Warning: This file is not a C or C++ file. It does not have highlighting.
1 | /*===------------- avx512bwintrin.h - AVX512BW intrinsics ------------------=== |
---|---|
2 | * |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | * |
8 | *===-----------------------------------------------------------------------=== |
9 | */ |
10 | #ifndef __IMMINTRIN_H |
11 | #error "Never use <avx512bwintrin.h> directly; include <immintrin.h> instead." |
12 | #endif |
13 | |
14 | #ifndef __AVX512BWINTRIN_H |
15 | #define __AVX512BWINTRIN_H |
16 | |
17 | typedef unsigned int __mmask32; |
18 | typedef unsigned long long __mmask64; |
19 | |
20 | /* Define the default attributes for the functions in this file. */ |
21 | #define __DEFAULT_FN_ATTRS512 \ |
22 | __attribute__((__always_inline__, __nodebug__, \ |
23 | __target__("avx512bw,evex512"), __min_vector_width__(512))) |
24 | #define __DEFAULT_FN_ATTRS \ |
25 | __attribute__((__always_inline__, __nodebug__, \ |
26 | __target__("avx512bw,no-evex512"))) |
27 | |
28 | static __inline __mmask32 __DEFAULT_FN_ATTRS |
29 | _knot_mask32(__mmask32 __M) |
30 | { |
31 | return __builtin_ia32_knotsi(__M); |
32 | } |
33 | |
34 | static __inline __mmask64 __DEFAULT_FN_ATTRS _knot_mask64(__mmask64 __M) { |
35 | return __builtin_ia32_knotdi(__M); |
36 | } |
37 | |
38 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
39 | _kand_mask32(__mmask32 __A, __mmask32 __B) |
40 | { |
41 | return (__mmask32)__builtin_ia32_kandsi((__mmask32)__A, (__mmask32)__B); |
42 | } |
43 | |
44 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kand_mask64(__mmask64 __A, |
45 | __mmask64 __B) { |
46 | return (__mmask64)__builtin_ia32_kanddi((__mmask64)__A, (__mmask64)__B); |
47 | } |
48 | |
49 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
50 | _kandn_mask32(__mmask32 __A, __mmask32 __B) |
51 | { |
52 | return (__mmask32)__builtin_ia32_kandnsi((__mmask32)__A, (__mmask32)__B); |
53 | } |
54 | |
55 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kandn_mask64(__mmask64 __A, |
56 | __mmask64 __B) { |
57 | return (__mmask64)__builtin_ia32_kandndi((__mmask64)__A, (__mmask64)__B); |
58 | } |
59 | |
60 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
61 | _kor_mask32(__mmask32 __A, __mmask32 __B) |
62 | { |
63 | return (__mmask32)__builtin_ia32_korsi((__mmask32)__A, (__mmask32)__B); |
64 | } |
65 | |
66 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kor_mask64(__mmask64 __A, |
67 | __mmask64 __B) { |
68 | return (__mmask64)__builtin_ia32_kordi((__mmask64)__A, (__mmask64)__B); |
69 | } |
70 | |
71 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
72 | _kxnor_mask32(__mmask32 __A, __mmask32 __B) |
73 | { |
74 | return (__mmask32)__builtin_ia32_kxnorsi((__mmask32)__A, (__mmask32)__B); |
75 | } |
76 | |
77 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kxnor_mask64(__mmask64 __A, |
78 | __mmask64 __B) { |
79 | return (__mmask64)__builtin_ia32_kxnordi((__mmask64)__A, (__mmask64)__B); |
80 | } |
81 | |
82 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
83 | _kxor_mask32(__mmask32 __A, __mmask32 __B) |
84 | { |
85 | return (__mmask32)__builtin_ia32_kxorsi((__mmask32)__A, (__mmask32)__B); |
86 | } |
87 | |
88 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kxor_mask64(__mmask64 __A, |
89 | __mmask64 __B) { |
90 | return (__mmask64)__builtin_ia32_kxordi((__mmask64)__A, (__mmask64)__B); |
91 | } |
92 | |
93 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
94 | _kortestc_mask32_u8(__mmask32 __A, __mmask32 __B) |
95 | { |
96 | return (unsigned char)__builtin_ia32_kortestcsi(__A, __B); |
97 | } |
98 | |
99 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
100 | _kortestz_mask32_u8(__mmask32 __A, __mmask32 __B) |
101 | { |
102 | return (unsigned char)__builtin_ia32_kortestzsi(__A, __B); |
103 | } |
104 | |
105 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
106 | _kortest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) { |
107 | *__C = (unsigned char)__builtin_ia32_kortestcsi(__A, __B); |
108 | return (unsigned char)__builtin_ia32_kortestzsi(__A, __B); |
109 | } |
110 | |
111 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
112 | _kortestc_mask64_u8(__mmask64 __A, __mmask64 __B) { |
113 | return (unsigned char)__builtin_ia32_kortestcdi(__A, __B); |
114 | } |
115 | |
116 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
117 | _kortestz_mask64_u8(__mmask64 __A, __mmask64 __B) { |
118 | return (unsigned char)__builtin_ia32_kortestzdi(__A, __B); |
119 | } |
120 | |
121 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
122 | _kortest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C) { |
123 | *__C = (unsigned char)__builtin_ia32_kortestcdi(__A, __B); |
124 | return (unsigned char)__builtin_ia32_kortestzdi(__A, __B); |
125 | } |
126 | |
127 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
128 | _ktestc_mask32_u8(__mmask32 __A, __mmask32 __B) |
129 | { |
130 | return (unsigned char)__builtin_ia32_ktestcsi(__A, __B); |
131 | } |
132 | |
133 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
134 | _ktestz_mask32_u8(__mmask32 __A, __mmask32 __B) |
135 | { |
136 | return (unsigned char)__builtin_ia32_ktestzsi(__A, __B); |
137 | } |
138 | |
139 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
140 | _ktest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) { |
141 | *__C = (unsigned char)__builtin_ia32_ktestcsi(__A, __B); |
142 | return (unsigned char)__builtin_ia32_ktestzsi(__A, __B); |
143 | } |
144 | |
145 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
146 | _ktestc_mask64_u8(__mmask64 __A, __mmask64 __B) { |
147 | return (unsigned char)__builtin_ia32_ktestcdi(__A, __B); |
148 | } |
149 | |
150 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
151 | _ktestz_mask64_u8(__mmask64 __A, __mmask64 __B) { |
152 | return (unsigned char)__builtin_ia32_ktestzdi(__A, __B); |
153 | } |
154 | |
155 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
156 | _ktest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C) { |
157 | *__C = (unsigned char)__builtin_ia32_ktestcdi(__A, __B); |
158 | return (unsigned char)__builtin_ia32_ktestzdi(__A, __B); |
159 | } |
160 | |
161 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
162 | _kadd_mask32(__mmask32 __A, __mmask32 __B) |
163 | { |
164 | return (__mmask32)__builtin_ia32_kaddsi((__mmask32)__A, (__mmask32)__B); |
165 | } |
166 | |
167 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _kadd_mask64(__mmask64 __A, |
168 | __mmask64 __B) { |
169 | return (__mmask64)__builtin_ia32_kadddi((__mmask64)__A, (__mmask64)__B); |
170 | } |
171 | |
172 | #define _kshiftli_mask32(A, I) \ |
173 | ((__mmask32)__builtin_ia32_kshiftlisi((__mmask32)(A), (unsigned int)(I))) |
174 | |
175 | #define _kshiftri_mask32(A, I) \ |
176 | ((__mmask32)__builtin_ia32_kshiftrisi((__mmask32)(A), (unsigned int)(I))) |
177 | |
178 | #define _kshiftli_mask64(A, I) \ |
179 | ((__mmask64)__builtin_ia32_kshiftlidi((__mmask64)(A), (unsigned int)(I))) |
180 | |
181 | #define _kshiftri_mask64(A, I) \ |
182 | ((__mmask64)__builtin_ia32_kshiftridi((__mmask64)(A), (unsigned int)(I))) |
183 | |
184 | static __inline__ unsigned int __DEFAULT_FN_ATTRS |
185 | _cvtmask32_u32(__mmask32 __A) { |
186 | return (unsigned int)__builtin_ia32_kmovd((__mmask32)__A); |
187 | } |
188 | |
189 | static __inline__ unsigned long long __DEFAULT_FN_ATTRS |
190 | _cvtmask64_u64(__mmask64 __A) { |
191 | return (unsigned long long)__builtin_ia32_kmovq((__mmask64)__A); |
192 | } |
193 | |
194 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
195 | _cvtu32_mask32(unsigned int __A) { |
196 | return (__mmask32)__builtin_ia32_kmovd((__mmask32)__A); |
197 | } |
198 | |
199 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
200 | _cvtu64_mask64(unsigned long long __A) { |
201 | return (__mmask64)__builtin_ia32_kmovq((__mmask64)__A); |
202 | } |
203 | |
204 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
205 | _load_mask32(__mmask32 *__A) { |
206 | return (__mmask32)__builtin_ia32_kmovd(*(__mmask32 *)__A); |
207 | } |
208 | |
209 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _load_mask64(__mmask64 *__A) { |
210 | return (__mmask64)__builtin_ia32_kmovq(*(__mmask64 *)__A); |
211 | } |
212 | |
213 | static __inline__ void __DEFAULT_FN_ATTRS |
214 | _store_mask32(__mmask32 *__A, __mmask32 __B) { |
215 | *(__mmask32 *)__A = __builtin_ia32_kmovd((__mmask32)__B); |
216 | } |
217 | |
218 | static __inline__ void __DEFAULT_FN_ATTRS _store_mask64(__mmask64 *__A, |
219 | __mmask64 __B) { |
220 | *(__mmask64 *)__A = __builtin_ia32_kmovq((__mmask64)__B); |
221 | } |
222 | |
223 | /* Integer compare */ |
224 | |
225 | #define _mm512_cmp_epi8_mask(a, b, p) \ |
226 | ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \ |
227 | (__v64qi)(__m512i)(b), (int)(p), \ |
228 | (__mmask64)-1)) |
229 | |
230 | #define _mm512_mask_cmp_epi8_mask(m, a, b, p) \ |
231 | ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \ |
232 | (__v64qi)(__m512i)(b), (int)(p), \ |
233 | (__mmask64)(m))) |
234 | |
235 | #define _mm512_cmp_epu8_mask(a, b, p) \ |
236 | ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \ |
237 | (__v64qi)(__m512i)(b), (int)(p), \ |
238 | (__mmask64)-1)) |
239 | |
240 | #define _mm512_mask_cmp_epu8_mask(m, a, b, p) \ |
241 | ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \ |
242 | (__v64qi)(__m512i)(b), (int)(p), \ |
243 | (__mmask64)(m))) |
244 | |
245 | #define _mm512_cmp_epi16_mask(a, b, p) \ |
246 | ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \ |
247 | (__v32hi)(__m512i)(b), (int)(p), \ |
248 | (__mmask32)-1)) |
249 | |
250 | #define _mm512_mask_cmp_epi16_mask(m, a, b, p) \ |
