Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 1 | /*===------------ avx512bf16intrin.h - AVX512_BF16 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 | #ifndef __IMMINTRIN_H |
| 10 | #error "Never use <avx512bf16intrin.h> directly; include <immintrin.h> instead." |
| 11 | #endif |
| 12 | |
| 13 | #ifndef __AVX512BF16INTRIN_H |
| 14 | #define __AVX512BF16INTRIN_H |
| 15 | |
| 16 | typedef short __m512bh __attribute__((__vector_size__(64), __aligned__(64))); |
| 17 | typedef short __m256bh __attribute__((__vector_size__(32), __aligned__(32))); |
| 18 | typedef unsigned short __bfloat16; |
| 19 | |
| 20 | #define __DEFAULT_FN_ATTRS512 \ |
| 21 | __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \ |
| 22 | __min_vector_width__(512))) |
| 23 | #define __DEFAULT_FN_ATTRS \ |
| 24 | __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"))) |
| 25 | |
| 26 | /// Convert One BF16 Data to One Single Float Data. |
| 27 | /// |
| 28 | /// \headerfile <x86intrin.h> |
| 29 | /// |
| 30 | /// This intrinsic does not correspond to a specific instruction. |
| 31 | /// |
| 32 | /// \param __A |
| 33 | /// A bfloat data. |
| 34 | /// \returns A float data whose sign field and exponent field keep unchanged, |
| 35 | /// and fraction field is extended to 23 bits. |
| 36 | static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss(__bfloat16 __A) { |
| 37 | return __builtin_ia32_cvtsbf162ss_32(__A); |
| 38 | } |
| 39 | |
| 40 | /// Convert Two Packed Single Data to One Packed BF16 Data. |
| 41 | /// |
| 42 | /// \headerfile <x86intrin.h> |
| 43 | /// |
| 44 | /// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 45 | /// |
| 46 | /// \param __A |
| 47 | /// A 512-bit vector of [16 x float]. |
| 48 | /// \param __B |
| 49 | /// A 512-bit vector of [16 x float]. |
| 50 | /// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 51 | /// conversion of __B, and higher 256 bits come from conversion of __A. |
| 52 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 53 | _mm512_cvtne2ps_pbh(__m512 __A, __m512 __B) { |
| 54 | return (__m512bh)__builtin_ia32_cvtne2ps2bf16_512((__v16sf) __A, |
| 55 | (__v16sf) __B); |
| 56 | } |
| 57 | |
| 58 | /// Convert Two Packed Single Data to One Packed BF16 Data. |
| 59 | /// |
| 60 | /// \headerfile <x86intrin.h> |
| 61 | /// |
| 62 | /// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 63 | /// |
| 64 | /// \param __A |
| 65 | /// A 512-bit vector of [16 x float]. |
| 66 | /// \param __B |
| 67 | /// A 512-bit vector of [16 x float]. |
| 68 | /// \param __W |
| 69 | /// A 512-bit vector of [32 x bfloat]. |
| 70 | /// \param __U |
| 71 | /// A 32-bit mask value specifying what is chosen for each element. |
| 72 | /// A 1 means conversion of __A or __B. A 0 means element from __W. |
| 73 | /// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 74 | /// conversion of __B, and higher 256 bits come from conversion of __A. |
| 75 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 76 | _mm512_mask_cvtne2ps_pbh(__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B) { |
| 77 | return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U, |
| 78 | (__v32hi)_mm512_cvtne2ps_pbh(__A, __B), |
| 79 | (__v32hi)__W); |
| 80 | } |
| 81 | |
| 82 | /// Convert Two Packed Single Data to One Packed BF16 Data. |
| 83 | /// |
| 84 | /// \headerfile <x86intrin.h> |
| 85 | /// |
| 86 | /// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 87 | /// |
| 88 | /// \param __A |
| 89 | /// A 512-bit vector of [16 x float]. |
| 90 | /// \param __B |
| 91 | /// A 512-bit vector of [16 x float]. |
| 92 | /// \param __U |
| 93 | /// A 32-bit mask value specifying what is chosen for each element. |
| 94 | /// A 1 means conversion of __A or __B. A 0 means element is zero. |
| 95 | /// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 96 | /// conversion of __B, and higher 256 bits come from conversion of __A. |
| 97 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 98 | _mm512_maskz_cvtne2ps_pbh(__mmask32 __U, __m512 __A, __m512 __B) { |
