| 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 |
#ifdef __SSE2__ |
| 14 |
|
| 15 |
#ifndef __AVX512BF16INTRIN_H |
| 16 |
#define __AVX512BF16INTRIN_H |
| 17 |
|
| 18 |
typedef __bf16 __v32bf __attribute__((__vector_size__(64), __aligned__(64))); |
| 19 |
typedef __bf16 __m512bh __attribute__((__vector_size__(64), __aligned__(64))); |
| 20 |
typedef __bf16 __bfloat16 __attribute__((deprecated("use __bf16 instead"))); |
| 21 |
|
| 22 |
#define __DEFAULT_FN_ATTRS512 \ |
| 23 |
__attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \ |
| 24 |
__min_vector_width__(512))) |
| 25 |
#define __DEFAULT_FN_ATTRS \ |
| 26 |
__attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"))) |
| 27 |
|
| 28 |
/// Convert One BF16 Data to One Single Float Data. |
| 29 |
/// |
| 30 |
/// \headerfile <x86intrin.h> |
| 31 |
/// |
| 32 |
/// This intrinsic does not correspond to a specific instruction. |
| 33 |
/// |
| 34 |
/// \param __A |
| 35 |
/// A bfloat data. |
| 36 |
/// \returns A float data whose sign field and exponent field keep unchanged, |
| 37 |
/// and fraction field is extended to 23 bits. |
| 38 |
static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss(__bf16 __A) { |
| 39 |
return __builtin_ia32_cvtsbf162ss_32(__A); |
| 40 |
} |
| 41 |
|
| 42 |
/// Convert Two Packed Single Data to One Packed BF16 Data. |
| 43 |
/// |
| 44 |
/// \headerfile <x86intrin.h> |
| 45 |
/// |
| 46 |
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 47 |
/// |
| 48 |
/// \param __A |
| 49 |
/// A 512-bit vector of [16 x float]. |
| 50 |
/// \param __B |
| 51 |
/// A 512-bit vector of [16 x float]. |
| 52 |
/// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 53 |
/// conversion of __B, and higher 256 bits come from conversion of __A. |
| 54 |
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 55 |
_mm512_cvtne2ps_pbh(__m512 __A, __m512 __B) { |
| 56 |
return (__m512bh)__builtin_ia32_cvtne2ps2bf16_512((__v16sf) __A, |
| 57 |
(__v16sf) __B); |
| 58 |
} |
| 59 |
|
| 60 |
/// Convert Two Packed Single Data to One Packed BF16 Data. |
| 61 |
/// |
| 62 |
/// \headerfile <x86intrin.h> |
| 63 |
/// |
| 64 |
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 65 |
/// |
| 66 |
/// \param __A |
| 67 |
/// A 512-bit vector of [16 x float]. |
| 68 |
/// \param __B |
| 69 |
/// A 512-bit vector of [16 x float]. |
| 70 |
/// \param __W |
| 71 |
/// A 512-bit vector of [32 x bfloat]. |
| 72 |
/// \param __U |
| 73 |
/// A 32-bit mask value specifying what is chosen for each element. |
| 74 |
/// A 1 means conversion of __A or __B. A 0 means element from __W. |
| 75 |
/// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 76 |
/// conversion of __B, and higher 256 bits come from conversion of __A. |
| 77 |
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 78 |
_mm512_mask_cvtne2ps_pbh(__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B) { |
| 79 |
return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, |
| 80 |
(__v32bf)_mm512_cvtne2ps_pbh(__A, __B), |
| 81 |
(__v32bf)__W); |
| 82 |
} |
| 83 |
|
| 84 |
/// Convert Two Packed Single Data to One Packed BF16 Data. |
| 85 |
/// |
| 86 |
/// \headerfile <x86intrin.h> |
| 87 |
/// |
| 88 |
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. |
| 89 |
/// |
| 90 |
/// \param __A |
| 91 |
/// A 512-bit vector of [16 x float]. |
| 92 |
/// \param __B |
| 93 |
/// A 512-bit vector of [16 x float]. |
| 94 |
/// \param __U |
| 95 |
/// A 32-bit mask value specifying what is chosen for each element. |
| 96 |
/// A 1 means conversion of __A or __B. A 0 means element is zero. |
| 97 |
/// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from |
| 98 |
/// conversion of __B, and higher 256 bits come from conversion of __A. |
| 99 |
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 100 |
