clang  19.0.0git
avx512vlbwintrin.h
Go to the documentation of this file.
1 /*===---- avx512vlbwintrin.h - AVX512VL and AVX512BW intrinsics ------------===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  */
9 
10 #ifndef __IMMINTRIN_H
11 #error "Never use <avx512vlbwintrin.h> directly; include <immintrin.h> instead."
12 #endif
13 
14 #ifndef __AVX512VLBWINTRIN_H
15 #define __AVX512VLBWINTRIN_H
16 
17 /* Define the default attributes for the functions in this file. */
18 #define __DEFAULT_FN_ATTRS128 \
19  __attribute__((__always_inline__, __nodebug__, \
20  __target__("avx512vl,avx512bw,no-evex512"), \
21  __min_vector_width__(128)))
22 #define __DEFAULT_FN_ATTRS256 \
23  __attribute__((__always_inline__, __nodebug__, \
24  __target__("avx512vl,avx512bw,no-evex512"), \
25  __min_vector_width__(256)))
26 
27 /* Integer compare */
28 
29 #define _mm_cmp_epi8_mask(a, b, p) \
30  ((__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \
31  (__v16qi)(__m128i)(b), (int)(p), \
32  (__mmask16)-1))
33 
34 #define _mm_mask_cmp_epi8_mask(m, a, b, p) \
35  ((__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \
36  (__v16qi)(__m128i)(b), (int)(p), \
37  (__mmask16)(m)))
38 
39 #define _mm_cmp_epu8_mask(a, b, p) \
40  ((__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \
41  (__v16qi)(__m128i)(b), (int)(p), \
42  (__mmask16)-1))
43 
44 #define _mm_mask_cmp_epu8_mask(m, a, b, p) \
45  ((__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \
46  (__v16qi)(__m128i)(b), (int)(p), \
47  (__mmask16)(m)))
48 
49 #define _mm256_cmp_epi8_mask(a, b, p) \
50  ((__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \
51  (__v32qi)(__m256i)(b), (int)(p), \
52  (__mmask32)-1))
53 
54 #define _mm256_mask_cmp_epi8_mask(m, a, b, p) \
55  ((__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \
56  (__v32qi)(__m256i)(b), (int)(p), \
57  (__mmask32)(m)))
58 
59 #define _mm256_cmp_epu8_mask(a, b, p) \
60  ((__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \
61  (__v32qi)(__m256i)(b), (int)(p), \
62  (__mmask32)-1))
63 
64 #define _mm256_mask_cmp_epu8_mask(m, a, b, p) \
65  ((__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \
66  (__v32qi)(__m256i)(b), (int)(p), \
67  (__mmask32)(m)))
68 
69 #define _mm_cmp_epi16_mask(a, b, p) \
70  ((__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \
71  (__v8hi)(__m128i)(b), (int)(p), \
72  (__mmask8)-1))
73 
74 #define _mm_mask_cmp_epi16_mask(m, a, b, p) \
75  ((__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \
76  (__v8hi)(__m128i)(b), (int)(p), \
77  (__mmask8)(m)))
78 
79 #define _mm_cmp_epu16_mask(a, b, p) \
80  ((__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \
81  (__v8hi)(__m128i)(b), (int)(p), \
82  (__mmask8)-1))
83 
84 #define _mm_mask_cmp_epu16_mask(m, a, b, p) \
85  ((__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \
86  (__v8hi)(__m128i)(b), (int)(p), \
87  (__mmask8)(m)))
88 
89 #define _mm256_cmp_epi16_mask(a, b, p) \
90  ((__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \
91  (__v16hi)(__m256i)(b), (int)(p), \
92  (__mmask16)-1))
93 
94 #define _mm256_mask_cmp_epi16_mask(m, a, b, p) \
95  ((__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \
96  (__v16hi)(__m256i)(b), (int)(p), \
97  (__mmask16)(m)))
98 
99 #define _mm256_cmp_epu16_mask(a, b, p) \
100  ((__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \
101  (__v16hi)(__m256i)(b), (int)(p), \
102  (__mmask16)-1))
103 
104 #define _mm256_mask_cmp_epu16_mask(m, a, b, p) \
105  ((__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \
106  (__v16hi)(__m256i)(b), (int)(p), \
107  (__mmask16)(m)))
108 
109 #define _mm_cmpeq_epi8_mask(A, B) \
110  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ)
111 #define _mm_mask_cmpeq_epi8_mask(k, A, B) \
112  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ)
113 #define _mm_cmpge_epi8_mask(A, B) \
114  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_GE)
115 #define _mm_mask_cmpge_epi8_mask(k, A, B) \
116  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE)
117 #define _mm_cmpgt_epi8_mask(A, B) \
118  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_GT)
119 #define _mm_mask_cmpgt_epi8_mask(k, A, B) \
120  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT)
121 #define _mm_cmple_epi8_mask(A, B) \
122  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_LE)
123 #define _mm_mask_cmple_epi8_mask(k, A, B) \
124  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE)
125 #define _mm_cmplt_epi8_mask(A, B) \
126  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_LT)
127 #define _mm_mask_cmplt_epi8_mask(k, A, B) \
128  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT)
129 #define _mm_cmpneq_epi8_mask(A, B) \
130  _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_NE)
131 #define _mm_mask_cmpneq_epi8_mask(k, A, B) \
132  _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE)
133 
134 #define _mm256_cmpeq_epi8_mask(A, B) \
135  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ)
136 #define _mm256_mask_cmpeq_epi8_mask(k, A, B) \
137  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ)
138 #define _mm256_cmpge_epi8_mask(A, B) \
139  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_GE)
140 #define _mm256_mask_cmpge_epi8_mask(k, A, B) \
141  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE)
142 #define _mm256_cmpgt_epi8_mask(A, B) \
143  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_GT)
144 #define _mm256_mask_cmpgt_epi8_mask(k, A, B) \
145  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT)
146 #define _mm256_cmple_epi8_mask(A, B) \
147  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_LE)
148 #define _mm256_mask_cmple_epi8_mask(k, A, B) \
149  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE)
150 #define _mm256_cmplt_epi8_mask(A, B) \
151  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_LT)
152 #define _mm256_mask_cmplt_epi8_mask(k, A, B) \
153  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT)
154 #define _mm256_cmpneq_epi8_mask(A, B) \
155  _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_NE)
156 #define _mm256_mask_cmpneq_epi8_mask(k, A, B) \
157  _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE)
158 
159 #define _mm_cmpeq_epu8_mask(A, B) \
160  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ)
161 #define _mm_mask_cmpeq_epu8_mask(k, A, B) \
162  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ)
163 #define _mm_cmpge_epu8_mask(A, B) \
164  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_GE)
165 #define _mm_mask_cmpge_epu8_mask(k, A, B) \
166  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE)
167 #define _mm_cmpgt_epu8_mask(A, B) \
168  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_GT)
169 #define _mm_mask_cmpgt_epu8_mask(k, A, B) \
170  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT)
171 #define _mm_cmple_epu8_mask(A, B) \
172  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_LE)
173 #define _mm_mask_cmple_epu8_mask(k, A, B) \
174  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE)
175 #define _mm_cmplt_epu8_mask(A, B) \
176  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_LT)
177 #define _mm_mask_cmplt_epu8_mask(k, A, B) \
178  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT)
179 #define _mm_cmpneq_epu8_mask(A, B) \
180  _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_NE)
181 #define _mm_mask_cmpneq_epu8_mask(k, A, B) \
182  _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE)
183 
184 #define _mm256_cmpeq_epu8_mask(A, B) \
185  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ)
186 #define _mm256_mask_cmpeq_epu8_mask(k, A, B) \
187  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ)
188 #define _mm256_cmpge_epu8_mask(A, B) \
189  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_GE)
190 #define _mm256_mask_cmpge_epu8_mask(k, A, B) \
191  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE)
192 #define _mm256_cmpgt_epu8_mask(A, B) \
193  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_GT)
194 #define _mm256_mask_cmpgt_epu8_mask(k, A, B) \
195  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT)
196 #define _mm256_cmple_epu8_mask(A, B) \
197  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_LE)
198 #define _mm256_mask_cmple_epu8_mask(k, A, B) \
199  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE)
200 #define _mm256_cmplt_epu8_mask(A, B) \
201  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_LT)
202 #define _mm256_mask_cmplt_epu8_mask(k, A, B) \
203  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT)
204 #define _mm256_cmpneq_epu8_mask(A, B) \
205  _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_NE)
206 #define _mm256_mask_cmpneq_epu8_mask(k, A, B) \
207  _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE)
208 
209 #define _mm_cmpeq_epi16_mask(A, B) \
210  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ)
211 #define _mm_mask_cmpeq_epi16_mask(k, A, B) \
212  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ)
213 #define _mm_cmpge_epi16_mask(A, B) \
214  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_GE)
215 #define _mm_mask_cmpge_epi16_mask(k, A, B) \
216  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE)
217 #define _mm_cmpgt_epi16_mask(A, B) \
218  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_GT)
219 #define _mm_mask_cmpgt_epi16_mask(k, A, B) \
220  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT)
221 #define _mm_cmple_epi16_mask(A, B) \
222  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_LE)
223 #define _mm_mask_cmple_epi16_mask(k, A, B) \
224  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE)
225 #define _mm_cmplt_epi16_mask(A, B) \
226  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_LT)
227 #define _mm_mask_cmplt_epi16_mask(k, A, B) \