251 | ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \ |
252 | (__v32hi)(__m512i)(b), (int)(p), \ |
253 | (__mmask32)(m))) |
254 | |
255 | #define _mm512_cmp_epu16_mask(a, b, p) \ |
256 | ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \ |
257 | (__v32hi)(__m512i)(b), (int)(p), \ |
258 | (__mmask32)-1)) |
259 | |
260 | #define _mm512_mask_cmp_epu16_mask(m, a, b, p) \ |
261 | ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \ |
262 | (__v32hi)(__m512i)(b), (int)(p), \ |
263 | (__mmask32)(m))) |
264 | |
265 | #define _mm512_cmpeq_epi8_mask(A, B) \ |
266 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ) |
267 | #define _mm512_mask_cmpeq_epi8_mask(k, A, B) \ |
268 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ) |
269 | #define _mm512_cmpge_epi8_mask(A, B) \ |
270 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GE) |
271 | #define _mm512_mask_cmpge_epi8_mask(k, A, B) \ |
272 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE) |
273 | #define _mm512_cmpgt_epi8_mask(A, B) \ |
274 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GT) |
275 | #define _mm512_mask_cmpgt_epi8_mask(k, A, B) \ |
276 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT) |
277 | #define _mm512_cmple_epi8_mask(A, B) \ |
278 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LE) |
279 | #define _mm512_mask_cmple_epi8_mask(k, A, B) \ |
280 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE) |
281 | #define _mm512_cmplt_epi8_mask(A, B) \ |
282 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LT) |
283 | #define _mm512_mask_cmplt_epi8_mask(k, A, B) \ |
284 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT) |
285 | #define _mm512_cmpneq_epi8_mask(A, B) \ |
286 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_NE) |
287 | #define _mm512_mask_cmpneq_epi8_mask(k, A, B) \ |
288 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE) |
289 | |
290 | #define _mm512_cmpeq_epu8_mask(A, B) \ |
291 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ) |
292 | #define _mm512_mask_cmpeq_epu8_mask(k, A, B) \ |
293 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ) |
294 | #define _mm512_cmpge_epu8_mask(A, B) \ |
295 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GE) |
296 | #define _mm512_mask_cmpge_epu8_mask(k, A, B) \ |
297 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE) |
298 | #define _mm512_cmpgt_epu8_mask(A, B) \ |
299 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GT) |
300 | #define _mm512_mask_cmpgt_epu8_mask(k, A, B) \ |
301 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT) |
302 | #define _mm512_cmple_epu8_mask(A, B) \ |
303 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LE) |
304 | #define _mm512_mask_cmple_epu8_mask(k, A, B) \ |
305 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE) |
306 | #define _mm512_cmplt_epu8_mask(A, B) \ |
307 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LT) |
308 | #define _mm512_mask_cmplt_epu8_mask(k, A, B) \ |
309 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT) |
310 | #define _mm512_cmpneq_epu8_mask(A, B) \ |
311 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_NE) |
312 | #define _mm512_mask_cmpneq_epu8_mask(k, A, B) \ |
313 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE) |
314 | |
315 | #define _mm512_cmpeq_epi16_mask(A, B) \ |
316 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ) |
317 | #define _mm512_mask_cmpeq_epi16_mask(k, A, B) \ |
318 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ) |
319 | #define _mm512_cmpge_epi16_mask(A, B) \ |
320 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GE) |
321 | #define _mm512_mask_cmpge_epi16_mask(k, A, B) \ |
322 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE) |
323 | #define _mm512_cmpgt_epi16_mask(A, B) \ |
324 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GT) |
325 | #define _mm512_mask_cmpgt_epi16_mask(k, A, B) \ |
326 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT) |
327 | #define _mm512_cmple_epi16_mask(A, B) \ |
328 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LE) |
329 | #define _mm512_mask_cmple_epi16_mask(k, A, B) \ |
330 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE) |
331 | #define _mm512_cmplt_epi16_mask(A, B) \ |
332 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LT) |
333 | #define _mm512_mask_cmplt_epi16_mask(k, A, B) \ |
334 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT) |
335 | #define _mm512_cmpneq_epi16_mask(A, B) \ |
336 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_NE) |
337 | #define _mm512_mask_cmpneq_epi16_mask(k, A, B) \ |
338 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE) |
339 | |
340 | #define _mm512_cmpeq_epu16_mask(A, B) \ |
341 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ) |
342 | #define _mm512_mask_cmpeq_epu16_mask(k, A, B) \ |
343 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ) |
344 | #define _mm512_cmpge_epu16_mask(A, B) \ |
345 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GE) |
346 | #define _mm512_mask_cmpge_epu16_mask(k, A, B) \ |
347 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE) |
348 | #define _mm512_cmpgt_epu16_mask(A, B) \ |
349 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GT) |
350 | #define _mm512_mask_cmpgt_epu16_mask(k, A, B) \ |
351 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT) |
352 | #define _mm512_cmple_epu16_mask(A, B) \ |
353 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LE) |
354 | #define _mm512_mask_cmple_epu16_mask(k, A, B) \ |
355 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE) |
356 | #define _mm512_cmplt_epu16_mask(A, B) \ |
357 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LT) |
358 | #define _mm512_mask_cmplt_epu16_mask(k, A, B) \ |
359 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT) |
360 | #define _mm512_cmpneq_epu16_mask(A, B) \ |
361 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_NE) |
362 | #define _mm512_mask_cmpneq_epu16_mask(k, A, B) \ |
363 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE) |
364 | |
365 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
366 | _mm512_add_epi8 (__m512i __A, __m512i __B) { |
367 | return (__m512i) ((__v64qu) __A + (__v64qu) __B); |
368 | } |
369 | |
370 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
371 | _mm512_mask_add_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
372 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
373 | (__v64qi)_mm512_add_epi8(__A, __B), |
374 | (__v64qi)__W); |
375 | } |
376 | |
377 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
378 | _mm512_maskz_add_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
379 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
380 | (__v64qi)_mm512_add_epi8(__A, __B), |
381 | (__v64qi)_mm512_setzero_si512()); |
382 | } |
383 | |
384 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
385 | _mm512_sub_epi8 (__m512i __A, __m512i __B) { |
386 | return (__m512i) ((__v64qu) __A - (__v64qu) __B); |
387 | } |
388 | |
389 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
390 | _mm512_mask_sub_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
391 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
392 | (__v64qi)_mm512_sub_epi8(__A, __B), |
393 | (__v64qi)__W); |
394 | } |
395 | |
396 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
397 | _mm512_maskz_sub_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
398 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
399 | (__v64qi)_mm512_sub_epi8(__A, __B), |
400 | (__v64qi)_mm512_setzero_si512()); |
401 | } |
402 | |
403 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
404 | _mm512_add_epi16 (__m512i __A, __m512i __B) { |
405 | return (__m512i) ((__v32hu) __A + (__v32hu) __B); |
406 | } |
407 | |
408 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
409 | _mm512_mask_add_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
410 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
411 | (__v32hi)_mm512_add_epi16(__A, __B), |
412 | (__v32hi)__W); |
413 | } |
414 | |
415 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
416 | _mm512_maskz_add_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
417 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
418 | (__v32hi)_mm512_add_epi16(__A, __B), |
419 | (__v32hi)_mm512_setzero_si512()); |
420 | } |
421 | |
422 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
423 | _mm512_sub_epi16 (__m512i __A, __m512i __B) { |
424 | return (__m512i) ((__v32hu) __A - (__v32hu) __B); |
425 | } |
426 | |
427 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
428 | _mm512_mask_sub_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
429 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
430 | (__v32hi)_mm512_sub_epi16(__A, __B), |
431 | (__v32hi)__W); |
432 | } |
433 | |
434 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
435 | _mm512_maskz_sub_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
436 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
437 | (__v32hi)_mm512_sub_epi16(__A, __B), |
438 | (__v32hi)_mm512_setzero_si512()); |
439 | } |
440 | |
441 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
442 | _mm512_mullo_epi16 (__m512i __A, __m512i __B) { |
443 | return (__m512i) ((__v32hu) __A * (__v32hu) __B); |
444 | } |
445 | |
446 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
447 | _mm512_mask_mullo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
448 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
449 | (__v32hi)_mm512_mullo_epi16(__A, __B), |
450 | (__v32hi)__W); |
451 | } |
452 | |
453 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
