]>
Commit | Line | Data |
---|---|---|
936c0fe4 AI |
1 | /* Copyright (C) 2014 |
2 | Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 3, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | Under Section 7 of GPL version 3, you are granted additional | |
17 | permissions described in the GCC Runtime Library Exception, version | |
18 | 3.1, as published by the Free Software Foundation. | |
19 | ||
20 | You should have received a copy of the GNU General Public License and | |
21 | a copy of the GCC Runtime Library Exception along with this program; | |
22 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see | |
23 | <http://www.gnu.org/licenses/>. */ | |
24 | ||
25 | #ifndef _IMMINTRIN_H_INCLUDED | |
26 | #error "Never use <avx512vlbwintrin.h> directly; include <immintrin.h> instead." | |
27 | #endif | |
28 | ||
29 | #ifndef _AVX512VLBWINTRIN_H_INCLUDED | |
30 | #define _AVX512VLBWINTRIN_H_INCLUDED | |
31 | ||
32 | #if !defined(__AVX512VL__) || !defined(__AVX512BW__) | |
33 | #pragma GCC push_options | |
34 | #pragma GCC target("avx512vl,avx512bw") | |
35 | #define __DISABLE_AVX512VLBW__ | |
36 | #endif /* __AVX512VLBW__ */ | |
37 | ||
38 | ||
39 | extern __inline __m256i | |
40 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
41 | _mm256_mask_mov_epi8 (__m256i __W, __mmask32 __U, __m256i __A) | |
42 | { | |
43 | return (__m256i) __builtin_ia32_movdquqi256_mask ((__v32qi) __A, | |
44 | (__v32qi) __W, | |
45 | (__mmask32) __U); | |
46 | } | |
47 | ||
48 | extern __inline __m256i | |
49 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
50 | _mm256_maskz_mov_epi8 (__mmask32 __U, __m256i __A) | |
51 | { | |
52 | return (__m256i) __builtin_ia32_movdquqi256_mask ((__v32qi) __A, | |
53 | (__v32qi) | |
54 | _mm256_setzero_si256 (), | |
55 | (__mmask32) __U); | |
56 | } | |
57 | ||
58 | extern __inline __m128i | |
59 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
60 | _mm_mask_mov_epi8 (__m128i __W, __mmask16 __U, __m128i __A) | |
61 | { | |
62 | return (__m128i) __builtin_ia32_movdquqi128_mask ((__v16qi) __A, | |
63 | (__v16qi) __W, | |
64 | (__mmask16) __U); | |
65 | } | |
66 | ||
67 | extern __inline __m128i | |
68 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
69 | _mm_maskz_mov_epi8 (__mmask16 __U, __m128i __A) | |
70 | { | |
71 | return (__m128i) __builtin_ia32_movdquqi128_mask ((__v16qi) __A, | |
72 | (__v16qi) | |
73 | _mm_setzero_hi (), | |
74 | (__mmask16) __U); | |
75 | } | |
76 | ||
77 | extern __inline void | |
78 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
79 | _mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A) | |
80 | { | |
81 | __builtin_ia32_storedquqi256_mask ((__v32qi *) __P, | |
82 | (__v32qi) __A, | |
83 | (__mmask32) __U); | |
84 | } | |
85 | ||
86 | extern __inline void | |
87 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
88 | _mm_mask_storeu_epi8 (void *__P, __mmask16 __U, __m128i __A) | |
89 | { | |
90 | __builtin_ia32_storedquqi128_mask ((__v16qi *) __P, | |
91 | (__v16qi) __A, | |
92 | (__mmask16) __U); | |
93 | } | |
94 | ||
95 | extern __inline __m256i | |
96 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
97 | _mm256_mask_loadu_epi16 (__m256i __W, __mmask16 __U, void const *__P) | |
98 | { | |
99 | return (__m256i) __builtin_ia32_loaddquhi256_mask ((__v16hi *) __P, | |
100 | (__v16hi) __W, | |
101 | (__mmask16) __U); | |
102 | } | |
103 | ||
104 | extern __inline __m256i | |
105 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
106 | _mm256_maskz_loadu_epi16 (__mmask16 __U, void const *__P) | |
107 | { | |
108 | return (__m256i) __builtin_ia32_loaddquhi256_mask ((__v16hi *) __P, | |
109 | (__v16hi) | |
110 | _mm256_setzero_si256 (), | |
111 | (__mmask16) __U); | |
112 | } | |
113 | ||
114 | extern __inline __m128i | |
115 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
116 | _mm_mask_loadu_epi16 (__m128i __W, __mmask8 __U, void const *__P) | |
117 | { | |
118 | return (__m128i) __builtin_ia32_loaddquhi128_mask ((__v8hi *) __P, | |
119 | (__v8hi) __W, | |
120 | (__mmask8) __U); | |
121 | } | |
122 | ||
123 | extern __inline __m128i | |
124 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
125 | _mm_maskz_loadu_epi16 (__mmask8 __U, void const *__P) | |
126 | { | |
127 | return (__m128i) __builtin_ia32_loaddquhi128_mask ((__v8hi *) __P, | |
128 | (__v8hi) | |
129 | _mm_setzero_hi (), | |
130 | (__mmask8) __U); | |
131 | } | |
132 | ||
133 | ||
134 | extern __inline __m256i | |
135 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
136 | _mm256_mask_mov_epi16 (__m256i __W, __mmask16 __U, __m256i __A) | |
137 | { | |
138 | return (__m256i) __builtin_ia32_movdquhi256_mask ((__v16hi) __A, | |
139 | (__v16hi) __W, | |
140 | (__mmask16) __U); | |
141 | } | |
142 | ||
143 | extern __inline __m256i | |
144 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
145 | _mm256_maskz_mov_epi16 (__mmask16 __U, __m256i __A) | |
146 | { | |
147 | return (__m256i) __builtin_ia32_movdquhi256_mask ((__v16hi) __A, | |
148 | (__v16hi) | |
149 | _mm256_setzero_si256 (), | |
150 | (__mmask16) __U); | |
151 | } | |
152 | ||
153 | extern __inline __m128i | |
154 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
155 | _mm_mask_mov_epi16 (__m128i __W, __mmask8 __U, __m128i __A) | |
156 | { | |
157 | return (__m128i) __builtin_ia32_movdquhi128_mask ((__v8hi) __A, | |
158 | (__v8hi) __W, | |
159 | (__mmask8) __U); | |
160 | } | |
161 | ||
162 | extern __inline __m128i | |
163 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
164 | _mm_maskz_mov_epi16 (__mmask8 __U, __m128i __A) | |
165 | { | |
166 | return (__m128i) __builtin_ia32_movdquhi128_mask ((__v8hi) __A, | |
167 | (__v8hi) | |
168 | _mm_setzero_hi (), | |
169 | (__mmask8) __U); | |
170 | } | |
171 | ||
172 | extern __inline __m256i | |
173 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
174 | _mm256_mask_loadu_epi8 (__m256i __W, __mmask32 __U, void const *__P) | |
175 | { | |
176 | return (__m256i) __builtin_ia32_loaddquqi256_mask ((__v32qi *) __P, | |
177 | (__v32qi) __W, | |
178 | (__mmask32) __U); | |
179 | } | |
180 | ||
181 | extern __inline __m256i | |
182 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
183 | _mm256_maskz_loadu_epi8 (__mmask32 __U, void const *__P) | |
184 | { | |
185 | return (__m256i) __builtin_ia32_loaddquqi256_mask ((__v32qi *) __P, | |
186 | (__v32qi) | |
187 | _mm256_setzero_si256 (), | |
188 | (__mmask32) __U); | |
189 | } | |
190 | ||
191 | extern __inline __m128i | |
192 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
193 | _mm_mask_loadu_epi8 (__m128i __W, __mmask16 __U, void const *__P) | |
194 | { | |
195 | return (__m128i) __builtin_ia32_loaddquqi128_mask ((__v16qi *) __P, | |
196 | (__v16qi) __W, | |
197 | (__mmask16) __U); | |
198 | } | |
199 | ||
200 | extern __inline __m128i | |
201 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
202 | _mm_maskz_loadu_epi8 (__mmask16 __U, void const *__P) | |
203 | { | |
204 | return (__m128i) __builtin_ia32_loaddquqi128_mask ((__v16qi *) __P, | |
205 | (__v16qi) | |
206 | _mm_setzero_hi (), | |
207 | (__mmask16) __U); | |
208 | } | |
209 | ||
210 | extern __inline __m128i | |
211 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
212 | _mm256_cvtepi16_epi8 (__m256i __A) | |
213 | { | |
214 | ||
215 | return (__m128i) __builtin_ia32_pmovwb256_mask ((__v16hi) __A, | |
216 | (__v16qi)_mm_undefined_si128(), | |
217 | (__mmask16) -1); | |
218 | } | |
219 | ||
220 | extern __inline __m128i | |
221 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
222 | _mm256_mask_cvtepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) | |
223 | { | |
224 | return (__m128i) __builtin_ia32_pmovwb256_mask ((__v16hi) __A, | |
225 | (__v16qi) __O, __M); | |
226 | } | |
227 | ||
228 | extern __inline __m128i | |
229 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
230 | _mm256_maskz_cvtepi16_epi8 (__mmask16 __M, __m256i __A) | |
231 | { | |
232 | return (__m128i) __builtin_ia32_pmovwb256_mask ((__v16hi) __A, | |
233 | (__v16qi) | |
234 | _mm_setzero_si128 (), | |
235 | __M); | |
236 | } | |
237 | ||
238 | extern __inline __m128i | |
239 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
240 | _mm_cvtsepi16_epi8 (__m128i __A) | |
241 | { | |
242 | ||
243 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, | |
244 | (__v16qi)_mm_undefined_si128(), | |
245 | (__mmask8) -1); | |
246 | } | |
247 | ||
248 | extern __inline __m128i | |
249 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
250 | _mm_mask_cvtsepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) | |
251 | { | |
252 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, | |
253 | (__v16qi) __O, __M); | |
254 | } | |
255 | ||
256 | extern __inline __m128i | |
257 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
258 | _mm_maskz_cvtsepi16_epi8 (__mmask8 __M, __m128i __A) | |
259 | { | |
260 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, | |
261 | (__v16qi) | |
262 | _mm_setzero_si128 (), | |
263 | __M); | |
264 | } | |
265 | ||
266 | extern __inline __m128i | |
267 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
268 | _mm256_cvtsepi16_epi8 (__m256i __A) | |
269 | { | |
270 | ||
271 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, | |
272 | (__v16qi)_mm_undefined_si128(), | |
273 | (__mmask16) -1); | |
274 | } | |
275 | ||
276 | extern __inline __m128i | |
277 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
278 | _mm256_mask_cvtsepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) | |
279 | { | |
280 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, | |
281 | (__v16qi) __O, __M); | |
282 | } | |
283 | ||
284 | extern __inline __m128i | |
285 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
286 | _mm256_maskz_cvtsepi16_epi8 (__mmask16 __M, __m256i __A) | |
287 | { | |
288 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, | |
289 | (__v16qi) | |
290 | _mm_setzero_si128 (), | |
291 | __M); | |
292 | } | |
293 | ||
294 | extern __inline __m128i | |
295 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
296 | _mm_cvtusepi16_epi8 (__m128i __A) | |
297 | { | |
298 | ||
299 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, | |
300 | (__v16qi)_mm_undefined_si128(), | |
301 | (__mmask8) -1); | |
302 | } | |
303 | ||
304 | extern __inline __m128i | |
305 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
306 | _mm_mask_cvtusepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) | |
307 | { | |
308 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, | |
309 | (__v16qi) __O, | |
310 | __M); | |
311 | } | |
312 | ||
313 | extern __inline __m128i | |
314 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
315 | _mm_maskz_cvtusepi16_epi8 (__mmask8 __M, __m128i __A) | |
316 | { | |
317 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, | |
318 | (__v16qi) | |
319 | _mm_setzero_si128 (), | |
320 | __M); | |
321 | } | |
322 | ||
323 | extern __inline __m128i | |
324 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
325 | _mm256_cvtusepi16_epi8 (__m256i __A) | |
326 | { | |
327 | ||
328 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, | |
329 | (__v16qi)_mm_undefined_si128(), | |
330 | (__mmask16) -1); | |
331 | } | |
332 | ||
333 | extern __inline __m128i | |
334 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
335 | _mm256_mask_cvtusepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) | |
336 | { | |
337 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, | |
338 | (__v16qi) __O, | |
339 | __M); | |
340 | } | |
341 | ||
342 | extern __inline __m128i | |
343 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
344 | _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) | |
345 | { | |
346 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, | |
347 | (__v16qi) | |
348 | _mm_setzero_si128 (), | |
349 | __M); | |
350 | } | |
351 | ||
352 | extern __inline __m256i | |
353 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
354 | _mm256_mask_broadcastb_epi8 (__m256i __O, __mmask32 __M, __m128i __A) | |
355 | { | |
356 | return (__m256i) __builtin_ia32_pbroadcastb256_mask ((__v16qi) __A, | |
357 | (__v32qi) __O, | |
358 | __M); | |
359 | } | |
360 | ||
361 | extern __inline __m256i | |
362 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
363 | _mm256_maskz_broadcastb_epi8 (__mmask32 __M, __m128i __A) | |
364 | { | |
365 | return (__m256i) __builtin_ia32_pbroadcastb256_mask ((__v16qi) __A, | |
366 | (__v32qi) | |
367 | _mm256_setzero_si256 (), | |
368 | __M); | |
369 | } | |
370 | ||
371 | extern __inline __m256i | |
372 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
373 | _mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A) | |
374 | { | |
375 | return (__m256i) __builtin_ia32_pbroadcastb256_gpr_mask (__A, | |
376 | (__v32qi) __O, | |
377 | __M); | |
378 | } | |
379 | ||
380 | extern __inline __m256i | |
381 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
382 | _mm256_maskz_set1_epi8 (__mmask32 __M, char __A) | |
383 | { | |
384 | return (__m256i) __builtin_ia32_pbroadcastb256_gpr_mask (__A, | |
385 | (__v32qi) | |
386 | _mm256_setzero_si256 (), | |
387 | __M); | |
388 | } | |
389 | ||
390 | extern __inline __m128i | |
391 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
392 | _mm_mask_broadcastb_epi8 (__m128i __O, __mmask16 __M, __m128i __A) | |
393 | { | |
394 | return (__m128i) __builtin_ia32_pbroadcastb128_mask ((__v16qi) __A, | |
395 | (__v16qi) __O, | |
396 | __M); | |
397 | } | |
398 | ||
399 | extern __inline __m128i | |
400 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
401 | _mm_maskz_broadcastb_epi8 (__mmask16 __M, __m128i __A) | |
402 | { | |
403 | return (__m128i) __builtin_ia32_pbroadcastb128_mask ((__v16qi) __A, | |
404 | (__v16qi) | |
405 | _mm_setzero_si128 (), | |
406 | __M); | |
407 | } | |
408 | ||
409 | extern __inline __m128i | |
410 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
411 | _mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A) | |
412 | { | |
413 | return (__m128i) __builtin_ia32_pbroadcastb128_gpr_mask (__A, | |
414 | (__v16qi) __O, | |
415 | __M); | |
416 | } | |
417 | ||
418 | extern __inline __m128i | |
419 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
420 | _mm_maskz_set1_epi8 (__mmask16 __M, char __A) | |
421 | { | |
422 | return (__m128i) __builtin_ia32_pbroadcastb128_gpr_mask (__A, | |
423 | (__v16qi) | |
424 | _mm_setzero_si128 (), | |
425 | __M); | |
426 | } | |
427 | ||
428 | extern __inline __m256i | |
429 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
430 | _mm256_mask_broadcastw_epi16 (__m256i __O, __mmask16 __M, __m128i __A) | |
431 | { | |
432 | return (__m256i) __builtin_ia32_pbroadcastw256_mask ((__v8hi) __A, | |
433 | (__v16hi) __O, | |
434 | __M); | |
435 | } | |
436 | ||
437 | extern __inline __m256i | |
438 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
439 | _mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A) | |
440 | { | |
441 | return (__m256i) __builtin_ia32_pbroadcastw256_mask ((__v8hi) __A, | |
442 | (__v16hi) | |
443 | _mm256_setzero_si256 (), | |
444 | __M); | |
445 | } | |
446 | ||
447 | extern __inline __m256i | |
448 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
449 | _mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A) | |
450 | { | |
451 | return (__m256i) __builtin_ia32_pbroadcastw256_gpr_mask (__A, | |
452 | (__v16hi) __O, | |
453 | __M); | |
454 | } | |
455 | ||
456 | extern __inline __m256i | |
457 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
458 | _mm256_maskz_set1_epi16 (__mmask16 __M, short __A) | |
459 | { | |
460 | return (__m256i) __builtin_ia32_pbroadcastw256_gpr_mask (__A, | |
461 | (__v16hi) | |
462 | _mm256_setzero_si256 (), | |
463 | __M); | |
464 | } | |
465 | ||
466 | extern __inline __m128i | |
467 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
468 | _mm_mask_broadcastw_epi16 (__m128i __O, __mmask8 __M, __m128i __A) | |
469 | { | |
470 | return (__m128i) __builtin_ia32_pbroadcastw128_mask ((__v8hi) __A, | |
471 | (__v8hi) __O, | |
472 | __M); | |
473 | } | |
474 | ||
475 | extern __inline __m128i | |
476 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
477 | _mm_maskz_broadcastw_epi16 (__mmask8 __M, __m128i __A) | |
478 | { | |
479 | return (__m128i) __builtin_ia32_pbroadcastw128_mask ((__v8hi) __A, | |
480 | (__v8hi) | |
481 | _mm_setzero_si128 (), | |
482 | __M); | |
483 | } | |
484 | ||
485 | extern __inline __m128i | |
486 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
487 | _mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A) | |
488 | { | |
489 | return (__m128i) __builtin_ia32_pbroadcastw128_gpr_mask (__A, | |
490 | (__v8hi) __O, | |
491 | __M); | |
492 | } | |
493 | ||
494 | extern __inline __m128i | |
495 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
496 | _mm_maskz_set1_epi16 (__mmask8 __M, short __A) | |
497 | { | |
498 | return (__m128i) __builtin_ia32_pbroadcastw128_gpr_mask (__A, | |
499 | (__v8hi) | |
500 | _mm_setzero_si128 (), | |
501 | __M); | |
502 | } | |
503 | ||
504 | extern __inline __m256i | |
505 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
506 | _mm256_permutexvar_epi16 (__m256i __A, __m256i __B) | |
507 | { | |
508 | return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B, | |
509 | (__v16hi) __A, | |
510 | (__v16hi) | |
511 | _mm256_setzero_si256 (), | |
512 | (__mmask16) -1); | |
513 | } | |
514 | ||
515 | extern __inline __m256i | |
516 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
517 | _mm256_maskz_permutexvar_epi16 (__mmask16 __M, __m256i __A, | |
518 | __m256i __B) | |
519 | { | |
520 | return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B, | |
521 | (__v16hi) __A, | |
522 | (__v16hi) | |
523 | _mm256_setzero_si256 (), | |
524 | (__mmask16) __M); | |
525 | } | |
526 | ||
527 | extern __inline __m256i | |
528 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
529 | _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A, | |
530 | __m256i __B) | |
531 | { | |
532 | return (__m256i) __builtin_ia32_permvarhi256_mask ((__v16hi) __B, | |
533 | (__v16hi) __A, | |
534 | (__v16hi) __W, | |
535 | (__mmask16) __M); | |
536 | } | |
537 | ||
538 | extern __inline __m128i | |
539 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
540 | _mm_permutexvar_epi16 (__m128i __A, __m128i __B) | |
541 | { | |
542 | return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B, | |
543 | (__v8hi) __A, | |
544 | (__v8hi) | |
545 | _mm_setzero_hi (), | |
546 | (__mmask8) -1); | |
547 | } | |
548 | ||
549 | extern __inline __m128i | |
550 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
551 | _mm_maskz_permutexvar_epi16 (__mmask8 __M, __m128i __A, __m128i __B) | |
552 | { | |
553 | return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B, | |
554 | (__v8hi) __A, | |