228  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT)
229 #define _mm_cmpneq_epi16_mask(A, B) \
230  _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_NE)
231 #define _mm_mask_cmpneq_epi16_mask(k, A, B) \
232  _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE)
233 
234 #define _mm256_cmpeq_epi16_mask(A, B) \
235  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ)
236 #define _mm256_mask_cmpeq_epi16_mask(k, A, B) \
237  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ)
238 #define _mm256_cmpge_epi16_mask(A, B) \
239  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_GE)
240 #define _mm256_mask_cmpge_epi16_mask(k, A, B) \
241  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE)
242 #define _mm256_cmpgt_epi16_mask(A, B) \
243  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_GT)
244 #define _mm256_mask_cmpgt_epi16_mask(k, A, B) \
245  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT)
246 #define _mm256_cmple_epi16_mask(A, B) \
247  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_LE)
248 #define _mm256_mask_cmple_epi16_mask(k, A, B) \
249  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE)
250 #define _mm256_cmplt_epi16_mask(A, B) \
251  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_LT)
252 #define _mm256_mask_cmplt_epi16_mask(k, A, B) \
253  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT)
254 #define _mm256_cmpneq_epi16_mask(A, B) \
255  _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_NE)
256 #define _mm256_mask_cmpneq_epi16_mask(k, A, B) \
257  _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE)
258 
259 #define _mm_cmpeq_epu16_mask(A, B) \
260  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ)
261 #define _mm_mask_cmpeq_epu16_mask(k, A, B) \
262  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ)
263 #define _mm_cmpge_epu16_mask(A, B) \
264  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_GE)
265 #define _mm_mask_cmpge_epu16_mask(k, A, B) \
266  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE)
267 #define _mm_cmpgt_epu16_mask(A, B) \
268  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_GT)
269 #define _mm_mask_cmpgt_epu16_mask(k, A, B) \
270  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT)
271 #define _mm_cmple_epu16_mask(A, B) \
272  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_LE)
273 #define _mm_mask_cmple_epu16_mask(k, A, B) \
274  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE)
275 #define _mm_cmplt_epu16_mask(A, B) \
276  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_LT)
277 #define _mm_mask_cmplt_epu16_mask(k, A, B) \
278  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT)
279 #define _mm_cmpneq_epu16_mask(A, B) \
280  _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_NE)
281 #define _mm_mask_cmpneq_epu16_mask(k, A, B) \
282  _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE)
283 
284 #define _mm256_cmpeq_epu16_mask(A, B) \
285  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ)
286 #define _mm256_mask_cmpeq_epu16_mask(k, A, B) \
287  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ)
288 #define _mm256_cmpge_epu16_mask(A, B) \
289  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_GE)
290 #define _mm256_mask_cmpge_epu16_mask(k, A, B) \
291  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE)
292 #define _mm256_cmpgt_epu16_mask(A, B) \
293  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_GT)
294 #define _mm256_mask_cmpgt_epu16_mask(k, A, B) \
295  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT)
296 #define _mm256_cmple_epu16_mask(A, B) \
297  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_LE)
298 #define _mm256_mask_cmple_epu16_mask(k, A, B) \
299  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE)
300 #define _mm256_cmplt_epu16_mask(A, B) \
301  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_LT)
302 #define _mm256_mask_cmplt_epu16_mask(k, A, B) \
303  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT)
304 #define _mm256_cmpneq_epu16_mask(A, B) \
305  _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_NE)
306 #define _mm256_mask_cmpneq_epu16_mask(k, A, B) \
307  _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE)
308 
309 static __inline__ __m256i __DEFAULT_FN_ATTRS256
310 _mm256_mask_add_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B){
311  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
312  (__v32qi)_mm256_add_epi8(__A, __B),
313  (__v32qi)__W);
314 }
315 
316 static __inline__ __m256i __DEFAULT_FN_ATTRS256
317 _mm256_maskz_add_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
318  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
319  (__v32qi)_mm256_add_epi8(__A, __B),
320  (__v32qi)_mm256_setzero_si256());
321 }
322 
323 static __inline__ __m256i __DEFAULT_FN_ATTRS256
324 _mm256_mask_add_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
325  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
326  (__v16hi)_mm256_add_epi16(__A, __B),
327  (__v16hi)__W);
328 }
329 
330 static __inline__ __m256i __DEFAULT_FN_ATTRS256
331 _mm256_maskz_add_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
332  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
333  (__v16hi)_mm256_add_epi16(__A, __B),
334  (__v16hi)_mm256_setzero_si256());
335 }
336 
337 static __inline__ __m256i __DEFAULT_FN_ATTRS256
338 _mm256_mask_sub_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
339  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
340  (__v32qi)_mm256_sub_epi8(__A, __B),
341  (__v32qi)__W);
342 }
343 
344 static __inline__ __m256i __DEFAULT_FN_ATTRS256
345 _mm256_maskz_sub_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
346  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
347  (__v32qi)_mm256_sub_epi8(__A, __B),
348  (__v32qi)_mm256_setzero_si256());
349 }
350 
351 static __inline__ __m256i __DEFAULT_FN_ATTRS256
352 _mm256_mask_sub_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
353  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
354  (__v16hi)_mm256_sub_epi16(__A, __B),
355  (__v16hi)__W);
356 }
357 
358 static __inline__ __m256i __DEFAULT_FN_ATTRS256
359 _mm256_maskz_sub_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
360  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
361  (__v16hi)_mm256_sub_epi16(__A, __B),
362  (__v16hi)_mm256_setzero_si256());
363 }
364 
365 static __inline__ __m128i __DEFAULT_FN_ATTRS128
366 _mm_mask_add_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
367  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
368  (__v16qi)_mm_add_epi8(__A, __B),
369  (__v16qi)__W);
370 }
371 
372 static __inline__ __m128i __DEFAULT_FN_ATTRS128
373 _mm_maskz_add_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
374  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
375  (__v16qi)_mm_add_epi8(__A, __B),
376  (__v16qi)_mm_setzero_si128());
377 }
378 
379 static __inline__ __m128i __DEFAULT_FN_ATTRS128
380 _mm_mask_add_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
381  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
382  (__v8hi)_mm_add_epi16(__A, __B),
383  (__v8hi)__W);
384 }
385 
386 static __inline__ __m128i __DEFAULT_FN_ATTRS128
387 _mm_maskz_add_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
388  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
389  (__v8hi)_mm_add_epi16(__A, __B),
390  (__v8hi)_mm_setzero_si128());
391 }
392 
393 static __inline__ __m128i __DEFAULT_FN_ATTRS128
394 _mm_mask_sub_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
395  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
396  (__v16qi)_mm_sub_epi8(__A, __B),
397  (__v16qi)__W);
398 }
399 
400 static __inline__ __m128i __DEFAULT_FN_ATTRS128
401 _mm_maskz_sub_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
402  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
403  (__v16qi)_mm_sub_epi8(__A, __B),
404  (__v16qi)_mm_setzero_si128());
405 }
406 
407 static __inline__ __m128i __DEFAULT_FN_ATTRS128
408 _mm_mask_sub_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
409  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
410  (__v8hi)_mm_sub_epi16(__A, __B),
411  (__v8hi)__W);
412 }
413 
414 static __inline__ __m128i __DEFAULT_FN_ATTRS128
415 _mm_maskz_sub_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
416  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
417  (__v8hi)_mm_sub_epi16(__A, __B),
418  (__v8hi)_mm_setzero_si128());
419 }
420 
421 static __inline__ __m256i __DEFAULT_FN_ATTRS256
422 _mm256_mask_mullo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
423  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
424  (__v16hi)_mm256_mullo_epi16(__A, __B),
425  (__v16hi)__W);
426 }
427 
428 static __inline__ __m256i __DEFAULT_FN_ATTRS256
429 _mm256_maskz_mullo_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
430  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
431  (__v16hi)_mm256_mullo_epi16(__A, __B),
432  (__v16hi)_mm256_setzero_si256());
433 }
434 
435 static __inline__ __m128i __DEFAULT_FN_ATTRS128
436 _mm_mask_mullo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
437  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
438  (__v8hi)_mm_mullo_epi16(__A, __B),
439  (__v8hi)__W);
440 }
441 
442 static __inline__ __m128i __DEFAULT_FN_ATTRS128
443 _mm_maskz_mullo_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
444  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
445  (__v8hi)_mm_mullo_epi16(__A, __B),
446  (__v8hi)_mm_setzero_si128());
447 }
448 
449 static __inline__ __m128i __DEFAULT_FN_ATTRS128
450 _mm_mask_blend_epi8 (__mmask16 __U, __m128i __A, __m128i __W)
451 {
452  return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
453  (__v16qi) __W,
454  (__v16qi) __A);
455 }
456 