454 | _mm512_maskz_mullo_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
455 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
456 | (__v32hi)_mm512_mullo_epi16(__A, __B), |
457 | (__v32hi)_mm512_setzero_si512()); |
458 | } |
459 | |
460 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
461 | _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W) |
462 | { |
463 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
464 | (__v64qi) __W, |
465 | (__v64qi) __A); |
466 | } |
467 | |
468 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
469 | _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) |
470 | { |
471 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
472 | (__v32hi) __W, |
473 | (__v32hi) __A); |
474 | } |
475 | |
476 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
477 | _mm512_abs_epi8 (__m512i __A) |
478 | { |
479 | return (__m512i)__builtin_elementwise_abs((__v64qs)__A); |
480 | } |
481 | |
482 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
483 | _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) |
484 | { |
485 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
486 | (__v64qi)_mm512_abs_epi8(__A), |
487 | (__v64qi)__W); |
488 | } |
489 | |
490 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
491 | _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) |
492 | { |
493 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
494 | (__v64qi)_mm512_abs_epi8(__A), |
495 | (__v64qi)_mm512_setzero_si512()); |
496 | } |
497 | |
498 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
499 | _mm512_abs_epi16 (__m512i __A) |
500 | { |
501 | return (__m512i)__builtin_elementwise_abs((__v32hi)__A); |
502 | } |
503 | |
504 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
505 | _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) |
506 | { |
507 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
508 | (__v32hi)_mm512_abs_epi16(__A), |
509 | (__v32hi)__W); |
510 | } |
511 | |
512 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
513 | _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) |
514 | { |
515 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
516 | (__v32hi)_mm512_abs_epi16(__A), |
517 | (__v32hi)_mm512_setzero_si512()); |
518 | } |
519 | |
520 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
521 | _mm512_packs_epi32(__m512i __A, __m512i __B) |
522 | { |
523 | return (__m512i)__builtin_ia32_packssdw512((__v16si)__A, (__v16si)__B); |
524 | } |
525 | |
526 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
527 | _mm512_maskz_packs_epi32(__mmask32 __M, __m512i __A, __m512i __B) |
528 | { |
529 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
530 | (__v32hi)_mm512_packs_epi32(__A, __B), |
531 | (__v32hi)_mm512_setzero_si512()); |
532 | } |
533 | |
534 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
535 | _mm512_mask_packs_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
536 | { |
537 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
538 | (__v32hi)_mm512_packs_epi32(__A, __B), |
539 | (__v32hi)__W); |
540 | } |
541 | |
542 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
543 | _mm512_packs_epi16(__m512i __A, __m512i __B) |
544 | { |
545 | return (__m512i)__builtin_ia32_packsswb512((__v32hi)__A, (__v32hi) __B); |
546 | } |
547 | |
548 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
549 | _mm512_mask_packs_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
550 | { |
551 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
552 | (__v64qi)_mm512_packs_epi16(__A, __B), |
553 | (__v64qi)__W); |
554 | } |
555 | |
556 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
557 | _mm512_maskz_packs_epi16(__mmask64 __M, __m512i __A, __m512i __B) |
558 | { |
559 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
560 | (__v64qi)_mm512_packs_epi16(__A, __B), |
561 | (__v64qi)_mm512_setzero_si512()); |
562 | } |
563 | |
564 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
565 | _mm512_packus_epi32(__m512i __A, __m512i __B) |
566 | { |
567 | return (__m512i)__builtin_ia32_packusdw512((__v16si) __A, (__v16si) __B); |
568 | } |
569 | |
570 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
571 | _mm512_maskz_packus_epi32(__mmask32 __M, __m512i __A, __m512i __B) |
572 | { |
573 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
574 | (__v32hi)_mm512_packus_epi32(__A, __B), |
575 | (__v32hi)_mm512_setzero_si512()); |
576 | } |
577 | |
578 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
579 | _mm512_mask_packus_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
580 | { |
581 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
582 | (__v32hi)_mm512_packus_epi32(__A, __B), |
583 | (__v32hi)__W); |
584 | } |
585 | |
586 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
587 | _mm512_packus_epi16(__m512i __A, __m512i __B) |
588 | { |
589 | return (__m512i)__builtin_ia32_packuswb512((__v32hi) __A, (__v32hi) __B); |
590 | } |
591 | |
592 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
593 | _mm512_mask_packus_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
594 | { |
595 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
596 | (__v64qi)_mm512_packus_epi16(__A, __B), |
597 | (__v64qi)__W); |
598 | } |
599 | |
600 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
601 | _mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B) |
602 | { |
603 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
604 | (__v64qi)_mm512_packus_epi16(__A, __B), |
605 | (__v64qi)_mm512_setzero_si512()); |
606 | } |
607 | |
608 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
609 | _mm512_adds_epi8 (__m512i __A, __m512i __B) |
610 | { |
611 | return (__m512i)__builtin_elementwise_add_sat((__v64qs)__A, (__v64qs)__B); |
612 | } |
613 | |
614 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
615 | _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
616 | { |
617 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
618 | (__v64qi)_mm512_adds_epi8(__A, __B), |
619 | (__v64qi)__W); |
620 | } |
621 | |
622 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
623 | _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) |
624 | { |
625 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
626 | (__v64qi)_mm512_adds_epi8(__A, __B), |
627 | (__v64qi)_mm512_setzero_si512()); |
628 | } |
629 | |
630 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
631 | _mm512_adds_epi16 (__m512i __A, __m512i __B) |
632 | { |
633 | return (__m512i)__builtin_elementwise_add_sat((__v32hi)__A, (__v32hi)__B); |
634 | } |
635 | |
636 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
637 | _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
638 | { |
639 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
640 | (__v32hi)_mm512_adds_epi16(__A, __B), |
641 | (__v32hi)__W); |
642 | } |
643 | |
644 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
645 | _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) |
646 | { |
647 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
648 | (__v32hi)_mm512_adds_epi16(__A, __B), |
649 | (__v32hi)_mm512_setzero_si512()); |
650 | } |
651 | |
652 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
653 | _mm512_adds_epu8 (__m512i __A, __m512i __B) |
654 | { |
655 | return (__m512i)__builtin_elementwise_add_sat((__v64qu) __A, (__v64qu) __B); |
656 | } |
657 | |
658 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
659 | _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
660 | { |
661 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
662 | (__v64qi)_mm512_adds_epu8(__A, __B), |
663 | (__v64qi)__W); |
664 | } |
665 | |
666 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
667 | _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
668 | { |
669 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
670 | (__v64qi)_mm512_adds_epu8(__A, __B), |
671 | (__v64qi)_mm512_setzero_si512()); |
672 | } |
673 | |
674 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
675 | _mm512_adds_epu16 (__m512i __A, __m512i __B) |
676 | { |
677 | return (__m512i)__builtin_elementwise_add_sat((__v32hu) __A, (__v32hu) __B); |
678 | } |
679 | |
680 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
681 | _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
682 | { |
683 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
684 | (__v32hi)_mm512_adds_epu16(__A, __B), |
685 | (__v32hi)__W); |
686 | } |
687 | |
688 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
689 | _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
690 | { |
691 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
692 | (__v32hi)_mm512_adds_epu16(__A, __B), |
693 | (__v32hi)_mm512_setzero_si512()); |
694 | } |
695 | |
696 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
697 | _mm512_avg_epu8 (__m512i __A, __m512i __B) |
698 | { |
699 | return (__m512i)__builtin_ia32_pavgb512((__v64qi)__A, (__v64qi)__B); |
700 | } |
701 | |
702 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
703 | _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A, |
704 | __m512i __B) |
705 | { |
706 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
707 | (__v64qi)_mm512_avg_epu8(__A, __B), |
708 | (__v64qi)__W); |
709 | } |
710 | |
711 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
712 | _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
713 | { |
714 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
715 | (__v64qi)_mm512_avg_epu8(__A, __B), |
716 | (__v64qi)_mm512_setzero_si512()); |
717 | } |
718 | |
719 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
720 | _mm512_avg_epu16 (__m512i __A, __m512i __B) |
721 | { |