_mm512_maskz_cvtne2ps_pbh(__mmask32 __U, __m512 __A, __m512 __B) { |
| 101 |
return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, |
| 102 |
(__v32bf)_mm512_cvtne2ps_pbh(__A, __B), |
| 103 |
(__v32bf)_mm512_setzero_si512()); |
| 104 |
} |
| 105 |
|
| 106 |
/// Convert Packed Single Data to Packed BF16 Data. |
| 107 |
/// |
| 108 |
/// \headerfile <x86intrin.h> |
| 109 |
/// |
| 110 |
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 111 |
/// |
| 112 |
/// \param __A |
| 113 |
/// A 512-bit vector of [16 x float]. |
| 114 |
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 115 |
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 116 |
_mm512_cvtneps_pbh(__m512 __A) { |
| 117 |
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 118 |
(__v16bf)_mm256_undefined_si256(), |
| 119 |
(__mmask16)-1); |
| 120 |
} |
| 121 |
|
| 122 |
/// Convert Packed Single Data to Packed BF16 Data. |
| 123 |
/// |
| 124 |
/// \headerfile <x86intrin.h> |
| 125 |
/// |
| 126 |
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 127 |
/// |
| 128 |
/// \param __A |
| 129 |
/// A 512-bit vector of [16 x float]. |
| 130 |
/// \param __W |
| 131 |
/// A 256-bit vector of [16 x bfloat]. |
| 132 |
/// \param __U |
| 133 |
/// A 16-bit mask value specifying what is chosen for each element. |
| 134 |
/// A 1 means conversion of __A. A 0 means element from __W. |
| 135 |
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 136 |
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 137 |
_mm512_mask_cvtneps_pbh(__m256bh __W, __mmask16 __U, __m512 __A) { |
| 138 |
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 139 |
(__v16bf)__W, |
| 140 |
(__mmask16)__U); |
| 141 |
} |
| 142 |
|
| 143 |
/// Convert Packed Single Data to Packed BF16 Data. |
| 144 |
/// |
| 145 |
/// \headerfile <x86intrin.h> |
| 146 |
/// |
| 147 |
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. |
| 148 |
/// |
| 149 |
/// \param __A |
| 150 |
/// A 512-bit vector of [16 x float]. |
| 151 |
/// \param __U |
| 152 |
/// A 16-bit mask value specifying what is chosen for each element. |
| 153 |
/// A 1 means conversion of __A. A 0 means element is zero. |
| 154 |
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A. |
| 155 |
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 156 |
_mm512_maskz_cvtneps_pbh(__mmask16 __U, __m512 __A) { |
| 157 |
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, |
| 158 |
(__v16bf)_mm256_setzero_si256(), |
| 159 |
(__mmask16)__U); |
| 160 |
} |
| 161 |
|
| 162 |
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 163 |
/// |
| 164 |
/// \headerfile <x86intrin.h> |
| 165 |
/// |
| 166 |
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 167 |
/// |
| 168 |
/// \param __A |
| 169 |
/// A 512-bit vector of [32 x bfloat]. |
| 170 |
/// \param __B |
| 171 |
/// A 512-bit vector of [32 x bfloat]. |
| 172 |
/// \param __D |
| 173 |
/// A 512-bit vector of [16 x float]. |
| 174 |
/// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 175 |
/// __A, __B and __D |
| 176 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 177 |
_mm512_dpbf16_ps(__m512 __D, __m512bh __A, __m512bh __B) { |
| 178 |
return (__m512)__builtin_ia32_dpbf16ps_512((__v16sf) __D, |
| 179 |
(__v32bf) __A, |
| 180 |
(__v32bf) __B); |
| 181 |
} |
| 182 |
|
| 183 |
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 184 |
/// |
| 185 |
/// \headerfile <x86intrin.h> |
| 186 |
/// |
| 187 |
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 188 |
/// |
| 189 |
/// \param __A |
| 190 |
/// A 512-bit vector of [32 x bfloat]. |
| 191 |
/// \param __B |
| 192 |
/// A 512-bit vector of [32 x bfloat]. |
| 193 |
/// \param __D |
| 194 |
/// A 512-bit vector of [16 x float]. |
| 195 |
/// \param __U |
| 196 |
/// A 16-bit mask value specifying what is chosen for each element. |
