __m128 190 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 191 _mm_cvtpbh_ps (__m128bh __A) 192 { 193 return (__m128)_mm_castsi128_ps ((__m128i)_mm_slli_epi32 ( 194 (__m128i)_mm_cvtepi16_epi32 ((__m128i)__A), 16)); 195 } 196 197 extern __inline __m256 198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 199 _mm256_cvtpbh_ps (__m128bh __A) 200 { 201 return (__m256)_mm256_castsi256_ps ((__m256i)_mm256_slli_epi32 ( 202 (__m256i)_mm256_cvtepi16_epi32 ((__m128i)__A), 16)); 203 } 204 205 extern __inline __m128 206 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 207 _mm_maskz_cvtpbh_ps (__mmask8 __U, __m128bh __A) 208 { 209 return (__m128)_mm_castsi128_ps ((__m128i)_mm_slli_epi32 ( 210 (__m128i)_mm_maskz_cvtepi16_epi32 ( 211 (__mmask8)__U, (__m128i)__A), 16)); 212 } 213 214 extern __inline __m256 215 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 216 _mm256_maskz_cvtpbh_ps (__mmask8 __U, __m128bh __A) 217 { 218 return (__m256)_mm256_castsi256_ps ((__m256i)_mm256_slli_epi32 ( 219 (__m256i)_mm256_maskz_cvtepi16_epi32 ( 220 (__mmask8)__U, (__m128i)__A), 16)); 221 } 222 223 extern __inline __m128 224 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 225 _mm_mask_cvtpbh_ps (__m128 __S, __mmask8 __U, __m128bh __A) 226 { 227 return (__m128)_mm_castsi128_ps ((__m128i)_mm_mask_slli_epi32 ( 228 (__m128i)__S, (__mmask8)__U, (__m128i)_mm_cvtepi16_epi32 ( 229 (__m128i)__A), 16)); 230 } 231 232 extern __inline __m256 233 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 234 _mm256_mask_cvtpbh_ps (__m256 __S, __mmask8 __U, __m128bh __A) 235 { 236 return (__m256)_mm256_castsi256_ps ((__m256i)_mm256_mask_slli_epi32 ( 237 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32 ( 238 (__m128i)__A), 16)); 239 } 240 241 #ifdef __DISABLE_AVX512BF16VL__ 242 #undef __DISABLE_AVX512BF16VL__ 243 #pragma GCC pop_options 244 #endif /* __DISABLE_AVX512BF16VL__ */ 245 246 #endif /* _AVX512BF16VLINTRIN_H_INCLUDED */ 247