722 | return (__m512i)__builtin_ia32_pavgw512((__v32hi)__A, (__v32hi)__B); |
723 | } |
724 | |
725 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
726 | _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A, |
727 | __m512i __B) |
728 | { |
729 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
730 | (__v32hi)_mm512_avg_epu16(__A, __B), |
731 | (__v32hi)__W); |
732 | } |
733 | |
734 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
735 | _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
736 | { |
737 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
738 | (__v32hi)_mm512_avg_epu16(__A, __B), |
739 | (__v32hi) _mm512_setzero_si512()); |
740 | } |
741 | |
742 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
743 | _mm512_max_epi8 (__m512i __A, __m512i __B) |
744 | { |
745 | return (__m512i)__builtin_elementwise_max((__v64qs) __A, (__v64qs) __B); |
746 | } |
747 | |
748 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
749 | _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) |
750 | { |
751 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
752 | (__v64qi)_mm512_max_epi8(__A, __B), |
753 | (__v64qi)_mm512_setzero_si512()); |
754 | } |
755 | |
756 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
757 | _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
758 | { |
759 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
760 | (__v64qi)_mm512_max_epi8(__A, __B), |
761 | (__v64qi)__W); |
762 | } |
763 | |
764 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
765 | _mm512_max_epi16 (__m512i __A, __m512i __B) |
766 | { |
767 | return (__m512i)__builtin_elementwise_max((__v32hi) __A, (__v32hi) __B); |
768 | } |
769 | |
770 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
771 | _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) |
772 | { |
773 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
774 | (__v32hi)_mm512_max_epi16(__A, __B), |
775 | (__v32hi)_mm512_setzero_si512()); |
776 | } |
777 | |
778 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
779 | _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, |
780 | __m512i __B) |
781 | { |
782 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
783 | (__v32hi)_mm512_max_epi16(__A, __B), |
784 | (__v32hi)__W); |
785 | } |
786 | |
787 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
788 | _mm512_max_epu8 (__m512i __A, __m512i __B) |
789 | { |
790 | return (__m512i)__builtin_elementwise_max((__v64qu)__A, (__v64qu)__B); |
791 | } |
792 | |
793 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
794 | _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) |
795 | { |
796 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
797 | (__v64qi)_mm512_max_epu8(__A, __B), |
798 | (__v64qi)_mm512_setzero_si512()); |
799 | } |
800 | |
801 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
802 | _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
803 | { |
804 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
805 | (__v64qi)_mm512_max_epu8(__A, __B), |
806 | (__v64qi)__W); |
807 | } |
808 | |
809 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
810 | _mm512_max_epu16 (__m512i __A, __m512i __B) |
811 | { |
812 | return (__m512i)__builtin_elementwise_max((__v32hu)__A, (__v32hu)__B); |
813 | } |
814 | |
815 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
816 | _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) |
817 | { |
818 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
819 | (__v32hi)_mm512_max_epu16(__A, __B), |
820 | (__v32hi)_mm512_setzero_si512()); |
821 | } |
822 | |
823 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
824 | _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
825 | { |
826 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
827 | (__v32hi)_mm512_max_epu16(__A, __B), |
828 | (__v32hi)__W); |
829 | } |
830 | |
831 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
832 | _mm512_min_epi8 (__m512i __A, __m512i __B) |
833 | { |
834 | return (__m512i)__builtin_elementwise_min((__v64qs) __A, (__v64qs) __B); |
835 | } |
836 | |
837 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
838 | _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) |
839 | { |
840 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
841 | (__v64qi)_mm512_min_epi8(__A, __B), |
842 | (__v64qi)_mm512_setzero_si512()); |
843 | } |
844 | |
845 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
846 | _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
847 | { |
848 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
849 | (__v64qi)_mm512_min_epi8(__A, __B), |
850 | (__v64qi)__W); |
851 | } |
852 | |
853 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
854 | _mm512_min_epi16 (__m512i __A, __m512i __B) |
855 | { |
856 | return (__m512i)__builtin_elementwise_min((__v32hi) __A, (__v32hi) __B); |
857 | } |
858 | |
859 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
860 | _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) |
861 | { |
862 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
863 | (__v32hi)_mm512_min_epi16(__A, __B), |
864 | (__v32hi)_mm512_setzero_si512()); |
865 | } |
866 | |
867 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
868 | _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
869 | { |
870 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
871 | (__v32hi)_mm512_min_epi16(__A, __B), |
872 | (__v32hi)__W); |
873 | } |
874 | |
875 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
876 | _mm512_min_epu8 (__m512i __A, __m512i __B) |
877 | { |
878 | return (__m512i)__builtin_elementwise_min((__v64qu)__A, (__v64qu)__B); |
879 | } |
880 | |
881 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
882 | _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) |
883 | { |
884 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
885 | (__v64qi)_mm512_min_epu8(__A, __B), |
886 | (__v64qi)_mm512_setzero_si512()); |
887 | } |
888 | |
889 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
890 | _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
891 | { |
892 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
893 | (__v64qi)_mm512_min_epu8(__A, __B), |
894 | (__v64qi)__W); |
895 | } |
896 | |
897 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
898 | _mm512_min_epu16 (__m512i __A, __m512i __B) |
899 | { |
900 | return (__m512i)__builtin_elementwise_min((__v32hu)__A, (__v32hu)__B); |
901 | } |
902 | |
903 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
904 | _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) |
905 | { |
906 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
907 | (__v32hi)_mm512_min_epu16(__A, __B), |
908 | (__v32hi)_mm512_setzero_si512()); |
909 | } |
910 | |
911 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
912 | _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
913 | { |
914 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
915 | (__v32hi)_mm512_min_epu16(__A, __B), |
916 | (__v32hi)__W); |
917 | } |
918 | |
919 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
920 | _mm512_shuffle_epi8(__m512i __A, __m512i __B) |
921 | { |
922 | return (__m512i)__builtin_ia32_pshufb512((__v64qi)__A,(__v64qi)__B); |
923 | } |
924 | |
925 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
926 | _mm512_mask_shuffle_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
927 | { |
928 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
929 | (__v64qi)_mm512_shuffle_epi8(__A, __B), |
930 | (__v64qi)__W); |
931 | } |
932 | |
933 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
934 | _mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B) |
935 | { |
936 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
937 | (__v64qi)_mm512_shuffle_epi8(__A, __B), |
938 | (__v64qi)_mm512_setzero_si512()); |
939 | } |
940 | |
941 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
942 | _mm512_subs_epi8 (__m512i __A, __m512i __B) |
943 | { |
944 | return (__m512i)__builtin_elementwise_sub_sat((__v64qs)__A, (__v64qs)__B); |
945 | } |
946 | |
947 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
948 | _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
949 | { |
950 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
951 | (__v64qi)_mm512_subs_epi8(__A, __B), |
952 | (__v64qi)__W); |
953 | } |
954 | |
955 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
956 | _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) |
957 | { |
958 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
959 | (__v64qi)_mm512_subs_epi8(__A, __B), |
960 | (__v64qi)_mm512_setzero_si512()); |
961 | } |
962 | |
963 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
964 | _mm512_subs_epi16 (__m512i __A, __m512i __B) |
965 | { |
966 | return (__m512i)__builtin_elementwise_sub_sat((__v32hi)__A, (__v32hi)__B); |
967 | } |
968 | |
969 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
970 | _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
971 | { |
972 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
973 | (__v32hi)_mm512_subs_epi16(__A, __B), |
974 | (__v32hi)__W); |
975 | } |
976 | |
977 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
978 | _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) |
979 | { |
980 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
981 | (__v32hi)_mm512_subs_epi16(__A, __B), |
982 | (__v32hi)_mm512_setzero_si512()); |
983 | } |
984 | |
985 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
986 | _mm512_subs_epu8 (__m512i __A, __m512i __B) |
987 | { |
988 | return (__m512i)__builtin_elementwise_sub_sat((__v64qu) __A, (__v64qu) __B); |
989 | } |
990 | |
991 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
