1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
|
/*===--------- avx512vlbf16intrin.h - AVX512_BF16 intrinsics ---------------===
*
* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
*
*===-----------------------------------------------------------------------===
*/
#ifndef __IMMINTRIN_H
#error "Never use <avx512vlbf16intrin.h> directly; include <immintrin.h> instead."
#endif
#ifdef __SSE2__
#ifndef __AVX512VLBF16INTRIN_H
#define __AVX512VLBF16INTRIN_H
#define __DEFAULT_FN_ATTRS128 \
__attribute__((__always_inline__, __nodebug__, \
__target__("avx512vl, avx512bf16"), __min_vector_width__(128)))
#define __DEFAULT_FN_ATTRS256 \
__attribute__((__always_inline__, __nodebug__, \
__target__("avx512vl, avx512bf16"), __min_vector_width__(256)))
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \param __B
/// A 128-bit vector of [4 x float].
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __B, and higher 64 bits come from conversion of __A.
static __inline__ __m128bh __DEFAULT_FN_ATTRS128
_mm_cvtne2ps_pbh(__m128 __A, __m128 __B) {
return (__m128bh)__builtin_ia32_cvtne2ps2bf16_128((__v4sf) __A,
(__v4sf) __B);
}
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \param __B
/// A 128-bit vector of [4 x float].
/// \param __W
/// A 128-bit vector of [8 x bfloat].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A or __B. A 0 means element from __W.
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __B, and higher 64 bits come from conversion of __A.
static __inline__ __m128bh __DEFAULT_FN_ATTRS128
_mm_mask_cvtne2ps_pbh(__m128bh __W, __mmask8 __U, __m128 __A, __m128 __B) {
return (__m128bh)__builtin_ia32_selectpbf_128((__mmask8)__U,
(__v8bf)_mm_cvtne2ps_pbh(__A, __B),
(__v8bf)__W);
}
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \param __B
/// A 128-bit vector of [4 x float].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A or __B. A 0 means element is zero.
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __B, and higher 64 bits come from conversion of __A.
static __inline__ __m128bh __DEFAULT_FN_ATTRS128
_mm_maskz_cvtne2ps_pbh(__mmask8 __U, __m128 __A, __m128 __B) {
return (__m128bh)__builtin_ia32_selectpbf_128((__mmask8)__U,
(__v8bf)_mm_cvtne2ps_pbh(__A, __B),
(__v8bf)_mm_setzero_si128());
}
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \param __B
/// A 256-bit vector of [8 x float].
/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from
/// conversion of __B, and higher 128 bits come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS256
_mm256_cvtne2ps_pbh(__m256 __A, __m256 __B) {
return (__m256bh)__builtin_ia32_cvtne2ps2bf16_256((__v8sf) __A,
(__v8sf) __B);
}
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \param __B
/// A 256-bit vector of [8 x float].
/// \param __W
/// A 256-bit vector of [16 x bfloat].
/// \param __U
/// A 16-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A or __B. A 0 means element from __W.
/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from
/// conversion of __B, and higher 128 bits come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS256
_mm256_mask_cvtne2ps_pbh(__m256bh __W, __mmask16 __U, __m256 __A, __m256 __B) {
return (__m256bh)__builtin_ia32_selectpbf_256((__mmask16)__U,
(__v16bf)_mm256_cvtne2ps_pbh(__A, __B),
(__v16bf)__W);
}
/// Convert Two Packed Single Data to One Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNE2PS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \param __B
/// A 256-bit vector of [8 x float].
/// \param __U
/// A 16-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A or __B. A 0 means element is zero.
/// \returns A 256-bit vector of [16 x bfloat] whose lower 128 bits come from
/// conversion of __B, and higher 128 bits come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS256
_mm256_maskz_cvtne2ps_pbh(__mmask16 __U, __m256 __A, __m256 __B) {
return (__m256bh)__builtin_ia32_selectpbf_256((__mmask16)__U,
(__v16bf)_mm256_cvtne2ps_pbh(__A, __B),
(__v16bf)_mm256_setzero_si256());
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __A, and higher 64 bits are 0.
#define _mm_cvtneps_pbh(A) \
((__m128bh)__builtin_ia32_vcvtneps2bf16128((__v4sf)(A)))
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \param __W
/// A 128-bit vector of [8 x bfloat].