555 | (__v8hi) | |
556 | _mm_setzero_si128 (), | |
557 | (__mmask8) __M); | |
558 | } | |
559 | ||
560 | extern __inline __m128i | |
561 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
562 | _mm_mask_permutexvar_epi16 (__m128i __W, __mmask8 __M, __m128i __A, | |
563 | __m128i __B) | |
564 | { | |
565 | return (__m128i) __builtin_ia32_permvarhi128_mask ((__v8hi) __B, | |
566 | (__v8hi) __A, | |
567 | (__v8hi) __W, | |
568 | (__mmask8) __M); | |
569 | } | |
570 | ||
571 | extern __inline __m256i | |
572 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
573 | _mm256_permutex2var_epi16 (__m256i __A, __m256i __I, __m256i __B) | |
574 | { | |
575 | return (__m256i) __builtin_ia32_vpermt2varhi256_mask ((__v16hi) __I | |
576 | /* idx */ , | |
577 | (__v16hi) __A, | |
578 | (__v16hi) __B, | |
579 | (__mmask16) - | |
580 | 1); | |
581 | } | |
582 | ||
583 | extern __inline __m256i | |
584 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
585 | _mm256_mask_permutex2var_epi16 (__m256i __A, __mmask16 __U, | |
586 | __m256i __I, __m256i __B) | |
587 | { | |
588 | return (__m256i) __builtin_ia32_vpermt2varhi256_mask ((__v16hi) __I | |
589 | /* idx */ , | |
590 | (__v16hi) __A, | |
591 | (__v16hi) __B, | |
592 | (__mmask16) | |
593 | __U); | |
594 | } | |
595 | ||
596 | extern __inline __m256i | |
597 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
598 | _mm256_mask2_permutex2var_epi16 (__m256i __A, __m256i __I, | |
599 | __mmask16 __U, __m256i __B) | |
600 | { | |
601 | return (__m256i) __builtin_ia32_vpermi2varhi256_mask ((__v16hi) __A, | |
602 | (__v16hi) __I | |
603 | /* idx */ , | |
604 | (__v16hi) __B, | |
605 | (__mmask16) | |
606 | __U); | |
607 | } | |
608 | ||
609 | extern __inline __m256i | |
610 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
611 | _mm256_maskz_permutex2var_epi16 (__mmask16 __U, __m256i __A, | |
612 | __m256i __I, __m256i __B) | |
613 | { | |
614 | return (__m256i) __builtin_ia32_vpermt2varhi256_maskz ((__v16hi) __I | |
615 | /* idx */ , | |
616 | (__v16hi) __A, | |
617 | (__v16hi) __B, | |
618 | (__mmask16) | |
619 | __U); | |
620 | } | |
621 | ||
622 | extern __inline __m128i | |
623 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
624 | _mm_permutex2var_epi16 (__m128i __A, __m128i __I, __m128i __B) | |
625 | { | |
626 | return (__m128i) __builtin_ia32_vpermt2varhi128_mask ((__v8hi) __I | |
627 | /* idx */ , | |
628 | (__v8hi) __A, | |
629 | (__v8hi) __B, | |
630 | (__mmask8) - | |
631 | 1); | |
632 | } | |
633 | ||
634 | extern __inline __m128i | |
635 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
636 | _mm_mask_permutex2var_epi16 (__m128i __A, __mmask8 __U, __m128i __I, | |
637 | __m128i __B) | |
638 | { | |
639 | return (__m128i) __builtin_ia32_vpermt2varhi128_mask ((__v8hi) __I | |
640 | /* idx */ , | |
641 | (__v8hi) __A, | |
642 | (__v8hi) __B, | |
643 | (__mmask8) | |
644 | __U); | |
645 | } | |
646 | ||
647 | extern __inline __m128i | |
648 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
649 | _mm_mask2_permutex2var_epi16 (__m128i __A, __m128i __I, __mmask8 __U, | |
650 | __m128i __B) | |
651 | { | |
652 | return (__m128i) __builtin_ia32_vpermi2varhi128_mask ((__v8hi) __A, | |
653 | (__v8hi) __I | |
654 | /* idx */ , | |
655 | (__v8hi) __B, | |
656 | (__mmask8) | |
657 | __U); | |
658 | } | |
659 | ||
660 | extern __inline __m128i | |
661 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
662 | _mm_maskz_permutex2var_epi16 (__mmask8 __U, __m128i __A, __m128i __I, | |
663 | __m128i __B) | |
664 | { | |
665 | return (__m128i) __builtin_ia32_vpermt2varhi128_maskz ((__v8hi) __I | |
666 | /* idx */ , | |
667 | (__v8hi) __A, | |
668 | (__v8hi) __B, | |
669 | (__mmask8) | |
670 | __U); | |
671 | } | |
672 | ||
673 | extern __inline __m256i | |
674 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
675 | _mm256_mask_maddubs_epi16 (__m256i __W, __mmask16 __U, __m256i __X, | |
676 | __m256i __Y) | |
677 | { | |
678 | return (__m256i) __builtin_ia32_pmaddubsw256_mask ((__v32qi) __X, | |
679 | (__v32qi) __Y, | |
680 | (__v16hi) __W, | |
681 | (__mmask16) __U); | |
682 | } | |
683 | ||
684 | extern __inline __m256i | |
685 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
686 | _mm256_maskz_maddubs_epi16 (__mmask16 __U, __m256i __X, __m256i __Y) | |
687 | { | |
688 | return (__m256i) __builtin_ia32_pmaddubsw256_mask ((__v32qi) __X, | |
689 | (__v32qi) __Y, | |
690 | (__v16hi) | |
691 | _mm256_setzero_si256 (), | |
692 | (__mmask16) __U); | |
693 | } | |
694 | ||
695 | extern __inline __m128i | |
696 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
697 | _mm_mask_maddubs_epi16 (__m128i __W, __mmask8 __U, __m128i __X, | |
698 | __m128i __Y) | |
699 | { | |
700 | return (__m128i) __builtin_ia32_pmaddubsw128_mask ((__v16qi) __X, | |
701 | (__v16qi) __Y, | |
702 | (__v8hi) __W, | |
703 | (__mmask8) __U); | |
704 | } | |
705 | ||
706 | extern __inline __m128i | |
707 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
708 | _mm_maskz_maddubs_epi16 (__mmask8 __U, __m128i __X, __m128i __Y) | |
709 | { | |
710 | return (__m128i) __builtin_ia32_pmaddubsw128_mask ((__v16qi) __X, | |
711 | (__v16qi) __Y, | |
712 | (__v8hi) | |
713 | _mm_setzero_hi (), | |
714 | (__mmask8) __U); | |
715 | } | |
716 | ||
717 | extern __inline __m256i | |
718 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
719 | _mm256_mask_madd_epi16 (__m256i __W, __mmask8 __U, __m256i __A, | |
720 | __m256i __B) | |
721 | { | |
722 | return (__m256i) __builtin_ia32_pmaddwd256_mask ((__v16hi) __A, | |
723 | (__v16hi) __B, | |
724 | (__v8si) __W, | |
725 | (__mmask8) __U); | |
726 | } | |
727 | ||
728 | extern __inline __m256i | |
729 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
730 | _mm256_maskz_madd_epi16 (__mmask8 __U, __m256i __A, __m256i __B) | |
731 | { | |
732 | return (__m256i) __builtin_ia32_pmaddwd256_mask ((__v16hi) __A, | |
733 | (__v16hi) __B, | |
734 | (__v8si) | |
735 | _mm256_setzero_si256 (), | |
736 | (__mmask8) __U); | |
737 | } | |
738 | ||
739 | extern __inline __m128i | |
740 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
741 | _mm_mask_madd_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
742 | __m128i __B) | |
743 | { | |
744 | return (__m128i) __builtin_ia32_pmaddwd128_mask ((__v8hi) __A, | |
745 | (__v8hi) __B, | |
746 | (__v4si) __W, | |
747 | (__mmask8) __U); | |
748 | } | |
749 | ||
750 | extern __inline __m128i | |
751 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
752 | _mm_maskz_madd_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
753 | { | |
754 | return (__m128i) __builtin_ia32_pmaddwd128_mask ((__v8hi) __A, | |
755 | (__v8hi) __B, | |
756 | (__v4si) | |
757 | _mm_setzero_si128 (), | |
758 | (__mmask8) __U); | |
759 | } | |
760 | ||
761 | extern __inline __mmask16 | |
762 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
763 | _mm_movepi8_mask (__m128i __A) | |
764 | { | |
765 | return (__mmask16) __builtin_ia32_cvtb2mask128 ((__v16qi) __A); | |
766 | } | |
767 | ||
768 | extern __inline __mmask32 | |
769 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
770 | _mm256_movepi8_mask (__m256i __A) | |
771 | { | |
772 | return (__mmask32) __builtin_ia32_cvtb2mask256 ((__v32qi) __A); | |
773 | } | |
774 | ||
775 | extern __inline __mmask8 | |
776 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
777 | _mm_movepi16_mask (__m128i __A) | |
778 | { | |
779 | return (__mmask8) __builtin_ia32_cvtw2mask128 ((__v8hi) __A); | |
780 | } | |
781 | ||
782 | extern __inline __mmask16 | |
783 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
784 | _mm256_movepi16_mask (__m256i __A) | |
785 | { | |
786 | return (__mmask16) __builtin_ia32_cvtw2mask256 ((__v16hi) __A); | |
787 | } | |
788 | ||
789 | extern __inline __m128i | |
790 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
791 | _mm_movm_epi8 (__mmask16 __A) | |
792 | { | |
793 | return (__m128i) __builtin_ia32_cvtmask2b128 (__A); | |
794 | } | |
795 | ||
796 | extern __inline __m256i | |
797 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
798 | _mm256_movm_epi8 (__mmask32 __A) | |
799 | { | |
800 | return (__m256i) __builtin_ia32_cvtmask2b256 (__A); | |
801 | } | |
802 | ||
803 | extern __inline __m128i | |
804 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
805 | _mm_movm_epi16 (__mmask8 __A) | |
806 | { | |
807 | return (__m128i) __builtin_ia32_cvtmask2w128 (__A); | |
808 | } | |
809 | ||
810 | extern __inline __m256i | |
811 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
812 | _mm256_movm_epi16 (__mmask16 __A) | |
813 | { | |
814 | return (__m256i) __builtin_ia32_cvtmask2w256 (__A); | |
815 | } | |
816 | ||
817 | extern __inline __mmask16 | |
818 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
819 | _mm_test_epi8_mask (__m128i __A, __m128i __B) | |
820 | { | |
821 | return (__mmask16) __builtin_ia32_ptestmb128 ((__v16qi) __A, | |
822 | (__v16qi) __B, | |
823 | (__mmask16) -1); | |
824 | } | |
825 | ||
826 | extern __inline __mmask16 | |
827 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
828 | _mm_mask_test_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) | |
829 | { | |
830 | return (__mmask16) __builtin_ia32_ptestmb128 ((__v16qi) __A, | |
831 | (__v16qi) __B, __U); | |
832 | } | |
833 | ||
834 | extern __inline __mmask32 | |
835 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
836 | _mm256_test_epi8_mask (__m256i __A, __m256i __B) | |
837 | { | |
838 | return (__mmask32) __builtin_ia32_ptestmb256 ((__v32qi) __A, | |
839 | (__v32qi) __B, | |
840 | (__mmask32) -1); | |
841 | } | |
842 | ||
843 | extern __inline __mmask32 | |
844 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
845 | _mm256_mask_test_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) | |
846 | { | |
847 | return (__mmask32) __builtin_ia32_ptestmb256 ((__v32qi) __A, | |
848 | (__v32qi) __B, __U); | |
849 | } | |
850 | ||
851 | extern __inline __mmask8 | |
852 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
853 | _mm_test_epi16_mask (__m128i __A, __m128i __B) | |
854 | { | |
855 | return (__mmask8) __builtin_ia32_ptestmw128 ((__v8hi) __A, | |
856 | (__v8hi) __B, | |
857 | (__mmask8) -1); | |
858 | } | |
859 | ||
860 | extern __inline __mmask8 | |
861 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
862 | _mm_mask_test_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) | |
863 | { | |
864 | return (__mmask8) __builtin_ia32_ptestmw128 ((__v8hi) __A, | |
865 | (__v8hi) __B, __U); | |
866 | } | |
867 | ||
868 | extern __inline __mmask16 | |
869 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
870 | _mm256_test_epi16_mask (__m256i __A, __m256i __B) | |
871 | { | |
872 | return (__mmask16) __builtin_ia32_ptestmw256 ((__v16hi) __A, | |
873 | (__v16hi) __B, | |
874 | (__mmask16) -1); | |
875 | } | |
876 | ||
877 | extern __inline __mmask16 | |
878 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
879 | _mm256_mask_test_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) | |
880 | { | |
881 | return (__mmask16) __builtin_ia32_ptestmw256 ((__v16hi) __A, | |
882 | (__v16hi) __B, __U); | |
883 | } | |
884 | ||
885 | extern __inline __m256i | |
886 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
887 | _mm256_maskz_min_epu16 (__mmask16 __M, __m256i __A, __m256i __B) | |
888 | { | |
889 | return (__m256i) __builtin_ia32_pminuw256_mask ((__v16hi) __A, | |
890 | (__v16hi) __B, | |
891 | (__v16hi) | |
892 | _mm256_setzero_si256 (), | |
893 | (__mmask16) __M); | |
894 | } | |
895 | ||
896 | extern __inline __m256i | |
897 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
898 | _mm256_mask_min_epu16 (__m256i __W, __mmask16 __M, __m256i __A, | |
899 | __m256i __B) | |
900 | { | |
901 | return (__m256i) __builtin_ia32_pminuw256_mask ((__v16hi) __A, | |
902 | (__v16hi) __B, | |
903 | (__v16hi) __W, | |
904 | (__mmask16) __M); | |
905 | } | |
906 | ||
907 | extern __inline __m128i | |
908 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
909 | _mm_maskz_min_epu16 (__mmask8 __M, __m128i __A, __m128i __B) | |
910 | { | |
911 | return (__m128i) __builtin_ia32_pminuw128_mask ((__v8hi) __A, | |
912 | (__v8hi) __B, | |
913 | (__v8hi) | |
914 | _mm_setzero_di (), | |
915 | (__mmask8) __M); | |
916 | } | |
917 | ||
918 | extern __inline __m128i | |
919 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
920 | _mm_mask_min_epu16 (__m128i __W, __mmask8 __M, __m128i __A, | |
921 | __m128i __B) | |
922 | { | |
923 | return (__m128i) __builtin_ia32_pminuw128_mask ((__v8hi) __A, | |
924 | (__v8hi) __B, | |
925 | (__v8hi) __W, | |
926 | (__mmask8) __M); | |
927 | } | |
928 | ||
929 | extern __inline __m256i | |
930 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
931 | _mm256_maskz_min_epi16 (__mmask16 __M, __m256i __A, __m256i __B) | |
932 | { | |
933 | return (__m256i) __builtin_ia32_pminsw256_mask ((__v16hi) __A, | |
934 | (__v16hi) __B, | |
935 | (__v16hi) | |
936 | _mm256_setzero_si256 (), | |
937 | (__mmask16) __M); | |
938 | } | |
939 | ||
940 | extern __inline __m256i | |
941 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
942 | _mm256_mask_min_epi16 (__m256i __W, __mmask16 __M, __m256i __A, | |
943 | __m256i __B) | |
944 | { | |
945 | return (__m256i) __builtin_ia32_pminsw256_mask ((__v16hi) __A, | |
946 | (__v16hi) __B, | |
947 | (__v16hi) __W, | |
948 | (__mmask16) __M); | |
949 | } | |
950 | ||
951 | extern __inline __m256i | |
952 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
953 | _mm256_maskz_max_epu8 (__mmask32 __M, __m256i __A, __m256i __B) | |
954 | { | |
955 | return (__m256i) __builtin_ia32_pmaxub256_mask ((__v32qi) __A, | |
956 | (__v32qi) __B, | |
957 | (__v32qi) | |
958 | _mm256_setzero_si256 (), | |
959 | (__mmask32) __M); | |
960 | } | |
961 | ||
962 | extern __inline __m256i | |
963 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
964 | _mm256_mask_max_epu8 (__m256i __W, __mmask32 __M, __m256i __A, | |
965 | __m256i __B) | |
966 | { | |
967 | return (__m256i) __builtin_ia32_pmaxub256_mask ((__v32qi) __A, | |
968 | (__v32qi) __B, | |
969 | (__v32qi) __W, | |
970 | (__mmask32) __M); | |
971 | } | |
972 | ||
973 | extern __inline __m128i | |
974 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
975 | _mm_maskz_max_epu8 (__mmask16 __M, __m128i __A, __m128i __B) | |
976 | { | |
977 | return (__m128i) __builtin_ia32_pmaxub128_mask ((__v16qi) __A, | |
978 | (__v16qi) __B, | |
979 | (__v16qi) | |
980 | _mm_setzero_di (), | |
981 | (__mmask16) __M); | |
982 | } | |
983 | ||
984 | extern __inline __m128i | |
985 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
986 | _mm_mask_max_epu8 (__m128i __W, __mmask16 __M, __m128i __A, | |
987 | __m128i __B) | |
988 | { | |
989 | return (__m128i) __builtin_ia32_pmaxub128_mask ((__v16qi) __A, | |
990 | (__v16qi) __B, | |
991 | (__v16qi) __W, | |
992 | (__mmask16) __M); | |
993 | } | |
994 | ||
995 | extern __inline __m256i | |
996 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
997 | _mm256_maskz_max_epi8 (__mmask32 __M, __m256i __A, __m256i __B) | |
998 | { | |
999 | return (__m256i) __builtin_ia32_pmaxsb256_mask ((__v32qi) __A, | |
1000 | (__v32qi) __B, | |
1001 | (__v32qi) | |
1002 | _mm256_setzero_si256 (), | |
1003 | (__mmask32) __M); | |
1004 | } | |
1005 | ||
1006 | extern __inline __m256i | |
1007 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1008 | _mm256_mask_max_epi8 (__m256i __W, __mmask32 __M, __m256i __A, | |
1009 | __m256i __B) | |
1010 | { | |
1011 | return (__m256i) __builtin_ia32_pmaxsb256_mask ((__v32qi) __A, | |
1012 | (__v32qi) __B, | |
1013 | (__v32qi) __W, | |
1014 | (__mmask32) __M); | |
1015 | } | |
1016 | ||
1017 | extern __inline __m128i | |
1018 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1019 | _mm_maskz_max_epi8 (__mmask16 __M, __m128i __A, __m128i __B) | |
1020 | { | |
1021 | return (__m128i) __builtin_ia32_pmaxsb128_mask ((__v16qi) __A, | |
1022 | (__v16qi) __B, | |
1023 | (__v16qi) | |
1024 | _mm_setzero_di (), | |
1025 | (__mmask16) __M); | |
1026 | } | |
1027 | ||
1028 | extern __inline __m128i | |
1029 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1030 | _mm_mask_max_epi8 (__m128i __W, __mmask16 __M, __m128i __A, | |
1031 | __m128i __B) | |
1032 | { | |
1033 | return (__m128i) __builtin_ia32_pmaxsb128_mask ((__v16qi) __A, | |
1034 | (__v16qi) __B, | |
1035 | (__v16qi) __W, | |
1036 | (__mmask16) __M); | |
1037 | } | |
1038 | ||
1039 | extern __inline __m256i | |
1040 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1041 | _mm256_maskz_min_epu8 (__mmask32 __M, __m256i __A, __m256i __B) | |
1042 | { | |
1043 | return (__m256i) __builtin_ia32_pminub256_mask ((__v32qi) __A, | |
1044 | (__v32qi) __B, | |
1045 | (__v32qi) | |
1046 | _mm256_setzero_si256 (), | |
1047 | (__mmask32) __M); | |
1048 | } | |
1049 | ||
1050 | extern __inline __m256i | |
1051 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1052 | _mm256_mask_min_epu8 (__m256i __W, __mmask32 __M, __m256i __A, | |
1053 | __m256i __B) | |
1054 | { | |
1055 | return (__m256i) __builtin_ia32_pminub256_mask ((__v32qi) __A, | |
1056 | (__v32qi) __B, | |
1057 | (__v32qi) __W, | |
1058 | (__mmask32) __M); | |
1059 | } | |
1060 | ||
1061 | extern __inline __m128i | |
1062 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1063 | _mm_maskz_min_epu8 (__mmask16 __M, __m128i __A, __m128i __B) | |
1064 | { | |
1065 | return (__m128i) __builtin_ia32_pminub128_mask ((__v16qi) __A, | |
1066 | (__v16qi) __B, | |
1067 | (__v16qi) | |
1068 | _mm_setzero_di (), | |
1069 | (__mmask16) __M); | |
1070 | } | |
1071 | ||
1072 | extern __inline __m128i | |
1073 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1074 | _mm_mask_min_epu8 (__m128i __W, __mmask16 __M, __m128i __A, | |
1075 | __m128i __B) | |
1076 | { | |
1077 | return (__m128i) __builtin_ia32_pminub128_mask ((__v16qi) __A, | |
1078 | (__v16qi) __B, | |
1079 | (__v16qi) __W, | |
1080 | (__mmask16) __M); | |
1081 | } | |
1082 | ||
1083 | extern __inline __m256i | |
1084 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1085 | _mm256_maskz_min_epi8 (__mmask32 __M, __m256i __A, __m256i __B) | |
1086 | { | |
1087 | return (__m256i) __builtin_ia32_pminsb256_mask ((__v32qi) __A, | |
1088 | (__v32qi) __B, | |
1089 | (__v32qi) | |
1090 | _mm256_setzero_si256 (), | |
1091 | (__mmask32) __M); | |
1092 | } | |
1093 | ||
1094 | extern __inline __m256i | |
1095 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1096 | _mm256_mask_min_epi8 (__m256i __W, __mmask32 __M, __m256i __A, | |
1097 | __m256i __B) | |
1098 | { | |
1099 | return (__m256i) __builtin_ia32_pminsb256_mask ((__v32qi) __A, | |
1100 | (__v32qi) __B, | |
1101 | (__v32qi) __W, | |
1102 | (__mmask32) __M); | |
1103 | } | |
1104 | ||
1105 | extern __inline __m128i | |