457 static __inline__ __m256i __DEFAULT_FN_ATTRS256
458 _mm256_mask_blend_epi8 (__mmask32 __U, __m256i __A, __m256i __W)
459 {
460  return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
461  (__v32qi) __W,
462  (__v32qi) __A);
463 }
464 
465 static __inline__ __m128i __DEFAULT_FN_ATTRS128
466 _mm_mask_blend_epi16 (__mmask8 __U, __m128i __A, __m128i __W)
467 {
468  return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
469  (__v8hi) __W,
470  (__v8hi) __A);
471 }
472 
473 static __inline__ __m256i __DEFAULT_FN_ATTRS256
474 _mm256_mask_blend_epi16 (__mmask16 __U, __m256i __A, __m256i __W)
475 {
476  return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
477  (__v16hi) __W,
478  (__v16hi) __A);
479 }
480 
481 static __inline__ __m128i __DEFAULT_FN_ATTRS128
482 _mm_mask_abs_epi8(__m128i __W, __mmask16 __U, __m128i __A)
483 {
484  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
485  (__v16qi)_mm_abs_epi8(__A),
486  (__v16qi)__W);
487 }
488 
489 static __inline__ __m128i __DEFAULT_FN_ATTRS128
490 _mm_maskz_abs_epi8(__mmask16 __U, __m128i __A)
491 {
492  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
493  (__v16qi)_mm_abs_epi8(__A),
494  (__v16qi)_mm_setzero_si128());
495 }
496 
497 static __inline__ __m256i __DEFAULT_FN_ATTRS256
498 _mm256_mask_abs_epi8(__m256i __W, __mmask32 __U, __m256i __A)
499 {
500  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
501  (__v32qi)_mm256_abs_epi8(__A),
502  (__v32qi)__W);
503 }
504 
505 static __inline__ __m256i __DEFAULT_FN_ATTRS256
506 _mm256_maskz_abs_epi8 (__mmask32 __U, __m256i __A)
507 {
508  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
509  (__v32qi)_mm256_abs_epi8(__A),
510  (__v32qi)_mm256_setzero_si256());
511 }
512 
513 static __inline__ __m128i __DEFAULT_FN_ATTRS128
514 _mm_mask_abs_epi16(__m128i __W, __mmask8 __U, __m128i __A)
515 {
516  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
517  (__v8hi)_mm_abs_epi16(__A),
518  (__v8hi)__W);
519 }
520 
521 static __inline__ __m128i __DEFAULT_FN_ATTRS128
522 _mm_maskz_abs_epi16(__mmask8 __U, __m128i __A)
523 {
524  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
525  (__v8hi)_mm_abs_epi16(__A),
526  (__v8hi)_mm_setzero_si128());
527 }
528 
529 static __inline__ __m256i __DEFAULT_FN_ATTRS256
530 _mm256_mask_abs_epi16(__m256i __W, __mmask16 __U, __m256i __A)
531 {
532  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
533  (__v16hi)_mm256_abs_epi16(__A),
534  (__v16hi)__W);
535 }
536 
537 static __inline__ __m256i __DEFAULT_FN_ATTRS256
538 _mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A)
539 {
540  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
541  (__v16hi)_mm256_abs_epi16(__A),
542  (__v16hi)_mm256_setzero_si256());
543 }
544 
545 static __inline__ __m128i __DEFAULT_FN_ATTRS128
546 _mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B) {
547  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
548  (__v8hi)_mm_packs_epi32(__A, __B),
549  (__v8hi)_mm_setzero_si128());
550 }
551 
552 static __inline__ __m128i __DEFAULT_FN_ATTRS128
553 _mm_mask_packs_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
554 {
555  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
556  (__v8hi)_mm_packs_epi32(__A, __B),
557  (__v8hi)__W);
558 }
559 
560 static __inline__ __m256i __DEFAULT_FN_ATTRS256
561 _mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B)
562 {
563  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
564  (__v16hi)_mm256_packs_epi32(__A, __B),
565  (__v16hi)_mm256_setzero_si256());
566 }
567 
568 static __inline__ __m256i __DEFAULT_FN_ATTRS256
569 _mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
570 {
571  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
572  (__v16hi)_mm256_packs_epi32(__A, __B),
573  (__v16hi)__W);
574 }
575 
576 static __inline__ __m128i __DEFAULT_FN_ATTRS128
577 _mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B)
578 {
579  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
580  (__v16qi)_mm_packs_epi16(__A, __B),
581  (__v16qi)_mm_setzero_si128());
582 }
583 
584 static __inline__ __m128i __DEFAULT_FN_ATTRS128
585 _mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
586 {
587  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
588  (__v16qi)_mm_packs_epi16(__A, __B),
589  (__v16qi)__W);
590 }
591 
592 static __inline__ __m256i __DEFAULT_FN_ATTRS256
593 _mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B)
594 {
595  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
596  (__v32qi)_mm256_packs_epi16(__A, __B),
597  (__v32qi)_mm256_setzero_si256());
598 }
599 
600 static __inline__ __m256i __DEFAULT_FN_ATTRS256
601 _mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
602 {
603  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
604  (__v32qi)_mm256_packs_epi16(__A, __B),
605  (__v32qi)__W);
606 }
607 
608 static __inline__ __m128i __DEFAULT_FN_ATTRS128
609 _mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B)
610 {
611  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
612  (__v8hi)_mm_packus_epi32(__A, __B),
613  (__v8hi)_mm_setzero_si128());
614 }
615 
616 static __inline__ __m128i __DEFAULT_FN_ATTRS128
617 _mm_mask_packus_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
618 {
619  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
620  (__v8hi)_mm_packus_epi32(__A, __B),
621  (__v8hi)__W);
622 }
623 
624 static __inline__ __m256i __DEFAULT_FN_ATTRS256
625 _mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B)
626 {
627  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
628  (__v16hi)_mm256_packus_epi32(__A, __B),
629  (__v16hi)_mm256_setzero_si256());
630 }
631 
632 static __inline__ __m256i __DEFAULT_FN_ATTRS256
633 _mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
634 {
635  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
636  (__v16hi)_mm256_packus_epi32(__A, __B),
637  (__v16hi)__W);
638 }
639 
640 static __inline__ __m128i __DEFAULT_FN_ATTRS128
641 _mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B)
642 {
643  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
644  (__v16qi)_mm_packus_epi16(__A, __B),
645  (__v16qi)_mm_setzero_si128());
646 }
647 
648 static __inline__ __m128i __DEFAULT_FN_ATTRS128
649 _mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
650 {
651  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
652  (__v16qi)_mm_packus_epi16(__A, __B),
653  (__v16qi)__W);
654 }
655 
656 static __inline__ __m256i __DEFAULT_FN_ATTRS256
657 _mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B)
658 {
659  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
660  (__v32qi)_mm256_packus_epi16(__A, __B),
661  (__v32qi)_mm256_setzero_si256());
662 }
663 
664 static __inline__ __m256i __DEFAULT_FN_ATTRS256
665 _mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
666 {
667  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
668  (__v32qi)_mm256_packus_epi16(__A, __B),
669  (__v32qi)__W);
670 }
671 
672 static __inline__ __m128i __DEFAULT_FN_ATTRS128
673 _mm_mask_adds_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
674 {
675  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
676  (__v16qi)_mm_adds_epi8(__A, __B),
677  (__v16qi)__W);
678 }
679 
680 static __inline__ __m128i __DEFAULT_FN_ATTRS128
681 _mm_maskz_adds_epi8(__mmask16 __U, __m128i __A, __m128i __B)
682 {
683  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
684  (__v16qi)_mm_adds_epi8(__A, __B),
685  (__v16qi)_mm_setzero_si128());
686 }
687 
688 static __inline__ __m256i __DEFAULT_FN_ATTRS256
689 _mm256_mask_adds_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
690 {
691  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
692  (__v32qi)_mm256_adds_epi8(__A, __B),
693  (__v32qi)__W);
694 }
695 
696 static __inline__ __m256i __DEFAULT_FN_ATTRS256
697 _mm256_maskz_adds_epi8(__mmask32 __U, __m256i __A, __m256i __B)
698 {
699  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
700  (__v32qi)_mm256_adds_epi8(__A, __B),
701  (__v32qi)_mm256_setzero_si256());
702 }
703 
704 static __inline__ __m128i __DEFAULT_FN_ATTRS128
705 _mm_mask_adds_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
706 {
707  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
708  (__v8hi)_mm_adds_epi16(__A, __B),
709  (__v8hi)__W);
710 }
711 
712 static __inline__ __m128i __DEFAULT_FN_ATTRS128
713 _mm_maskz_adds_epi16(__mmask8 __U, __m128i __A, __m128i __B)
714 {
715  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
716  (__v8hi)_mm_adds_epi16(__A, __B),
717  (__v8hi)_mm_setzero_si128());
718 }
719 
720 static __inline__ __m256i __DEFAULT_FN_ATTRS256
721 _mm256_mask_adds_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
722 {
723  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
724  (__v16hi)_mm256_adds_epi16(__A, __B),
725  (__v16hi)__W);
726 }
727 
728 static __inline__ __m256i __DEFAULT_FN_ATTRS256
729 _mm256_maskz_adds_epi16(__mmask16 __U, __m256i __A, __m256i __B)
730 {
731  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
732  (__v16hi)_mm256_adds_epi16(__A, __B),
733  (__v16hi)_mm256_setzero_si256());
734 }
735 
736 static __inline__ __m128i __DEFAULT_FN_ATTRS128
737 _mm_mask_adds_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
738 {
739  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
740  (__v16qi)_mm_adds_epu8(__A, __B),
741  (__v16qi)__W);
742 }
743 
744 static __inline__ __m128i __DEFAULT_FN_ATTRS128
745 _mm_maskz_adds_epu8(__mmask16 __U, __m128i __A, __m128i __B)
746 {
747  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
748  (__v16qi)_mm_adds_epu8(__A, __B),
749  (__v16qi)_mm_setzero_si128());
750 }
751 
752 static __inline__ __m256i __DEFAULT_FN_ATTRS256
753 _mm256_mask_adds_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
754 {
755  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