/// \param __U
/// A 4-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A. A 0 means element from __W.
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __A, and higher 64 bits are 0.
static __inline__ __m128bh __DEFAULT_FN_ATTRS128
_mm_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m128 __A) {
return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A,
(__v8bf)__W,
(__mmask8)__U);
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 128-bit vector of [4 x float].
/// \param __U
/// A 4-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A. A 0 means element is zero.
/// \returns A 128-bit vector of [8 x bfloat] whose lower 64 bits come from
/// conversion of __A, and higher 64 bits are 0.
static __inline__ __m128bh __DEFAULT_FN_ATTRS128
_mm_maskz_cvtneps_pbh(__mmask8 __U, __m128 __A) {
return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A,
(__v8bf)_mm_setzero_si128(),
(__mmask8)__U);
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
#define _mm256_cvtneps_pbh(A) \
((__m128bh)__builtin_ia32_vcvtneps2bf16256((__v8sf)(A)))
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \param __W
/// A 256-bit vector of [8 x bfloat].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A. A 0 means element from __W.
/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
static __inline__ __m128bh __DEFAULT_FN_ATTRS256
_mm256_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m256 __A) {
return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A,
(__v8bf)__W,
(__mmask8)__U);
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 256-bit vector of [8 x float].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A. A 0 means element is zero.
/// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
static __inline__ __m128bh __DEFAULT_FN_ATTRS256
_mm256_maskz_cvtneps_pbh(__mmask8 __U, __m256 __A) {
return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A,
(__v8bf)_mm_setzero_si128(),
(__mmask8)__U);
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \param __B
/// A 128-bit vector of [8 x bfloat].
/// \param __D
/// A 128-bit vector of [4 x float].
/// \returns A 128-bit vector of [4 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m128 __DEFAULT_FN_ATTRS128
_mm_dpbf16_ps(__m128 __D, __m128bh __A, __m128bh __B) {
return (__m128)__builtin_ia32_dpbf16ps_128((__v4sf)__D,
(__v8bf)__A,
(__v8bf)__B);
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \param __B
/// A 128-bit vector of [8 x bfloat].
/// \param __D
/// A 128-bit vector of [4 x float].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D.
/// \returns A 128-bit vector of [4 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m128 __DEFAULT_FN_ATTRS128
_mm_mask_dpbf16_ps(__m128 __D, __mmask8 __U, __m128bh __A, __m128bh __B) {
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
(__v4sf)_mm_dpbf16_ps(__D, __A, __B),
(__v4sf)__D);
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \param __B
/// A 128-bit vector of [8 x bfloat].
/// \param __D
/// A 128-bit vector of [4 x float].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0.
/// \returns A 128-bit vector of [4 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m128 __DEFAULT_FN_ATTRS128
_mm_maskz_dpbf16_ps(__mmask8 __U, __m128 __D, __m128bh __A, __m128bh __B) {
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
(__v4sf)_mm_dpbf16_ps(__D, __A, __B),
(__v4sf)_mm_setzero_si128());
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \param __B
/// A 256-bit vector of [16 x bfloat].
/// \param __D
/// A 256-bit vector of [8 x float].
/// \returns A 256-bit vector of [8 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m256 __DEFAULT_FN_ATTRS256
_mm256_dpbf16_ps(__m256 __D, __m256bh __A, __m256bh __B) {
return (__m256)__builtin_ia32_dpbf16ps_256((__v8sf)__D,
(__v16bf)__A,
(__v16bf)__B);
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \param __B
/// A 256-bit vector of [16 x bfloat].
/// \param __D
/// A 256-bit vector of [8 x float].
/// \param __U
/// A 16-bit mask value specifying what is chosen for each element.
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D.
/// \returns A 256-bit vector of [8 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m256 __DEFAULT_FN_ATTRS256
_mm256_mask_dpbf16_ps(__m256 __D, __mmask8 __U, __m256bh __A, __m256bh __B) {
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
(__v8sf)_mm256_dpbf16_ps(__D, __A, __B),
(__v8sf)__D);
}
/// Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VDPBF16PS </c> instructions.
///
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \param __B
/// A 256-bit vector of [16 x bfloat].
/// \param __D
/// A 256-bit vector of [8 x float].
/// \param __U
/// A 8-bit mask value specifying what is chosen for each element.
/// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0.