1106 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1107 | _mm_maskz_min_epi8 (__mmask16 __M, __m128i __A, __m128i __B) | |
1108 | { | |
1109 | return (__m128i) __builtin_ia32_pminsb128_mask ((__v16qi) __A, | |
1110 | (__v16qi) __B, | |
1111 | (__v16qi) | |
1112 | _mm_setzero_di (), | |
1113 | (__mmask16) __M); | |
1114 | } | |
1115 | ||
1116 | extern __inline __m128i | |
1117 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1118 | _mm_mask_min_epi8 (__m128i __W, __mmask16 __M, __m128i __A, | |
1119 | __m128i __B) | |
1120 | { | |
1121 | return (__m128i) __builtin_ia32_pminsb128_mask ((__v16qi) __A, | |
1122 | (__v16qi) __B, | |
1123 | (__v16qi) __W, | |
1124 | (__mmask16) __M); | |
1125 | } | |
1126 | ||
1127 | extern __inline __m256i | |
1128 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1129 | _mm256_maskz_max_epi16 (__mmask16 __M, __m256i __A, __m256i __B) | |
1130 | { | |
1131 | return (__m256i) __builtin_ia32_pmaxsw256_mask ((__v16hi) __A, | |
1132 | (__v16hi) __B, | |
1133 | (__v16hi) | |
1134 | _mm256_setzero_si256 (), | |
1135 | (__mmask16) __M); | |
1136 | } | |
1137 | ||
1138 | extern __inline __m256i | |
1139 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1140 | _mm256_mask_max_epi16 (__m256i __W, __mmask16 __M, __m256i __A, | |
1141 | __m256i __B) | |
1142 | { | |
1143 | return (__m256i) __builtin_ia32_pmaxsw256_mask ((__v16hi) __A, | |
1144 | (__v16hi) __B, | |
1145 | (__v16hi) __W, | |
1146 | (__mmask16) __M); | |
1147 | } | |
1148 | ||
1149 | extern __inline __m128i | |
1150 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1151 | _mm_maskz_max_epi16 (__mmask8 __M, __m128i __A, __m128i __B) | |
1152 | { | |
1153 | return (__m128i) __builtin_ia32_pmaxsw128_mask ((__v8hi) __A, | |
1154 | (__v8hi) __B, | |
1155 | (__v8hi) | |
1156 | _mm_setzero_di (), | |
1157 | (__mmask8) __M); | |
1158 | } | |
1159 | ||
1160 | extern __inline __m128i | |
1161 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1162 | _mm_mask_max_epi16 (__m128i __W, __mmask8 __M, __m128i __A, | |
1163 | __m128i __B) | |
1164 | { | |
1165 | return (__m128i) __builtin_ia32_pmaxsw128_mask ((__v8hi) __A, | |
1166 | (__v8hi) __B, | |
1167 | (__v8hi) __W, | |
1168 | (__mmask8) __M); | |
1169 | } | |
1170 | ||
1171 | extern __inline __m256i | |
1172 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1173 | _mm256_maskz_max_epu16 (__mmask16 __M, __m256i __A, __m256i __B) | |
1174 | { | |
1175 | return (__m256i) __builtin_ia32_pmaxuw256_mask ((__v16hi) __A, | |
1176 | (__v16hi) __B, | |
1177 | (__v16hi) | |
1178 | _mm256_setzero_si256 (), | |
1179 | (__mmask16) __M); | |
1180 | } | |
1181 | ||
1182 | extern __inline __m256i | |
1183 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1184 | _mm256_mask_max_epu16 (__m256i __W, __mmask16 __M, __m256i __A, | |
1185 | __m256i __B) | |
1186 | { | |
1187 | return (__m256i) __builtin_ia32_pmaxuw256_mask ((__v16hi) __A, | |
1188 | (__v16hi) __B, | |
1189 | (__v16hi) __W, | |
1190 | (__mmask16) __M); | |
1191 | } | |
1192 | ||
1193 | extern __inline __m128i | |
1194 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1195 | _mm_maskz_max_epu16 (__mmask8 __M, __m128i __A, __m128i __B) | |
1196 | { | |
1197 | return (__m128i) __builtin_ia32_pmaxuw128_mask ((__v8hi) __A, | |
1198 | (__v8hi) __B, | |
1199 | (__v8hi) | |
1200 | _mm_setzero_di (), | |
1201 | (__mmask8) __M); | |
1202 | } | |
1203 | ||
1204 | extern __inline __m128i | |
1205 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1206 | _mm_mask_max_epu16 (__m128i __W, __mmask8 __M, __m128i __A, | |
1207 | __m128i __B) | |
1208 | { | |
1209 | return (__m128i) __builtin_ia32_pmaxuw128_mask ((__v8hi) __A, | |
1210 | (__v8hi) __B, | |
1211 | (__v8hi) __W, | |
1212 | (__mmask8) __M); | |
1213 | } | |
1214 | ||
1215 | extern __inline __m128i | |
1216 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1217 | _mm_maskz_min_epi16 (__mmask8 __M, __m128i __A, __m128i __B) | |
1218 | { | |
1219 | return (__m128i) __builtin_ia32_pminsw128_mask ((__v8hi) __A, | |
1220 | (__v8hi) __B, | |
1221 | (__v8hi) | |
1222 | _mm_setzero_di (), | |
1223 | (__mmask8) __M); | |
1224 | } | |
1225 | ||
1226 | extern __inline __m128i | |
1227 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1228 | _mm_mask_min_epi16 (__m128i __W, __mmask8 __M, __m128i __A, | |
1229 | __m128i __B) | |
1230 | { | |
1231 | return (__m128i) __builtin_ia32_pminsw128_mask ((__v8hi) __A, | |
1232 | (__v8hi) __B, | |
1233 | (__v8hi) __W, | |
1234 | (__mmask8) __M); | |
1235 | } | |
1236 | ||
1237 | #ifdef __OPTIMIZE__ | |
1238 | extern __inline __m256i | |
1239 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1240 | _mm256_mask_alignr_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
1241 | __m256i __B, const int __N) | |
1242 | { | |
1243 | return (__m256i) __builtin_ia32_palignr256_mask ((__v4di) __A, | |
1244 | (__v4di) __B, | |
1245 | __N * 8, | |
1246 | (__v4di) __W, | |
1247 | (__mmask32) __U); | |
1248 | } | |
1249 | ||
1250 | extern __inline __m256i | |
1251 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1252 | _mm256_maskz_alignr_epi8 (__mmask32 __U, __m256i __A, __m256i __B, | |
1253 | const int __N) | |
1254 | { | |
1255 | return (__m256i) __builtin_ia32_palignr256_mask ((__v4di) __A, | |
1256 | (__v4di) __B, | |
1257 | __N * 8, | |
1258 | (__v4di) | |
1259 | _mm256_setzero_si256 (), | |
1260 | (__mmask32) __U); | |
1261 | } | |
1262 | ||
1263 | extern __inline __m128i | |
1264 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1265 | _mm_mask_alignr_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
1266 | __m128i __B, const int __N) | |
1267 | { | |
1268 | return (__m128i) __builtin_ia32_palignr128_mask ((__v2di) __A, | |
1269 | (__v2di) __B, | |
1270 | __N * 8, | |
1271 | (__v2di) __W, | |
1272 | (__mmask16) __U); | |
1273 | } | |
1274 | ||
1275 | extern __inline __m128i | |
1276 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1277 | _mm_maskz_alignr_epi8 (__mmask16 __U, __m128i __A, __m128i __B, | |
1278 | const int __N) | |
1279 | { | |
1280 | return (__m128i) __builtin_ia32_palignr128_mask ((__v2di) __A, | |
1281 | (__v2di) __B, | |
1282 | __N * 8, | |
1283 | (__v2di) | |
1284 | _mm_setzero_si128 (), | |
1285 | (__mmask16) __U); | |
1286 | } | |
1287 | ||
1288 | extern __inline __m256i | |
1289 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1290 | _mm256_dbsad_epu8 (__m256i __A, __m256i __B, const int __imm) | |
1291 | { | |
1292 | return (__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi) __A, | |
1293 | (__v32qi) __B, | |
1294 | __imm, | |
1295 | (__v16hi) | |
1296 | _mm256_setzero_si256 (), | |
1297 | (__mmask16) -1); | |
1298 | } | |
1299 | ||
1300 | extern __inline __m256i | |
1301 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1302 | _mm256_mask_dbsad_epu8 (__m256i __W, __mmask16 __U, __m256i __A, | |
1303 | __m256i __B, const int __imm) | |
1304 | { | |
1305 | return (__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi) __A, | |
1306 | (__v32qi) __B, | |
1307 | __imm, | |
1308 | (__v16hi) __W, | |
1309 | (__mmask16) __U); | |
1310 | } | |
1311 | ||
1312 | extern __inline __m256i | |
1313 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1314 | _mm256_maskz_dbsad_epu8 (__mmask16 __U, __m256i __A, __m256i __B, | |
1315 | const int __imm) | |
1316 | { | |
1317 | return (__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi) __A, | |
1318 | (__v32qi) __B, | |
1319 | __imm, | |
1320 | (__v16hi) | |
1321 | _mm256_setzero_si256 (), | |
1322 | (__mmask16) __U); | |
1323 | } | |
1324 | ||
1325 | extern __inline __m128i | |
1326 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1327 | _mm_dbsad_epu8 (__m128i __A, __m128i __B, const int __imm) | |
1328 | { | |
1329 | return (__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi) __A, | |
1330 | (__v16qi) __B, | |
1331 | __imm, | |
1332 | (__v8hi) | |
1333 | _mm_setzero_hi (), | |
1334 | (__mmask8) -1); | |
1335 | } | |
1336 | ||
1337 | extern __inline __m128i | |
1338 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1339 | _mm_mask_dbsad_epu8 (__m128i __W, __mmask8 __U, __m128i __A, | |
1340 | __m128i __B, const int __imm) | |
1341 | { | |
1342 | return (__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi) __A, | |
1343 | (__v16qi) __B, | |
1344 | __imm, | |
1345 | (__v8hi) __W, | |
1346 | (__mmask8) __U); | |
1347 | } | |
1348 | ||
1349 | extern __inline __m128i | |
1350 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1351 | _mm_maskz_dbsad_epu8 (__mmask8 __U, __m128i __A, __m128i __B, | |
1352 | const int __imm) | |
1353 | { | |
1354 | return (__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi) __A, | |
1355 | (__v16qi) __B, | |
1356 | __imm, | |
1357 | (__v8hi) | |
1358 | _mm_setzero_si128 (), | |
1359 | (__mmask8) __U); | |
1360 | } | |
1361 | ||
1362 | extern __inline __m128i | |
1363 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1364 | _mm_mask_blend_epi16 (__mmask8 __U, __m128i __A, __m128i __W) | |
1365 | { | |
1366 | return (__m128i) __builtin_ia32_blendmw_128_mask ((__v8hi) __A, | |
1367 | (__v8hi) __W, | |
1368 | (__mmask8) __U); | |
1369 | } | |
1370 | ||
1371 | extern __inline __m128i | |
1372 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1373 | _mm_mask_blend_epi8 (__mmask16 __U, __m128i __A, __m128i __W) | |
1374 | { | |
1375 | return (__m128i) __builtin_ia32_blendmb_128_mask ((__v16qi) __A, | |
1376 | (__v16qi) __W, | |
1377 | (__mmask16) __U); | |
1378 | } | |
1379 | ||
1380 | extern __inline __m256i | |
1381 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1382 | _mm256_mask_blend_epi16 (__mmask16 __U, __m256i __A, __m256i __W) | |
1383 | { | |
1384 | return (__m256i) __builtin_ia32_blendmw_256_mask ((__v16hi) __A, | |
1385 | (__v16hi) __W, | |
1386 | (__mmask16) __U); | |
1387 | } | |
1388 | ||
1389 | extern __inline __m256i | |
1390 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1391 | _mm256_mask_blend_epi8 (__mmask32 __U, __m256i __A, __m256i __W) | |
1392 | { | |
1393 | return (__m256i) __builtin_ia32_blendmb_256_mask ((__v32qi) __A, | |
1394 | (__v32qi) __W, | |
1395 | (__mmask32) __U); | |
1396 | } | |
1397 | ||
1398 | extern __inline __mmask8 | |
1399 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1400 | _mm_mask_cmp_epi16_mask (__mmask8 __U, __m128i __X, __m128i __Y, | |
1401 | const int __P) | |
1402 | { | |
1403 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
1404 | (__v8hi) __Y, __P, | |
1405 | (__mmask8) __U); | |
1406 | } | |
1407 | ||
1408 | extern __inline __mmask8 | |
1409 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1410 | _mm_cmp_epi16_mask (__m128i __X, __m128i __Y, const int __P) | |
1411 | { | |
1412 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
1413 | (__v8hi) __Y, __P, | |
1414 | (__mmask8) -1); | |
1415 | } | |
1416 | ||
1417 | extern __inline __mmask16 | |
1418 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1419 | _mm256_mask_cmp_epi16_mask (__mmask16 __U, __m256i __X, __m256i __Y, | |
1420 | const int __P) | |
1421 | { | |
1422 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
1423 | (__v16hi) __Y, __P, | |
1424 | (__mmask16) __U); | |
1425 | } | |
1426 | ||
1427 | extern __inline __mmask16 | |
1428 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1429 | _mm256_cmp_epi16_mask (__m256i __X, __m256i __Y, const int __P) | |
1430 | { | |
1431 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
1432 | (__v16hi) __Y, __P, | |
1433 | (__mmask16) -1); | |
1434 | } | |
1435 | ||
1436 | extern __inline __mmask16 | |
1437 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1438 | _mm_mask_cmp_epi8_mask (__mmask8 __U, __m128i __X, __m128i __Y, | |
1439 | const int __P) | |
1440 | { | |
1441 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
1442 | (__v16qi) __Y, __P, | |
1443 | (__mmask16) __U); | |
1444 | } | |
1445 | ||
1446 | extern __inline __mmask16 | |
1447 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1448 | _mm_cmp_epi8_mask (__m128i __X, __m128i __Y, const int __P) | |
1449 | { | |
1450 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
1451 | (__v16qi) __Y, __P, | |
1452 | (__mmask16) -1); | |
1453 | } | |
1454 | ||
1455 | extern __inline __mmask32 | |
1456 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1457 | _mm256_mask_cmp_epi8_mask (__mmask16 __U, __m256i __X, __m256i __Y, | |
1458 | const int __P) | |
1459 | { | |
1460 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
1461 | (__v32qi) __Y, __P, | |
1462 | (__mmask32) __U); | |
1463 | } | |
1464 | ||
1465 | extern __inline __mmask16 | |
1466 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1467 | _mm256_cmp_epi8_mask (__m256i __X, __m256i __Y, const int __P) | |
1468 | { | |
1469 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
1470 | (__v32qi) __Y, __P, | |
1471 | (__mmask32) -1); | |
1472 | } | |
1473 | ||
1474 | extern __inline __mmask8 | |
1475 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1476 | _mm_mask_cmp_epu16_mask (__mmask8 __U, __m128i __X, __m128i __Y, | |
1477 | const int __P) | |
1478 | { | |
1479 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
1480 | (__v8hi) __Y, __P, | |
1481 | (__mmask8) __U); | |
1482 | } | |
1483 | ||
1484 | extern __inline __mmask8 | |
1485 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1486 | _mm_cmp_epu16_mask (__m128i __X, __m128i __Y, const int __P) | |
1487 | { | |
1488 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
1489 | (__v8hi) __Y, __P, | |
1490 | (__mmask8) -1); | |
1491 | } | |
1492 | ||
1493 | extern __inline __mmask16 | |
1494 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1495 | _mm256_mask_cmp_epu16_mask (__mmask16 __U, __m256i __X, __m256i __Y, | |
1496 | const int __P) | |
1497 | { | |
1498 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
1499 | (__v16hi) __Y, __P, | |
1500 | (__mmask16) __U); | |
1501 | } | |
1502 | ||
1503 | extern __inline __mmask16 | |
1504 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1505 | _mm256_cmp_epu16_mask (__m256i __X, __m256i __Y, const int __P) | |
1506 | { | |
1507 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
1508 | (__v16hi) __Y, __P, | |
1509 | (__mmask16) -1); | |
1510 | } | |
1511 | ||
1512 | extern __inline __mmask16 | |
1513 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1514 | _mm_mask_cmp_epu8_mask (__mmask8 __U, __m128i __X, __m128i __Y, | |
1515 | const int __P) | |
1516 | { | |
1517 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
1518 | (__v16qi) __Y, __P, | |
1519 | (__mmask16) __U); | |
1520 | } | |
1521 | ||
1522 | extern __inline __mmask16 | |
1523 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1524 | _mm_cmp_epu8_mask (__m128i __X, __m128i __Y, const int __P) | |
1525 | { | |
1526 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
1527 | (__v16qi) __Y, __P, | |
1528 | (__mmask16) -1); | |
1529 | } | |
1530 | ||
1531 | extern __inline __mmask32 | |
1532 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1533 | _mm256_mask_cmp_epu8_mask (__mmask16 __U, __m256i __X, __m256i __Y, | |
1534 | const int __P) | |
1535 | { | |
1536 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
1537 | (__v32qi) __Y, __P, | |
1538 | (__mmask32) __U); | |
1539 | } | |
1540 | ||
1541 | extern __inline __mmask16 | |
1542 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1543 | _mm256_cmp_epu8_mask (__m256i __X, __m256i __Y, const int __P) | |
1544 | { | |
1545 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
1546 | (__v32qi) __Y, __P, | |
1547 | (__mmask32) -1); | |
1548 | } | |
1549 | ||
1550 | extern __inline __m256i | |
1551 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1552 | _mm256_mask_srli_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
1553 | const int __imm) | |
1554 | { | |
1555 | return (__m256i) __builtin_ia32_psrlwi256_mask ((__v16hi) __A, __imm, | |
1556 | (__v16hi) __W, | |
1557 | (__mmask16) __U); | |
1558 | } | |
1559 | ||
1560 | extern __inline __m256i | |
1561 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1562 | _mm256_maskz_srli_epi16 (__mmask16 __U, __m256i __A, const int __imm) | |
1563 | { | |
1564 | return (__m256i) __builtin_ia32_psrlwi256_mask ((__v16hi) __A, __imm, | |
1565 | (__v16hi) | |
1566 | _mm256_setzero_si256 (), | |
1567 | (__mmask16) __U); | |
1568 | } | |
1569 | ||
1570 | extern __inline __m128i | |
1571 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1572 | _mm_mask_srli_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
1573 | const int __imm) | |
1574 | { | |
1575 | return (__m128i) __builtin_ia32_psrlwi128_mask ((__v8hi) __A, __imm, | |
1576 | (__v8hi) __W, | |
1577 | (__mmask8) __U); | |
1578 | } | |
1579 | ||
1580 | extern __inline __m128i | |
1581 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1582 | _mm_maskz_srli_epi16 (__mmask8 __U, __m128i __A, const int __imm) | |
1583 | { | |
1584 | return (__m128i) __builtin_ia32_psrlwi128_mask ((__v8hi) __A, __imm, | |
1585 | (__v8hi) | |
1586 | _mm_setzero_si128 (), | |
1587 | (__mmask8) __U); | |
1588 | } | |
1589 | ||
1590 | extern __inline __m256i | |
1591 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1592 | _mm256_mask_shufflehi_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
1593 | const int __imm) | |
1594 | { | |
1595 | return (__m256i) __builtin_ia32_pshufhw256_mask ((__v16hi) __A, | |
1596 | __imm, | |
1597 | (__v16hi) __W, | |
1598 | (__mmask16) __U); | |
1599 | } | |
1600 | ||
1601 | extern __inline __m256i | |
1602 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1603 | _mm256_maskz_shufflehi_epi16 (__mmask16 __U, __m256i __A, | |
1604 | const int __imm) | |
1605 | { | |
1606 | return (__m256i) __builtin_ia32_pshufhw256_mask ((__v16hi) __A, | |
1607 | __imm, | |
1608 | (__v16hi) | |
1609 | _mm256_setzero_si256 (), | |
1610 | (__mmask16) __U); | |
1611 | } | |
1612 | ||
1613 | extern __inline __m128i | |
1614 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1615 | _mm_mask_shufflehi_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
1616 | const int __imm) | |
1617 | { | |
1618 | return (__m128i) __builtin_ia32_pshufhw128_mask ((__v8hi) __A, __imm, | |
1619 | (__v8hi) __W, | |
1620 | (__mmask8) __U); | |
1621 | } | |
1622 | ||
1623 | extern __inline __m128i | |
1624 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1625 | _mm_maskz_shufflehi_epi16 (__mmask8 __U, __m128i __A, const int __imm) | |
1626 | { | |
1627 | return (__m128i) __builtin_ia32_pshufhw128_mask ((__v8hi) __A, __imm, | |
1628 | (__v8hi) | |
1629 | _mm_setzero_hi (), | |
1630 | (__mmask8) __U); | |
1631 | } | |
1632 | ||
1633 | extern __inline __m256i | |
1634 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1635 | _mm256_mask_shufflelo_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
1636 | const int __imm) | |
1637 | { | |