756  (__v32qi)_mm256_adds_epu8(__A, __B),
757  (__v32qi)__W);
758 }
759 
760 static __inline__ __m256i __DEFAULT_FN_ATTRS256
761 _mm256_maskz_adds_epu8(__mmask32 __U, __m256i __A, __m256i __B)
762 {
763  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
764  (__v32qi)_mm256_adds_epu8(__A, __B),
765  (__v32qi)_mm256_setzero_si256());
766 }
767 
768 static __inline__ __m128i __DEFAULT_FN_ATTRS128
769 _mm_mask_adds_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
770 {
771  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
772  (__v8hi)_mm_adds_epu16(__A, __B),
773  (__v8hi)__W);
774 }
775 
776 static __inline__ __m128i __DEFAULT_FN_ATTRS128
777 _mm_maskz_adds_epu16(__mmask8 __U, __m128i __A, __m128i __B)
778 {
779  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
780  (__v8hi)_mm_adds_epu16(__A, __B),
781  (__v8hi)_mm_setzero_si128());
782 }
783 
784 static __inline__ __m256i __DEFAULT_FN_ATTRS256
785 _mm256_mask_adds_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
786 {
787  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
788  (__v16hi)_mm256_adds_epu16(__A, __B),
789  (__v16hi)__W);
790 }
791 
792 static __inline__ __m256i __DEFAULT_FN_ATTRS256
793 _mm256_maskz_adds_epu16(__mmask16 __U, __m256i __A, __m256i __B)
794 {
795  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
796  (__v16hi)_mm256_adds_epu16(__A, __B),
797  (__v16hi)_mm256_setzero_si256());
798 }
799 
800 static __inline__ __m128i __DEFAULT_FN_ATTRS128
801 _mm_mask_avg_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
802 {
803  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
804  (__v16qi)_mm_avg_epu8(__A, __B),
805  (__v16qi)__W);
806 }
807 
808 static __inline__ __m128i __DEFAULT_FN_ATTRS128
809 _mm_maskz_avg_epu8(__mmask16 __U, __m128i __A, __m128i __B)
810 {
811  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
812  (__v16qi)_mm_avg_epu8(__A, __B),
813  (__v16qi)_mm_setzero_si128());
814 }
815 
816 static __inline__ __m256i __DEFAULT_FN_ATTRS256
817 _mm256_mask_avg_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
818 {
819  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
820  (__v32qi)_mm256_avg_epu8(__A, __B),
821  (__v32qi)__W);
822 }
823 
824 static __inline__ __m256i __DEFAULT_FN_ATTRS256
825 _mm256_maskz_avg_epu8(__mmask32 __U, __m256i __A, __m256i __B)
826 {
827  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
828  (__v32qi)_mm256_avg_epu8(__A, __B),
829  (__v32qi)_mm256_setzero_si256());
830 }
831 
832 static __inline__ __m128i __DEFAULT_FN_ATTRS128
833 _mm_mask_avg_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
834 {
835  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
836  (__v8hi)_mm_avg_epu16(__A, __B),
837  (__v8hi)__W);
838 }
839 
840 static __inline__ __m128i __DEFAULT_FN_ATTRS128
841 _mm_maskz_avg_epu16(__mmask8 __U, __m128i __A, __m128i __B)
842 {
843  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
844  (__v8hi)_mm_avg_epu16(__A, __B),
845  (__v8hi)_mm_setzero_si128());
846 }
847 
848 static __inline__ __m256i __DEFAULT_FN_ATTRS256
849 _mm256_mask_avg_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
850 {
851  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
852  (__v16hi)_mm256_avg_epu16(__A, __B),
853  (__v16hi)__W);
854 }
855 
856 static __inline__ __m256i __DEFAULT_FN_ATTRS256
857 _mm256_maskz_avg_epu16(__mmask16 __U, __m256i __A, __m256i __B)
858 {
859  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
860  (__v16hi)_mm256_avg_epu16(__A, __B),
861  (__v16hi)_mm256_setzero_si256());
862 }
863 
864 static __inline__ __m128i __DEFAULT_FN_ATTRS128
865 _mm_maskz_max_epi8(__mmask16 __M, __m128i __A, __m128i __B)
866 {
867  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
868  (__v16qi)_mm_max_epi8(__A, __B),
869  (__v16qi)_mm_setzero_si128());
870 }
871 
872 static __inline__ __m128i __DEFAULT_FN_ATTRS128
873 _mm_mask_max_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
874 {
875  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
876  (__v16qi)_mm_max_epi8(__A, __B),
877  (__v16qi)__W);
878 }
879 
880 static __inline__ __m256i __DEFAULT_FN_ATTRS256
881 _mm256_maskz_max_epi8(__mmask32 __M, __m256i __A, __m256i __B)
882 {
883  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
884  (__v32qi)_mm256_max_epi8(__A, __B),
885  (__v32qi)_mm256_setzero_si256());
886 }
887 
888 static __inline__ __m256i __DEFAULT_FN_ATTRS256
889 _mm256_mask_max_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
890 {
891  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
892  (__v32qi)_mm256_max_epi8(__A, __B),
893  (__v32qi)__W);
894 }
895 
896 static __inline__ __m128i __DEFAULT_FN_ATTRS128
897 _mm_maskz_max_epi16(__mmask8 __M, __m128i __A, __m128i __B)
898 {
899  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
900  (__v8hi)_mm_max_epi16(__A, __B),
901  (__v8hi)_mm_setzero_si128());
902 }
903 
904 static __inline__ __m128i __DEFAULT_FN_ATTRS128
905 _mm_mask_max_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
906 {
907  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
908  (__v8hi)_mm_max_epi16(__A, __B),
909  (__v8hi)__W);
910 }
911 
912 static __inline__ __m256i __DEFAULT_FN_ATTRS256
913 _mm256_maskz_max_epi16(__mmask16 __M, __m256i __A, __m256i __B)
914 {
915  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
916  (__v16hi)_mm256_max_epi16(__A, __B),
917  (__v16hi)_mm256_setzero_si256());
918 }
919 
920 static __inline__ __m256i __DEFAULT_FN_ATTRS256
921 _mm256_mask_max_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
922 {
923  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
924  (__v16hi)_mm256_max_epi16(__A, __B),
925  (__v16hi)__W);
926 }
927 
928 static __inline__ __m128i __DEFAULT_FN_ATTRS128
929 _mm_maskz_max_epu8(__mmask16 __M, __m128i __A, __m128i __B)
930 {
931  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
932  (__v16qi)_mm_max_epu8(__A, __B),
933  (__v16qi)_mm_setzero_si128());
934 }
935 
936 static __inline__ __m128i __DEFAULT_FN_ATTRS128
937 _mm_mask_max_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
938 {
939  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
940  (__v16qi)_mm_max_epu8(__A, __B),
941  (__v16qi)__W);
942 }
943 
944 static __inline__ __m256i __DEFAULT_FN_ATTRS256
945 _mm256_maskz_max_epu8 (__mmask32 __M, __m256i __A, __m256i __B)
946 {
947  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
948  (__v32qi)_mm256_max_epu8(__A, __B),
949  (__v32qi)_mm256_setzero_si256());
950 }
951 
952 static __inline__ __m256i __DEFAULT_FN_ATTRS256
953 _mm256_mask_max_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
954 {
955  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
956  (__v32qi)_mm256_max_epu8(__A, __B),
957  (__v32qi)__W);
958 }
959 
960 static __inline__ __m128i __DEFAULT_FN_ATTRS128
961 _mm_maskz_max_epu16(__mmask8 __M, __m128i __A, __m128i __B)
962 {
963  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
964  (__v8hi)_mm_max_epu16(__A, __B),
965  (__v8hi)_mm_setzero_si128());
966 }
967 
968 static __inline__ __m128i __DEFAULT_FN_ATTRS128
969 _mm_mask_max_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
970 {
971  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
972  (__v8hi)_mm_max_epu16(__A, __B),
973  (__v8hi)__W);
974 }
975 
976 static __inline__ __m256i __DEFAULT_FN_ATTRS256
977 _mm256_maskz_max_epu16(__mmask16 __M, __m256i __A, __m256i __B)
978 {
979  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
980  (__v16hi)_mm256_max_epu16(__A, __B),
981  (__v16hi)_mm256_setzero_si256());
982 }
983 
984 static __inline__ __m256i __DEFAULT_FN_ATTRS256
985 _mm256_mask_max_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
986 {
987  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
988  (__v16hi)_mm256_max_epu16(__A, __B),
989  (__v16hi)__W);
990 }
991 
992 static __inline__ __m128i __DEFAULT_FN_ATTRS128
993 _mm_maskz_min_epi8(__mmask16 __M, __m128i __A, __m128i __B)
994 {
995  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
996  (__v16qi)_mm_min_epi8(__A, __B),
997  (__v16qi)_mm_setzero_si128());
998 }
999 
1000 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1001 _mm_mask_min_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
1002 {
1003  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
1004  (__v16qi)_mm_min_epi8(__A, __B),
1005  (__v16qi)__W);
1006 }
1007 
1008 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1009 _mm256_maskz_min_epi8(__mmask32 __M, __m256i __A, __m256i __B)
1010 {
1011  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
1012  (__v32qi)_mm256_min_epi8(__A, __B),
1013  (__v32qi)_mm256_setzero_si256());
1014 }
1015 
1016 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1017 _mm256_mask_min_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
1018 {
1019  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
1020  (__v32qi)_mm256_min_epi8(__A, __B),
1021  (__v32qi)__W);
1022 }
1023 
1024 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1025 _mm_maskz_min_epi16(__mmask8 __M, __m128i __A, __m128i __B)
1026 {
1027  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
1028  (__v8hi)_mm_min_epi16(__A, __B),
1029  (__v8hi)_mm_setzero_si128());
1030 }
1031 
1032 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1033 _mm_mask_min_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
1034 {
1035  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
1036  (__v8hi)_mm_min_epi16(__A, __B),
1037  (__v8hi)__W);
1038 }
1039 
1040 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1041 _mm256_maskz_min_epi16(__mmask16 __M, __m256i __A, __m256i __B)
1042 {
1043  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
1044  (__v16hi)_mm256_min_epi16(__A, __B),
1045  (__v16hi)_mm256_setzero_si256());
1046 }
1047 
