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
| /*===------------ avx512bf16intrin.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 <avx512bf16intrin.h> directly; include <immintrin.h> instead."
#endif
#ifndef __AVX512BF16INTRIN_H
#define __AVX512BF16INTRIN_H
typedef short __m512bh __attribute__((__vector_size__(64), __aligned__(64)));
typedef short __m256bh __attribute__((__vector_size__(32), __aligned__(32)));
typedef unsigned short __bfloat16;
#define __DEFAULT_FN_ATTRS512 \
__attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \
__min_vector_width__(512)))
#define __DEFAULT_FN_ATTRS \
__attribute__((__always_inline__, __nodebug__, __target__("avx512bf16")))
/// Convert One BF16 Data to One Single Float Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic does not correspond to a specific instruction.
///
/// \param __A
/// A bfloat data.
/// \returns A float data whose sign field and exponent field keep unchanged,
/// and fraction field is extended to 23 bits.
static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss(__bfloat16 __A) {
return __builtin_ia32_cvtsbf162ss_32(__A);
}
/// 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 512-bit vector of [16 x float].
/// \param __B
/// A 512-bit vector of [16 x float].
/// \returns A 512-bit vector of [32 x bfloat] whose lower 256 bits come from
/// conversion of __B, and higher 256 bits come from conversion of __A.
static __inline__ __m512bh __DEFAULT_FN_ATTRS512
_mm512_cvtne2ps_pbh(__m512 __A, __m512 __B) {
return (__m512bh)__builtin_ia32_cvtne2ps2bf16_512((__v16sf) __A,
(__v16sf) __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 512-bit vector of [16 x float].
/// \param __B
/// A 512-bit vector of [16 x float].
/// \param __W
/// A 512-bit vector of [32 x bfloat].
/// \param __U
/// A 32-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 512-bit vector of [32 x bfloat] whose lower 256 bits come from
/// conversion of __B, and higher 256 bits come from conversion of __A.
static __inline__ __m512bh __DEFAULT_FN_ATTRS512
_mm512_mask_cvtne2ps_pbh(__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B) {
return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U,
(__v32hi)_mm512_cvtne2ps_pbh(__A, __B),
(__v32hi)__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 512-bit vector of [16 x float].
/// \param __B
/// A 512-bit vector of [16 x float].
/// \param __U
/// A 32-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 512-bit vector of [32 x bfloat] whose lower 256 bits come from
/// conversion of __B, and higher 256 bits come from conversion of __A.
static __inline__ __m512bh __DEFAULT_FN_ATTRS512
_mm512_maskz_cvtne2ps_pbh(__mmask32 __U, __m512 __A, __m512 __B) {
return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U,
(__v32hi)_mm512_cvtne2ps_pbh(__A, __B),
(__v32hi)_mm512_setzero_si512());
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 512-bit vector of [16 x float].
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS512
_mm512_cvtneps_pbh(__m512 __A) {
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A,
(__v16hi)_mm256_undefined_si256(),
(__mmask16)-1);
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 512-bit vector of [16 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. A 0 means element from __W.
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS512
_mm512_mask_cvtneps_pbh(__m256bh __W, __mmask16 __U, __m512 __A) {
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A,
(__v16hi)__W,
(__mmask16)__U);
}
/// Convert Packed Single Data to Packed BF16 Data.
///
/// \headerfile <x86intrin.h>
///
/// This intrinsic corresponds to the <c> VCVTNEPS2BF16 </c> instructions.
///
/// \param __A
/// A 512-bit vector of [16 x float].
/// \param __U
/// A 16-bit mask value specifying what is chosen for each element.
/// A 1 means conversion of __A. A 0 means element is zero.
/// \returns A 256-bit vector of [16 x bfloat] come from conversion of __A.
static __inline__ __m256bh __DEFAULT_FN_ATTRS512
_mm512_maskz_cvtneps_pbh(__mmask16 __U, __m512 __A) {
return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A,
(__v16hi)_mm256_setzero_si256(),
(__mmask16)__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 512-bit vector of [32 x bfloat].
/// \param __B
/// A 512-bit vector of [32 x bfloat].
/// \param __D
/// A 512-bit vector of [16 x float].
/// \returns A 512-bit vector of [16 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m512 __DEFAULT_FN_ATTRS512
_mm512_dpbf16_ps(__m512 __D, __m512bh __A, __m512bh __B) {
return (__m512)__builtin_ia32_dpbf16ps_512((__v16sf) __D,
(__v16si) __A,
(__v16si) __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 512-bit vector of [32 x bfloat].
/// \param __B
/// A 512-bit vector of [32 x bfloat].
/// \param __D
/// A 512-bit vector of [16 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 512-bit vector of [16 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m512 __DEFAULT_FN_ATTRS512
_mm512_mask_dpbf16_ps(__m512 __D, __mmask16 __U, __m512bh __A, __m512bh __B) {
return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
(__v16sf)_mm512_dpbf16_ps(__D, __A, __B),
(__v16sf)__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 512-bit vector of [32 x bfloat].
/// \param __B
/// A 512-bit vector of [32 x bfloat].
/// \param __D
/// A 512-bit vector of [16 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 0.
/// \returns A 512-bit vector of [16 x float] comes from Dot Product of
/// __A, __B and __D
static __inline__ __m512 __DEFAULT_FN_ATTRS512
_mm512_maskz_dpbf16_ps(__mmask16 __U, __m512 __D, __m512bh __A, __m512bh __B) {
return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
(__v16sf)_mm512_dpbf16_ps(__D, __A, __B),
(__v16sf)_mm512_setzero_si512());
}
/// Convert Packed BF16 Data to Packed float Data.
///
/// \headerfile <x86intrin.h>
///
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \returns A 512-bit vector of [16 x float] come from convertion of __A
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps(__m256bh __A) {
return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32(
(__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data using zeroing mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __U
/// A 16-bit mask. Elements are zeroed out when the corresponding mask
/// bit is not set.
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \returns A 512-bit vector of [16 x float] come from convertion of __A
static __inline__ __m512 __DEFAULT_FN_ATTRS512
_mm512_maskz_cvtpbh_ps(__mmask16 __U, __m256bh __A) {
return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32(
(__m512i)_mm512_maskz_cvtepi16_epi32((__mmask16)__U, (__m256i)__A), 16));
}
/// Convert Packed BF16 Data to Packed float Data using merging mask.
///
/// \headerfile <x86intrin.h>
///
/// \param __S
/// A 512-bit vector of [16 x float]. Elements are copied from __S when
/// the corresponding mask bit is not set.
/// \param __U
/// A 16-bit mask.
/// \param __A
/// A 256-bit vector of [16 x bfloat].
/// \returns A 512-bit vector of [16 x float] come from convertion of __A
static __inline__ __m512 __DEFAULT_FN_ATTRS512
_mm512_mask_cvtpbh_ps(__m512 __S, __mmask16 __U, __m256bh __A) {
return _mm512_castsi512_ps((__m512i)_mm512_mask_slli_epi32(
(__m512i)__S, (__mmask16)__U,
(__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16));
}
#undef __DEFAULT_FN_ATTRS
#undef __DEFAULT_FN_ATTRS512
#endif
|