1638 | return (__m256i) __builtin_ia32_pshuflw256_mask ((__v16hi) __A, | |
1639 | __imm, | |
1640 | (__v16hi) __W, | |
1641 | (__mmask16) __U); | |
1642 | } | |
1643 | ||
1644 | extern __inline __m256i | |
1645 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1646 | _mm256_maskz_shufflelo_epi16 (__mmask16 __U, __m256i __A, | |
1647 | const int __imm) | |
1648 | { | |
1649 | return (__m256i) __builtin_ia32_pshuflw256_mask ((__v16hi) __A, | |
1650 | __imm, | |
1651 | (__v16hi) | |
1652 | _mm256_setzero_si256 (), | |
1653 | (__mmask16) __U); | |
1654 | } | |
1655 | ||
1656 | extern __inline __m128i | |
1657 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1658 | _mm_mask_shufflelo_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
1659 | const int __imm) | |
1660 | { | |
1661 | return (__m128i) __builtin_ia32_pshuflw128_mask ((__v8hi) __A, __imm, | |
1662 | (__v8hi) __W, | |
1663 | (__mmask8) __U); | |
1664 | } | |
1665 | ||
1666 | extern __inline __m128i | |
1667 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1668 | _mm_maskz_shufflelo_epi16 (__mmask8 __U, __m128i __A, const int __imm) | |
1669 | { | |
1670 | return (__m128i) __builtin_ia32_pshuflw128_mask ((__v8hi) __A, __imm, | |
1671 | (__v8hi) | |
1672 | _mm_setzero_hi (), | |
1673 | (__mmask8) __U); | |
1674 | } | |
1675 | ||
1676 | extern __inline __m256i | |
1677 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1678 | _mm256_mask_srai_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
1679 | const int __imm) | |
1680 | { | |
1681 | return (__m256i) __builtin_ia32_psrawi256_mask ((__v16hi) __A, __imm, | |
1682 | (__v16hi) __W, | |
1683 | (__mmask16) __U); | |
1684 | } | |
1685 | ||
1686 | extern __inline __m256i | |
1687 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1688 | _mm256_maskz_srai_epi16 (__mmask16 __U, __m256i __A, const int __imm) | |
1689 | { | |
1690 | return (__m256i) __builtin_ia32_psrawi256_mask ((__v16hi) __A, __imm, | |
1691 | (__v16hi) | |
1692 | _mm256_setzero_si256 (), | |
1693 | (__mmask16) __U); | |
1694 | } | |
1695 | ||
1696 | extern __inline __m128i | |
1697 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1698 | _mm_mask_srai_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
1699 | const int __imm) | |
1700 | { | |
1701 | return (__m128i) __builtin_ia32_psrawi128_mask ((__v8hi) __A, __imm, | |
1702 | (__v8hi) __W, | |
1703 | (__mmask8) __U); | |
1704 | } | |
1705 | ||
1706 | extern __inline __m128i | |
1707 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1708 | _mm_maskz_srai_epi16 (__mmask8 __U, __m128i __A, const int __imm) | |
1709 | { | |
1710 | return (__m128i) __builtin_ia32_psrawi128_mask ((__v8hi) __A, __imm, | |
1711 | (__v8hi) | |
1712 | _mm_setzero_si128 (), | |
1713 | (__mmask8) __U); | |
1714 | } | |
1715 | ||
1716 | extern __inline __m256i | |
1717 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1718 | _mm256_mask_slli_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
1719 | int __B) | |
1720 | { | |
1721 | return (__m256i) __builtin_ia32_psllwi256_mask ((__v16hi) __A, __B, | |
1722 | (__v16hi) __W, | |
1723 | (__mmask16) __U); | |
1724 | } | |
1725 | ||
1726 | extern __inline __m256i | |
1727 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1728 | _mm256_maskz_slli_epi16 (__mmask16 __U, __m256i __A, int __B) | |
1729 | { | |
1730 | return (__m256i) __builtin_ia32_psllwi256_mask ((__v16hi) __A, __B, | |
1731 | (__v16hi) | |
1732 | _mm256_setzero_si256 (), | |
1733 | (__mmask16) __U); | |
1734 | } | |
1735 | ||
1736 | extern __inline __m128i | |
1737 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1738 | _mm_mask_slli_epi16 (__m128i __W, __mmask8 __U, __m128i __A, int __B) | |
1739 | { | |
1740 | return (__m128i) __builtin_ia32_psllwi128_mask ((__v8hi) __A, __B, | |
1741 | (__v8hi) __W, | |
1742 | (__mmask8) __U); | |
1743 | } | |
1744 | ||
1745 | extern __inline __m128i | |
1746 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1747 | _mm_maskz_slli_epi16 (__mmask8 __U, __m128i __A, int __B) | |
1748 | { | |
1749 | return (__m128i) __builtin_ia32_psllwi128_mask ((__v8hi) __A, __B, | |
1750 | (__v8hi) | |
1751 | _mm_setzero_si128 (), | |
1752 | (__mmask8) __U); | |
1753 | } | |
1754 | ||
1755 | #else | |
1756 | #define _mm256_mask_alignr_epi8(W, U, X, Y, N) \ | |
1757 | ((__m256i) __builtin_ia32_palignr256_mask ((__v4di)(__m256i)(X), \ | |
1758 | (__v4di)(__m256i)(Y), (int)(N * 8), \ | |
1759 | (__v4di)(__m256i)(X), (__mmask32)(U))) | |
1760 | ||
1761 | #define _mm256_mask_srli_epi16(W, U, A, B) \ | |
1762 | ((__m256i) __builtin_ia32_psrlwi256_mask ((__v16hi)(__m256i)(A), \ | |
1763 | (int)(B), (__v16hi)(__m256i)(W), (__mmask16)(U))) | |
1764 | ||
1765 | #define _mm256_maskz_srli_epi16(U, A, B) \ | |
1766 | ((__m256i) __builtin_ia32_psrlwi256_mask ((__v16hi)(__m256i)(A), \ | |
1767 | (int)(B), (__v16hi)_mm256_setzero_si256 (), (__mmask16)(U))) | |
1768 | ||
1769 | #define _mm_mask_srli_epi16(W, U, A, B) \ | |
1770 | ((__m128i) __builtin_ia32_psrlwi128_mask ((__v8hi)(__m128i)(A), \ | |
1771 | (int)(B), (__v8hi)(__m128i)(W), (__mmask8)(U))) | |
1772 | ||
1773 | #define _mm_maskz_srli_epi16(U, A, B) \ | |
1774 | ((__m128i) __builtin_ia32_psrlwi128_mask ((__v8hi)(__m128i)(A), \ | |
1775 | (int)(B), (__v8hi)_mm_setzero_si128(), (__mmask8)(U))) | |
1776 | ||
1777 | #define _mm256_mask_srai_epi16(W, U, A, B) \ | |
1778 | ((__m256i) __builtin_ia32_psrawi256_mask ((__v16hi)(__m256i)(A), \ | |
1779 | (int)(B), (__v16hi)(__m256i)(W), (__mmask16)(U))) | |
1780 | ||
1781 | #define _mm256_maskz_srai_epi16(U, A, B) \ | |
1782 | ((__m256i) __builtin_ia32_psrawi256_mask ((__v16hi)(__m256i)(A), \ | |
1783 | (int)(B), (__v16hi)_mm256_setzero_si256 (), (__mmask16)(U))) | |
1784 | ||
1785 | #define _mm_mask_srai_epi16(W, U, A, B) \ | |
1786 | ((__m128i) __builtin_ia32_psrawi128_mask ((__v8hi)(__m128i)(A), \ | |
1787 | (int)(B), (__v8hi)(__m128i)(W), (__mmask8)(U))) | |
1788 | ||
1789 | #define _mm_maskz_srai_epi16(U, A, B) \ | |
1790 | ((__m128i) __builtin_ia32_psrawi128_mask ((__v8hi)(__m128i)(A), \ | |
1791 | (int)(B), (__v8hi)_mm_setzero_si128(), (__mmask8)(U))) | |
1792 | ||
1793 | #define _mm256_mask_shufflehi_epi16(W, U, A, B) \ | |
1794 | ((__m256i) __builtin_ia32_pshufhw256_mask ((__v16hi)(__m256i)(A), (int)(B), \ | |
1795 | (__v16hi)(__m256i)(W), \ | |
1796 | (__mmask16)(U))) | |
1797 | ||
1798 | #define _mm256_maskz_shufflehi_epi16(U, A, B) \ | |
1799 | ((__m256i) __builtin_ia32_pshufhw256_mask ((__v16hi)(__m256i)(A), (int)(B), \ | |
1800 | (__v16hi)(__m256i)_mm256_setzero_si256 (), \ | |
1801 | (__mmask16)(U))) | |
1802 | ||
1803 | #define _mm_mask_shufflehi_epi16(W, U, A, B) \ | |
1804 | ((__m128i) __builtin_ia32_pshufhw128_mask ((__v8hi)(__m128i)(A), (int)(B), \ | |
1805 | (__v8hi)(__m128i)(W), \ | |
1806 | (__mmask8)(U))) | |
1807 | ||
1808 | #define _mm_maskz_shufflehi_epi16(U, A, B) \ | |
1809 | ((__m128i) __builtin_ia32_pshufhw128_mask ((__v8hi)(__m128i)(A), (int)(B), \ | |
1810 | (__v8hi)(__m128i)_mm_setzero_hi(), \ | |
1811 | (__mmask8)(U))) | |
1812 | ||
1813 | #define _mm256_mask_shufflelo_epi16(W, U, A, B) \ | |
1814 | ((__m256i) __builtin_ia32_pshuflw256_mask ((__v16hi)(__m256i)(A), (int)(B), \ | |
1815 | (__v16hi)(__m256i)(W), \ | |
1816 | (__mmask16)(U))) | |
1817 | ||
1818 | #define _mm256_maskz_shufflelo_epi16(U, A, B) \ | |
1819 | ((__m256i) __builtin_ia32_pshuflw256_mask ((__v16hi)(__m256i)(A), (int)(B), \ | |
1820 | (__v16hi)(__m256i)_mm256_setzero_si256 (), \ | |
1821 | (__mmask16)(U))) | |
1822 | ||
1823 | #define _mm_mask_shufflelo_epi16(W, U, A, B) \ | |
1824 | ((__m128i) __builtin_ia32_pshuflw128_mask ((__v8hi)(__m128i)(A), (int)(B), \ | |
1825 | (__v8hi)(__m128i)(W), \ | |
1826 | (__mmask8)(U))) | |
1827 | ||
1828 | #define _mm_maskz_shufflelo_epi16(U, A, B) \ | |
1829 | ((__m128i) __builtin_ia32_pshuflw128_mask ((__v8hi)(__m128i)(A), (int)(B), \ | |
1830 | (__v8hi)(__m128i)_mm_setzero_hi(), \ | |
1831 | (__mmask8)(U))) | |
1832 | ||
1833 | #define _mm256_maskz_alignr_epi8(U, X, Y, N) \ | |
1834 | ((__m256i) __builtin_ia32_palignr256_mask ((__v4di)(__m256i)(X), \ | |
1835 | (__v4di)(__m256i)(Y), (int)(N * 8), \ | |
1836 | (__v4di)(__m256i)_mm256_setzero_si256 (), \ | |
1837 | (__mmask32)(U))) | |
1838 | ||
1839 | #define _mm_mask_alignr_epi8(W, U, X, Y, N) \ | |
1840 | ((__m128i) __builtin_ia32_palignr128_mask ((__v2di)(__m128i)(X), \ | |
1841 | (__v2di)(__m128i)(Y), (int)(N * 8), \ | |
1842 | (__v2di)(__m128i)(X), (__mmask16)(U))) | |
1843 | ||
1844 | #define _mm_maskz_alignr_epi8(U, X, Y, N) \ | |
1845 | ((__m128i) __builtin_ia32_palignr128_mask ((__v2di)(__m128i)(X), \ | |
1846 | (__v2di)(__m128i)(Y), (int)(N * 8), \ | |
1847 | (__v2di)(__m128i)_mm_setzero_di(), \ | |
1848 | (__mmask16)(U))) | |
1849 | ||
1850 | #define _mm_mask_slli_epi16(W, U, X, C) \ | |
1851 | ((__m128i)__builtin_ia32_psllwi128_mask ((__v8hi)(__m128i)(X), (int)(C),\ | |
1852 | (__v8hi)(__m128i)(W),\ | |
1853 | (__mmask8)(U))) | |
1854 | ||
1855 | #define _mm_maskz_slli_epi16(U, X, C) \ | |
1856 | ((__m128i)__builtin_ia32_psllwi128_mask ((__v8hi)(__m128i)(X), (int)(C),\ | |
1857 | (__v8hi)(__m128i)_mm_setzero_hi(),\ | |
1858 | (__mmask8)(U))) | |
1859 | ||
1860 | #define _mm256_dbsad_epu8(X, Y, C) \ | |
1861 | ((__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi)(__m256i) (X), \ | |
1862 | (__v32qi)(__m256i) (Y), (int) (C), \ | |
1863 | (__v16hi)(__m256i)_mm256_setzero_si256(),\ | |
1864 | (__mmask16)-1)) | |
1865 | ||
1866 | #define _mm256_mask_slli_epi16(W, U, X, C) \ | |
1867 | ((__m256i)__builtin_ia32_psllwi256_mask ((__v16hi)(__m256i)(X), (int)(C),\ | |
1868 | (__v16hi)(__m256i)(W),\ | |
1869 | (__mmask16)(U))) | |
1870 | ||
1871 | #define _mm256_maskz_slli_epi16(U, X, C) \ | |
1872 | ((__m256i)__builtin_ia32_psllwi256_mask ((__v16hi)(__m256i)(X), (int)(C),\ | |
1873 | (__v16hi)(__m256i)_mm256_setzero_si256 (),\ | |
1874 | (__mmask16)(U))) | |
1875 | ||
1876 | #define _mm256_mask_dbsad_epu8(W, U, X, Y, C) \ | |
1877 | ((__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi)(__m256i) (X), \ | |
1878 | (__v32qi)(__m256i) (Y), (int) (C), \ | |
1879 | (__v16hi)(__m256i)(W), \ | |
1880 | (__mmask16)(U))) | |
1881 | ||
1882 | #define _mm256_maskz_dbsad_epu8(U, X, Y, C) \ | |
1883 | ((__m256i) __builtin_ia32_dbpsadbw256_mask ((__v32qi)(__m256i) (X), \ | |
1884 | (__v32qi)(__m256i) (Y), (int) (C), \ | |
1885 | (__v16hi)(__m256i)_mm256_setzero_si256(),\ | |
1886 | (__mmask16)(U))) | |
1887 | ||
1888 | #define _mm_dbsad_epu8(X, Y, C) \ | |
1889 | ((__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi)(__m128i) (X), \ | |
1890 | (__v16qi)(__m128i) (Y), (int) (C), \ | |
1891 | (__v8hi)(__m128i)_mm_setzero_si128(), \ | |
1892 | (__mmask8)-1)) | |
1893 | ||
1894 | #define _mm_mask_dbsad_epu8(W, U, X, Y, C) \ | |
1895 | ((__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi)(__m128i) (X), \ | |
1896 | (__v16qi)(__m128i) (Y), (int) (C), \ | |
1897 | (__v8hi)(__m128i)(W), \ | |
1898 | (__mmask8)(U))) | |
1899 | ||
1900 | #define _mm_maskz_dbsad_epu8(U, X, Y, C) \ | |
1901 | ((__m128i) __builtin_ia32_dbpsadbw128_mask ((__v16qi)(__m128i) (X), \ | |
1902 | (__v16qi)(__m128i) (Y), (int) (C), \ | |
1903 | (__v8hi)(__m128i)_mm_setzero_si128(), \ | |
1904 | (__mmask8)(U))) | |
1905 | ||
1906 | #define _mm_mask_blend_epi16(__U, __A, __W) \ | |
1907 | ((__m128i) __builtin_ia32_blendmw_128_mask ((__v8hi) (__A), \ | |
1908 | (__v8hi) (__W), \ | |
1909 | (__mmask8) (__U))) | |
1910 | ||
1911 | #define _mm_mask_blend_epi8(__U, __A, __W) \ | |
1912 | ((__m128i) __builtin_ia32_blendmb_128_mask ((__v16qi) (__A), \ | |
1913 | (__v16qi) (__W), \ | |
1914 | (__mmask16) (__U))) | |
1915 | ||
1916 | #define _mm256_mask_blend_epi16(__U, __A, __W) \ | |
1917 | ((__m256i) __builtin_ia32_blendmw_256_mask ((__v16hi) (__A), \ | |
1918 | (__v16hi) (__W), \ | |
1919 | (__mmask16) (__U))) | |
1920 | ||
1921 | #define _mm256_mask_blend_epi8(__U, __A, __W) \ | |
1922 | ((__m256i) __builtin_ia32_blendmb_256_mask ((__v32qi) (__A), \ | |
1923 | (__v32qi) (__W), \ | |
1924 | (__mmask32) (__U))) | |
1925 | ||
1926 | #define _mm_cmp_epi16_mask(X, Y, P) \ | |
1927 | ((__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi)(__m128i)(X), \ | |
1928 | (__v8hi)(__m128i)(Y), (int)(P),\ | |
1929 | (__mmask8)(-1))) | |
1930 | ||
1931 | #define _mm_cmp_epi8_mask(X, Y, P) \ | |
1932 | ((__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi)(__m128i)(X), \ | |
1933 | (__v16qi)(__m128i)(Y), (int)(P),\ | |
1934 | (__mmask16)(-1))) | |
1935 | ||
1936 | #define _mm256_cmp_epi16_mask(X, Y, P) \ | |
1937 | ((__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi)(__m256i)(X), \ | |
1938 | (__v16hi)(__m256i)(Y), (int)(P),\ | |
1939 | (__mmask16)(-1))) | |
1940 | ||
1941 | #define _mm256_cmp_epi8_mask(X, Y, P) \ | |
1942 | ((__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi)(__m256i)(X), \ | |
1943 | (__v32qi)(__m256i)(Y), (int)(P),\ | |
1944 | (__mmask32)(-1))) | |
1945 | ||
1946 | #define _mm_cmp_epu16_mask(X, Y, P) \ | |
1947 | ((__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi)(__m128i)(X), \ | |
1948 | (__v8hi)(__m128i)(Y), (int)(P),\ | |
1949 | (__mmask8)(-1))) | |
1950 | ||
1951 | #define _mm_cmp_epu8_mask(X, Y, P) \ | |
1952 | ((__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi)(__m128i)(X), \ | |
1953 | (__v16qi)(__m128i)(Y), (int)(P),\ | |
1954 | (__mmask16)(-1))) | |
1955 | ||
1956 | #define _mm256_cmp_epu16_mask(X, Y, P) \ | |
1957 | ((__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi)(__m256i)(X), \ | |
1958 | (__v16hi)(__m256i)(Y), (int)(P),\ | |
1959 | (__mmask16)(-1))) | |
1960 | ||
1961 | #define _mm256_cmp_epu8_mask(X, Y, P) \ | |
1962 | ((__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi)(__m256i)(X), \ | |
1963 | (__v32qi)(__m256i)(Y), (int)(P),\ | |
1964 | (__mmask32)-1)) | |
1965 | ||
1966 | #define _mm_mask_cmp_epi16_mask(M, X, Y, P) \ | |
1967 | ((__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi)(__m128i)(X), \ | |
1968 | (__v8hi)(__m128i)(Y), (int)(P),\ | |
1969 | (__mmask16)(M))) | |
1970 | ||
1971 | #define _mm_mask_cmp_epi8_mask(M, X, Y, P) \ | |
1972 | ((__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi)(__m128i)(X), \ | |
1973 | (__v16qi)(__m128i)(Y), (int)(P),\ | |
1974 | (__mmask16)(M))) | |
1975 | ||
1976 | #define _mm256_mask_cmp_epi16_mask(M, X, Y, P) \ | |
1977 | ((__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi)(__m256i)(X), \ | |
1978 | (__v16hi)(__m256i)(Y), (int)(P),\ | |
1979 | (__mmask16)(M))) | |
1980 | ||
1981 | #define _mm256_mask_cmp_epi8_mask(M, X, Y, P) \ | |
1982 | ((__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi)(__m256i)(X), \ | |
1983 | (__v32qi)(__m256i)(Y), (int)(P),\ | |
1984 | (__mmask32)(M))) | |
1985 | ||
1986 | #define _mm_mask_cmp_epu16_mask(M, X, Y, P) \ | |
1987 | ((__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi)(__m128i)(X), \ | |
1988 | (__v8hi)(__m128i)(Y), (int)(P),\ | |
1989 | (__mmask8)(M))) | |
1990 | ||
1991 | #define _mm_mask_cmp_epu8_mask(M, X, Y, P) \ | |
1992 | ((__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi)(__m128i)(X), \ | |
1993 | (__v16qi)(__m128i)(Y), (int)(P),\ | |
1994 | (__mmask16)(M))) | |
1995 | ||
1996 | #define _mm256_mask_cmp_epu16_mask(M, X, Y, P) \ | |
1997 | ((__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi)(__m256i)(X), \ | |
1998 | (__v16hi)(__m256i)(Y), (int)(P),\ | |
1999 | (__mmask16)(M))) | |
2000 | ||
2001 | #define _mm256_mask_cmp_epu8_mask(M, X, Y, P) \ | |
2002 | ((__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi)(__m256i)(X), \ | |
2003 | (__v32qi)(__m256i)(Y), (int)(P),\ | |
2004 | (__mmask32)M)) | |
2005 | #endif | |
2006 | ||
2007 | extern __inline __mmask32 | |
2008 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2009 | _mm256_cmpneq_epi8_mask (__m256i __X, __m256i __Y) | |
2010 | { | |
2011 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
2012 | (__v32qi) __Y, 4, | |
2013 | (__mmask32) - 1); | |
2014 | } | |
2015 | ||
2016 | extern __inline __mmask32 | |
2017 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2018 | _mm256_cmplt_epi8_mask (__m256i __X, __m256i __Y) | |
2019 | { | |
2020 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
2021 | (__v32qi) __Y, 1, | |
2022 | (__mmask32) - 1); | |
2023 | } | |
2024 | ||
2025 | extern __inline __mmask32 | |
2026 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2027 | _mm256_cmpge_epi8_mask (__m256i __X, __m256i __Y) | |
2028 | { | |
2029 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
2030 | (__v32qi) __Y, 5, | |
2031 | (__mmask32) - 1); | |
2032 | } | |
2033 | ||
2034 | extern __inline __mmask32 | |
2035 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2036 | _mm256_cmple_epi8_mask (__m256i __X, __m256i __Y) | |
2037 | { | |
2038 | return (__mmask32) __builtin_ia32_cmpb256_mask ((__v32qi) __X, | |
2039 | (__v32qi) __Y, 2, | |
2040 | (__mmask32) - 1); | |
2041 | } | |
2042 | ||
2043 | extern __inline __mmask16 | |
2044 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2045 | _mm256_cmpneq_epi16_mask (__m256i __X, __m256i __Y) | |
2046 | { | |
2047 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
2048 | (__v16hi) __Y, 4, | |
2049 | (__mmask16) - 1); | |
2050 | } | |
2051 | ||
2052 | extern __inline __mmask16 | |
2053 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2054 | _mm256_cmplt_epi16_mask (__m256i __X, __m256i __Y) | |
2055 | { | |
2056 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
2057 | (__v16hi) __Y, 1, | |
2058 | (__mmask16) - 1); | |
2059 | } | |
2060 | ||
2061 | extern __inline __mmask16 | |
2062 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2063 | _mm256_cmpge_epi16_mask (__m256i __X, __m256i __Y) | |
2064 | { | |
2065 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
2066 | (__v16hi) __Y, 5, | |
2067 | (__mmask16) - 1); | |
2068 | } | |
2069 | ||
2070 | extern __inline __mmask16 | |
2071 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2072 | _mm256_cmple_epi16_mask (__m256i __X, __m256i __Y) | |
2073 | { | |
2074 | return (__mmask16) __builtin_ia32_cmpw256_mask ((__v16hi) __X, | |
2075 | (__v16hi) __Y, 2, | |
2076 | (__mmask16) - 1); | |
2077 | } | |
2078 | ||
2079 | extern __inline __mmask16 | |
2080 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2081 | _mm_cmpneq_epu8_mask (__m128i __X, __m128i __Y) | |
2082 | { | |
2083 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
2084 | (__v16qi) __Y, 4, | |
2085 | (__mmask16) - 1); | |
2086 | } | |
2087 | ||
2088 | extern __inline __mmask16 | |
2089 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2090 | _mm_cmplt_epu8_mask (__m128i __X, __m128i __Y) | |
2091 | { | |
2092 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
2093 | (__v16qi) __Y, 1, | |
2094 | (__mmask16) - 1); | |