| 99 | return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U, |
| 100 | (__v32hi)_mm512_cvtne2ps_pbh(__A, __B), |
| 101 | (__v32hi)_mm512_setzero_si512()); |
| 102 | } |
| 103 | |
| 104 | /// Convert Packed Single Data to Packed BF16 Data. |
| 105 | /// |
| 106 | /// \headerfile <x86intrin.h> |
| 107 | /// |
| 108 | /// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 109 | /// |
| 110 | /// \param __A |
| 111 | /// A 512-bit vector of [16 x float]. |
| 112 | /// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 113 | static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 114 | _mm512_cvtneps_pbh(__m512 __A) { |
| 115 | return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 116 | (__v16hi)_mm256_undefined_si256(), |
| 117 | (__mmask16)-1); |
| 118 | } |
| 119 | |
| 120 | /// Convert Packed Single Data to Packed BF16 Data. |
| 121 | /// |
| 122 | /// \headerfile <x86intrin.h> |
| 123 | /// |
| 124 | /// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 125 | /// |
| 126 | /// \param __A |
| 127 | /// A 512-bit vector of [16 x float]. |
| 128 | /// \param __W |
| 129 | /// A 256-bit vector of [16 x bfloat]. |
| 130 | /// \param __U |
| 131 | /// A 16-bit mask value specifying what is chosen for each element. |
| 132 | /// A 1 means conversion of __A. A 0 means element from __W. |
| 133 | /// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 134 | static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 135 | _mm512_mask_cvtneps_pbh(__m256bh __W, __mmask16 __U, __m512 __A) { |
| 136 | return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 137 | (__v16hi)__W, |
| 138 | (__mmask16)__U); |
| 139 | } |
| 140 | |
| 141 | /// Convert Packed Single Data to Packed BF16 Data. |
| 142 | /// |
| 143 | /// \headerfile <x86intrin.h> |
| 144 | /// |
| 145 | /// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 146 | /// |
| 147 | /// \param __A |
| 148 | /// A 512-bit vector of [16 x float]. |
| 149 | /// \param __U |
| 150 | /// A 16-bit mask value specifying what is chosen for each element. |
| 151 | /// A 1 means conversion of __A. A 0 means element is zero. |
| 152 | /// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 153 | static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 154 | _mm512_maskz_cvtneps_pbh(__mmask16 __U, __m512 __A) { |
| 155 | return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 156 | (__v16hi)_mm256_setzero_si256(), |
| 157 | (__mmask16)__U); |
| 158 | } |
| 159 | |
| 160 | /// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 161 | /// |
| 162 | /// \headerfile <x86intrin.h> |
| 163 | /// |
| 164 | /// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 165 | /// |
| 166 | /// \param __A |
| 167 | /// A 512-bit vector of [32 x bfloat]. |
| 168 | /// \param __B |
| 169 | /// A 512-bit vector of [32 x bfloat]. |
| 170 | /// \param __D |
| 171 | /// A 512-bit vector of [16 x float]. |
| 172 | /// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 173 | /// __A, __B and __D |
| 174 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 175 | _mm512_dpbf16_ps(__m512 __D, __m512bh __A, __m512bh __B) { |
| 176 | return (__m512)__builtin_ia32_dpbf16ps_512((__v16sf) __D, |
| 177 | (__v16si) __A, |
| 178 | (__v16si) __B); |
| 179 | } |
| 180 | |
| 181 | /// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 182 | /// |
| 183 | /// \headerfile <x86intrin.h> |
| 184 | /// |
| 185 | /// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 186 | /// |
| 187 | /// \param __A |
| 188 | /// A 512-bit vector of [32 x bfloat]. |
| 189 | /// \param __B |
| 190 | /// A 512-bit vector of [32 x bfloat]. |
| 191 | /// \param __D |
| 192 | /// A 512-bit vector of [16 x float]. |
| 193 | /// \param __U |
| 194 | /// A 16-bit mask value specifying what is chosen for each element. |
| 195 | /// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D. |
| 196 | /// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 197 | /// __A, __B and __D |