/// \returns A 256-bit vector of [8 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m256 __DEFAULT_FN_ATTRS256
_mm256_maskz_dpbf16_ps(__mmask8 __U, __m256 __D, __m256bh __A, __m256bh __B) {
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
(__v8sf)_mm256_dpbf16_ps(__D, __A, __B),
(__v8sf)_mm256_setzero_si256());
}
/// Convert One Single float Data to One BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A float data.
/// \returns A bf16 data whose sign field and exponent field keep unchanged,
/// and fraction field is truncated to 7 bits.
static __inline__ __bf16 __DEFAULT_FN_ATTRS128 _mm_cvtness_sbh(float __A) {
__v4sf __V = {__A, 0, 0, 0};
__v8bf __R = __builtin_ia32_cvtneps2bf16_128_mask(
(__v4sf)__V, (__v8bf)_mm_undefined_si128(), (__mmask8)-1);
return (__bf16)__R[0];
}
/// Convert Packed BF16 Data to Packed float Data.
///
/// \headerfile <x86intrin.h>
///
/// \param __A
/// A 128-bit vector of [4 x bfloat].
/// \returns A 128-bit vector of [4 x float] come from conversion of __A
static __inline__ __m128 __DEFAULT_FN_ATTRS128 _mm_cvtpbh_ps(__m128bh __A) {
return _mm_castsi128_ps(
(__m128i)_mm_slli_epi32((__m128i)_mm_cvtepi16_epi32((__m128i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data.
///
/// \headerfile <x86intrin.h>
///
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \returns A 256-bit vector of [8 x float] come from conversion of __A
static __inline__ __m256 __DEFAULT_FN_ATTRS256 _mm256_cvtpbh_ps(__m128bh __A) {
return _mm256_castsi256_ps((__m256i)_mm256_slli_epi32(
(__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data using zeroing mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __U
/// A 4-bit mask. Elements are zeroed out when the corresponding mask
/// bit is not set.
/// \param __A
/// A 128-bit vector of [4 x bfloat].
/// \returns A 128-bit vector of [4 x float] come from conversion of __A
static __inline__ __m128 __DEFAULT_FN_ATTRS128
_mm_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) {
return _mm_castsi128_ps((__m128i)_mm_slli_epi32(
(__m128i)_mm_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data using zeroing mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __U
/// A 8-bit mask. Elements are zeroed out when the corresponding mask
/// bit is not set.
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \returns A 256-bit vector of [8 x float] come from conversion of __A
static __inline__ __m256 __DEFAULT_FN_ATTRS256
_mm256_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) {
return _mm256_castsi256_ps((__m256i)_mm256_slli_epi32(
(__m256i)_mm256_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data using merging mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __S
/// A 128-bit vector of [4 x float]. Elements are copied from __S when
/// the corresponding mask bit is not set.
/// \param __U
/// A 4-bit mask. Elements are zeroed out when the corresponding mask
/// bit is not set.
/// \param __A
/// A 128-bit vector of [4 x bfloat].
/// \returns A 128-bit vector of [4 x float] come from conversion of __A
static __inline__ __m128 __DEFAULT_FN_ATTRS128
_mm_mask_cvtpbh_ps(__m128 __S, __mmask8 __U, __m128bh __A) {
return _mm_castsi128_ps((__m128i)_mm_mask_slli_epi32(
(__m128i)__S, (__mmask8)__U, (__m128i)_mm_cvtepi16_epi32((__m128i)__A),
16));
}
/// Convert Packed BF16 Data to Packed float Data using merging mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __S
/// A 256-bit vector of [8 x float]. Elements are copied from __S when
/// the corresponding mask bit is not set.
/// \param __U
/// A 8-bit mask. Elements are zeroed out when the corresponding mask
/// bit is not set.
/// \param __A
/// A 128-bit vector of [8 x bfloat].
/// \returns A 256-bit vector of [8 x float] come from conversion of __A
static __inline__ __m256 __DEFAULT_FN_ATTRS256
_mm256_mask_cvtpbh_ps(__m256 __S, __mmask8 __U, __m128bh __A) {
return _mm256_castsi256_ps((__m256i)_mm256_mask_slli_epi32(
(__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A),
16));
}
#undef __DEFAULT_FN_ATTRS128
#undef __DEFAULT_FN_ATTRS256
#endif
#endif
|
