1/*===--------- avx512vlbf16intrin.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 <avx512vlbf16intrin.h> directly; include <immintrin.h> instead." 11#endif 12 13#ifndef __AVX512VLBF16INTRIN_H 14#define __AVX512VLBF16INTRIN_H 15 16typedef short __m128bh __attribute__((__vector_size__(16), __aligned__(16))); 17 18#define __DEFAULT_FN_ATTRS128 \ 19 __attribute__((__always_inline__, __nodebug__, \ 20 __target__("avx512vl, avx512bf16"), __min_vector_width__(128))) 21#define __DEFAULT_FN_ATTRS256 \ 22 __attribute__((__always_inline__, __nodebug__, \ 23 __target__("avx512vl, avx512bf16"), __min_vector_width__(256))) 24 25/// Convert Two Packed Single Data to One Packed BF16 Data. 26/// 27/// \headerfile <x86intrin.h> 28/// 29/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 30/// 31/// \param __A 32/// A 128-bit vector of [4 x float]. 33/// \param __B 34/// A 128-bit vector of [4 x float]. 35/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 36/// conversion of __B, and higher 64 bits come from conversion of __A. 37static __inline__ __m128bh __DEFAULT_FN_ATTRS128 38_mm_cvtne2ps_pbh(__m128 __A, __m128 __B) { 39 return (__m128bh)__builtin_ia32_cvtne2ps2bf16_128((__v4sf) __A, 40 (__v4sf) __B); 41} 42 43/// Convert Two Packed Single Data to One Packed BF16 Data. 44/// 45/// \headerfile <x86intrin.h> 46/// 47/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 48/// 49/// \param __A 50/// A 128-bit vector of [4 x float]. 51/// \param __B 52/// A 128-bit vector of [4 x float]. 53/// \param __W 54/// A 128-bit vector of [8 x bfloat]. 55/// \param __U 56/// A 8-bit mask value specifying what is chosen for each element. 57/// A 1 means conversion of __A or __B. A 0 means element from __W. 58/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 59/// conversion of __B, and higher 64 bits come from conversion of __A. 60static __inline__ __m128bh __DEFAULT_FN_ATTRS128 61_mm_mask_cvtne2ps_pbh(__m128bh __W, __mmask8 __U, __m128 __A, __m128 __B) { 62 return (__m128bh)__builtin_ia32_selectw_128((__mmask8)__U, 63 (__v8hi)_mm_cvtne2ps_pbh(__A, __B), 64 (__v8hi)__W); 65} 66 67/// Convert Two Packed Single Data to One Packed BF16 Data. 68/// 69/// \headerfile <x86intrin.h> 70/// 71/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 72/// 73/// \param __A 74/// A 128-bit vector of [4 x float]. 75/// \param __B 76/// A 128-bit vector of [4 x float]. 77/// \param __U 78/// A 8-bit mask value specifying what is chosen for each element. 79/// A 1 means conversion of __A or __B. A 0 means element is zero. 80/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 81/// conversion of __B, and higher 64 bits come from conversion of __A. 82static __inline__ __m128bh __DEFAULT_FN_ATTRS128 83_mm_maskz_cvtne2ps_pbh(__mmask8 __U, __m128 __A, __m128 __B) { 84 return (__m128bh)__builtin_ia32_selectw_128((__mmask8)__U, 85 (__v8hi)_mm_cvtne2ps_pbh(__A, __B), 86 (__v8hi)_mm_setzero_si128()); 87} 88 89/// Convert Two Packed Single Data to One Packed BF16 Data. 90/// 91/// \headerfile <x86intrin.h> 92/// 93/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 94/// 95/// \param __A 96/// A 256-bit vector of [8 x float]. 97/// \param __B 98/// A 256-bit vector of [8 x float]. 99/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from 100/// conversion of __B, and higher 128 bits come from conversion of __A. 101static __inline__ __m256bh __DEFAULT_FN_ATTRS256 102_mm256_cvtne2ps_pbh(__m256 __A, __m256 __B) { 103 return (__m256bh)__builtin_ia32_cvtne2ps2bf16_256((__v8sf) __A, 104 (__v8sf) __B); 105} 106 107/// Convert Two Packed Single Data to One Packed BF16 Data. 108/// 109/// \headerfile <x86intrin.h> 110/// 111/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 112/// 113/// \param __A 114/// A 256-bit vector of [8 x float]. 115/// \param __B 116/// A 256-bit vector of [8 x float]. 117/// \param __W 118/// A 256-bit vector of [16 x bfloat]. 119/// \param __U 120/// A 16-bit mask value specifying what is chosen for each element. 