2095 | } | |
2096 | ||
2097 | extern __inline __mmask16 | |
2098 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2099 | _mm_cmpge_epu8_mask (__m128i __X, __m128i __Y) | |
2100 | { | |
2101 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
2102 | (__v16qi) __Y, 5, | |
2103 | (__mmask16) - 1); | |
2104 | } | |
2105 | ||
2106 | extern __inline __mmask16 | |
2107 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2108 | _mm_cmple_epu8_mask (__m128i __X, __m128i __Y) | |
2109 | { | |
2110 | return (__mmask16) __builtin_ia32_ucmpb128_mask ((__v16qi) __X, | |
2111 | (__v16qi) __Y, 2, | |
2112 | (__mmask16) - 1); | |
2113 | } | |
2114 | ||
2115 | extern __inline __mmask8 | |
2116 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2117 | _mm_cmpneq_epu16_mask (__m128i __X, __m128i __Y) | |
2118 | { | |
2119 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
2120 | (__v8hi) __Y, 4, | |
2121 | (__mmask8) - 1); | |
2122 | } | |
2123 | ||
2124 | extern __inline __mmask8 | |
2125 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2126 | _mm_cmplt_epu16_mask (__m128i __X, __m128i __Y) | |
2127 | { | |
2128 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
2129 | (__v8hi) __Y, 1, | |
2130 | (__mmask8) - 1); | |
2131 | } | |
2132 | ||
2133 | extern __inline __mmask8 | |
2134 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2135 | _mm_cmpge_epu16_mask (__m128i __X, __m128i __Y) | |
2136 | { | |
2137 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
2138 | (__v8hi) __Y, 5, | |
2139 | (__mmask8) - 1); | |
2140 | } | |
2141 | ||
2142 | extern __inline __mmask8 | |
2143 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2144 | _mm_cmple_epu16_mask (__m128i __X, __m128i __Y) | |
2145 | { | |
2146 | return (__mmask8) __builtin_ia32_ucmpw128_mask ((__v8hi) __X, | |
2147 | (__v8hi) __Y, 2, | |
2148 | (__mmask8) - 1); | |
2149 | } | |
2150 | ||
2151 | extern __inline __mmask16 | |
2152 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2153 | _mm_cmpneq_epi8_mask (__m128i __X, __m128i __Y) | |
2154 | { | |
2155 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
2156 | (__v16qi) __Y, 4, | |
2157 | (__mmask16) - 1); | |
2158 | } | |
2159 | ||
2160 | extern __inline __mmask16 | |
2161 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2162 | _mm_cmplt_epi8_mask (__m128i __X, __m128i __Y) | |
2163 | { | |
2164 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
2165 | (__v16qi) __Y, 1, | |
2166 | (__mmask16) - 1); | |
2167 | } | |
2168 | ||
2169 | extern __inline __mmask16 | |
2170 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2171 | _mm_cmpge_epi8_mask (__m128i __X, __m128i __Y) | |
2172 | { | |
2173 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
2174 | (__v16qi) __Y, 5, | |
2175 | (__mmask16) - 1); | |
2176 | } | |
2177 | ||
2178 | extern __inline __mmask16 | |
2179 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2180 | _mm_cmple_epi8_mask (__m128i __X, __m128i __Y) | |
2181 | { | |
2182 | return (__mmask16) __builtin_ia32_cmpb128_mask ((__v16qi) __X, | |
2183 | (__v16qi) __Y, 2, | |
2184 | (__mmask16) - 1); | |
2185 | } | |
2186 | ||
2187 | extern __inline __mmask8 | |
2188 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2189 | _mm_cmpneq_epi16_mask (__m128i __X, __m128i __Y) | |
2190 | { | |
2191 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
2192 | (__v8hi) __Y, 4, | |
2193 | (__mmask8) - 1); | |
2194 | } | |
2195 | ||
2196 | extern __inline __mmask8 | |
2197 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2198 | _mm_cmplt_epi16_mask (__m128i __X, __m128i __Y) | |
2199 | { | |
2200 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
2201 | (__v8hi) __Y, 1, | |
2202 | (__mmask8) - 1); | |
2203 | } | |
2204 | ||
2205 | extern __inline __mmask8 | |
2206 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2207 | _mm_cmpge_epi16_mask (__m128i __X, __m128i __Y) | |
2208 | { | |
2209 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
2210 | (__v8hi) __Y, 5, | |
2211 | (__mmask8) - 1); | |
2212 | } | |
2213 | ||
2214 | extern __inline __mmask8 | |
2215 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2216 | _mm_cmple_epi16_mask (__m128i __X, __m128i __Y) | |
2217 | { | |
2218 | return (__mmask8) __builtin_ia32_cmpw128_mask ((__v8hi) __X, | |
2219 | (__v8hi) __Y, 2, | |
2220 | (__mmask8) - 1); | |
2221 | } | |
2222 | ||
2223 | extern __inline __m256i | |
2224 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2225 | _mm256_mask_mulhrs_epi16 (__m256i __W, __mmask16 __U, __m256i __X, | |
2226 | __m256i __Y) | |
2227 | { | |
2228 | return (__m256i) __builtin_ia32_pmulhrsw256_mask ((__v16hi) __X, | |
2229 | (__v16hi) __Y, | |
2230 | (__v16hi) __W, | |
2231 | (__mmask16) __U); | |
2232 | } | |
2233 | ||
2234 | extern __inline __m256i | |
2235 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2236 | _mm256_maskz_mulhrs_epi16 (__mmask16 __U, __m256i __X, __m256i __Y) | |
2237 | { | |
2238 | return (__m256i) __builtin_ia32_pmulhrsw256_mask ((__v16hi) __X, | |
2239 | (__v16hi) __Y, | |
2240 | (__v16hi) | |
2241 | _mm256_setzero_si256 (), | |
2242 | (__mmask16) __U); | |
2243 | } | |
2244 | ||
2245 | extern __inline __m256i | |
2246 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2247 | _mm256_mask_mulhi_epu16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2248 | __m256i __B) | |
2249 | { | |
2250 | return (__m256i) __builtin_ia32_pmulhuw256_mask ((__v16hi) __A, | |
2251 | (__v16hi) __B, | |
2252 | (__v16hi) __W, | |
2253 | (__mmask16) __U); | |
2254 | } | |
2255 | ||
2256 | extern __inline __m256i | |
2257 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2258 | _mm256_maskz_mulhi_epu16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2259 | { | |
2260 | return (__m256i) __builtin_ia32_pmulhuw256_mask ((__v16hi) __A, | |
2261 | (__v16hi) __B, | |
2262 | (__v16hi) | |
2263 | _mm256_setzero_si256 (), | |
2264 | (__mmask16) __U); | |
2265 | } | |
2266 | ||
2267 | extern __inline __m256i | |
2268 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2269 | _mm256_mask_mulhi_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2270 | __m256i __B) | |
2271 | { | |
2272 | return (__m256i) __builtin_ia32_pmulhw256_mask ((__v16hi) __A, | |
2273 | (__v16hi) __B, | |
2274 | (__v16hi) __W, | |
2275 | (__mmask16) __U); | |
2276 | } | |
2277 | ||
2278 | extern __inline __m256i | |
2279 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2280 | _mm256_maskz_mulhi_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2281 | { | |
2282 | return (__m256i) __builtin_ia32_pmulhw256_mask ((__v16hi) __A, | |
2283 | (__v16hi) __B, | |
2284 | (__v16hi) | |
2285 | _mm256_setzero_si256 (), | |
2286 | (__mmask16) __U); | |
2287 | } | |
2288 | ||
2289 | extern __inline __m128i | |
2290 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2291 | _mm_mask_mulhi_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2292 | __m128i __B) | |
2293 | { | |
2294 | return (__m128i) __builtin_ia32_pmulhw128_mask ((__v8hi) __A, | |
2295 | (__v8hi) __B, | |
2296 | (__v8hi) __W, | |
2297 | (__mmask8) __U); | |
2298 | } | |
2299 | ||
2300 | extern __inline __m128i | |
2301 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2302 | _mm_maskz_mulhi_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2303 | { | |
2304 | return (__m128i) __builtin_ia32_pmulhw128_mask ((__v8hi) __A, | |
2305 | (__v8hi) __B, | |
2306 | (__v8hi) | |
2307 | _mm_setzero_hi (), | |
2308 | (__mmask8) __U); | |
2309 | } | |
2310 | ||
2311 | extern __inline __m128i | |
2312 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2313 | _mm_mask_mulhi_epu16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2314 | __m128i __B) | |
2315 | { | |
2316 | return (__m128i) __builtin_ia32_pmulhuw128_mask ((__v8hi) __A, | |
2317 | (__v8hi) __B, | |
2318 | (__v8hi) __W, | |
2319 | (__mmask8) __U); | |
2320 | } | |
2321 | ||
2322 | extern __inline __m128i | |
2323 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2324 | _mm_maskz_mulhi_epu16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2325 | { | |
2326 | return (__m128i) __builtin_ia32_pmulhuw128_mask ((__v8hi) __A, | |
2327 | (__v8hi) __B, | |
2328 | (__v8hi) | |
2329 | _mm_setzero_hi (), | |
2330 | (__mmask8) __U); | |
2331 | } | |
2332 | ||
2333 | extern __inline __m128i | |
2334 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2335 | _mm_mask_mulhrs_epi16 (__m128i __W, __mmask8 __U, __m128i __X, | |
2336 | __m128i __Y) | |
2337 | { | |
2338 | return (__m128i) __builtin_ia32_pmulhrsw128_mask ((__v8hi) __X, | |
2339 | (__v8hi) __Y, | |
2340 | (__v8hi) __W, | |
2341 | (__mmask8) __U); | |
2342 | } | |
2343 | ||
2344 | extern __inline __m128i | |
2345 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2346 | _mm_maskz_mulhrs_epi16 (__mmask8 __U, __m128i __X, __m128i __Y) | |
2347 | { | |
2348 | return (__m128i) __builtin_ia32_pmulhrsw128_mask ((__v8hi) __X, | |
2349 | (__v8hi) __Y, | |
2350 | (__v8hi) | |
2351 | _mm_setzero_hi (), | |
2352 | (__mmask8) __U); | |
2353 | } | |
2354 | ||
2355 | extern __inline __m256i | |
2356 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2357 | _mm256_mask_mullo_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2358 | __m256i __B) | |
2359 | { | |
2360 | return (__m256i) __builtin_ia32_pmullw256_mask ((__v16hi) __A, | |
2361 | (__v16hi) __B, | |
2362 | (__v16hi) __W, | |
2363 | (__mmask16) __U); | |
2364 | } | |
2365 | ||
2366 | extern __inline __m256i | |
2367 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2368 | _mm256_maskz_mullo_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2369 | { | |
2370 | return (__m256i) __builtin_ia32_pmullw256_mask ((__v16hi) __A, | |
2371 | (__v16hi) __B, | |
2372 | (__v16hi) | |
2373 | _mm256_setzero_si256 (), | |
2374 | (__mmask16) __U); | |
2375 | } | |
2376 | ||
2377 | extern __inline __m128i | |
2378 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2379 | _mm_mask_mullo_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2380 | __m128i __B) | |
2381 | { | |
2382 | return (__m128i) __builtin_ia32_pmullw128_mask ((__v8hi) __A, | |
2383 | (__v8hi) __B, | |
2384 | (__v8hi) __W, | |
2385 | (__mmask8) __U); | |
2386 | } | |
2387 | ||
2388 | extern __inline __m128i | |
2389 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2390 | _mm_maskz_mullo_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2391 | { | |
2392 | return (__m128i) __builtin_ia32_pmullw128_mask ((__v8hi) __A, | |
2393 | (__v8hi) __B, | |
2394 | (__v8hi) | |
2395 | _mm_setzero_hi (), | |
2396 | (__mmask8) __U); | |
2397 | } | |
2398 | ||
2399 | extern __inline __m256i | |
2400 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2401 | _mm256_mask_cvtepi8_epi16 (__m256i __W, __mmask32 __U, __m128i __A) | |
2402 | { | |
2403 | return (__m256i) __builtin_ia32_pmovsxbw256_mask ((__v16qi) __A, | |
2404 | (__v16hi) __W, | |
2405 | (__mmask16) __U); | |
2406 | } | |
2407 | ||
2408 | extern __inline __m256i | |
2409 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2410 | _mm256_maskz_cvtepi8_epi16 (__mmask16 __U, __m128i __A) | |
2411 | { | |
2412 | return (__m256i) __builtin_ia32_pmovsxbw256_mask ((__v16qi) __A, | |
2413 | (__v16hi) | |
2414 | _mm256_setzero_si256 (), | |
2415 | (__mmask16) __U); | |
2416 | } | |
2417 | ||
2418 | extern __inline __m128i | |
2419 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2420 | _mm_mask_cvtepi8_epi16 (__m128i __W, __mmask32 __U, __m128i __A) | |
2421 | { | |
2422 | return (__m128i) __builtin_ia32_pmovsxbw128_mask ((__v16qi) __A, | |
2423 | (__v8hi) __W, | |
2424 | (__mmask8) __U); | |
2425 | } | |
2426 | ||
2427 | extern __inline __m128i | |
2428 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2429 | _mm_maskz_cvtepi8_epi16 (__mmask8 __U, __m128i __A) | |
2430 | { | |
2431 | return (__m128i) __builtin_ia32_pmovsxbw128_mask ((__v16qi) __A, | |
2432 | (__v8hi) | |
2433 | _mm_setzero_si128 (), | |
2434 | (__mmask8) __U); | |
2435 | } | |
2436 | ||
2437 | extern __inline __m256i | |
2438 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2439 | _mm256_mask_cvtepu8_epi16 (__m256i __W, __mmask32 __U, __m128i __A) | |
2440 | { | |
2441 | return (__m256i) __builtin_ia32_pmovzxbw256_mask ((__v16qi) __A, | |
2442 | (__v16hi) __W, | |
2443 | (__mmask16) __U); | |
2444 | } | |
2445 | ||
2446 | extern __inline __m256i | |
2447 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2448 | _mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A) | |
2449 | { | |
2450 | return (__m256i) __builtin_ia32_pmovzxbw256_mask ((__v16qi) __A, | |
2451 | (__v16hi) | |
2452 | _mm256_setzero_si256 (), | |
2453 | (__mmask16) __U); | |
2454 | } | |
2455 | ||
2456 | extern __inline __m128i | |
2457 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2458 | _mm_mask_cvtepu8_epi16 (__m128i __W, __mmask32 __U, __m128i __A) | |
2459 | { | |
2460 | return (__m128i) __builtin_ia32_pmovzxbw128_mask ((__v16qi) __A, | |
2461 | (__v8hi) __W, | |
2462 | (__mmask8) __U); | |
2463 | } | |
2464 | ||
2465 | extern __inline __m128i | |
2466 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2467 | _mm_maskz_cvtepu8_epi16 (__mmask8 __U, __m128i __A) | |
2468 | { | |
2469 | return (__m128i) __builtin_ia32_pmovzxbw128_mask ((__v16qi) __A, | |
2470 | (__v8hi) | |
2471 | _mm_setzero_si128 (), | |
2472 | (__mmask8) __U); | |
2473 | } | |
2474 | ||
2475 | extern __inline __m256i | |
2476 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2477 | _mm256_mask_avg_epu8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2478 | __m256i __B) | |
2479 | { | |
2480 | return (__m256i) __builtin_ia32_pavgb256_mask ((__v32qi) __A, | |
2481 | (__v32qi) __B, | |
2482 | (__v32qi) __W, | |
2483 | (__mmask32) __U); | |
2484 | } | |
2485 | ||
2486 | extern __inline __m256i | |
2487 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2488 | _mm256_maskz_avg_epu8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2489 | { | |
2490 | return (__m256i) __builtin_ia32_pavgb256_mask ((__v32qi) __A, | |
2491 | (__v32qi) __B, | |
2492 | (__v32qi) | |
2493 | _mm256_setzero_si256 (), | |
2494 | (__mmask32) __U); | |
2495 | } | |
2496 | ||
2497 | extern __inline __m128i | |
2498 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2499 | _mm_mask_avg_epu8 (__m128i __W, __mmask16 __U, __m128i __A, | |
2500 | __m128i __B) | |
2501 | { | |
2502 | return (__m128i) __builtin_ia32_pavgb128_mask ((__v16qi) __A, | |
2503 | (__v16qi) __B, | |
2504 | (__v16qi) __W, | |
2505 | (__mmask16) __U); | |
2506 | } | |
2507 | ||
2508 | extern __inline __m128i | |
2509 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2510 | _mm_maskz_avg_epu8 (__mmask16 __U, __m128i __A, __m128i __B) | |
2511 | { | |
2512 | return (__m128i) __builtin_ia32_pavgb128_mask ((__v16qi) __A, | |
2513 | (__v16qi) __B, | |
2514 | (__v16qi) | |
2515 | _mm_setzero_si128 (), | |
2516 | (__mmask16) __U); | |
2517 | } | |
2518 | ||
2519 | extern __inline __m256i | |
2520 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2521 | _mm256_mask_avg_epu16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2522 | __m256i __B) | |
2523 | { | |
2524 | return (__m256i) __builtin_ia32_pavgw256_mask ((__v16hi) __A, | |
2525 | (__v16hi) __B, | |
2526 | (__v16hi) __W, | |
2527 | (__mmask16) __U); | |
2528 | } | |
2529 | ||
2530 | extern __inline __m256i | |
2531 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2532 | _mm256_maskz_avg_epu16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2533 | { | |
2534 | return (__m256i) __builtin_ia32_pavgw256_mask ((__v16hi) __A, | |
2535 | (__v16hi) __B, | |
2536 | (__v16hi) | |
2537 | _mm256_setzero_si256 (), | |
2538 | (__mmask16) __U); | |
2539 | } | |
2540 | ||
2541 | extern __inline __m128i | |
2542 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2543 | _mm_mask_avg_epu16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2544 | __m128i __B) | |
2545 | { | |
2546 | return (__m128i) __builtin_ia32_pavgw128_mask ((__v8hi) __A, | |
2547 | (__v8hi) __B, | |
2548 | (__v8hi) __W, | |
2549 | (__mmask8) __U); | |
2550 | } | |
2551 | ||
2552 | extern __inline __m128i | |
2553 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2554 | _mm_maskz_avg_epu16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2555 | { | |
2556 | return (__m128i) __builtin_ia32_pavgw128_mask ((__v8hi) __A, | |
2557 | (__v8hi) __B, | |
2558 | (__v8hi) | |
2559 | _mm_setzero_si128 (), | |
2560 | (__mmask8) __U); | |
2561 | } | |
2562 | ||
2563 | extern __inline __m256i | |
2564 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2565 | _mm256_mask_add_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2566 | __m256i __B) | |
2567 | { | |
2568 | return (__m256i) __builtin_ia32_paddb256_mask ((__v32qi) __A, | |
2569 | (__v32qi) __B, | |
2570 | (__v32qi) __W, | |
2571 | (__mmask32) __U); | |
2572 | } | |
2573 | ||
2574 | extern __inline __m256i | |
2575 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2576 | _mm256_maskz_add_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2577 | { | |
2578 | return (__m256i) __builtin_ia32_paddb256_mask ((__v32qi) __A, | |
2579 | (__v32qi) __B, | |
2580 | (__v32qi) | |
2581 | _mm256_setzero_si256 (), | |
2582 | (__mmask32) __U); | |
2583 | } | |
2584 | ||
2585 | extern __inline __m256i | |
2586 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2587 | _mm256_mask_add_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2588 | __m256i __B) | |
2589 | { | |
2590 | return (__m256i) __builtin_ia32_paddw256_mask ((__v16hi) __A, | |
2591 | (__v16hi) __B, | |
2592 | (__v16hi) __W, | |
2593 | (__mmask16) __U); | |
2594 | } | |
2595 | ||
2596 | extern __inline __m256i | |
2597 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2598 | _mm256_maskz_add_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2599 | { | |
2600 | return (__m256i) __builtin_ia32_paddw256_mask ((__v16hi) __A, | |
2601 | (__v16hi) __B, | |
2602 | (__v16hi) | |
2603 | _mm256_setzero_si256 (), | |
2604 | (__mmask16) __U); | |
2605 | } | |
2606 | ||
2607 | extern __inline __m256i | |
2608 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2609 | _mm256_mask_adds_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2610 | __m256i __B) | |
2611 | { | |
2612 | return (__m256i) __builtin_ia32_paddsb256_mask ((__v32qi) __A, | |
2613 | (__v32qi) __B, | |
2614 | (__v32qi) __W, | |
2615 | (__mmask32) __U); | |
2616 | } | |
2617 | ||
2618 | extern __inline __m256i | |
2619 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2620 | _mm256_maskz_adds_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2621 | { | |
2622 | return (__m256i) __builtin_ia32_paddsb256_mask ((__v32qi) __A, | |
2623 | (__v32qi) __B, | |
2624 | (__v32qi) | |
2625 | _mm256_setzero_si256 (), | |
2626 | (__mmask32) __U); | |
2627 | } | |
2628 | ||