| 198 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 199 | _mm512_mask_dpbf16_ps(__m512 __D, __mmask16 __U, __m512bh __A, __m512bh __B) { |
| 200 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 201 | (__v16sf)_mm512_dpbf16_ps(__D, __A, __B), |
| 202 | (__v16sf)__D); |
| 203 | } |
| 204 | |
| 205 | /// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 206 | /// |
| 207 | /// \headerfile <x86intrin.h> |
| 208 | /// |
| 209 | /// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 210 | /// |
| 211 | /// \param __A |
| 212 | /// A 512-bit vector of [32 x bfloat]. |
| 213 | /// \param __B |
| 214 | /// A 512-bit vector of [32 x bfloat]. |
| 215 | /// \param __D |
| 216 | /// A 512-bit vector of [16 x float]. |
| 217 | /// \param __U |
| 218 | /// A 16-bit mask value specifying what is chosen for each element. |
| 219 | /// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0. |
| 220 | /// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 221 | /// __A, __B and __D |
| 222 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 223 | _mm512_maskz_dpbf16_ps(__mmask16 __U, __m512 __D, __m512bh __A, __m512bh __B) { |
| 224 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 225 | (__v16sf)_mm512_dpbf16_ps(__D, __A, __B), |
| 226 | (__v16sf)_mm512_setzero_si512()); |
| 227 | } |
| 228 | |
| 229 | /// Convert Packed BF16 Data to Packed float Data. |
| 230 | /// |
| 231 | /// \headerfile <x86intrin.h> |
| 232 | /// |
| 233 | /// \param __A |
| 234 | /// A 256-bit vector of [16 x bfloat]. |
Pirama Arumuga Nainar | ec8c89d | 2022-02-23 09:26:16 -0800 | [diff] [blame] | 235 | /// \returns A 512-bit vector of [16 x float] come from conversion of __A |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 236 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps(__m256bh __A) { |
| 237 | return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( |
| 238 | (__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16)); |
| 239 | } |
| 240 | |
| 241 | /// Convert Packed BF16 Data to Packed float Data using zeroing mask. |
| 242 | /// |
| 243 | /// \headerfile <x86intrin.h> |
| 244 | /// |
| 245 | /// \param __U |
| 246 | /// A 16-bit mask. Elements are zeroed out when the corresponding mask |
| 247 | /// bit is not set. |
| 248 | /// \param __A |
| 249 | /// A 256-bit vector of [16 x bfloat]. |
Pirama Arumuga Nainar | ec8c89d | 2022-02-23 09:26:16 -0800 | [diff] [blame] | 250 | /// \returns A 512-bit vector of [16 x float] come from conversion of __A |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 251 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 252 | _mm512_maskz_cvtpbh_ps(__mmask16 __U, __m256bh __A) { |
| 253 | return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( |
| 254 | (__m512i)_mm512_maskz_cvtepi16_epi32((__mmask16)__U, (__m256i)__A), 16)); |
| 255 | } |
| 256 | |
| 257 | /// Convert Packed BF16 Data to Packed float Data using merging mask. |
| 258 | /// |
| 259 | /// \headerfile <x86intrin.h> |
| 260 | /// |
| 261 | /// \param __S |
| 262 | /// A 512-bit vector of [16 x float]. Elements are copied from __S when |
| 263 | /// the corresponding mask bit is not set. |
| 264 | /// \param __U |
| 265 | /// A 16-bit mask. |
| 266 | /// \param __A |
| 267 | /// A 256-bit vector of [16 x bfloat]. |
Pirama Arumuga Nainar | ec8c89d | 2022-02-23 09:26:16 -0800 | [diff] [blame] | 268 | /// \returns A 512-bit vector of [16 x float] come from conversion of __A |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 269 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 270 | _mm512_mask_cvtpbh_ps(__m512 __S, __mmask16 __U, __m256bh __A) { |
| 271 | return _mm512_castsi512_ps((__m512i)_mm512_mask_slli_epi32( |
| 272 | (__m512i)__S, (__mmask16)__U, |
| 273 | (__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16)); |
| 274 | } |
| 275 | |
| 276 | #undef __DEFAULT_FN_ATTRS |
| 277 | #undef __DEFAULT_FN_ATTRS512 |
| 278 | |
| 279 | #endif |