1048 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1049 _mm256_mask_min_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
1050 {
1051  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
1052  (__v16hi)_mm256_min_epi16(__A, __B),
1053  (__v16hi)__W);
1054 }
1055 
1056 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1057 _mm_maskz_min_epu8(__mmask16 __M, __m128i __A, __m128i __B)
1058 {
1059  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
1060  (__v16qi)_mm_min_epu8(__A, __B),
1061  (__v16qi)_mm_setzero_si128());
1062 }
1063 
1064 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1065 _mm_mask_min_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
1066 {
1067  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
1068  (__v16qi)_mm_min_epu8(__A, __B),
1069  (__v16qi)__W);
1070 }
1071 
1072 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1073 _mm256_maskz_min_epu8 (__mmask32 __M, __m256i __A, __m256i __B)
1074 {
1075  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
1076  (__v32qi)_mm256_min_epu8(__A, __B),
1077  (__v32qi)_mm256_setzero_si256());
1078 }
1079 
1080 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1081 _mm256_mask_min_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
1082 {
1083  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
1084  (__v32qi)_mm256_min_epu8(__A, __B),
1085  (__v32qi)__W);
1086 }
1087 
1088 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1089 _mm_maskz_min_epu16(__mmask8 __M, __m128i __A, __m128i __B)
1090 {
1091  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
1092  (__v8hi)_mm_min_epu16(__A, __B),
1093  (__v8hi)_mm_setzero_si128());
1094 }
1095 
1096 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1097 _mm_mask_min_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
1098 {
1099  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
1100  (__v8hi)_mm_min_epu16(__A, __B),
1101  (__v8hi)__W);
1102 }
1103 
1104 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1105 _mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B)
1106 {
1107  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
1108  (__v16hi)_mm256_min_epu16(__A, __B),
1109  (__v16hi)_mm256_setzero_si256());
1110 }
1111 
1112 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1113 _mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
1114 {
1115  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
1116  (__v16hi)_mm256_min_epu16(__A, __B),
1117  (__v16hi)__W);
1118 }
1119 
1120 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1121 _mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
1122 {
1123  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1124  (__v16qi)_mm_shuffle_epi8(__A, __B),
1125  (__v16qi)__W);
1126 }
1127 
1128 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1129 _mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B)
1130 {
1131  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1132  (__v16qi)_mm_shuffle_epi8(__A, __B),
1133  (__v16qi)_mm_setzero_si128());
1134 }
1135 
1136 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1137 _mm256_mask_shuffle_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
1138 {
1139  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1140  (__v32qi)_mm256_shuffle_epi8(__A, __B),
1141  (__v32qi)__W);
1142 }
1143 
1144 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1145 _mm256_maskz_shuffle_epi8(__mmask32 __U, __m256i __A, __m256i __B)
1146 {
1147  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1148  (__v32qi)_mm256_shuffle_epi8(__A, __B),
1149  (__v32qi)_mm256_setzero_si256());
1150 }
1151 
1152 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1153 _mm_mask_subs_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
1154 {
1155  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1156  (__v16qi)_mm_subs_epi8(__A, __B),
1157  (__v16qi)__W);
1158 }
1159 
1160 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1161 _mm_maskz_subs_epi8(__mmask16 __U, __m128i __A, __m128i __B)
1162 {
1163  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1164  (__v16qi)_mm_subs_epi8(__A, __B),
1165  (__v16qi)_mm_setzero_si128());
1166 }
1167 
1168 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1169 _mm256_mask_subs_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
1170 {
1171  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1172  (__v32qi)_mm256_subs_epi8(__A, __B),
1173  (__v32qi)__W);
1174 }
1175 
1176 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1177 _mm256_maskz_subs_epi8(__mmask32 __U, __m256i __A, __m256i __B)
1178 {
1179  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1180  (__v32qi)_mm256_subs_epi8(__A, __B),
1181  (__v32qi)_mm256_setzero_si256());
1182 }
1183 
1184 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1185 _mm_mask_subs_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
1186 {
1187  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1188  (__v8hi)_mm_subs_epi16(__A, __B),
1189  (__v8hi)__W);
1190 }
1191 
1192 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1193 _mm_maskz_subs_epi16(__mmask8 __U, __m128i __A, __m128i __B)
1194 {
1195  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1196  (__v8hi)_mm_subs_epi16(__A, __B),
1197  (__v8hi)_mm_setzero_si128());
1198 }
1199 
1200 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1201 _mm256_mask_subs_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
1202 {
1203  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1204  (__v16hi)_mm256_subs_epi16(__A, __B),
1205  (__v16hi)__W);
1206 }
1207 
1208 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1209 _mm256_maskz_subs_epi16(__mmask16 __U, __m256i __A, __m256i __B)
1210 {
1211  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1212  (__v16hi)_mm256_subs_epi16(__A, __B),
1213  (__v16hi)_mm256_setzero_si256());
1214 }
1215 
1216 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1217 _mm_mask_subs_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
1218 {
1219  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1220  (__v16qi)_mm_subs_epu8(__A, __B),
1221  (__v16qi)__W);
1222 }
1223 
1224 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1225 _mm_maskz_subs_epu8(__mmask16 __U, __m128i __A, __m128i __B)
1226 {
1227  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1228  (__v16qi)_mm_subs_epu8(__A, __B),
1229  (__v16qi)_mm_setzero_si128());
1230 }
1231 
1232 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1233 _mm256_mask_subs_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
1234 {
1235  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1236  (__v32qi)_mm256_subs_epu8(__A, __B),
1237  (__v32qi)__W);
1238 }
1239 
1240 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1241 _mm256_maskz_subs_epu8(__mmask32 __U, __m256i __A, __m256i __B)
1242 {
1243  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1244  (__v32qi)_mm256_subs_epu8(__A, __B),
1245  (__v32qi)_mm256_setzero_si256());
1246 }
1247 
1248 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1249 _mm_mask_subs_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
1250 {
1251  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1252  (__v8hi)_mm_subs_epu16(__A, __B),
1253  (__v8hi)__W);
1254 }
1255 
1256 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1257 _mm_maskz_subs_epu16(__mmask8 __U, __m128i __A, __m128i __B)
1258 {
1259  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1260  (__v8hi)_mm_subs_epu16(__A, __B),
1261  (__v8hi)_mm_setzero_si128());
1262 }
1263 
1264 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1265 _mm256_mask_subs_epu16(__m256i __W, __mmask16 __U, __m256i __A,
1266  __m256i __B) {
1267  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1268  (__v16hi)_mm256_subs_epu16(__A, __B),
1269  (__v16hi)__W);
1270 }
1271 
1272 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1273 _mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B)
1274 {
1275  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1276  (__v16hi)_mm256_subs_epu16(__A, __B),
1277  (__v16hi)_mm256_setzero_si256());
1278 }
1279 
1280 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1281 _mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B)
1282 {
1283  return (__m128i)__builtin_ia32_vpermi2varhi128((__v8hi)__A, (__v8hi)__I,
1284  (__v8hi) __B);
1285 }
1286 
1287 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1288 _mm_mask_permutex2var_epi16(__m128i __A, __mmask8 __U, __m128i __I,
1289  __m128i __B)
1290 {
1291  return (__m128i)__builtin_ia32_selectw_128(__U,
1292  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
1293  (__v8hi)__A);
1294 }
1295 
1296 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1297 _mm_mask2_permutex2var_epi16(__m128i __A, __m128i __I, __mmask8 __U,
1298  __m128i __B)
1299 {
1300  return (__m128i)__builtin_ia32_selectw_128(__U,
1301  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
1302  (__v8hi)__I);
1303 }
1304 
1305 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1306 _mm_maskz_permutex2var_epi16 (__mmask8 __U, __m128i __A, __m128i __I,
1307  __m128i __B)
1308 {
1309  return (__m128i)__builtin_ia32_selectw_128(__U,
1310  (__v8hi)_mm_permutex2var_epi16(__A, __I, __B),
1311  (__v8hi)_mm_setzero_si128());
1312 }
1313 
1314 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1315 _mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B)
1316 {
1317  return (__m256i)__builtin_ia32_vpermi2varhi256((__v16hi)__A, (__v16hi)__I,
1318  (__v16hi)__B);
1319 }
1320 
1321 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1322 _mm256_mask_permutex2var_epi16(__m256i __A, __mmask16 __U, __m256i __I,
1323  __m256i __B)
1324 {
1325  return (__m256i)__builtin_ia32_selectw_256(__U,
1326  (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
1327  (__v16hi)__A);
1328 }
1329 