121/// A 1 means conversion of __A or __B. A 0 means element from __W. 122/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from 123/// conversion of __B, and higher 128 bits come from conversion of __A. 124static __inline__ __m256bh __DEFAULT_FN_ATTRS256 125_mm256_mask_cvtne2ps_pbh(__m256bh __W, __mmask16 __U, __m256 __A, __m256 __B) { 126 return (__m256bh)__builtin_ia32_selectw_256((__mmask16)__U, 127 (__v16hi)_mm256_cvtne2ps_pbh(__A, __B), 128 (__v16hi)__W); 129} 130 131/// Convert Two Packed Single Data to One Packed BF16 Data. 132/// 133/// \headerfile <x86intrin.h> 134/// 135/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions. 136/// 137/// \param __A 138/// A 256-bit vector of [8 x float]. 139/// \param __B 140/// A 256-bit vector of [8 x float]. 141/// \param __U 142/// A 16-bit mask value specifying what is chosen for each element. 143/// A 1 means conversion of __A or __B. A 0 means element is zero. 144/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from 145/// conversion of __B, and higher 128 bits come from conversion of __A. 146static __inline__ __m256bh __DEFAULT_FN_ATTRS256 147_mm256_maskz_cvtne2ps_pbh(__mmask16 __U, __m256 __A, __m256 __B) { 148 return (__m256bh)__builtin_ia32_selectw_256((__mmask16)__U, 149 (__v16hi)_mm256_cvtne2ps_pbh(__A, __B), 150 (__v16hi)_mm256_setzero_si256()); 151} 152 153/// Convert Packed Single Data to Packed BF16 Data. 154/// 155/// \headerfile <x86intrin.h> 156/// 157/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 158/// 159/// \param __A 160/// A 128-bit vector of [4 x float]. 161/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 162/// conversion of __A, and higher 64 bits are 0. 163static __inline__ __m128bh __DEFAULT_FN_ATTRS128 164_mm_cvtneps_pbh(__m128 __A) { 165 return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A, 166 (__v8hi)_mm_undefined_si128(), 167 (__mmask8)-1); 168} 169 170/// Convert Packed Single Data to Packed BF16 Data. 171/// 172/// \headerfile <x86intrin.h> 173/// 174/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 175/// 176/// \param __A 177/// A 128-bit vector of [4 x float]. 178/// \param __W 179/// A 128-bit vector of [8 x bfloat]. 180/// \param __U 181/// A 4-bit mask value specifying what is chosen for each element. 182/// A 1 means conversion of __A. A 0 means element from __W. 183/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 184/// conversion of __A, and higher 64 bits are 0. 185static __inline__ __m128bh __DEFAULT_FN_ATTRS128 186_mm_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m128 __A) { 187 return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A, 188 (__v8hi)__W, 189 (__mmask8)__U); 190} 191 192/// Convert Packed Single Data to Packed BF16 Data. 193/// 194/// \headerfile <x86intrin.h> 195/// 196/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 197/// 198/// \param __A 199/// A 128-bit vector of [4 x float]. 200/// \param __U 201/// A 4-bit mask value specifying what is chosen for each element. 202/// A 1 means conversion of __A. A 0 means element is zero. 203/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from 204/// conversion of __A, and higher 64 bits are 0. 205static __inline__ __m128bh __DEFAULT_FN_ATTRS128 206_mm_maskz_cvtneps_pbh(__mmask8 __U, __m128 __A) { 207 return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A, 208 (__v8hi)_mm_setzero_si128(), 209 (__mmask8)__U); 210} 211 212/// Convert Packed Single Data to Packed BF16 Data. 213/// 214/// \headerfile <x86intrin.h> 215/// 216/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 217/// 218/// \param __A 219/// A 256-bit vector of [8 x float]. 220/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A. 221static __inline__ __m128bh __DEFAULT_FN_ATTRS256 222_mm256_cvtneps_pbh(__m256 __A) { 223 return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A, 224 (__v8hi)_mm_undefined_si128(), 225 (__mmask8)-1); 226} 227 228/// Convert Packed Single Data to Packed BF16 Data. 229/// 230/// \headerfile <x86intrin.h> 231/// 232/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 233/// 234/// \param __A 235/// A 256-bit vector of [8 x float]. 236/// \param __W 237/// A 256-bit vector of [8 x bfloat]. 238/// \param __U 239/// A 8-bit mask value specifying what is chosen for each element. 