2629 | extern __inline __m256i | |
2630 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2631 | _mm256_mask_adds_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2632 | __m256i __B) | |
2633 | { | |
2634 | return (__m256i) __builtin_ia32_paddsw256_mask ((__v16hi) __A, | |
2635 | (__v16hi) __B, | |
2636 | (__v16hi) __W, | |
2637 | (__mmask16) __U); | |
2638 | } | |
2639 | ||
2640 | extern __inline __m256i | |
2641 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2642 | _mm256_maskz_adds_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2643 | { | |
2644 | return (__m256i) __builtin_ia32_paddsw256_mask ((__v16hi) __A, | |
2645 | (__v16hi) __B, | |
2646 | (__v16hi) | |
2647 | _mm256_setzero_si256 (), | |
2648 | (__mmask16) __U); | |
2649 | } | |
2650 | ||
2651 | extern __inline __m256i | |
2652 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2653 | _mm256_mask_adds_epu8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2654 | __m256i __B) | |
2655 | { | |
2656 | return (__m256i) __builtin_ia32_paddusb256_mask ((__v32qi) __A, | |
2657 | (__v32qi) __B, | |
2658 | (__v32qi) __W, | |
2659 | (__mmask32) __U); | |
2660 | } | |
2661 | ||
2662 | extern __inline __m256i | |
2663 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2664 | _mm256_maskz_adds_epu8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2665 | { | |
2666 | return (__m256i) __builtin_ia32_paddusb256_mask ((__v32qi) __A, | |
2667 | (__v32qi) __B, | |
2668 | (__v32qi) | |
2669 | _mm256_setzero_si256 (), | |
2670 | (__mmask32) __U); | |
2671 | } | |
2672 | ||
2673 | extern __inline __m256i | |
2674 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2675 | _mm256_mask_adds_epu16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2676 | __m256i __B) | |
2677 | { | |
2678 | return (__m256i) __builtin_ia32_paddusw256_mask ((__v16hi) __A, | |
2679 | (__v16hi) __B, | |
2680 | (__v16hi) __W, | |
2681 | (__mmask16) __U); | |
2682 | } | |
2683 | ||
2684 | extern __inline __m256i | |
2685 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2686 | _mm256_maskz_adds_epu16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2687 | { | |
2688 | return (__m256i) __builtin_ia32_paddusw256_mask ((__v16hi) __A, | |
2689 | (__v16hi) __B, | |
2690 | (__v16hi) | |
2691 | _mm256_setzero_si256 (), | |
2692 | (__mmask16) __U); | |
2693 | } | |
2694 | ||
2695 | extern __inline __m256i | |
2696 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2697 | _mm256_mask_sub_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2698 | __m256i __B) | |
2699 | { | |
2700 | return (__m256i) __builtin_ia32_psubb256_mask ((__v32qi) __A, | |
2701 | (__v32qi) __B, | |
2702 | (__v32qi) __W, | |
2703 | (__mmask32) __U); | |
2704 | } | |
2705 | ||
2706 | extern __inline __m256i | |
2707 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2708 | _mm256_maskz_sub_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2709 | { | |
2710 | return (__m256i) __builtin_ia32_psubb256_mask ((__v32qi) __A, | |
2711 | (__v32qi) __B, | |
2712 | (__v32qi) | |
2713 | _mm256_setzero_si256 (), | |
2714 | (__mmask32) __U); | |
2715 | } | |
2716 | ||
2717 | extern __inline __m256i | |
2718 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2719 | _mm256_mask_sub_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2720 | __m256i __B) | |
2721 | { | |
2722 | return (__m256i) __builtin_ia32_psubw256_mask ((__v16hi) __A, | |
2723 | (__v16hi) __B, | |
2724 | (__v16hi) __W, | |
2725 | (__mmask16) __U); | |
2726 | } | |
2727 | ||
2728 | extern __inline __m256i | |
2729 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2730 | _mm256_maskz_sub_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2731 | { | |
2732 | return (__m256i) __builtin_ia32_psubw256_mask ((__v16hi) __A, | |
2733 | (__v16hi) __B, | |
2734 | (__v16hi) | |
2735 | _mm256_setzero_si256 (), | |
2736 | (__mmask16) __U); | |
2737 | } | |
2738 | ||
2739 | extern __inline __m256i | |
2740 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2741 | _mm256_mask_subs_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2742 | __m256i __B) | |
2743 | { | |
2744 | return (__m256i) __builtin_ia32_psubsb256_mask ((__v32qi) __A, | |
2745 | (__v32qi) __B, | |
2746 | (__v32qi) __W, | |
2747 | (__mmask32) __U); | |
2748 | } | |
2749 | ||
2750 | extern __inline __m256i | |
2751 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2752 | _mm256_maskz_subs_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2753 | { | |
2754 | return (__m256i) __builtin_ia32_psubsb256_mask ((__v32qi) __A, | |
2755 | (__v32qi) __B, | |
2756 | (__v32qi) | |
2757 | _mm256_setzero_si256 (), | |
2758 | (__mmask32) __U); | |
2759 | } | |
2760 | ||
2761 | extern __inline __m256i | |
2762 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2763 | _mm256_mask_subs_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2764 | __m256i __B) | |
2765 | { | |
2766 | return (__m256i) __builtin_ia32_psubsw256_mask ((__v16hi) __A, | |
2767 | (__v16hi) __B, | |
2768 | (__v16hi) __W, | |
2769 | (__mmask16) __U); | |
2770 | } | |
2771 | ||
2772 | extern __inline __m256i | |
2773 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2774 | _mm256_maskz_subs_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2775 | { | |
2776 | return (__m256i) __builtin_ia32_psubsw256_mask ((__v16hi) __A, | |
2777 | (__v16hi) __B, | |
2778 | (__v16hi) | |
2779 | _mm256_setzero_si256 (), | |
2780 | (__mmask16) __U); | |
2781 | } | |
2782 | ||
2783 | extern __inline __m256i | |
2784 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2785 | _mm256_mask_subs_epu8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2786 | __m256i __B) | |
2787 | { | |
2788 | return (__m256i) __builtin_ia32_psubusb256_mask ((__v32qi) __A, | |
2789 | (__v32qi) __B, | |
2790 | (__v32qi) __W, | |
2791 | (__mmask32) __U); | |
2792 | } | |
2793 | ||
2794 | extern __inline __m256i | |
2795 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2796 | _mm256_maskz_subs_epu8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2797 | { | |
2798 | return (__m256i) __builtin_ia32_psubusb256_mask ((__v32qi) __A, | |
2799 | (__v32qi) __B, | |
2800 | (__v32qi) | |
2801 | _mm256_setzero_si256 (), | |
2802 | (__mmask32) __U); | |
2803 | } | |
2804 | ||
2805 | extern __inline __m256i | |
2806 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2807 | _mm256_mask_subs_epu16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2808 | __m256i __B) | |
2809 | { | |
2810 | return (__m256i) __builtin_ia32_psubusw256_mask ((__v16hi) __A, | |
2811 | (__v16hi) __B, | |
2812 | (__v16hi) __W, | |
2813 | (__mmask16) __U); | |
2814 | } | |
2815 | ||
2816 | extern __inline __m256i | |
2817 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2818 | _mm256_maskz_subs_epu16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2819 | { | |
2820 | return (__m256i) __builtin_ia32_psubusw256_mask ((__v16hi) __A, | |
2821 | (__v16hi) __B, | |
2822 | (__v16hi) | |
2823 | _mm256_setzero_si256 (), | |
2824 | (__mmask16) __U); | |
2825 | } | |
2826 | ||
2827 | extern __inline __m128i | |
2828 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2829 | _mm_mask_add_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
2830 | __m128i __B) | |
2831 | { | |
2832 | return (__m128i) __builtin_ia32_paddb128_mask ((__v16qi) __A, | |
2833 | (__v16qi) __B, | |
2834 | (__v16qi) __W, | |
2835 | (__mmask16) __U); | |
2836 | } | |
2837 | ||
2838 | extern __inline __m128i | |
2839 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2840 | _mm_maskz_add_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
2841 | { | |
2842 | return (__m128i) __builtin_ia32_paddb128_mask ((__v16qi) __A, | |
2843 | (__v16qi) __B, | |
2844 | (__v16qi) | |
2845 | _mm_setzero_si128 (), | |
2846 | (__mmask16) __U); | |
2847 | } | |
2848 | ||
2849 | extern __inline __m128i | |
2850 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2851 | _mm_mask_add_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2852 | __m128i __B) | |
2853 | { | |
2854 | return (__m128i) __builtin_ia32_paddw128_mask ((__v8hi) __A, | |
2855 | (__v8hi) __B, | |
2856 | (__v8hi) __W, | |
2857 | (__mmask8) __U); | |
2858 | } | |
2859 | ||
2860 | extern __inline __m128i | |
2861 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2862 | _mm_maskz_add_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2863 | { | |
2864 | return (__m128i) __builtin_ia32_paddw128_mask ((__v8hi) __A, | |
2865 | (__v8hi) __B, | |
2866 | (__v8hi) | |
2867 | _mm_setzero_si128 (), | |
2868 | (__mmask8) __U); | |
2869 | } | |
2870 | ||
2871 | extern __inline __m256i | |
2872 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2873 | _mm256_mask_unpackhi_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2874 | __m256i __B) | |
2875 | { | |
2876 | return (__m256i) __builtin_ia32_punpckhbw256_mask ((__v32qi) __A, | |
2877 | (__v32qi) __B, | |
2878 | (__v32qi) __W, | |
2879 | (__mmask32) __U); | |
2880 | } | |
2881 | ||
2882 | extern __inline __m256i | |
2883 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2884 | _mm256_maskz_unpackhi_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2885 | { | |
2886 | return (__m256i) __builtin_ia32_punpckhbw256_mask ((__v32qi) __A, | |
2887 | (__v32qi) __B, | |
2888 | (__v32qi) | |
2889 | _mm256_setzero_si256 (), | |
2890 | (__mmask32) __U); | |
2891 | } | |
2892 | ||
2893 | extern __inline __m128i | |
2894 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2895 | _mm_mask_unpackhi_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
2896 | __m128i __B) | |
2897 | { | |
2898 | return (__m128i) __builtin_ia32_punpckhbw128_mask ((__v16qi) __A, | |
2899 | (__v16qi) __B, | |
2900 | (__v16qi) __W, | |
2901 | (__mmask16) __U); | |
2902 | } | |
2903 | ||
2904 | extern __inline __m128i | |
2905 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2906 | _mm_maskz_unpackhi_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
2907 | { | |
2908 | return (__m128i) __builtin_ia32_punpckhbw128_mask ((__v16qi) __A, | |
2909 | (__v16qi) __B, | |
2910 | (__v16qi) | |
2911 | _mm_setzero_si128 (), | |
2912 | (__mmask16) __U); | |
2913 | } | |
2914 | ||
2915 | extern __inline __m256i | |
2916 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2917 | _mm256_mask_unpackhi_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
2918 | __m256i __B) | |
2919 | { | |
2920 | return (__m256i) __builtin_ia32_punpckhwd256_mask ((__v16hi) __A, | |
2921 | (__v16hi) __B, | |
2922 | (__v16hi) __W, | |
2923 | (__mmask16) __U); | |
2924 | } | |
2925 | ||
2926 | extern __inline __m256i | |
2927 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2928 | _mm256_maskz_unpackhi_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
2929 | { | |
2930 | return (__m256i) __builtin_ia32_punpckhwd256_mask ((__v16hi) __A, | |
2931 | (__v16hi) __B, | |
2932 | (__v16hi) | |
2933 | _mm256_setzero_si256 (), | |
2934 | (__mmask16) __U); | |
2935 | } | |
2936 | ||
2937 | extern __inline __m128i | |
2938 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2939 | _mm_mask_unpackhi_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
2940 | __m128i __B) | |
2941 | { | |
2942 | return (__m128i) __builtin_ia32_punpckhwd128_mask ((__v8hi) __A, | |
2943 | (__v8hi) __B, | |
2944 | (__v8hi) __W, | |
2945 | (__mmask8) __U); | |
2946 | } | |
2947 | ||
2948 | extern __inline __m128i | |
2949 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2950 | _mm_maskz_unpackhi_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
2951 | { | |
2952 | return (__m128i) __builtin_ia32_punpckhwd128_mask ((__v8hi) __A, | |
2953 | (__v8hi) __B, | |
2954 | (__v8hi) | |
2955 | _mm_setzero_si128 (), | |
2956 | (__mmask8) __U); | |
2957 | } | |
2958 | ||
2959 | extern __inline __m256i | |
2960 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2961 | _mm256_mask_unpacklo_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
2962 | __m256i __B) | |
2963 | { | |
2964 | return (__m256i) __builtin_ia32_punpcklbw256_mask ((__v32qi) __A, | |
2965 | (__v32qi) __B, | |
2966 | (__v32qi) __W, | |
2967 | (__mmask32) __U); | |
2968 | } | |
2969 | ||
2970 | extern __inline __m256i | |
2971 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2972 | _mm256_maskz_unpacklo_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
2973 | { | |
2974 | return (__m256i) __builtin_ia32_punpcklbw256_mask ((__v32qi) __A, | |
2975 | (__v32qi) __B, | |
2976 | (__v32qi) | |
2977 | _mm256_setzero_si256 (), | |
2978 | (__mmask32) __U); | |
2979 | } | |
2980 | ||
2981 | extern __inline __m128i | |
2982 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2983 | _mm_mask_unpacklo_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
2984 | __m128i __B) | |
2985 | { | |
2986 | return (__m128i) __builtin_ia32_punpcklbw128_mask ((__v16qi) __A, | |
2987 | (__v16qi) __B, | |
2988 | (__v16qi) __W, | |
2989 | (__mmask16) __U); | |
2990 | } | |
2991 | ||
2992 | extern __inline __m128i | |
2993 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2994 | _mm_maskz_unpacklo_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
2995 | { | |
2996 | return (__m128i) __builtin_ia32_punpcklbw128_mask ((__v16qi) __A, | |
2997 | (__v16qi) __B, | |
2998 | (__v16qi) | |
2999 | _mm_setzero_si128 (), | |
3000 | (__mmask16) __U); | |
3001 | } | |
3002 | ||
3003 | extern __inline __m256i | |
3004 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3005 | _mm256_mask_unpacklo_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
3006 | __m256i __B) | |
3007 | { | |
3008 | return (__m256i) __builtin_ia32_punpcklwd256_mask ((__v16hi) __A, | |
3009 | (__v16hi) __B, | |
3010 | (__v16hi) __W, | |
3011 | (__mmask16) __U); | |
3012 | } | |
3013 | ||
3014 | extern __inline __m256i | |
3015 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3016 | _mm256_maskz_unpacklo_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
3017 | { | |
3018 | return (__m256i) __builtin_ia32_punpcklwd256_mask ((__v16hi) __A, | |
3019 | (__v16hi) __B, | |
3020 | (__v16hi) | |
3021 | _mm256_setzero_si256 (), | |
3022 | (__mmask16) __U); | |
3023 | } | |
3024 | ||
3025 | extern __inline __m128i | |
3026 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3027 | _mm_mask_unpacklo_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3028 | __m128i __B) | |
3029 | { | |
3030 | return (__m128i) __builtin_ia32_punpcklwd128_mask ((__v8hi) __A, | |
3031 | (__v8hi) __B, | |
3032 | (__v8hi) __W, | |
3033 | (__mmask8) __U); | |
3034 | } | |
3035 | ||
3036 | extern __inline __m128i | |
3037 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3038 | _mm_maskz_unpacklo_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3039 | { | |
3040 | return (__m128i) __builtin_ia32_punpcklwd128_mask ((__v8hi) __A, | |
3041 | (__v8hi) __B, | |
3042 | (__v8hi) | |
3043 | _mm_setzero_si128 (), | |
3044 | (__mmask8) __U); | |
3045 | } | |
3046 | ||
3047 | extern __inline __mmask16 | |
3048 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3049 | _mm_cmpeq_epi8_mask (__m128i __A, __m128i __B) | |
3050 | { | |
3051 | return (__mmask16) __builtin_ia32_pcmpeqb128_mask ((__v16qi) __A, | |
3052 | (__v16qi) __B, | |
3053 | (__mmask16) -1); | |
3054 | } | |
3055 | ||
3056 | extern __inline __mmask16 | |
3057 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3058 | _mm_mask_cmpeq_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) | |
3059 | { | |
3060 | return (__mmask16) __builtin_ia32_pcmpeqb128_mask ((__v16qi) __A, | |
3061 | (__v16qi) __B, | |
3062 | __U); | |
3063 | } | |
3064 | ||
3065 | extern __inline __mmask32 | |
3066 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3067 | _mm256_cmpeq_epi8_mask (__m256i __A, __m256i __B) | |
3068 | { | |
3069 | return (__mmask32) __builtin_ia32_pcmpeqb256_mask ((__v32qi) __A, | |
3070 | (__v32qi) __B, | |
3071 | (__mmask32) -1); | |
3072 | } | |
3073 | ||
3074 | extern __inline __mmask32 | |
3075 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3076 | _mm256_mask_cmpeq_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) | |
3077 | { | |
3078 | return (__mmask32) __builtin_ia32_pcmpeqb256_mask ((__v32qi) __A, | |
3079 | (__v32qi) __B, | |
3080 | __U); | |
3081 | } | |
3082 | ||
3083 | extern __inline __mmask8 | |
3084 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3085 | _mm_cmpeq_epi16_mask (__m128i __A, __m128i __B) | |
3086 | { | |
3087 | return (__mmask8) __builtin_ia32_pcmpeqw128_mask ((__v8hi) __A, | |
3088 | (__v8hi) __B, | |
3089 | (__mmask8) -1); | |
3090 | } | |
3091 | ||
3092 | extern __inline __mmask8 | |
3093 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3094 | _mm_mask_cmpeq_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) | |
3095 | { | |
3096 | return (__mmask8) __builtin_ia32_pcmpeqw128_mask ((__v8hi) __A, | |
3097 | (__v8hi) __B, __U); | |
3098 | } | |
3099 | ||
3100 | extern __inline __mmask16 | |
3101 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3102 | _mm256_cmpeq_epi16_mask (__m256i __A, __m256i __B) | |
3103 | { | |
3104 | return (__mmask16) __builtin_ia32_pcmpeqw256_mask ((__v16hi) __A, | |
3105 | (__v16hi) __B, | |
3106 | (__mmask16) -1); | |
3107 | } | |
3108 | ||
3109 | extern __inline __mmask16 | |
3110 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3111 | _mm256_mask_cmpeq_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) | |
3112 | { | |
3113 | return (__mmask16) __builtin_ia32_pcmpeqw256_mask ((__v16hi) __A, | |
3114 | (__v16hi) __B, | |
3115 | __U); | |
3116 | } | |
3117 | ||
3118 | extern __inline __mmask16 | |
3119 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3120 | _mm_cmpgt_epi8_mask (__m128i __A, __m128i __B) | |
3121 | { | |
3122 | return (__mmask16) __builtin_ia32_pcmpgtb128_mask ((__v16qi) __A, | |
3123 | (__v16qi) __B, | |
3124 | (__mmask16) -1); | |
3125 | } | |
3126 | ||
3127 | extern __inline __mmask16 | |
3128 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3129 | _mm_mask_cmpgt_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) | |
3130 | { | |
3131 | return (__mmask16) __builtin_ia32_pcmpgtb128_mask ((__v16qi) __A, | |
3132 | (__v16qi) __B, | |
3133 | __U); | |
3134 | } | |
3135 | ||
3136 | extern __inline __mmask32 | |
3137 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3138 | _mm256_cmpgt_epi8_mask (__m256i __A, __m256i __B) | |
3139 | { | |
3140 | return (__mmask32) __builtin_ia32_pcmpgtb256_mask ((__v32qi) __A, | |
3141 | (__v32qi) __B, | |
3142 | (__mmask32) -1); | |
3143 | } | |
3144 | ||
3145 | extern __inline __mmask32 | |
3146 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3147 | _mm256_mask_cmpgt_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) | |
3148 | { | |
3149 | return (__mmask32) __builtin_ia32_pcmpgtb256_mask ((__v32qi) __A, | |
3150 | (__v32qi) __B, | |
3151 | __U); | |
3152 | } | |