1330 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1331 _mm256_mask2_permutex2var_epi16(__m256i __A, __m256i __I, __mmask16 __U,
1332  __m256i __B)
1333 {
1334  return (__m256i)__builtin_ia32_selectw_256(__U,
1335  (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
1336  (__v16hi)__I);
1337 }
1338 
1339 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1340 _mm256_maskz_permutex2var_epi16 (__mmask16 __U, __m256i __A, __m256i __I,
1341  __m256i __B)
1342 {
1343  return (__m256i)__builtin_ia32_selectw_256(__U,
1344  (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B),
1345  (__v16hi)_mm256_setzero_si256());
1346 }
1347 
1348 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1349 _mm_mask_maddubs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) {
1350  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1351  (__v8hi)_mm_maddubs_epi16(__X, __Y),
1352  (__v8hi)__W);
1353 }
1354 
1355 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1356 _mm_maskz_maddubs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) {
1357  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1358  (__v8hi)_mm_maddubs_epi16(__X, __Y),
1359  (__v8hi)_mm_setzero_si128());
1360 }
1361 
1362 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1363 _mm256_mask_maddubs_epi16(__m256i __W, __mmask16 __U, __m256i __X,
1364  __m256i __Y) {
1365  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1366  (__v16hi)_mm256_maddubs_epi16(__X, __Y),
1367  (__v16hi)__W);
1368 }
1369 
1370 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1371 _mm256_maskz_maddubs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) {
1372  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1373  (__v16hi)_mm256_maddubs_epi16(__X, __Y),
1374  (__v16hi)_mm256_setzero_si256());
1375 }
1376 
1377 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1378 _mm_mask_madd_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
1379  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
1380  (__v4si)_mm_madd_epi16(__A, __B),
1381  (__v4si)__W);
1382 }
1383 
1384 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1385 _mm_maskz_madd_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
1386  return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
1387  (__v4si)_mm_madd_epi16(__A, __B),
1388  (__v4si)_mm_setzero_si128());
1389 }
1390 
1391 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1392 _mm256_mask_madd_epi16(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
1393  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
1394  (__v8si)_mm256_madd_epi16(__A, __B),
1395  (__v8si)__W);
1396 }
1397 
1398 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1399 _mm256_maskz_madd_epi16(__mmask8 __U, __m256i __A, __m256i __B) {
1400  return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
1401  (__v8si)_mm256_madd_epi16(__A, __B),
1402  (__v8si)_mm256_setzero_si256());
1403 }
1404 
1405 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1406 _mm_cvtsepi16_epi8 (__m128i __A) {
1407  return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A,
1408  (__v16qi) _mm_setzero_si128(),
1409  (__mmask8) -1);
1410 }
1411 
1412 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1413 _mm_mask_cvtsepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) {
1414  return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A,
1415  (__v16qi) __O,
1416  __M);
1417 }
1418 
1419 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1420 _mm_maskz_cvtsepi16_epi8 (__mmask8 __M, __m128i __A) {
1421  return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A,
1422  (__v16qi) _mm_setzero_si128(),
1423  __M);
1424 }
1425 
1426 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1427 _mm256_cvtsepi16_epi8 (__m256i __A) {
1428  return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A,
1429  (__v16qi) _mm_setzero_si128(),
1430  (__mmask16) -1);
1431 }
1432 
1433 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1434 _mm256_mask_cvtsepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) {
1435  return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A,
1436  (__v16qi) __O,
1437  __M);
1438 }
1439 
1440 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1441 _mm256_maskz_cvtsepi16_epi8 (__mmask16 __M, __m256i __A) {
1442  return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A,
1443  (__v16qi) _mm_setzero_si128(),
1444  __M);
1445 }
1446 
1447 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1448 _mm_cvtusepi16_epi8 (__m128i __A) {
1449  return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A,
1450  (__v16qi) _mm_setzero_si128(),
1451  (__mmask8) -1);
1452 }
1453 
1454 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1455 _mm_mask_cvtusepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) {
1456  return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A,
1457  (__v16qi) __O,
1458  __M);
1459 }
1460 
1461 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1462 _mm_maskz_cvtusepi16_epi8 (__mmask8 __M, __m128i __A) {
1463  return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A,
1464  (__v16qi) _mm_setzero_si128(),
1465  __M);
1466 }
1467 
1468 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1469 _mm256_cvtusepi16_epi8 (__m256i __A) {
1470  return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A,
1471  (__v16qi) _mm_setzero_si128(),
1472  (__mmask16) -1);
1473 }
1474 
1475 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1476 _mm256_mask_cvtusepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) {
1477  return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A,
1478  (__v16qi) __O,
1479  __M);
1480 }
1481 
1482 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1483 _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) {
1484  return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A,
1485  (__v16qi) _mm_setzero_si128(),
1486  __M);
1487 }
1488 
1489 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1490 _mm_cvtepi16_epi8 (__m128i __A) {
1491  return (__m128i)__builtin_shufflevector(
1492  __builtin_convertvector((__v8hi)__A, __v8qi),
1493  (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
1494  12, 13, 14, 15);
1495 }
1496 
1497 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1498 _mm_mask_cvtepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) {
1499  return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A,
1500  (__v16qi) __O,
1501  __M);
1502 }
1503 
1504 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1505 _mm_maskz_cvtepi16_epi8 (__mmask8 __M, __m128i __A) {
1506  return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A,
1507  (__v16qi) _mm_setzero_si128(),
1508  __M);
1509 }
1510 
1511 static __inline__ void __DEFAULT_FN_ATTRS128
1512 _mm_mask_cvtepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1513 {
1514  __builtin_ia32_pmovwb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M);
1515 }
1516 
1517 
1518 static __inline__ void __DEFAULT_FN_ATTRS128
1519 _mm_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1520 {
1521  __builtin_ia32_pmovswb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M);
1522 }
1523 
1524 static __inline__ void __DEFAULT_FN_ATTRS128
1525 _mm_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1526 {
1527  __builtin_ia32_pmovuswb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M);
1528 }
1529 
1530 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1531 _mm256_cvtepi16_epi8 (__m256i __A) {
1532  return (__m128i)__builtin_convertvector((__v16hi) __A, __v16qi);
1533 }
1534 
1535 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1536 _mm256_mask_cvtepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) {
1537  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
1538  (__v16qi)_mm256_cvtepi16_epi8(__A),
1539  (__v16qi)__O);
1540 }
1541 
1542 static __inline__ __m128i __DEFAULT_FN_ATTRS256
1543 _mm256_maskz_cvtepi16_epi8 (__mmask16 __M, __m256i __A) {
1544  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
1545  (__v16qi)_mm256_cvtepi16_epi8(__A),
1546  (__v16qi)_mm_setzero_si128());
1547 }
1548 
1549 static __inline__ void __DEFAULT_FN_ATTRS256
1550 _mm256_mask_cvtepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A)
1551 {
1552  __builtin_ia32_pmovwb256mem_mask ((__v16qi *) __P, (__v16hi) __A, __M);
1553 }
1554 
1555 static __inline__ void __DEFAULT_FN_ATTRS256
1556 _mm256_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A)
1557 {
1558  __builtin_ia32_pmovswb256mem_mask ((__v16qi *) __P, (__v16hi) __A, __M);
1559 }
1560 
1561 static __inline__ void __DEFAULT_FN_ATTRS256
1562 _mm256_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A)
1563 {
1564  __builtin_ia32_pmovuswb256mem_mask ((__v16qi*) __P, (__v16hi) __A, __M);
1565 }
1566 
1567 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1568 _mm_mask_mulhrs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) {
1569  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1570  (__v8hi)_mm_mulhrs_epi16(__X, __Y),
1571  (__v8hi)__W);
1572 }
1573 
1574 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1575 _mm_maskz_mulhrs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) {
1576  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1577  (__v8hi)_mm_mulhrs_epi16(__X, __Y),
1578  (__v8hi)_mm_setzero_si128());
1579 }
1580 
1581 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1582 _mm256_mask_mulhrs_epi16(__m256i __W, __mmask16 __U, __m256i __X, __m256i __Y) {
1583  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1584  (__v16hi)_mm256_mulhrs_epi16(__X, __Y),
1585  (__v16hi)__W);
1586 }
1587 
1588 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1589 _mm256_maskz_mulhrs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) {
1590  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1591  (__v16hi)_mm256_mulhrs_epi16(__X, __Y),
1592  (__v16hi)_mm256_setzero_si256());