240/// A 1 means conversion of __A. A 0 means element from __W. 241/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A. 242static __inline__ __m128bh __DEFAULT_FN_ATTRS256 243_mm256_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m256 __A) { 244 return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A, 245 (__v8hi)__W, 246 (__mmask8)__U); 247} 248 249/// Convert Packed Single Data to Packed BF16 Data. 250/// 251/// \headerfile <x86intrin.h> 252/// 253/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 254/// 255/// \param __A 256/// A 256-bit vector of [8 x float]. 257/// \param __U 258/// A 8-bit mask value specifying what is chosen for each element. 259/// A 1 means conversion of __A. A 0 means element is zero. 260/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A. 261static __inline__ __m128bh __DEFAULT_FN_ATTRS256 262_mm256_maskz_cvtneps_pbh(__mmask8 __U, __m256 __A) { 263 return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A, 264 (__v8hi)_mm_setzero_si128(), 265 (__mmask8)__U); 266} 267 268/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 269/// 270/// \headerfile <x86intrin.h> 271/// 272/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 273/// 274/// \param __A 275/// A 128-bit vector of [8 x bfloat]. 276/// \param __B 277/// A 128-bit vector of [8 x bfloat]. 278/// \param __D 279/// A 128-bit vector of [4 x float]. 280/// \returns A 128-bit vector of [4 x float] comes from Dot Product of 281/// __A, __B and __D 282static __inline__ __m128 __DEFAULT_FN_ATTRS128 283_mm_dpbf16_ps(__m128 __D, __m128bh __A, __m128bh __B) { 284 return (__m128)__builtin_ia32_dpbf16ps_128((__v4sf)__D, 285 (__v4si)__A, 286 (__v4si)__B); 287} 288 289/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 290/// 291/// \headerfile <x86intrin.h> 292/// 293/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 294/// 295/// \param __A 296/// A 128-bit vector of [8 x bfloat]. 297/// \param __B 298/// A 128-bit vector of [8 x bfloat]. 299/// \param __D 300/// A 128-bit vector of [4 x float]. 301/// \param __U 302/// A 8-bit mask value specifying what is chosen for each element. 303/// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D. 304/// \returns A 128-bit vector of [4 x float] comes from Dot Product of 305/// __A, __B and __D 306static __inline__ __m128 __DEFAULT_FN_ATTRS128 307_mm_mask_dpbf16_ps(__m128 __D, __mmask8 __U, __m128bh __A, __m128bh __B) { 308 return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, 309 (__v4sf)_mm_dpbf16_ps(__D, __A, __B), 310 (__v4sf)__D); 311} 312 313/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 314/// 315/// \headerfile <x86intrin.h> 316/// 317/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 318/// 319/// \param __A 320/// A 128-bit vector of [8 x bfloat]. 321/// \param __B 322/// A 128-bit vector of [8 x bfloat]. 323/// \param __D 324/// A 128-bit vector of [4 x float]. 325/// \param __U 326/// A 8-bit mask value specifying what is chosen for each element. 327/// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0. 328/// \returns A 128-bit vector of [4 x float] comes from Dot Product of 329/// __A, __B and __D 330static __inline__ __m128 __DEFAULT_FN_ATTRS128 331_mm_maskz_dpbf16_ps(__mmask8 __U, __m128 __D, __m128bh __A, __m128bh __B) { 332 return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, 333 (__v4sf)_mm_dpbf16_ps(__D, __A, __B), 334 (__v4sf)_mm_setzero_si128()); 335} 336 337/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 338/// 339/// \headerfile <x86intrin.h> 340/// 341/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 342/// 343/// \param __A 344/// A 256-bit vector of [16 x bfloat]. 345/// \param __B 346/// A 256-bit vector of [16 x bfloat]. 347/// \param __D 348/// A 256-bit vector of [8 x float]. 349/// \returns A 256-bit vector of [8 x float] comes from Dot Product of 350/// __A, __B and __D 351static __inline__ __m256 __DEFAULT_FN_ATTRS256 352_mm256_dpbf16_ps(__m256 __D, __m256bh __A, __m256bh __B) { 353 return (__m256)__builtin_ia32_dpbf16ps_256((__v8sf)__D, 354 (__v8si)__A, 355 (__v8si)__B); 356} 357 358/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 359/// 360/// \headerfile <x86intrin.h> 361/// 362/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 363/// 364/// \param __A 365/// A 256-bit vector of [16 x bfloat]. 