3153 | ||
3154 | extern __inline __mmask8 | |
3155 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3156 | _mm_cmpgt_epi16_mask (__m128i __A, __m128i __B) | |
3157 | { | |
3158 | return (__mmask8) __builtin_ia32_pcmpgtw128_mask ((__v8hi) __A, | |
3159 | (__v8hi) __B, | |
3160 | (__mmask8) -1); | |
3161 | } | |
3162 | ||
3163 | extern __inline __mmask8 | |
3164 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3165 | _mm_mask_cmpgt_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) | |
3166 | { | |
3167 | return (__mmask8) __builtin_ia32_pcmpgtw128_mask ((__v8hi) __A, | |
3168 | (__v8hi) __B, __U); | |
3169 | } | |
3170 | ||
3171 | extern __inline __mmask16 | |
3172 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3173 | _mm256_cmpgt_epi16_mask (__m256i __A, __m256i __B) | |
3174 | { | |
3175 | return (__mmask16) __builtin_ia32_pcmpgtw256_mask ((__v16hi) __A, | |
3176 | (__v16hi) __B, | |
3177 | (__mmask16) -1); | |
3178 | } | |
3179 | ||
3180 | extern __inline __mmask16 | |
3181 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3182 | _mm256_mask_cmpgt_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) | |
3183 | { | |
3184 | return (__mmask16) __builtin_ia32_pcmpgtw256_mask ((__v16hi) __A, | |
3185 | (__v16hi) __B, | |
3186 | __U); | |
3187 | } | |
3188 | ||
3189 | extern __inline __mmask16 | |
3190 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3191 | _mm_testn_epi8_mask (__m128i __A, __m128i __B) | |
3192 | { | |
3193 | return (__mmask16) __builtin_ia32_ptestnmb128 ((__v16qi) __A, | |
3194 | (__v16qi) __B, | |
3195 | (__mmask16) -1); | |
3196 | } | |
3197 | ||
3198 | extern __inline __mmask16 | |
3199 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3200 | _mm_mask_testn_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) | |
3201 | { | |
3202 | return (__mmask16) __builtin_ia32_ptestnmb128 ((__v16qi) __A, | |
3203 | (__v16qi) __B, __U); | |
3204 | } | |
3205 | ||
3206 | extern __inline __mmask32 | |
3207 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3208 | _mm256_testn_epi8_mask (__m256i __A, __m256i __B) | |
3209 | { | |
3210 | return (__mmask32) __builtin_ia32_ptestnmb256 ((__v32qi) __A, | |
3211 | (__v32qi) __B, | |
3212 | (__mmask32) -1); | |
3213 | } | |
3214 | ||
3215 | extern __inline __mmask32 | |
3216 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3217 | _mm256_mask_testn_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) | |
3218 | { | |
3219 | return (__mmask32) __builtin_ia32_ptestnmb256 ((__v32qi) __A, | |
3220 | (__v32qi) __B, __U); | |
3221 | } | |
3222 | ||
3223 | extern __inline __mmask8 | |
3224 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3225 | _mm_testn_epi16_mask (__m128i __A, __m128i __B) | |
3226 | { | |
3227 | return (__mmask8) __builtin_ia32_ptestnmw128 ((__v8hi) __A, | |
3228 | (__v8hi) __B, | |
3229 | (__mmask8) -1); | |
3230 | } | |
3231 | ||
3232 | extern __inline __mmask8 | |
3233 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3234 | _mm_mask_testn_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) | |
3235 | { | |
3236 | return (__mmask8) __builtin_ia32_ptestnmw128 ((__v8hi) __A, | |
3237 | (__v8hi) __B, __U); | |
3238 | } | |
3239 | ||
3240 | extern __inline __mmask16 | |
3241 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3242 | _mm256_testn_epi16_mask (__m256i __A, __m256i __B) | |
3243 | { | |
3244 | return (__mmask16) __builtin_ia32_ptestnmw256 ((__v16hi) __A, | |
3245 | (__v16hi) __B, | |
3246 | (__mmask16) -1); | |
3247 | } | |
3248 | ||
3249 | extern __inline __mmask16 | |
3250 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3251 | _mm256_mask_testn_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) | |
3252 | { | |
3253 | return (__mmask16) __builtin_ia32_ptestnmw256 ((__v16hi) __A, | |
3254 | (__v16hi) __B, __U); | |
3255 | } | |
3256 | ||
3257 | extern __inline __m256i | |
3258 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3259 | _mm256_mask_shuffle_epi8 (__m256i __W, __mmask32 __U, __m256i __A, | |
3260 | __m256i __B) | |
3261 | { | |
3262 | return (__m256i) __builtin_ia32_pshufb256_mask ((__v32qi) __A, | |
3263 | (__v32qi) __B, | |
3264 | (__v32qi) __W, | |
3265 | (__mmask32) __U); | |
3266 | } | |
3267 | ||
3268 | extern __inline __m256i | |
3269 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3270 | _mm256_maskz_shuffle_epi8 (__mmask32 __U, __m256i __A, __m256i __B) | |
3271 | { | |
3272 | return (__m256i) __builtin_ia32_pshufb256_mask ((__v32qi) __A, | |
3273 | (__v32qi) __B, | |
3274 | (__v32qi) | |
3275 | _mm256_setzero_si256 (), | |
3276 | (__mmask32) __U); | |
3277 | } | |
3278 | ||
3279 | extern __inline __m128i | |
3280 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3281 | _mm_mask_shuffle_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3282 | __m128i __B) | |
3283 | { | |
3284 | return (__m128i) __builtin_ia32_pshufb128_mask ((__v16qi) __A, | |
3285 | (__v16qi) __B, | |
3286 | (__v16qi) __W, | |
3287 | (__mmask16) __U); | |
3288 | } | |
3289 | ||
3290 | extern __inline __m128i | |
3291 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3292 | _mm_maskz_shuffle_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3293 | { | |
3294 | return (__m128i) __builtin_ia32_pshufb128_mask ((__v16qi) __A, | |
3295 | (__v16qi) __B, | |
3296 | (__v16qi) | |
3297 | _mm_setzero_si128 (), | |
3298 | (__mmask16) __U); | |
3299 | } | |
3300 | ||
3301 | extern __inline __m256i | |
3302 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3303 | _mm256_maskz_packs_epi16 (__mmask32 __M, __m256i __A, __m256i __B) | |
3304 | { | |
3305 | return (__m256i) __builtin_ia32_packsswb256_mask ((__v16hi) __A, | |
3306 | (__v16hi) __B, | |
3307 | (__v32qi) | |
3308 | _mm256_setzero_si256 (), | |
3309 | __M); | |
3310 | } | |
3311 | ||
3312 | extern __inline __m256i | |
3313 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3314 | _mm256_mask_packs_epi16 (__m256i __W, __mmask32 __M, __m256i __A, | |
3315 | __m256i __B) | |
3316 | { | |
3317 | return (__m256i) __builtin_ia32_packsswb256_mask ((__v16hi) __A, | |
3318 | (__v16hi) __B, | |
3319 | (__v32qi) __W, | |
3320 | __M); | |
3321 | } | |
3322 | ||
3323 | extern __inline __m128i | |
3324 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3325 | _mm_maskz_packs_epi16 (__mmask16 __M, __m128i __A, __m128i __B) | |
3326 | { | |
3327 | return (__m128i) __builtin_ia32_packsswb128_mask ((__v8hi) __A, | |
3328 | (__v8hi) __B, | |
3329 | (__v16qi) | |
3330 | _mm_setzero_si128 (), | |
3331 | __M); | |
3332 | } | |
3333 | ||
3334 | extern __inline __m128i | |
3335 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3336 | _mm_mask_packs_epi16 (__m128i __W, __mmask16 __M, __m128i __A, | |
3337 | __m128i __B) | |
3338 | { | |
3339 | return (__m128i) __builtin_ia32_packsswb128_mask ((__v8hi) __A, | |
3340 | (__v8hi) __B, | |
3341 | (__v16qi) __W, | |
3342 | __M); | |
3343 | } | |
3344 | ||
3345 | extern __inline __m256i | |
3346 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3347 | _mm256_maskz_packus_epi16 (__mmask32 __M, __m256i __A, __m256i __B) | |
3348 | { | |
3349 | return (__m256i) __builtin_ia32_packuswb256_mask ((__v16hi) __A, | |
3350 | (__v16hi) __B, | |
3351 | (__v32qi) | |
3352 | _mm256_setzero_si256 (), | |
3353 | __M); | |
3354 | } | |
3355 | ||
3356 | extern __inline __m256i | |
3357 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3358 | _mm256_mask_packus_epi16 (__m256i __W, __mmask32 __M, __m256i __A, | |
3359 | __m256i __B) | |
3360 | { | |
3361 | return (__m256i) __builtin_ia32_packuswb256_mask ((__v16hi) __A, | |
3362 | (__v16hi) __B, | |
3363 | (__v32qi) __W, | |
3364 | __M); | |
3365 | } | |
3366 | ||
3367 | extern __inline __m128i | |
3368 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3369 | _mm_maskz_packus_epi16 (__mmask16 __M, __m128i __A, __m128i __B) | |
3370 | { | |
3371 | return (__m128i) __builtin_ia32_packuswb128_mask ((__v8hi) __A, | |
3372 | (__v8hi) __B, | |
3373 | (__v16qi) | |
3374 | _mm_setzero_si128 (), | |
3375 | __M); | |
3376 | } | |
3377 | ||
3378 | extern __inline __m128i | |
3379 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3380 | _mm_mask_packus_epi16 (__m128i __W, __mmask16 __M, __m128i __A, | |
3381 | __m128i __B) | |
3382 | { | |
3383 | return (__m128i) __builtin_ia32_packuswb128_mask ((__v8hi) __A, | |
3384 | (__v8hi) __B, | |
3385 | (__v16qi) __W, | |
3386 | __M); | |
3387 | } | |
3388 | ||
3389 | extern __inline __m256i | |
3390 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3391 | _mm256_mask_abs_epi8 (__m256i __W, __mmask32 __U, __m256i __A) | |
3392 | { | |
3393 | return (__m256i) __builtin_ia32_pabsb256_mask ((__v32qi) __A, | |
3394 | (__v32qi) __W, | |
3395 | (__mmask32) __U); | |
3396 | } | |
3397 | ||
3398 | extern __inline __m256i | |
3399 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3400 | _mm256_maskz_abs_epi8 (__mmask32 __U, __m256i __A) | |
3401 | { | |
3402 | return (__m256i) __builtin_ia32_pabsb256_mask ((__v32qi) __A, | |
3403 | (__v32qi) | |
3404 | _mm256_setzero_si256 (), | |
3405 | (__mmask32) __U); | |
3406 | } | |
3407 | ||
3408 | extern __inline __m128i | |
3409 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3410 | _mm_mask_abs_epi8 (__m128i __W, __mmask16 __U, __m128i __A) | |
3411 | { | |
3412 | return (__m128i) __builtin_ia32_pabsb128_mask ((__v16qi) __A, | |
3413 | (__v16qi) __W, | |
3414 | (__mmask16) __U); | |
3415 | } | |
3416 | ||
3417 | extern __inline __m128i | |
3418 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3419 | _mm_maskz_abs_epi8 (__mmask16 __U, __m128i __A) | |
3420 | { | |
3421 | return (__m128i) __builtin_ia32_pabsb128_mask ((__v16qi) __A, | |
3422 | (__v16qi) | |
3423 | _mm_setzero_si128 (), | |
3424 | (__mmask16) __U); | |
3425 | } | |
3426 | ||
3427 | extern __inline __m256i | |
3428 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3429 | _mm256_mask_abs_epi16 (__m256i __W, __mmask16 __U, __m256i __A) | |
3430 | { | |
3431 | return (__m256i) __builtin_ia32_pabsw256_mask ((__v16hi) __A, | |
3432 | (__v16hi) __W, | |
3433 | (__mmask16) __U); | |
3434 | } | |
3435 | ||
3436 | extern __inline __m256i | |
3437 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3438 | _mm256_maskz_abs_epi16 (__mmask16 __U, __m256i __A) | |
3439 | { | |
3440 | return (__m256i) __builtin_ia32_pabsw256_mask ((__v16hi) __A, | |
3441 | (__v16hi) | |
3442 | _mm256_setzero_si256 (), | |
3443 | (__mmask16) __U); | |
3444 | } | |
3445 | ||
3446 | extern __inline __m128i | |
3447 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3448 | _mm_mask_abs_epi16 (__m128i __W, __mmask8 __U, __m128i __A) | |
3449 | { | |
3450 | return (__m128i) __builtin_ia32_pabsw128_mask ((__v8hi) __A, | |
3451 | (__v8hi) __W, | |
3452 | (__mmask8) __U); | |
3453 | } | |
3454 | ||
3455 | extern __inline __m128i | |
3456 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3457 | _mm_maskz_abs_epi16 (__mmask8 __U, __m128i __A) | |
3458 | { | |
3459 | return (__m128i) __builtin_ia32_pabsw128_mask ((__v8hi) __A, | |
3460 | (__v8hi) | |
3461 | _mm_setzero_si128 (), | |
3462 | (__mmask8) __U); | |
3463 | } | |
3464 | ||
3465 | extern __inline __mmask32 | |
3466 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3467 | _mm256_cmpneq_epu8_mask (__m256i __X, __m256i __Y) | |
3468 | { | |
3469 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
3470 | (__v32qi) __Y, 4, | |
3471 | (__mmask32) - 1); | |
3472 | } | |
3473 | ||
3474 | extern __inline __mmask32 | |
3475 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3476 | _mm256_cmplt_epu8_mask (__m256i __X, __m256i __Y) | |
3477 | { | |
3478 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
3479 | (__v32qi) __Y, 1, | |
3480 | (__mmask32) - 1); | |
3481 | } | |
3482 | ||
3483 | extern __inline __mmask32 | |
3484 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3485 | _mm256_cmpge_epu8_mask (__m256i __X, __m256i __Y) | |
3486 | { | |
3487 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
3488 | (__v32qi) __Y, 5, | |
3489 | (__mmask32) - 1); | |
3490 | } | |
3491 | ||
3492 | extern __inline __mmask32 | |
3493 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3494 | _mm256_cmple_epu8_mask (__m256i __X, __m256i __Y) | |
3495 | { | |
3496 | return (__mmask32) __builtin_ia32_ucmpb256_mask ((__v32qi) __X, | |
3497 | (__v32qi) __Y, 2, | |
3498 | (__mmask32) - 1); | |
3499 | } | |
3500 | ||
3501 | extern __inline __mmask16 | |
3502 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3503 | _mm256_cmpneq_epu16_mask (__m256i __X, __m256i __Y) | |
3504 | { | |
3505 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
3506 | (__v16hi) __Y, 4, | |
3507 | (__mmask16) - 1); | |
3508 | } | |
3509 | ||
3510 | extern __inline __mmask16 | |
3511 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3512 | _mm256_cmplt_epu16_mask (__m256i __X, __m256i __Y) | |
3513 | { | |
3514 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
3515 | (__v16hi) __Y, 1, | |
3516 | (__mmask16) - 1); | |
3517 | } | |
3518 | ||
3519 | extern __inline __mmask16 | |
3520 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3521 | _mm256_cmpge_epu16_mask (__m256i __X, __m256i __Y) | |
3522 | { | |
3523 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
3524 | (__v16hi) __Y, 5, | |
3525 | (__mmask16) - 1); | |
3526 | } | |
3527 | ||
3528 | extern __inline __mmask16 | |
3529 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3530 | _mm256_cmple_epu16_mask (__m256i __X, __m256i __Y) | |
3531 | { | |
3532 | return (__mmask16) __builtin_ia32_ucmpw256_mask ((__v16hi) __X, | |
3533 | (__v16hi) __Y, 2, | |
3534 | (__mmask16) - 1); | |
3535 | } | |
3536 | ||
3537 | extern __inline void | |
3538 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3539 | _mm256_mask_storeu_epi16 (void *__P, __mmask16 __U, __m256i __A) | |
3540 | { | |
3541 | __builtin_ia32_storedquhi256_mask ((__v16hi *) __P, | |
3542 | (__v16hi) __A, | |
3543 | (__mmask16) __U); | |
3544 | } | |
3545 | ||
3546 | extern __inline void | |
3547 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3548 | _mm_mask_storeu_epi16 (void *__P, __mmask8 __U, __m128i __A) | |
3549 | { | |
3550 | __builtin_ia32_storedquhi128_mask ((__v8hi *) __P, | |
3551 | (__v8hi) __A, | |
3552 | (__mmask8) __U); | |
3553 | } | |
3554 | ||
3555 | extern __inline __m128i | |
3556 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3557 | _mm_mask_adds_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3558 | __m128i __B) | |
3559 | { | |
3560 | return (__m128i) __builtin_ia32_paddsw128_mask ((__v8hi) __A, | |
3561 | (__v8hi) __B, | |
3562 | (__v8hi) __W, | |
3563 | (__mmask8) __U); | |
3564 | } | |
3565 | ||
3566 | extern __inline __m128i | |
3567 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3568 | _mm_mask_subs_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3569 | __m128i __B) | |
3570 | { | |
3571 | return (__m128i) __builtin_ia32_psubsb128_mask ((__v16qi) __A, | |
3572 | (__v16qi) __B, | |
3573 | (__v16qi) __W, | |
3574 | (__mmask16) __U); | |
3575 | } | |
3576 | ||
3577 | extern __inline __m128i | |
3578 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3579 | _mm_maskz_subs_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3580 | { | |
3581 | return (__m128i) __builtin_ia32_psubsb128_mask ((__v16qi) __A, | |
3582 | (__v16qi) __B, | |
3583 | (__v16qi) | |
3584 | _mm_setzero_si128 (), | |
3585 | (__mmask16) __U); | |
3586 | } | |
3587 | ||
3588 | extern __inline __m128i | |
3589 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3590 | _mm_mask_subs_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3591 | __m128i __B) | |
3592 | { | |
3593 | return (__m128i) __builtin_ia32_psubsw128_mask ((__v8hi) __A, | |
3594 | (__v8hi) __B, | |
3595 | (__v8hi) __W, | |
3596 | (__mmask8) __U); | |
3597 | } | |
3598 | ||
3599 | extern __inline __m128i | |
3600 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3601 | _mm_maskz_subs_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3602 | { | |
3603 | return (__m128i) __builtin_ia32_psubsw128_mask ((__v8hi) __A, | |
3604 | (__v8hi) __B, | |
3605 | (__v8hi) | |
3606 | _mm_setzero_si128 (), | |
3607 | (__mmask8) __U); | |
3608 | } | |
3609 | ||
3610 | extern __inline __m128i | |
3611 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3612 | _mm_mask_subs_epu8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3613 | __m128i __B) | |
3614 | { | |
3615 | return (__m128i) __builtin_ia32_psubusb128_mask ((__v16qi) __A, | |
3616 | (__v16qi) __B, | |
3617 | (__v16qi) __W, | |
3618 | (__mmask16) __U); | |
3619 | } | |
3620 | ||
3621 | extern __inline __m128i | |
3622 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3623 | _mm_maskz_subs_epu8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3624 | { | |
3625 | return (__m128i) __builtin_ia32_psubusb128_mask ((__v16qi) __A, | |
3626 | (__v16qi) __B, | |
3627 | (__v16qi) | |
3628 | _mm_setzero_si128 (), | |
3629 | (__mmask16) __U); | |
3630 | } | |
3631 | ||
3632 | extern __inline __m128i | |
3633 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3634 | _mm_mask_subs_epu16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3635 | __m128i __B) | |
3636 | { | |
3637 | return (__m128i) __builtin_ia32_psubusw128_mask ((__v8hi) __A, | |
3638 | (__v8hi) __B, | |
3639 | (__v8hi) __W, | |
3640 | (__mmask8) __U); | |
3641 | } | |
3642 | ||
3643 | extern __inline __m128i | |
3644 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3645 | _mm_maskz_subs_epu16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3646 | { | |
3647 | return (__m128i) __builtin_ia32_psubusw128_mask ((__v8hi) __A, | |
3648 | (__v8hi) __B, | |
3649 | (__v8hi) | |
3650 | _mm_setzero_si128 (), | |
3651 | (__mmask8) __U); | |
3652 | } | |
3653 | ||
3654 | extern __inline __m256i | |
3655 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3656 | _mm256_mask_srl_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
3657 | __m128i __B) | |
3658 | { | |
3659 | return (__m256i) __builtin_ia32_psrlw256_mask ((__v16hi) __A, | |
3660 | (__v8hi) __B, | |
3661 | (__v16hi) __W, | |
3662 | (__mmask16) __U); | |
3663 | } | |
3664 | ||
3665 | extern __inline __m256i | |
3666 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3667 | _mm256_maskz_srl_epi16 (__mmask16 __U, __m256i __A, __m128i __B) | |
3668 | { | |
3669 | return (__m256i) __builtin_ia32_psrlw256_mask ((__v16hi) __A, | |
3670 | (__v8hi) __B, | |
3671 | (__v16hi) | |
3672 | _mm256_setzero_si256 (), | |
3673 | (__mmask16) __U); | |
3674 | } | |
3675 | ||
3676 | extern __inline __m128i | |
3677 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3678 | _mm_mask_srl_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3679 | __m128i __B) | |
3680 | { | |
3681 | return (__m128i) __builtin_ia32_psrlw128_mask ((__v8hi) __A, | |