1593 }
1594 
1595 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1596 _mm_mask_mulhi_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
1597  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1598  (__v8hi)_mm_mulhi_epu16(__A, __B),
1599  (__v8hi)__W);
1600 }
1601 
1602 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1603 _mm_maskz_mulhi_epu16(__mmask8 __U, __m128i __A, __m128i __B) {
1604  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1605  (__v8hi)_mm_mulhi_epu16(__A, __B),
1606  (__v8hi)_mm_setzero_si128());
1607 }
1608 
1609 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1610 _mm256_mask_mulhi_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
1611  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1612  (__v16hi)_mm256_mulhi_epu16(__A, __B),
1613  (__v16hi)__W);
1614 }
1615 
1616 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1617 _mm256_maskz_mulhi_epu16(__mmask16 __U, __m256i __A, __m256i __B) {
1618  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1619  (__v16hi)_mm256_mulhi_epu16(__A, __B),
1620  (__v16hi)_mm256_setzero_si256());
1621 }
1622 
1623 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1624 _mm_mask_mulhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
1625  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1626  (__v8hi)_mm_mulhi_epi16(__A, __B),
1627  (__v8hi)__W);
1628 }
1629 
1630 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1631 _mm_maskz_mulhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
1632  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1633  (__v8hi)_mm_mulhi_epi16(__A, __B),
1634  (__v8hi)_mm_setzero_si128());
1635 }
1636 
1637 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1638 _mm256_mask_mulhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
1639  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1640  (__v16hi)_mm256_mulhi_epi16(__A, __B),
1641  (__v16hi)__W);
1642 }
1643 
1644 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1645 _mm256_maskz_mulhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
1646  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1647  (__v16hi)_mm256_mulhi_epi16(__A, __B),
1648  (__v16hi)_mm256_setzero_si256());
1649 }
1650 
1651 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1652 _mm_mask_unpackhi_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
1653  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1654  (__v16qi)_mm_unpackhi_epi8(__A, __B),
1655  (__v16qi)__W);
1656 }
1657 
1658 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1659 _mm_maskz_unpackhi_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
1660  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1661  (__v16qi)_mm_unpackhi_epi8(__A, __B),
1662  (__v16qi)_mm_setzero_si128());
1663 }
1664 
1665 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1666 _mm256_mask_unpackhi_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
1667  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1668  (__v32qi)_mm256_unpackhi_epi8(__A, __B),
1669  (__v32qi)__W);
1670 }
1671 
1672 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1673 _mm256_maskz_unpackhi_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
1674  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1675  (__v32qi)_mm256_unpackhi_epi8(__A, __B),
1676  (__v32qi)_mm256_setzero_si256());
1677 }
1678 
1679 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1680 _mm_mask_unpackhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
1681  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1682  (__v8hi)_mm_unpackhi_epi16(__A, __B),
1683  (__v8hi)__W);
1684 }
1685 
1686 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1687 _mm_maskz_unpackhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
1688  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1689  (__v8hi)_mm_unpackhi_epi16(__A, __B),
1690  (__v8hi) _mm_setzero_si128());
1691 }
1692 
1693 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1694 _mm256_mask_unpackhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
1695  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1696  (__v16hi)_mm256_unpackhi_epi16(__A, __B),
1697  (__v16hi)__W);
1698 }
1699 
1700 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1701 _mm256_maskz_unpackhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
1702  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1703  (__v16hi)_mm256_unpackhi_epi16(__A, __B),
1704  (__v16hi)_mm256_setzero_si256());
1705 }
1706 
1707 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1708 _mm_mask_unpacklo_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) {
1709  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1710  (__v16qi)_mm_unpacklo_epi8(__A, __B),
1711  (__v16qi)__W);
1712 }
1713 
1714 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1715 _mm_maskz_unpacklo_epi8(__mmask16 __U, __m128i __A, __m128i __B) {
1716  return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U,
1717  (__v16qi)_mm_unpacklo_epi8(__A, __B),
1718  (__v16qi)_mm_setzero_si128());
1719 }
1720 
1721 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1722 _mm256_mask_unpacklo_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) {
1723  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1724  (__v32qi)_mm256_unpacklo_epi8(__A, __B),
1725  (__v32qi)__W);
1726 }
1727 
1728 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1729 _mm256_maskz_unpacklo_epi8(__mmask32 __U, __m256i __A, __m256i __B) {
1730  return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U,
1731  (__v32qi)_mm256_unpacklo_epi8(__A, __B),
1732  (__v32qi)_mm256_setzero_si256());
1733 }
1734 
1735 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1736 _mm_mask_unpacklo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
1737  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1738  (__v8hi)_mm_unpacklo_epi16(__A, __B),
1739  (__v8hi)__W);
1740 }
1741 
1742 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1743 _mm_maskz_unpacklo_epi16(__mmask8 __U, __m128i __A, __m128i __B) {
1744  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1745  (__v8hi)_mm_unpacklo_epi16(__A, __B),
1746  (__v8hi) _mm_setzero_si128());
1747 }
1748 
1749 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1750 _mm256_mask_unpacklo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) {
1751  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1752  (__v16hi)_mm256_unpacklo_epi16(__A, __B),
1753  (__v16hi)__W);
1754 }
1755 
1756 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1757 _mm256_maskz_unpacklo_epi16(__mmask16 __U, __m256i __A, __m256i __B) {
1758  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1759  (__v16hi)_mm256_unpacklo_epi16(__A, __B),
1760  (__v16hi)_mm256_setzero_si256());
1761 }
1762 
1763 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1764 _mm_mask_cvtepi8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
1765 {
1766  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1767  (__v8hi)_mm_cvtepi8_epi16(__A),
1768  (__v8hi)__W);
1769 }
1770 
1771 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1772 _mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A)
1773 {
1774  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1775  (__v8hi)_mm_cvtepi8_epi16(__A),
1776  (__v8hi)_mm_setzero_si128());
1777 }
1778 
1779 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1780 _mm256_mask_cvtepi8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
1781 {
1782  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1783  (__v16hi)_mm256_cvtepi8_epi16(__A),
1784  (__v16hi)__W);
1785 }
1786 
1787 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1788 _mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A)
1789 {
1790  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1791  (__v16hi)_mm256_cvtepi8_epi16(__A),
1792  (__v16hi)_mm256_setzero_si256());
1793 }
1794 
1795 
1796 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1797 _mm_mask_cvtepu8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
1798 {
1799  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1800  (__v8hi)_mm_cvtepu8_epi16(__A),
1801  (__v8hi)__W);
1802 }
1803 
1804 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1805 _mm_maskz_cvtepu8_epi16(__mmask8 __U, __m128i __A)
1806 {
1807  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1808  (__v8hi)_mm_cvtepu8_epi16(__A),
1809  (__v8hi)_mm_setzero_si128());
1810 }
1811 
1812 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1813 _mm256_mask_cvtepu8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
1814 {
1815  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1816  (__v16hi)_mm256_cvtepu8_epi16(__A),
1817  (__v16hi)__W);
1818 }
1819 
1820 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1821 _mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A)
1822 {
1823  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1824  (__v16hi)_mm256_cvtepu8_epi16(__A),
1825  (__v16hi)_mm256_setzero_si256());
1826 }
1827 
1828 
1829 #define _mm_mask_shufflehi_epi16(W, U, A, imm) \
1830  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
1831  (__v8hi)_mm_shufflehi_epi16((A), (imm)), \
1832  (__v8hi)(__m128i)(W)))
1833 
1834 #define _mm_maskz_shufflehi_epi16(U, A, imm) \
1835  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
1836  (__v8hi)_mm_shufflehi_epi16((A), (imm)), \
1837  (__v8hi)_mm_setzero_si128()))
1838 
1839 #define _mm256_mask_shufflehi_epi16(W, U, A, imm) \
1840  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
1841  (__v16hi)_mm256_shufflehi_epi16((A), (imm)), \
1842  (__v16hi)(__m256i)(W)))
1843 
1844 #define _mm256_maskz_shufflehi_epi16(U, A, imm) \
1845  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
1846  (__v16hi)_mm256_shufflehi_epi16((A), (imm)), \