992 | _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
993 | { |
994 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
995 | (__v64qi)_mm512_subs_epu8(__A, __B), |
996 | (__v64qi)__W); |
997 | } |
998 | |
999 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1000 | _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
1001 | { |
1002 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
1003 | (__v64qi)_mm512_subs_epu8(__A, __B), |
1004 | (__v64qi)_mm512_setzero_si512()); |
1005 | } |
1006 | |
1007 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1008 | _mm512_subs_epu16 (__m512i __A, __m512i __B) |
1009 | { |
1010 | return (__m512i)__builtin_elementwise_sub_sat((__v32hu) __A, (__v32hu) __B); |
1011 | } |
1012 | |
1013 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1014 | _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1015 | { |
1016 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1017 | (__v32hi)_mm512_subs_epu16(__A, __B), |
1018 | (__v32hi)__W); |
1019 | } |
1020 | |
1021 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1022 | _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
1023 | { |
1024 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1025 | (__v32hi)_mm512_subs_epu16(__A, __B), |
1026 | (__v32hi)_mm512_setzero_si512()); |
1027 | } |
1028 | |
1029 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1030 | _mm512_permutex2var_epi16(__m512i __A, __m512i __I, __m512i __B) |
1031 | { |
1032 | return (__m512i)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I, |
1033 | (__v32hi)__B); |
1034 | } |
1035 | |
1036 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1037 | _mm512_mask_permutex2var_epi16(__m512i __A, __mmask32 __U, __m512i __I, |
1038 | __m512i __B) |
1039 | { |
1040 | return (__m512i)__builtin_ia32_selectw_512(__U, |
1041 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
1042 | (__v32hi)__A); |
1043 | } |
1044 | |
1045 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1046 | _mm512_mask2_permutex2var_epi16(__m512i __A, __m512i __I, __mmask32 __U, |
1047 | __m512i __B) |
1048 | { |
1049 | return (__m512i)__builtin_ia32_selectw_512(__U, |
1050 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
1051 | (__v32hi)__I); |
1052 | } |
1053 | |
1054 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1055 | _mm512_maskz_permutex2var_epi16(__mmask32 __U, __m512i __A, __m512i __I, |
1056 | __m512i __B) |
1057 | { |
1058 | return (__m512i)__builtin_ia32_selectw_512(__U, |
1059 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
1060 | (__v32hi)_mm512_setzero_si512()); |
1061 | } |
1062 | |
1063 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1064 | _mm512_mulhrs_epi16(__m512i __A, __m512i __B) |
1065 | { |
1066 | return (__m512i)__builtin_ia32_pmulhrsw512((__v32hi)__A, (__v32hi)__B); |
1067 | } |
1068 | |
1069 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1070 | _mm512_mask_mulhrs_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1071 | { |
1072 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1073 | (__v32hi)_mm512_mulhrs_epi16(__A, __B), |
1074 | (__v32hi)__W); |
1075 | } |
1076 | |
1077 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1078 | _mm512_maskz_mulhrs_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
1079 | { |
1080 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1081 | (__v32hi)_mm512_mulhrs_epi16(__A, __B), |
1082 | (__v32hi)_mm512_setzero_si512()); |
1083 | } |
1084 | |
1085 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1086 | _mm512_mulhi_epi16(__m512i __A, __m512i __B) |
1087 | { |
1088 | return (__m512i)__builtin_ia32_pmulhw512((__v32hi) __A, (__v32hi) __B); |
1089 | } |
1090 | |
1091 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1092 | _mm512_mask_mulhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
1093 | __m512i __B) |
1094 | { |
1095 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1096 | (__v32hi)_mm512_mulhi_epi16(__A, __B), |
1097 | (__v32hi)__W); |
1098 | } |
1099 | |
1100 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1101 | _mm512_maskz_mulhi_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
1102 | { |
1103 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1104 | (__v32hi)_mm512_mulhi_epi16(__A, __B), |
1105 | (__v32hi)_mm512_setzero_si512()); |
1106 | } |
1107 | |
1108 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1109 | _mm512_mulhi_epu16(__m512i __A, __m512i __B) |
1110 | { |
1111 | return (__m512i)__builtin_ia32_pmulhuw512((__v32hi) __A, (__v32hi) __B); |
1112 | } |
1113 | |
1114 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1115 | _mm512_mask_mulhi_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1116 | { |
1117 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1118 | (__v32hi)_mm512_mulhi_epu16(__A, __B), |
1119 | (__v32hi)__W); |
1120 | } |
1121 | |
1122 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1123 | _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
1124 | { |
1125 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1126 | (__v32hi)_mm512_mulhi_epu16(__A, __B), |
1127 | (__v32hi)_mm512_setzero_si512()); |
1128 | } |
1129 | |
1130 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1131 | _mm512_maddubs_epi16(__m512i __X, __m512i __Y) { |
1132 | return (__m512i)__builtin_ia32_pmaddubsw512((__v64qi)__X, (__v64qi)__Y); |
1133 | } |
1134 | |
1135 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1136 | _mm512_mask_maddubs_epi16(__m512i __W, __mmask32 __U, __m512i __X, |
1137 | __m512i __Y) { |
1138 | return (__m512i)__builtin_ia32_selectw_512((__mmask32) __U, |
1139 | (__v32hi)_mm512_maddubs_epi16(__X, __Y), |
1140 | (__v32hi)__W); |
1141 | } |
1142 | |
1143 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1144 | _mm512_maskz_maddubs_epi16(__mmask32 __U, __m512i __X, __m512i __Y) { |
1145 | return (__m512i)__builtin_ia32_selectw_512((__mmask32) __U, |
1146 | (__v32hi)_mm512_maddubs_epi16(__X, __Y), |
1147 | (__v32hi)_mm512_setzero_si512()); |
1148 | } |
1149 | |
1150 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1151 | _mm512_madd_epi16(__m512i __A, __m512i __B) { |
1152 | return (__m512i)__builtin_ia32_pmaddwd512((__v32hi)__A, (__v32hi)__B); |
1153 | } |
1154 | |
1155 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1156 | _mm512_mask_madd_epi16(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) { |
1157 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__U, |
1158 | (__v16si)_mm512_madd_epi16(__A, __B), |
1159 | (__v16si)__W); |
1160 | } |
1161 | |
1162 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1163 | _mm512_maskz_madd_epi16(__mmask16 __U, __m512i __A, __m512i __B) { |
1164 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__U, |
1165 | (__v16si)_mm512_madd_epi16(__A, __B), |
1166 | (__v16si)_mm512_setzero_si512()); |
1167 | } |
1168 | |
1169 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1170 | _mm512_cvtsepi16_epi8 (__m512i __A) { |
1171 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
1172 | (__v32qi)_mm256_setzero_si256(), |
1173 | (__mmask32) -1); |
1174 | } |
1175 | |
1176 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1177 | _mm512_mask_cvtsepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
1178 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
1179 | (__v32qi)__O, |
1180 | __M); |
1181 | } |
1182 | |
1183 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1184 | _mm512_maskz_cvtsepi16_epi8 (__mmask32 __M, __m512i __A) { |
1185 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
1186 | (__v32qi) _mm256_setzero_si256(), |
1187 | __M); |
1188 | } |
1189 | |
1190 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1191 | _mm512_cvtusepi16_epi8 (__m512i __A) { |
1192 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
1193 | (__v32qi) _mm256_setzero_si256(), |
1194 | (__mmask32) -1); |
1195 | } |
1196 | |
1197 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1198 | _mm512_mask_cvtusepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
1199 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
1200 | (__v32qi) __O, |
1201 | __M); |
1202 | } |
1203 | |
1204 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1205 | _mm512_maskz_cvtusepi16_epi8 (__mmask32 __M, __m512i __A) { |
1206 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
1207 | (__v32qi) _mm256_setzero_si256(), |
1208 | __M); |
1209 | } |
1210 | |
1211 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1212 | _mm512_cvtepi16_epi8 (__m512i __A) { |
1213 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
1214 | (__v32qi) _mm256_undefined_si256(), |
1215 | (__mmask32) -1); |
1216 | } |
1217 | |
1218 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1219 | _mm512_mask_cvtepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
1220 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
1221 | (__v32qi) __O, |
1222 | __M); |
1223 | } |
1224 | |
1225 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
1226 | _mm512_maskz_cvtepi16_epi8 (__mmask32 __M, __m512i __A) { |
1227 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
1228 | (__v32qi) _mm256_setzero_si256(), |
1229 | __M); |
1230 | } |
1231 | |
1232 | static __inline__ void __DEFAULT_FN_ATTRS512 |
1233 | _mm512_mask_cvtepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
1234 | { |
1235 | __builtin_ia32_pmovwb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
1236 | } |
1237 | |
1238 | static __inline__ void __DEFAULT_FN_ATTRS512 |
1239 | _mm512_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
1240 | { |
1241 | __builtin_ia32_pmovswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
1242 | } |
1243 | |