3682 | (__v8hi) __B, | |
3683 | (__v8hi) __W, | |
3684 | (__mmask8) __U); | |
3685 | } | |
3686 | ||
3687 | extern __inline __m128i | |
3688 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3689 | _mm_maskz_srl_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3690 | { | |
3691 | return (__m128i) __builtin_ia32_psrlw128_mask ((__v8hi) __A, | |
3692 | (__v8hi) __B, | |
3693 | (__v8hi) | |
3694 | _mm_setzero_si128 (), | |
3695 | (__mmask8) __U); | |
3696 | } | |
3697 | ||
3698 | extern __inline __m256i | |
3699 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3700 | _mm256_mask_sra_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
3701 | __m128i __B) | |
3702 | { | |
3703 | return (__m256i) __builtin_ia32_psraw256_mask ((__v16hi) __A, | |
3704 | (__v8hi) __B, | |
3705 | (__v16hi) __W, | |
3706 | (__mmask16) __U); | |
3707 | } | |
3708 | ||
3709 | extern __inline __m256i | |
3710 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3711 | _mm256_maskz_sra_epi16 (__mmask16 __U, __m256i __A, __m128i __B) | |
3712 | { | |
3713 | return (__m256i) __builtin_ia32_psraw256_mask ((__v16hi) __A, | |
3714 | (__v8hi) __B, | |
3715 | (__v16hi) | |
3716 | _mm256_setzero_si256 (), | |
3717 | (__mmask16) __U); | |
3718 | } | |
3719 | ||
3720 | extern __inline __m128i | |
3721 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3722 | _mm_mask_sra_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3723 | __m128i __B) | |
3724 | { | |
3725 | return (__m128i) __builtin_ia32_psraw128_mask ((__v8hi) __A, | |
3726 | (__v8hi) __B, | |
3727 | (__v8hi) __W, | |
3728 | (__mmask8) __U); | |
3729 | } | |
3730 | ||
3731 | extern __inline __m128i | |
3732 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3733 | _mm_maskz_sra_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3734 | { | |
3735 | return (__m128i) __builtin_ia32_psraw128_mask ((__v8hi) __A, | |
3736 | (__v8hi) __B, | |
3737 | (__v8hi) | |
3738 | _mm_setzero_si128 (), | |
3739 | (__mmask8) __U); | |
3740 | } | |
3741 | ||
3742 | extern __inline __m128i | |
3743 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3744 | _mm_maskz_adds_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3745 | { | |
3746 | return (__m128i) __builtin_ia32_paddsw128_mask ((__v8hi) __A, | |
3747 | (__v8hi) __B, | |
3748 | (__v8hi) | |
3749 | _mm_setzero_si128 (), | |
3750 | (__mmask8) __U); | |
3751 | } | |
3752 | ||
3753 | extern __inline __m128i | |
3754 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3755 | _mm_mask_adds_epu8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3756 | __m128i __B) | |
3757 | { | |
3758 | return (__m128i) __builtin_ia32_paddusb128_mask ((__v16qi) __A, | |
3759 | (__v16qi) __B, | |
3760 | (__v16qi) __W, | |
3761 | (__mmask16) __U); | |
3762 | } | |
3763 | ||
3764 | extern __inline __m128i | |
3765 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3766 | _mm_maskz_adds_epu8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3767 | { | |
3768 | return (__m128i) __builtin_ia32_paddusb128_mask ((__v16qi) __A, | |
3769 | (__v16qi) __B, | |
3770 | (__v16qi) | |
3771 | _mm_setzero_si128 (), | |
3772 | (__mmask16) __U); | |
3773 | } | |
3774 | ||
3775 | extern __inline __m128i | |
3776 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3777 | _mm_mask_adds_epu16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3778 | __m128i __B) | |
3779 | { | |
3780 | return (__m128i) __builtin_ia32_paddusw128_mask ((__v8hi) __A, | |
3781 | (__v8hi) __B, | |
3782 | (__v8hi) __W, | |
3783 | (__mmask8) __U); | |
3784 | } | |
3785 | ||
3786 | extern __inline __m128i | |
3787 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3788 | _mm_maskz_adds_epu16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3789 | { | |
3790 | return (__m128i) __builtin_ia32_paddusw128_mask ((__v8hi) __A, | |
3791 | (__v8hi) __B, | |
3792 | (__v8hi) | |
3793 | _mm_setzero_si128 (), | |
3794 | (__mmask8) __U); | |
3795 | } | |
3796 | ||
3797 | extern __inline __m128i | |
3798 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3799 | _mm_mask_sub_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3800 | __m128i __B) | |
3801 | { | |
3802 | return (__m128i) __builtin_ia32_psubb128_mask ((__v16qi) __A, | |
3803 | (__v16qi) __B, | |
3804 | (__v16qi) __W, | |
3805 | (__mmask16) __U); | |
3806 | } | |
3807 | ||
3808 | extern __inline __m128i | |
3809 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3810 | _mm_maskz_sub_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3811 | { | |
3812 | return (__m128i) __builtin_ia32_psubb128_mask ((__v16qi) __A, | |
3813 | (__v16qi) __B, | |
3814 | (__v16qi) | |
3815 | _mm_setzero_si128 (), | |
3816 | (__mmask16) __U); | |
3817 | } | |
3818 | ||
3819 | extern __inline __m128i | |
3820 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3821 | _mm_mask_sub_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3822 | __m128i __B) | |
3823 | { | |
3824 | return (__m128i) __builtin_ia32_psubw128_mask ((__v8hi) __A, | |
3825 | (__v8hi) __B, | |
3826 | (__v8hi) __W, | |
3827 | (__mmask8) __U); | |
3828 | } | |
3829 | ||
3830 | extern __inline __m128i | |
3831 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3832 | _mm_maskz_sub_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3833 | { | |
3834 | return (__m128i) __builtin_ia32_psubw128_mask ((__v8hi) __A, | |
3835 | (__v8hi) __B, | |
3836 | (__v8hi) | |
3837 | _mm_setzero_si128 (), | |
3838 | (__mmask8) __U); | |
3839 | } | |
3840 | ||
3841 | extern __inline __m128i | |
3842 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3843 | _mm_mask_adds_epi8 (__m128i __W, __mmask16 __U, __m128i __A, | |
3844 | __m128i __B) | |
3845 | { | |
3846 | return (__m128i) __builtin_ia32_paddsb128_mask ((__v16qi) __A, | |
3847 | (__v16qi) __B, | |
3848 | (__v16qi) __W, | |
3849 | (__mmask16) __U); | |
3850 | } | |
3851 | ||
3852 | extern __inline __m128i | |
3853 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3854 | _mm_maskz_adds_epi8 (__mmask16 __U, __m128i __A, __m128i __B) | |
3855 | { | |
3856 | return (__m128i) __builtin_ia32_paddsb128_mask ((__v16qi) __A, | |
3857 | (__v16qi) __B, | |
3858 | (__v16qi) | |
3859 | _mm_setzero_si128 (), | |
3860 | (__mmask16) __U); | |
3861 | } | |
3862 | ||
3863 | extern __inline __m128i | |
3864 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3865 | _mm_cvtepi16_epi8 (__m128i __A) | |
3866 | { | |
3867 | ||
3868 | return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A, | |
3869 | (__v16qi)_mm_undefined_si128(), | |
3870 | (__mmask8) -1); | |
3871 | } | |
3872 | ||
3873 | extern __inline __m128i | |
3874 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3875 | _mm_mask_cvtepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) | |
3876 | { | |
3877 | return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A, | |
3878 | (__v16qi) __O, __M); | |
3879 | } | |
3880 | ||
3881 | extern __inline __m128i | |
3882 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3883 | _mm_maskz_cvtepi16_epi8 (__mmask8 __M, __m128i __A) | |
3884 | { | |
3885 | return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A, | |
3886 | (__v16qi) | |
3887 | _mm_setzero_si128 (), | |
3888 | __M); | |
3889 | } | |
3890 | ||
3891 | extern __inline __m256i | |
3892 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3893 | _mm256_srav_epi16 (__m256i __A, __m256i __B) | |
3894 | { | |
3895 | return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, | |
3896 | (__v16hi) __B, | |
3897 | (__v16hi) | |
3898 | _mm256_setzero_si256 (), | |
3899 | (__mmask16) -1); | |
3900 | } | |
3901 | ||
3902 | extern __inline __m256i | |
3903 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3904 | _mm256_mask_srav_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
3905 | __m256i __B) | |
3906 | { | |
3907 | return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, | |
3908 | (__v16hi) __B, | |
3909 | (__v16hi) __W, | |
3910 | (__mmask16) __U); | |
3911 | } | |
3912 | ||
3913 | extern __inline __m256i | |
3914 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3915 | _mm256_maskz_srav_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
3916 | { | |
3917 | return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, | |
3918 | (__v16hi) __B, | |
3919 | (__v16hi) | |
3920 | _mm256_setzero_si256 (), | |
3921 | (__mmask16) __U); | |
3922 | } | |
3923 | ||
3924 | extern __inline __m128i | |
3925 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3926 | _mm_srav_epi16 (__m128i __A, __m128i __B) | |
3927 | { | |
3928 | return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, | |
3929 | (__v8hi) __B, | |
3930 | (__v8hi) | |
3931 | _mm_setzero_hi (), | |
3932 | (__mmask8) -1); | |
3933 | } | |
3934 | ||
3935 | extern __inline __m128i | |
3936 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3937 | _mm_mask_srav_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
3938 | __m128i __B) | |
3939 | { | |
3940 | return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, | |
3941 | (__v8hi) __B, | |
3942 | (__v8hi) __W, | |
3943 | (__mmask8) __U); | |
3944 | } | |
3945 | ||
3946 | extern __inline __m128i | |
3947 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3948 | _mm_maskz_srav_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
3949 | { | |
3950 | return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, | |
3951 | (__v8hi) __B, | |
3952 | (__v8hi) | |
3953 | _mm_setzero_si128 (), | |
3954 | (__mmask8) __U); | |
3955 | } | |
3956 | ||
3957 | extern __inline __m256i | |
3958 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3959 | _mm256_srlv_epi16 (__m256i __A, __m256i __B) | |
3960 | { | |
3961 | return (__m256i) __builtin_ia32_psrlv16hi_mask ((__v16hi) __A, | |
3962 | (__v16hi) __B, | |
3963 | (__v16hi) | |
3964 | _mm256_setzero_si256 (), | |
3965 | (__mmask16) -1); | |
3966 | } | |
3967 | ||
3968 | extern __inline __m256i | |
3969 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3970 | _mm256_mask_srlv_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
3971 | __m256i __B) | |
3972 | { | |
3973 | return (__m256i) __builtin_ia32_psrlv16hi_mask ((__v16hi) __A, | |
3974 | (__v16hi) __B, | |
3975 | (__v16hi) __W, | |
3976 | (__mmask16) __U); | |
3977 | } | |
3978 | ||
3979 | extern __inline __m256i | |
3980 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3981 | _mm256_maskz_srlv_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
3982 | { | |
3983 | return (__m256i) __builtin_ia32_psrlv16hi_mask ((__v16hi) __A, | |
3984 | (__v16hi) __B, | |
3985 | (__v16hi) | |
3986 | _mm256_setzero_si256 (), | |
3987 | (__mmask16) __U); | |
3988 | } | |
3989 | ||
3990 | extern __inline __m128i | |
3991 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3992 | _mm_srlv_epi16 (__m128i __A, __m128i __B) | |
3993 | { | |
3994 | return (__m128i) __builtin_ia32_psrlv8hi_mask ((__v8hi) __A, | |
3995 | (__v8hi) __B, | |
3996 | (__v8hi) | |
3997 | _mm_setzero_hi (), | |
3998 | (__mmask8) -1); | |
3999 | } | |
4000 | ||
4001 | extern __inline __m128i | |
4002 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4003 | _mm_mask_srlv_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
4004 | __m128i __B) | |
4005 | { | |
4006 | return (__m128i) __builtin_ia32_psrlv8hi_mask ((__v8hi) __A, | |
4007 | (__v8hi) __B, | |
4008 | (__v8hi) __W, | |
4009 | (__mmask8) __U); | |
4010 | } | |
4011 | ||
4012 | extern __inline __m128i | |
4013 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4014 | _mm_maskz_srlv_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
4015 | { | |
4016 | return (__m128i) __builtin_ia32_psrlv8hi_mask ((__v8hi) __A, | |
4017 | (__v8hi) __B, | |
4018 | (__v8hi) | |
4019 | _mm_setzero_si128 (), | |
4020 | (__mmask8) __U); | |
4021 | } | |
4022 | ||
4023 | extern __inline __m256i | |
4024 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4025 | _mm256_sllv_epi16 (__m256i __A, __m256i __B) | |
4026 | { | |
4027 | return (__m256i) __builtin_ia32_psllv16hi_mask ((__v16hi) __A, | |
4028 | (__v16hi) __B, | |
4029 | (__v16hi) | |
4030 | _mm256_setzero_si256 (), | |
4031 | (__mmask16) -1); | |
4032 | } | |
4033 | ||
4034 | extern __inline __m256i | |
4035 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4036 | _mm256_mask_sllv_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
4037 | __m256i __B) | |
4038 | { | |
4039 | return (__m256i) __builtin_ia32_psllv16hi_mask ((__v16hi) __A, | |
4040 | (__v16hi) __B, | |
4041 | (__v16hi) __W, | |
4042 | (__mmask16) __U); | |
4043 | } | |
4044 | ||
4045 | extern __inline __m256i | |
4046 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4047 | _mm256_maskz_sllv_epi16 (__mmask16 __U, __m256i __A, __m256i __B) | |
4048 | { | |
4049 | return (__m256i) __builtin_ia32_psllv16hi_mask ((__v16hi) __A, | |
4050 | (__v16hi) __B, | |
4051 | (__v16hi) | |
4052 | _mm256_setzero_si256 (), | |
4053 | (__mmask16) __U); | |
4054 | } | |
4055 | ||
4056 | extern __inline __m128i | |
4057 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4058 | _mm_sllv_epi16 (__m128i __A, __m128i __B) | |
4059 | { | |
4060 | return (__m128i) __builtin_ia32_psllv8hi_mask ((__v8hi) __A, | |
4061 | (__v8hi) __B, | |
4062 | (__v8hi) | |
4063 | _mm_setzero_hi (), | |
4064 | (__mmask8) -1); | |
4065 | } | |
4066 | ||
4067 | extern __inline __m128i | |
4068 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4069 | _mm_mask_sllv_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
4070 | __m128i __B) | |
4071 | { | |
4072 | return (__m128i) __builtin_ia32_psllv8hi_mask ((__v8hi) __A, | |
4073 | (__v8hi) __B, | |
4074 | (__v8hi) __W, | |
4075 | (__mmask8) __U); | |
4076 | } | |
4077 | ||
4078 | extern __inline __m128i | |
4079 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4080 | _mm_maskz_sllv_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
4081 | { | |
4082 | return (__m128i) __builtin_ia32_psllv8hi_mask ((__v8hi) __A, | |
4083 | (__v8hi) __B, | |
4084 | (__v8hi) | |
4085 | _mm_setzero_si128 (), | |
4086 | (__mmask8) __U); | |
4087 | } | |
4088 | ||
4089 | extern __inline __m128i | |
4090 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4091 | _mm_mask_sll_epi16 (__m128i __W, __mmask8 __U, __m128i __A, | |
4092 | __m128i __B) | |
4093 | { | |
4094 | return (__m128i) __builtin_ia32_psllw128_mask ((__v8hi) __A, | |
4095 | (__v8hi) __B, | |
4096 | (__v8hi) __W, | |
4097 | (__mmask8) __U); | |
4098 | } | |
4099 | ||
4100 | extern __inline __m128i | |
4101 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4102 | _mm_maskz_sll_epi16 (__mmask8 __U, __m128i __A, __m128i __B) | |
4103 | { | |
4104 | return (__m128i) __builtin_ia32_psllw128_mask ((__v8hi) __A, | |
4105 | (__v8hi) __B, | |
4106 | (__v8hi) | |
4107 | _mm_setzero_si128 (), | |
4108 | (__mmask8) __U); | |
4109 | } | |
4110 | ||
4111 | extern __inline __m256i | |
4112 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4113 | _mm256_mask_sll_epi16 (__m256i __W, __mmask16 __U, __m256i __A, | |
4114 | __m128i __B) | |
4115 | { | |
4116 | return (__m256i) __builtin_ia32_psllw256_mask ((__v16hi) __A, | |
4117 | (__v8hi) __B, | |
4118 | (__v16hi) __W, | |
4119 | (__mmask16) __U); | |
4120 | } | |
4121 | ||
4122 | extern __inline __m256i | |
4123 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4124 | _mm256_maskz_sll_epi16 (__mmask16 __U, __m256i __A, __m128i __B) | |
4125 | { | |
4126 | return (__m256i) __builtin_ia32_psllw256_mask ((__v16hi) __A, | |
4127 | (__v8hi) __B, | |
4128 | (__v16hi) | |
4129 | _mm256_setzero_si256 (), | |
4130 | (__mmask16) __U); | |
4131 | } | |
4132 | ||
4133 | extern __inline __m256i | |
4134 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4135 | _mm256_maskz_packus_epi32 (__mmask16 __M, __m256i __A, __m256i __B) | |
4136 | { | |
4137 | return (__m256i) __builtin_ia32_packusdw256_mask ((__v8si) __A, | |
4138 | (__v8si) __B, | |
4139 | (__v16hi) | |
4140 | _mm256_setzero_si256 (), | |
4141 | __M); | |
4142 | } | |
4143 | ||
4144 | extern __inline __m256i | |
4145 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4146 | _mm256_mask_packus_epi32 (__m256i __W, __mmask16 __M, __m256i __A, | |
4147 | __m256i __B) | |
4148 | { | |
4149 | return (__m256i) __builtin_ia32_packusdw256_mask ((__v8si) __A, | |
4150 | (__v8si) __B, | |
4151 | (__v16hi) __W, | |
4152 | __M); | |
4153 | } | |
4154 | ||
4155 | extern __inline __m128i | |
4156 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4157 | _mm_maskz_packus_epi32 (__mmask8 __M, __m128i __A, __m128i __B) | |
4158 | { | |
4159 | return (__m128i) __builtin_ia32_packusdw128_mask ((__v4si) __A, | |
4160 | (__v4si) __B, | |
4161 | (__v8hi) | |
4162 | _mm_setzero_si128 (), | |
4163 | __M); | |
4164 | } | |
4165 | ||
4166 | extern __inline __m128i | |
4167 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4168 | _mm_mask_packus_epi32 (__m128i __W, __mmask16 __M, __m128i __A, | |
4169 | __m128i __B) | |
4170 | { | |
4171 | return (__m128i) __builtin_ia32_packusdw128_mask ((__v4si) __A, | |
4172 | (__v4si) __B, | |
4173 | (__v8hi) __W, __M); | |
4174 | } | |
4175 | ||
4176 | extern __inline __m256i | |
4177 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4178 | _mm256_maskz_packs_epi32 (__mmask16 __M, __m256i __A, __m256i __B) | |
4179 | { | |
4180 | return (__m256i) __builtin_ia32_packssdw256_mask ((__v8si) __A, | |
4181 | (__v8si) __B, | |
4182 | (__v16hi) | |
4183 | _mm256_setzero_si256 (), | |
4184 | __M); | |
4185 | } | |
4186 | ||
4187 | extern __inline __m256i | |
4188 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4189 | _mm256_mask_packs_epi32 (__m256i __W, __mmask16 __M, __m256i __A, | |
4190 | __m256i __B) | |
4191 | { | |
4192 | return (__m256i) __builtin_ia32_packssdw256_mask ((__v8si) __A, | |
4193 | (__v8si) __B, | |
4194 | (__v16hi) __W, | |
4195 | __M); | |
4196 | } | |
4197 | ||
4198 | extern __inline __m128i | |
4199 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4200 | _mm_maskz_packs_epi32 (__mmask8 __M, __m128i __A, __m128i __B) | |
4201 | { | |
4202 | return (__m128i) __builtin_ia32_packssdw128_mask ((__v4si) __A, | |
4203 | (__v4si) __B, | |
4204 | (__v8hi) | |
4205 | _mm_setzero_si128 (), | |
4206 | __M); | |
4207 | } | |
4208 | ||
4209 | extern __inline __m128i | |
4210 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4211 | _mm_mask_packs_epi32 (__m128i __W, __mmask16 __M, __m128i __A, | |
4212 | __m128i __B) | |
4213 | { | |
4214 | return (__m128i) __builtin_ia32_packssdw128_mask ((__v4si) __A, | |
4215 | (__v4si) __B, | |
4216 | (__v8hi) __W, __M); | |
4217 | } | |
4218 | ||
4219 | #ifdef __DISABLE_AVX512VLBW__ | |
4220 | #undef __DISABLE_AVX512VLBW__ | |
4221 | #pragma GCC pop_options | |
4222 | #endif /* __DISABLE_AVX512VLBW__ */ | |
4223 | ||
4224 | #endif /* _AVX512VLBWINTRIN_H_INCLUDED */ |