366/// \param __B 367/// A 256-bit vector of [16 x bfloat]. 368/// \param __D 369/// A 256-bit vector of [8 x float]. 370/// \param __U 371/// A 16-bit mask value specifying what is chosen for each element. 372/// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D. 373/// \returns A 256-bit vector of [8 x float] comes from Dot Product of 374/// __A, __B and __D 375static __inline__ __m256 __DEFAULT_FN_ATTRS256 376_mm256_mask_dpbf16_ps(__m256 __D, __mmask8 __U, __m256bh __A, __m256bh __B) { 377 return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, 378 (__v8sf)_mm256_dpbf16_ps(__D, __A, __B), 379 (__v8sf)__D); 380} 381 382/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. 383/// 384/// \headerfile <x86intrin.h> 385/// 386/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions. 387/// 388/// \param __A 389/// A 256-bit vector of [16 x bfloat]. 390/// \param __B 391/// A 256-bit vector of [16 x bfloat]. 392/// \param __D 393/// A 256-bit vector of [8 x float]. 394/// \param __U 395/// A 8-bit mask value specifying what is chosen for each element. 396/// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0. 397/// \returns A 256-bit vector of [8 x float] comes from Dot Product of 398/// __A, __B and __D 399static __inline__ __m256 __DEFAULT_FN_ATTRS256 400_mm256_maskz_dpbf16_ps(__mmask8 __U, __m256 __D, __m256bh __A, __m256bh __B) { 401 return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, 402 (__v8sf)_mm256_dpbf16_ps(__D, __A, __B), 403 (__v8sf)_mm256_setzero_si256()); 404} 405 406/// Convert One Single float Data to One BF16 Data. 407/// 408/// \headerfile <x86intrin.h> 409/// 410/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions. 411/// 412/// \param __A 413/// A float data. 414/// \returns A bf16 data whose sign field and exponent field keep unchanged, 415/// and fraction field is truncated to 7 bits. 416static __inline__ __bfloat16 __DEFAULT_FN_ATTRS128 _mm_cvtness_sbh(float __A) { 417 __v4sf __V = {__A, 0, 0, 0}; 418 __v8hi __R = __builtin_ia32_cvtneps2bf16_128_mask( 419 (__v4sf)__V, (__v8hi)_mm_undefined_si128(), (__mmask8)-1); 420 return __R[0]; 421} 422 423/// Convert Packed BF16 Data to Packed float Data. 424/// 425/// \headerfile <x86intrin.h> 426/// 427/// \param __A 428/// A 128-bit vector of [8 x bfloat]. 429/// \returns A 256-bit vector of [8 x float] come from convertion of __A 430static __inline__ __m256 __DEFAULT_FN_ATTRS256 _mm256_cvtpbh_ps(__m128bh __A) { 431 return _mm256_castsi256_ps((__m256i)_mm256_slli_epi32( 432 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); 433} 434 435/// Convert Packed BF16 Data to Packed float Data using zeroing mask. 436/// 437/// \headerfile <x86intrin.h> 438/// 439/// \param __U 440/// A 8-bit mask. Elements are zeroed out when the corresponding mask 441/// bit is not set. 442/// \param __A 443/// A 128-bit vector of [8 x bfloat]. 444/// \returns A 256-bit vector of [8 x float] come from convertion of __A 445static __inline__ __m256 __DEFAULT_FN_ATTRS256 446_mm256_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) { 447 return _mm256_castsi256_ps((__m256i)_mm256_slli_epi32( 448 (__m256i)_mm256_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16)); 449} 450 451/// Convert Packed BF16 Data to Packed float Data using merging mask. 452/// 453/// \headerfile <x86intrin.h> 454/// 455/// \param __S 456/// A 256-bit vector of [8 x float]. Elements are copied from __S when 457/// the corresponding mask bit is not set. 458/// \param __U 459/// A 8-bit mask. Elements are zeroed out when the corresponding mask 460/// bit is not set. 461/// \param __A 462/// A 128-bit vector of [8 x bfloat]. 463/// \returns A 256-bit vector of [8 x float] come from convertion of __A 464static __inline__ __m256 __DEFAULT_FN_ATTRS256 465_mm256_mask_cvtpbh_ps(__m256 __S, __mmask8 __U, __m128bh __A) { 466 return _mm256_castsi256_ps((__m256i)_mm256_mask_slli_epi32( 467 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 468 16)); 469} 470 471#undef __DEFAULT_FN_ATTRS128 472#undef __DEFAULT_FN_ATTRS256 473 474#endif 475