1244 | static __inline__ void __DEFAULT_FN_ATTRS512 |
1245 | _mm512_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
1246 | { |
1247 | __builtin_ia32_pmovuswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
1248 | } |
1249 | |
1250 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1251 | _mm512_unpackhi_epi8(__m512i __A, __m512i __B) { |
1252 | return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B, |
1253 | 8, 64+8, 9, 64+9, |
1254 | 10, 64+10, 11, 64+11, |
1255 | 12, 64+12, 13, 64+13, |
1256 | 14, 64+14, 15, 64+15, |
1257 | 24, 64+24, 25, 64+25, |
1258 | 26, 64+26, 27, 64+27, |
1259 | 28, 64+28, 29, 64+29, |
1260 | 30, 64+30, 31, 64+31, |
1261 | 40, 64+40, 41, 64+41, |
1262 | 42, 64+42, 43, 64+43, |
1263 | 44, 64+44, 45, 64+45, |
1264 | 46, 64+46, 47, 64+47, |
1265 | 56, 64+56, 57, 64+57, |
1266 | 58, 64+58, 59, 64+59, |
1267 | 60, 64+60, 61, 64+61, |
1268 | 62, 64+62, 63, 64+63); |
1269 | } |
1270 | |
1271 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1272 | _mm512_mask_unpackhi_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
1273 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
1274 | (__v64qi)_mm512_unpackhi_epi8(__A, __B), |
1275 | (__v64qi)__W); |
1276 | } |
1277 | |
1278 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1279 | _mm512_maskz_unpackhi_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
1280 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
1281 | (__v64qi)_mm512_unpackhi_epi8(__A, __B), |
1282 | (__v64qi)_mm512_setzero_si512()); |
1283 | } |
1284 | |
1285 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1286 | _mm512_unpackhi_epi16(__m512i __A, __m512i __B) { |
1287 | return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B, |
1288 | 4, 32+4, 5, 32+5, |
1289 | 6, 32+6, 7, 32+7, |
1290 | 12, 32+12, 13, 32+13, |
1291 | 14, 32+14, 15, 32+15, |
1292 | 20, 32+20, 21, 32+21, |
1293 | 22, 32+22, 23, 32+23, |
1294 | 28, 32+28, 29, 32+29, |
1295 | 30, 32+30, 31, 32+31); |
1296 | } |
1297 | |
1298 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1299 | _mm512_mask_unpackhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
1300 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1301 | (__v32hi)_mm512_unpackhi_epi16(__A, __B), |
1302 | (__v32hi)__W); |
1303 | } |
1304 | |
1305 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1306 | _mm512_maskz_unpackhi_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
1307 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1308 | (__v32hi)_mm512_unpackhi_epi16(__A, __B), |
1309 | (__v32hi)_mm512_setzero_si512()); |
1310 | } |
1311 | |
1312 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1313 | _mm512_unpacklo_epi8(__m512i __A, __m512i __B) { |
1314 | return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B, |
1315 | 0, 64+0, 1, 64+1, |
1316 | 2, 64+2, 3, 64+3, |
1317 | 4, 64+4, 5, 64+5, |
1318 | 6, 64+6, 7, 64+7, |
1319 | 16, 64+16, 17, 64+17, |
1320 | 18, 64+18, 19, 64+19, |
1321 | 20, 64+20, 21, 64+21, |
1322 | 22, 64+22, 23, 64+23, |
1323 | 32, 64+32, 33, 64+33, |
1324 | 34, 64+34, 35, 64+35, |
1325 | 36, 64+36, 37, 64+37, |
1326 | 38, 64+38, 39, 64+39, |
1327 | 48, 64+48, 49, 64+49, |
1328 | 50, 64+50, 51, 64+51, |
1329 | 52, 64+52, 53, 64+53, |
1330 | 54, 64+54, 55, 64+55); |
1331 | } |
1332 | |
1333 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1334 | _mm512_mask_unpacklo_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
1335 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
1336 | (__v64qi)_mm512_unpacklo_epi8(__A, __B), |
1337 | (__v64qi)__W); |
1338 | } |
1339 | |
1340 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1341 | _mm512_maskz_unpacklo_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
1342 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
1343 | (__v64qi)_mm512_unpacklo_epi8(__A, __B), |
1344 | (__v64qi)_mm512_setzero_si512()); |
1345 | } |
1346 | |
1347 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1348 | _mm512_unpacklo_epi16(__m512i __A, __m512i __B) { |
1349 | return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B, |
1350 | 0, 32+0, 1, 32+1, |
1351 | 2, 32+2, 3, 32+3, |
1352 | 8, 32+8, 9, 32+9, |
1353 | 10, 32+10, 11, 32+11, |
1354 | 16, 32+16, 17, 32+17, |
1355 | 18, 32+18, 19, 32+19, |
1356 | 24, 32+24, 25, 32+25, |
1357 | 26, 32+26, 27, 32+27); |
1358 | } |
1359 | |
1360 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1361 | _mm512_mask_unpacklo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
1362 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1363 | (__v32hi)_mm512_unpacklo_epi16(__A, __B), |
1364 | (__v32hi)__W); |
1365 | } |
1366 | |
1367 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1368 | _mm512_maskz_unpacklo_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
1369 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1370 | (__v32hi)_mm512_unpacklo_epi16(__A, __B), |
1371 | (__v32hi)_mm512_setzero_si512()); |
1372 | } |
1373 | |
1374 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1375 | _mm512_cvtepi8_epi16(__m256i __A) |
1376 | { |
1377 | /* This function always performs a signed extension, but __v32qi is a char |
1378 | which may be signed or unsigned, so use __v32qs. */ |
1379 | return (__m512i)__builtin_convertvector((__v32qs)__A, __v32hi); |
1380 | } |
1381 | |
1382 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1383 | _mm512_mask_cvtepi8_epi16(__m512i __W, __mmask32 __U, __m256i __A) |
1384 | { |
1385 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1386 | (__v32hi)_mm512_cvtepi8_epi16(__A), |
1387 | (__v32hi)__W); |
1388 | } |
1389 | |
1390 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1391 | _mm512_maskz_cvtepi8_epi16(__mmask32 __U, __m256i __A) |
1392 | { |
1393 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1394 | (__v32hi)_mm512_cvtepi8_epi16(__A), |
1395 | (__v32hi)_mm512_setzero_si512()); |
1396 | } |
1397 | |
1398 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1399 | _mm512_cvtepu8_epi16(__m256i __A) |
1400 | { |
1401 | return (__m512i)__builtin_convertvector((__v32qu)__A, __v32hi); |
1402 | } |
1403 | |
1404 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1405 | _mm512_mask_cvtepu8_epi16(__m512i __W, __mmask32 __U, __m256i __A) |
1406 | { |
1407 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1408 | (__v32hi)_mm512_cvtepu8_epi16(__A), |
1409 | (__v32hi)__W); |
1410 | } |
1411 | |
1412 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1413 | _mm512_maskz_cvtepu8_epi16(__mmask32 __U, __m256i __A) |
1414 | { |
1415 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1416 | (__v32hi)_mm512_cvtepu8_epi16(__A), |
1417 | (__v32hi)_mm512_setzero_si512()); |
1418 | } |
1419 | |
1420 | |
1421 | #define _mm512_shufflehi_epi16(A, imm) \ |
1422 | ((__m512i)__builtin_ia32_pshufhw512((__v32hi)(__m512i)(A), (int)(imm))) |
1423 | |
1424 | #define _mm512_mask_shufflehi_epi16(W, U, A, imm) \ |
1425 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
1426 | (__v32hi)_mm512_shufflehi_epi16((A), \ |
1427 | (imm)), \ |
1428 | (__v32hi)(__m512i)(W))) |
1429 | |
1430 | #define _mm512_maskz_shufflehi_epi16(U, A, imm) \ |
1431 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
1432 | (__v32hi)_mm512_shufflehi_epi16((A), \ |
1433 | (imm)), \ |
1434 | (__v32hi)_mm512_setzero_si512())) |
1435 | |
1436 | #define _mm512_shufflelo_epi16(A, imm) \ |
1437 | ((__m512i)__builtin_ia32_pshuflw512((__v32hi)(__m512i)(A), (int)(imm))) |
1438 | |
1439 | |
1440 | #define _mm512_mask_shufflelo_epi16(W, U, A, imm) \ |
1441 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
1442 | (__v32hi)_mm512_shufflelo_epi16((A), \ |
1443 | (imm)), \ |
1444 | (__v32hi)(__m512i)(W))) |
1445 | |
1446 | |
1447 | #define _mm512_maskz_shufflelo_epi16(U, A, imm) \ |
1448 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
1449 | (__v32hi)_mm512_shufflelo_epi16((A), \ |
1450 | (imm)), \ |
1451 | (__v32hi)_mm512_setzero_si512())) |
1452 | |
1453 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1454 | _mm512_sllv_epi16(__m512i __A, __m512i __B) |
1455 | { |
1456 | return (__m512i)__builtin_ia32_psllv32hi((__v32hi) __A, (__v32hi) __B); |
1457 | } |
1458 | |
1459 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1460 | _mm512_mask_sllv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1461 | { |
1462 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1463 | (__v32hi)_mm512_sllv_epi16(__A, __B), |
1464 | (__v32hi)__W); |
1465 | } |
1466 | |
1467 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1468 | _mm512_maskz_sllv_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
1469 | { |
1470 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1471 | (__v32hi)_mm512_sllv_epi16(__A, __B), |
1472 | (__v32hi)_mm512_setzero_si512()); |
1473 | } |
1474 | |
1475 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1476 | _mm512_sll_epi16(__m512i __A, __m128i __B) |
1477 | { |
1478 | return (__m512i)__builtin_ia32_psllw512((__v32hi) __A, (__v8hi) __B); |
1479 | } |
1480 | |
1481 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1482 | _mm512_mask_sll_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
1483 | { |
1484 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1485 | (__v32hi)_mm512_sll_epi16(__A, __B), |
1486 | (__v32hi)__W); |
1487 | } |
1488 | |
1489 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1490 | _mm512_maskz_sll_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
1491 | { |
1492 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1493 | (__v32hi)_mm512_sll_epi16(__A, __B), |
1494 | (__v32hi)_mm512_setzero_si512()); |
1495 | } |
1496 | |