| 197 |
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D. |
| 198 |
/// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 199 |
/// __A, __B and __D |
| 200 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 201 |
_mm512_mask_dpbf16_ps(__m512 __D, __mmask16 __U, __m512bh __A, __m512bh __B) { |
| 202 |
return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 203 |
(__v16sf)_mm512_dpbf16_ps(__D, __A, __B), |
| 204 |
(__v16sf)__D); |
| 205 |
} |
| 206 |
|
| 207 |
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. |
| 208 |
/// |
| 209 |
/// \headerfile <x86intrin.h> |
| 210 |
/// |
| 211 |
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. |
| 212 |
/// |
| 213 |
/// \param __A |
| 214 |
/// A 512-bit vector of [32 x bfloat]. |
| 215 |
/// \param __B |
| 216 |
/// A 512-bit vector of [32 x bfloat]. |
| 217 |
/// \param __D |
| 218 |
/// A 512-bit vector of [16 x float]. |
| 219 |
/// \param __U |
| 220 |
/// A 16-bit mask value specifying what is chosen for each element. |
| 221 |
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0. |
| 222 |
/// \returns A 512-bit vector of [16 x float] comes from Dot Product of |
| 223 |
/// __A, __B and __D |
| 224 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 225 |
_mm512_maskz_dpbf16_ps(__mmask16 __U, __m512 __D, __m512bh __A, __m512bh __B) { |
| 226 |
return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 227 |
(__v16sf)_mm512_dpbf16_ps(__D, __A, __B), |
| 228 |
(__v16sf)_mm512_setzero_si512()); |
| 229 |
} |
| 230 |
|
| 231 |
/// Convert Packed BF16 Data to Packed float Data. |
| 232 |
/// |
| 233 |
/// \headerfile <x86intrin.h> |
| 234 |
/// |
| 235 |
/// \param __A |
| 236 |
/// A 256-bit vector of [16 x bfloat]. |
| 237 |
/// \returns A 512-bit vector of [16 x float] come from conversion of __A |
| 238 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps(__m256bh __A) { |
| 239 |
return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( |
| 240 |
(__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16)); |
| 241 |
} |
| 242 |
|
| 243 |
/// Convert Packed BF16 Data to Packed float Data using zeroing mask. |
| 244 |
/// |
| 245 |
/// \headerfile <x86intrin.h> |
| 246 |
/// |
| 247 |
/// \param __U |
| 248 |
/// A 16-bit mask. Elements are zeroed out when the corresponding mask |
| 249 |
/// bit is not set. |
| 250 |
/// \param __A |
| 251 |
/// A 256-bit vector of [16 x bfloat]. |
| 252 |
/// \returns A 512-bit vector of [16 x float] come from conversion of __A |
| 253 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 254 |
_mm512_maskz_cvtpbh_ps(__mmask16 __U, __m256bh __A) { |
| 255 |
return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( |
| 256 |
(__m512i)_mm512_maskz_cvtepi16_epi32((__mmask16)__U, (__m256i)__A), 16)); |
| 257 |
} |
| 258 |
|
| 259 |
/// Convert Packed BF16 Data to Packed float Data using merging mask. |
| 260 |
/// |
| 261 |
/// \headerfile <x86intrin.h> |
| 262 |
/// |
| 263 |
/// \param __S |
| 264 |
/// A 512-bit vector of [16 x float]. Elements are copied from __S when |
| 265 |
/// the corresponding mask bit is not set. |
| 266 |
/// \param __U |
| 267 |
/// A 16-bit mask. |
| 268 |
/// \param __A |
| 269 |
/// A 256-bit vector of [16 x bfloat]. |
| 270 |
/// \returns A 512-bit vector of [16 x float] come from conversion of __A |
| 271 |
static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 272 |
_mm512_mask_cvtpbh_ps(__m512 __S, __mmask16 __U, __m256bh __A) { |
| 273 |
return _mm512_castsi512_ps((__m512i)_mm512_mask_slli_epi32( |
| 274 |
(__m512i)__S, (__mmask16)__U, |
| 275 |
(__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16)); |
| 276 |
} |
| 277 |
|
| 278 |
#undef __DEFAULT_FN_ATTRS |
| 279 |
#undef __DEFAULT_FN_ATTRS512 |
| 280 |
|
| 281 |
#endif |
| 282 |
#endif |