1497 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1498 | _mm512_slli_epi16(__m512i __A, unsigned int __B) |
1499 | { |
1500 | return (__m512i)__builtin_ia32_psllwi512((__v32hi)__A, (int)__B); |
1501 | } |
1502 | |
1503 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1504 | _mm512_mask_slli_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
1505 | unsigned int __B) |
1506 | { |
1507 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1508 | (__v32hi)_mm512_slli_epi16(__A, __B), |
1509 | (__v32hi)__W); |
1510 | } |
1511 | |
1512 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1513 | _mm512_maskz_slli_epi16(__mmask32 __U, __m512i __A, unsigned int __B) |
1514 | { |
1515 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1516 | (__v32hi)_mm512_slli_epi16(__A, __B), |
1517 | (__v32hi)_mm512_setzero_si512()); |
1518 | } |
1519 | |
1520 | #define _mm512_bslli_epi128(a, imm) \ |
1521 | ((__m512i)__builtin_ia32_pslldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm))) |
1522 | |
1523 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1524 | _mm512_srlv_epi16(__m512i __A, __m512i __B) |
1525 | { |
1526 | return (__m512i)__builtin_ia32_psrlv32hi((__v32hi)__A, (__v32hi)__B); |
1527 | } |
1528 | |
1529 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1530 | _mm512_mask_srlv_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1531 | { |
1532 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1533 | (__v32hi)_mm512_srlv_epi16(__A, __B), |
1534 | (__v32hi)__W); |
1535 | } |
1536 | |
1537 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1538 | _mm512_maskz_srlv_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
1539 | { |
1540 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1541 | (__v32hi)_mm512_srlv_epi16(__A, __B), |
1542 | (__v32hi)_mm512_setzero_si512()); |
1543 | } |
1544 | |
1545 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1546 | _mm512_srav_epi16(__m512i __A, __m512i __B) |
1547 | { |
1548 | return (__m512i)__builtin_ia32_psrav32hi((__v32hi)__A, (__v32hi)__B); |
1549 | } |
1550 | |
1551 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1552 | _mm512_mask_srav_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
1553 | { |
1554 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1555 | (__v32hi)_mm512_srav_epi16(__A, __B), |
1556 | (__v32hi)__W); |
1557 | } |
1558 | |
1559 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1560 | _mm512_maskz_srav_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
1561 | { |
1562 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1563 | (__v32hi)_mm512_srav_epi16(__A, __B), |
1564 | (__v32hi)_mm512_setzero_si512()); |
1565 | } |
1566 | |
1567 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1568 | _mm512_sra_epi16(__m512i __A, __m128i __B) |
1569 | { |
1570 | return (__m512i)__builtin_ia32_psraw512((__v32hi) __A, (__v8hi) __B); |
1571 | } |
1572 | |
1573 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1574 | _mm512_mask_sra_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
1575 | { |
1576 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1577 | (__v32hi)_mm512_sra_epi16(__A, __B), |
1578 | (__v32hi)__W); |
1579 | } |
1580 | |
1581 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1582 | _mm512_maskz_sra_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
1583 | { |
1584 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1585 | (__v32hi)_mm512_sra_epi16(__A, __B), |
1586 | (__v32hi)_mm512_setzero_si512()); |
1587 | } |
1588 | |
1589 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1590 | _mm512_srai_epi16(__m512i __A, unsigned int __B) |
1591 | { |
1592 | return (__m512i)__builtin_ia32_psrawi512((__v32hi)__A, (int)__B); |
1593 | } |
1594 | |
1595 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1596 | _mm512_mask_srai_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
1597 | unsigned int __B) |
1598 | { |
1599 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1600 | (__v32hi)_mm512_srai_epi16(__A, __B), |
1601 | (__v32hi)__W); |
1602 | } |
1603 | |
1604 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1605 | _mm512_maskz_srai_epi16(__mmask32 __U, __m512i __A, unsigned int __B) |
1606 | { |
1607 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1608 | (__v32hi)_mm512_srai_epi16(__A, __B), |
1609 | (__v32hi)_mm512_setzero_si512()); |
1610 | } |
1611 | |
1612 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1613 | _mm512_srl_epi16(__m512i __A, __m128i __B) |
1614 | { |
1615 | return (__m512i)__builtin_ia32_psrlw512((__v32hi) __A, (__v8hi) __B); |
1616 | } |
1617 | |
1618 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1619 | _mm512_mask_srl_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
1620 | { |
1621 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1622 | (__v32hi)_mm512_srl_epi16(__A, __B), |
1623 | (__v32hi)__W); |
1624 | } |
1625 | |
1626 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1627 | _mm512_maskz_srl_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
1628 | { |
1629 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1630 | (__v32hi)_mm512_srl_epi16(__A, __B), |
1631 | (__v32hi)_mm512_setzero_si512()); |
1632 | } |
1633 | |
1634 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1635 | _mm512_srli_epi16(__m512i __A, unsigned int __B) |
1636 | { |
1637 | return (__m512i)__builtin_ia32_psrlwi512((__v32hi)__A, (int)__B); |
1638 | } |
1639 | |
1640 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1641 | _mm512_mask_srli_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
1642 | unsigned int __B) |
1643 | { |
1644 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1645 | (__v32hi)_mm512_srli_epi16(__A, __B), |
1646 | (__v32hi)__W); |
1647 | } |
1648 | |
1649 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1650 | _mm512_maskz_srli_epi16(__mmask32 __U, __m512i __A, int __B) |
1651 | { |
1652 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
1653 | (__v32hi)_mm512_srli_epi16(__A, (unsigned int)__B), |
1654 | (__v32hi)_mm512_setzero_si512()); |
1655 | } |
1656 | |
1657 | #define _mm512_bsrli_epi128(a, imm) \ |
1658 | ((__m512i)__builtin_ia32_psrldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm))) |
1659 | |
1660 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1661 | _mm512_mask_mov_epi16 (__m512i __W, __mmask32 __U, __m512i __A) |
1662 | { |
1663 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
1664 | (__v32hi) __A, |
1665 | (__v32hi) __W); |
1666 | } |
1667 | |
1668 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1669 | _mm512_maskz_mov_epi16 (__mmask32 __U, __m512i __A) |
1670 | { |
1671 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
1672 | (__v32hi) __A, |
1673 | (__v32hi) _mm512_setzero_si512 ()); |
1674 | } |
1675 | |
1676 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1677 | _mm512_mask_mov_epi8 (__m512i __W, __mmask64 __U, __m512i __A) |
1678 | { |
1679 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
1680 | (__v64qi) __A, |
1681 | (__v64qi) __W); |
1682 | } |
1683 | |
1684 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1685 | _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A) |
1686 | { |
1687 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
1688 | (__v64qi) __A, |
1689 | (__v64qi) _mm512_setzero_si512 ()); |
1690 | } |
1691 | |
1692 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1693 | _mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A) |
1694 | { |
1695 | return (__m512i) __builtin_ia32_selectb_512(__M, |
1696 | (__v64qi)_mm512_set1_epi8(__A), |
1697 | (__v64qi) __O); |
1698 | } |
1699 | |
1700 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1701 | _mm512_maskz_set1_epi8 (__mmask64 __M, char __A) |
1702 | { |
1703 | return (__m512i) __builtin_ia32_selectb_512(__M, |
1704 | (__v64qi) _mm512_set1_epi8(__A), |
1705 | (__v64qi) _mm512_setzero_si512()); |
1706 | } |
1707 | |
1708 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS _mm512_kunpackd(__mmask64 __A, |
1709 | __mmask64 __B) { |
1710 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, |
1711 | (__mmask64) __B); |
1712 | } |
1713 | |
1714 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
1715 | _mm512_kunpackw (__mmask32 __A, __mmask32 __B) |
1716 | { |
1717 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, |
1718 | (__mmask32) __B); |
1719 | } |
1720 | |
1721 | static __inline __m512i __DEFAULT_FN_ATTRS512 |
1722 | _mm512_loadu_epi16 (void const *__P) |
1723 | { |
1724 | struct __loadu_epi16 { |
1725 | __m512i_u __v; |
1726 | } __attribute__((__packed__, __may_alias__)); |
1727 | return ((const struct __loadu_epi16*)__P)->__v; |
1728 | } |
1729 | |
1730 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1731 | _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P) |
1732 | { |
1733 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const __v32hi *) __P, |
1734 | (__v32hi) __W, |
1735 | (__mmask32) __U); |
1736 | } |
1737 | |
1738 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1739 | _mm512_maskz_loadu_epi16 (__mmask32 __U, void const *__P) |
1740 | { |
1741 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const __v32hi *) __P, |
1742 | (__v32hi) |
1743 | _mm512_setzero_si512 (), |
1744 | (__mmask32) __U); |
1745 | } |
1746 | |
1747 | static __inline __m512i __DEFAULT_FN_ATTRS512 |
1748 | _mm512_loadu_epi8 (void const *__P) |
1749 | { |
1750 | struct __loadu_epi8 { |
1751 | __m512i_u __v; |
1752 | } __attribute__((__packed__, __may_alias__)); |
1753 | return ((const struct __loadu_epi8*)__P)->__v; |
1754 | } |
1755 | |
1756 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1757 | _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P) |
1758 | { |
1759 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const __v64qi *) __P, |
1760 | (__v64qi) __W, |
1761 | (__mmask64) __U); |
1762 | } |
1763 | |
1764 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1765 | _mm512_maskz_loadu_epi8 (__mmask64 __U, void const *__P) |
1766 | { |
1767 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const __v64qi *) __P, |
1768 | (__v64qi) |
1769 | _mm512_setzero_si512 (), |
1770 | (__mmask64) __U); |
1771 | } |
1772 | |
1773 | static __inline void __DEFAULT_FN_ATTRS512 |
1774 | _mm512_storeu_epi16 (void *__P, __m512i __A) |
1775 | { |
1776 | struct __storeu_epi16 { |
1777 | __m512i_u __v; |
1778 | } __attribute__((__packed__, __may_alias__)); |
1779 | ((struct __storeu_epi16*)__P)->__v = __A; |
1780 | } |
1781 | |
1782 | static __inline__ void __DEFAULT_FN_ATTRS512 |
1783 | _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A) |
1784 | { |
1785 | __builtin_ia32_storedquhi512_mask ((__v32hi *) __P, |
1786 | (__v32hi) __A, |
1787 | (__mmask32) __U); |
1788 | } |
1789 | |
1790 | static __inline void __DEFAULT_FN_ATTRS512 |
1791 | _mm512_storeu_epi8 (void *__P, __m512i __A) |
1792 | { |
1793 | struct __storeu_epi8 { |
1794 | __m512i_u __v; |
1795 | } __attribute__((__packed__, __may_alias__)); |
1796 | ((struct __storeu_epi8*)__P)->__v = __A; |
1797 | } |
1798 | |
1799 | static __inline__ void __DEFAULT_FN_ATTRS512 |
1800 | _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A) |
1801 | { |
1802 | __builtin_ia32_storedquqi512_mask ((__v64qi *) __P, |
1803 | (__v64qi) __A, |
1804 | (__mmask64) __U); |
1805 | } |
1806 | |
1807 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
1808 | _mm512_test_epi8_mask (__m512i __A, __m512i __B) |
1809 | { |
1810 | return _mm512_cmpneq_epi8_mask (_mm512_and_epi32 (__A, __B), |
1811 | _mm512_setzero_si512()); |
1812 | } |
1813 | |
1814 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
1815 | _mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) |
1816 | { |
1817 | return _mm512_mask_cmpneq_epi8_mask (__U, _mm512_and_epi32 (__A, __B), |
1818 | _mm512_setzero_si512()); |
1819 | } |
1820 | |
1821 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
1822 | _mm512_test_epi16_mask (__m512i __A, __m512i __B) |
1823 | { |
1824 | return _mm512_cmpneq_epi16_mask (_mm512_and_epi32 (__A, __B), |
1825 | _mm512_setzero_si512()); |
1826 | } |
1827 | |
1828 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
1829 | _mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) |
1830 | { |
1831 | return _mm512_mask_cmpneq_epi16_mask (__U, _mm512_and_epi32 (__A, __B), |
1832 | _mm512_setzero_si512()); |
1833 | } |
1834 | |
1835 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
1836 | _mm512_testn_epi8_mask (__m512i __A, __m512i __B) |
1837 | { |
1838 | return _mm512_cmpeq_epi8_mask (_mm512_and_epi32 (__A, __B), _mm512_setzero_si512()); |
1839 | } |
1840 | |
1841 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
1842 | _mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) |
1843 | { |
1844 | return _mm512_mask_cmpeq_epi8_mask (__U, _mm512_and_epi32 (__A, __B), |
1845 | _mm512_setzero_si512()); |
1846 | } |
1847 | |
1848 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
1849 | _mm512_testn_epi16_mask (__m512i __A, __m512i __B) |
1850 | { |
1851 | return _mm512_cmpeq_epi16_mask (_mm512_and_epi32 (__A, __B), |
1852 | _mm512_setzero_si512()); |
1853 | } |
1854 | |
1855 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
1856 | _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) |
1857 | { |
1858 | return _mm512_mask_cmpeq_epi16_mask (__U, _mm512_and_epi32 (__A, __B), |
1859 | _mm512_setzero_si512()); |
1860 | } |
1861 | |
1862 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
1863 | _mm512_movepi8_mask (__m512i __A) |
1864 | { |
1865 | return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A); |
1866 | } |
1867 | |
1868 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
1869 | _mm512_movepi16_mask (__m512i __A) |
1870 | { |
1871 | return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); |
1872 | } |
1873 | |
1874 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1875 | _mm512_movm_epi8 (__mmask64 __A) |
1876 | { |
1877 | return (__m512i) __builtin_ia32_cvtmask2b512 (__A); |
1878 | } |
1879 | |
1880 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1881 | _mm512_movm_epi16 (__mmask32 __A) |
1882 | { |
1883 | return (__m512i) __builtin_ia32_cvtmask2w512 (__A); |
1884 | } |
1885 | |
1886 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1887 | _mm512_broadcastb_epi8 (__m128i __A) |
1888 | { |
1889 | return (__m512i)__builtin_shufflevector((__v16qi) __A, (__v16qi) __A, |
1890 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
1891 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
1892 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
1893 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); |
1894 | } |
1895 | |
1896 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1897 | _mm512_mask_broadcastb_epi8 (__m512i __O, __mmask64 __M, __m128i __A) |
1898 | { |
1899 | return (__m512i)__builtin_ia32_selectb_512(__M, |
1900 | (__v64qi) _mm512_broadcastb_epi8(__A), |
1901 | (__v64qi) __O); |
1902 | } |
1903 | |
1904 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1905 | _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A) |
1906 | { |
1907 | return (__m512i)__builtin_ia32_selectb_512(__M, |
1908 | (__v64qi) _mm512_broadcastb_epi8(__A), |
1909 | (__v64qi) _mm512_setzero_si512()); |
1910 | } |
1911 | |
1912 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1913 | _mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A) |
1914 | { |
1915 | return (__m512i) __builtin_ia32_selectw_512(__M, |
1916 | (__v32hi) _mm512_set1_epi16(__A), |
1917 | (__v32hi) __O); |
1918 | } |
1919 | |
1920 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1921 | _mm512_maskz_set1_epi16 (__mmask32 __M, short __A) |
1922 | { |
1923 | return (__m512i) __builtin_ia32_selectw_512(__M, |
1924 | (__v32hi) _mm512_set1_epi16(__A), |
1925 | (__v32hi) _mm512_setzero_si512()); |
1926 | } |
1927 | |
1928 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1929 | _mm512_broadcastw_epi16 (__m128i __A) |
1930 | { |
1931 | return (__m512i)__builtin_shufflevector((__v8hi) __A, (__v8hi) __A, |
1932 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
1933 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); |
1934 | } |
1935 | |
1936 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1937 | _mm512_mask_broadcastw_epi16 (__m512i __O, __mmask32 __M, __m128i __A) |
1938 | { |
1939 | return (__m512i)__builtin_ia32_selectw_512(__M, |
1940 | (__v32hi) _mm512_broadcastw_epi16(__A), |
1941 | (__v32hi) __O); |
1942 | } |
1943 | |
1944 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1945 | _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A) |
1946 | { |
1947 | return (__m512i)__builtin_ia32_selectw_512(__M, |
1948 | (__v32hi) _mm512_broadcastw_epi16(__A), |
1949 | (__v32hi) _mm512_setzero_si512()); |
1950 | } |
1951 | |
1952 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1953 | _mm512_permutexvar_epi16 (__m512i __A, __m512i __B) |
1954 | { |
1955 | return (__m512i)__builtin_ia32_permvarhi512((__v32hi)__B, (__v32hi)__A); |
1956 | } |
1957 | |
1958 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1959 | _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A, |
1960 | __m512i __B) |
1961 | { |
1962 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
1963 | (__v32hi)_mm512_permutexvar_epi16(__A, __B), |
1964 | (__v32hi)_mm512_setzero_si512()); |
1965 | } |
1966 | |
1967 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
1968 | _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A, |
1969 | __m512i __B) |
1970 | { |
1971 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
1972 | (__v32hi)_mm512_permutexvar_epi16(__A, __B), |
1973 | (__v32hi)__W); |
1974 | } |
1975 | |
1976 | #define _mm512_alignr_epi8(A, B, N) \ |
1977 | ((__m512i)__builtin_ia32_palignr512((__v64qi)(__m512i)(A), \ |
1978 | (__v64qi)(__m512i)(B), (int)(N))) |
1979 | |
1980 | #define _mm512_mask_alignr_epi8(W, U, A, B, N) \ |
1981 | ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \ |
1982 | (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \ |
1983 | (__v64qi)(__m512i)(W))) |
1984 | |
1985 | #define _mm512_maskz_alignr_epi8(U, A, B, N) \ |
1986 | ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \ |
1987 | (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \ |
1988 | (__v64qi)(__m512i)_mm512_setzero_si512())) |
1989 | |
1990 | #define _mm512_dbsad_epu8(A, B, imm) \ |
1991 | ((__m512i)__builtin_ia32_dbpsadbw512((__v64qi)(__m512i)(A), \ |
1992 | (__v64qi)(__m512i)(B), (int)(imm))) |
1993 | |
1994 | #define _mm512_mask_dbsad_epu8(W, U, A, B, imm) \ |
1995 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
1996 | (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \ |
1997 | (__v32hi)(__m512i)(W))) |
1998 | |
1999 | #define _mm512_maskz_dbsad_epu8(U, A, B, imm) \ |
2000 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
2001 | (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \ |
2002 | (__v32hi)_mm512_setzero_si512())) |
2003 | |
2004 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
2005 | _mm512_sad_epu8 (__m512i __A, __m512i __B) |
2006 | { |
2007 | return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A, |
2008 | (__v64qi) __B); |
2009 | } |
2010 | |
2011 | #undef __DEFAULT_FN_ATTRS512 |
2012 | #undef __DEFAULT_FN_ATTRS |
2013 | |
2014 | #endif |
2015 |
Warning: This file is not a C or C++ file. It does not have highlighting.