1847  (__v16hi)_mm256_setzero_si256()))
1848 
1849 #define _mm_mask_shufflelo_epi16(W, U, A, imm) \
1850  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
1851  (__v8hi)_mm_shufflelo_epi16((A), (imm)), \
1852  (__v8hi)(__m128i)(W)))
1853 
1854 #define _mm_maskz_shufflelo_epi16(U, A, imm) \
1855  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
1856  (__v8hi)_mm_shufflelo_epi16((A), (imm)), \
1857  (__v8hi)_mm_setzero_si128()))
1858 
1859 #define _mm256_mask_shufflelo_epi16(W, U, A, imm) \
1860  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
1861  (__v16hi)_mm256_shufflelo_epi16((A), \
1862  (imm)), \
1863  (__v16hi)(__m256i)(W)))
1864 
1865 #define _mm256_maskz_shufflelo_epi16(U, A, imm) \
1866  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
1867  (__v16hi)_mm256_shufflelo_epi16((A), \
1868  (imm)), \
1869  (__v16hi)_mm256_setzero_si256()))
1870 
1871 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1872 _mm256_sllv_epi16(__m256i __A, __m256i __B)
1873 {
1874  return (__m256i)__builtin_ia32_psllv16hi((__v16hi)__A, (__v16hi)__B);
1875 }
1876 
1877 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1878 _mm256_mask_sllv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
1879 {
1880  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1881  (__v16hi)_mm256_sllv_epi16(__A, __B),
1882  (__v16hi)__W);
1883 }
1884 
1885 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1886 _mm256_maskz_sllv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
1887 {
1888  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1889  (__v16hi)_mm256_sllv_epi16(__A, __B),
1890  (__v16hi)_mm256_setzero_si256());
1891 }
1892 
1893 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1894 _mm_sllv_epi16(__m128i __A, __m128i __B)
1895 {
1896  return (__m128i)__builtin_ia32_psllv8hi((__v8hi)__A, (__v8hi)__B);
1897 }
1898 
1899 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1900 _mm_mask_sllv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
1901 {
1902  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1903  (__v8hi)_mm_sllv_epi16(__A, __B),
1904  (__v8hi)__W);
1905 }
1906 
1907 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1908 _mm_maskz_sllv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
1909 {
1910  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1911  (__v8hi)_mm_sllv_epi16(__A, __B),
1912  (__v8hi)_mm_setzero_si128());
1913 }
1914 
1915 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1916 _mm_mask_sll_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
1917 {
1918  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1919  (__v8hi)_mm_sll_epi16(__A, __B),
1920  (__v8hi)__W);
1921 }
1922 
1923 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1924 _mm_maskz_sll_epi16 (__mmask8 __U, __m128i __A, __m128i __B)
1925 {
1926  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1927  (__v8hi)_mm_sll_epi16(__A, __B),
1928  (__v8hi)_mm_setzero_si128());
1929 }
1930 
1931 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1932 _mm256_mask_sll_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
1933 {
1934  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1935  (__v16hi)_mm256_sll_epi16(__A, __B),
1936  (__v16hi)__W);
1937 }
1938 
1939 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1940 _mm256_maskz_sll_epi16(__mmask16 __U, __m256i __A, __m128i __B)
1941 {
1942  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1943  (__v16hi)_mm256_sll_epi16(__A, __B),
1944  (__v16hi)_mm256_setzero_si256());
1945 }
1946 
1947 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1948 _mm_mask_slli_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
1949 {
1950  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1951  (__v8hi)_mm_slli_epi16(__A, (int)__B),
1952  (__v8hi)__W);
1953 }
1954 
1955 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1956 _mm_maskz_slli_epi16 (__mmask8 __U, __m128i __A, unsigned int __B)
1957 {
1958  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
1959  (__v8hi)_mm_slli_epi16(__A, (int)__B),
1960  (__v8hi)_mm_setzero_si128());
1961 }
1962 
1963 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1964 _mm256_mask_slli_epi16(__m256i __W, __mmask16 __U, __m256i __A,
1965  unsigned int __B)
1966 {
1967  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1968  (__v16hi)_mm256_slli_epi16(__A, (int)__B),
1969  (__v16hi)__W);
1970 }
1971 
1972 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1973 _mm256_maskz_slli_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
1974 {
1975  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1976  (__v16hi)_mm256_slli_epi16(__A, (int)__B),
1977  (__v16hi)_mm256_setzero_si256());
1978 }
1979 
1980 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1981 _mm256_srlv_epi16(__m256i __A, __m256i __B)
1982 {
1983  return (__m256i)__builtin_ia32_psrlv16hi((__v16hi)__A, (__v16hi)__B);
1984 }
1985 
1986 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1987 _mm256_mask_srlv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
1988 {
1989  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1990  (__v16hi)_mm256_srlv_epi16(__A, __B),
1991  (__v16hi)__W);
1992 }
1993 
1994 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1995 _mm256_maskz_srlv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
1996 {
1997  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
1998  (__v16hi)_mm256_srlv_epi16(__A, __B),
1999  (__v16hi)_mm256_setzero_si256());
2000 }
2001 
2002 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2003 _mm_srlv_epi16(__m128i __A, __m128i __B)
2004 {
2005  return (__m128i)__builtin_ia32_psrlv8hi((__v8hi)__A, (__v8hi)__B);
2006 }
2007 
2008 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2009 _mm_mask_srlv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
2010 {
2011  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2012  (__v8hi)_mm_srlv_epi16(__A, __B),
2013  (__v8hi)__W);
2014 }
2015 
2016 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2017 _mm_maskz_srlv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
2018 {
2019  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2020  (__v8hi)_mm_srlv_epi16(__A, __B),
2021  (__v8hi)_mm_setzero_si128());
2022 }
2023 
2024 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2025 _mm256_srav_epi16(__m256i __A, __m256i __B)
2026 {
2027  return (__m256i)__builtin_ia32_psrav16hi((__v16hi)__A, (__v16hi)__B);
2028 }
2029 
2030 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2031 _mm256_mask_srav_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
2032 {
2033  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2034  (__v16hi)_mm256_srav_epi16(__A, __B),
2035  (__v16hi)__W);
2036 }
2037 
2038 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2039 _mm256_maskz_srav_epi16(__mmask16 __U, __m256i __A, __m256i __B)
2040 {
2041  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2042  (__v16hi)_mm256_srav_epi16(__A, __B),
2043  (__v16hi)_mm256_setzero_si256());
2044 }
2045 
2046 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2047 _mm_srav_epi16(__m128i __A, __m128i __B)
2048 {
2049  return (__m128i)__builtin_ia32_psrav8hi((__v8hi)__A, (__v8hi)__B);
2050 }
2051 
2052 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2053 _mm_mask_srav_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
2054 {
2055  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2056  (__v8hi)_mm_srav_epi16(__A, __B),
2057  (__v8hi)__W);
2058 }
2059 
2060 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2061 _mm_maskz_srav_epi16(__mmask8 __U, __m128i __A, __m128i __B)
2062 {
2063  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2064  (__v8hi)_mm_srav_epi16(__A, __B),
2065  (__v8hi)_mm_setzero_si128());
2066 }
2067 
2068 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2069 _mm_mask_sra_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
2070 {
2071  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2072  (__v8hi)_mm_sra_epi16(__A, __B),
2073  (__v8hi)__W);
2074 }
2075 
2076 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2077 _mm_maskz_sra_epi16(__mmask8 __U, __m128i __A, __m128i __B)
2078 {
2079  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2080  (__v8hi)_mm_sra_epi16(__A, __B),
2081  (__v8hi)_mm_setzero_si128());
2082 }
2083 
2084 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2085 _mm256_mask_sra_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
2086 {
2087  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2088  (__v16hi)_mm256_sra_epi16(__A, __B),
2089  (__v16hi)__W);
2090 }
2091 
2092 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2093 _mm256_maskz_sra_epi16(__mmask16 __U, __m256i __A, __m128i __B)
2094 {
2095  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2096  (__v16hi)_mm256_sra_epi16(__A, __B),
2097  (__v16hi)_mm256_setzero_si256());
2098 }
2099 
2100 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2101 _mm_mask_srai_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
2102 {
2103  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2104  (__v8hi)_mm_srai_epi16(__A, (int)__B),
2105  (__v8hi)__W);
2106 }
2107 
2108 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2109 _mm_maskz_srai_epi16(__mmask8 __U, __m128i __A, unsigned int __B)
2110 {
2111  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2112  (__v8hi)_mm_srai_epi16(__A, (int)__B),
2113  (__v8hi)_mm_setzero_si128());
2114 }
2115 
2116 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2117 _mm256_mask_srai_epi16(__m256i __W, __mmask16 __U, __m256i __A,
2118  unsigned int __B)
2119 {
2120  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2121  (__v16hi)_mm256_srai_epi16(__A, (int)__B),
2122  (__v16hi)__W);
2123 }
2124 
2125 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2126 _mm256_maskz_srai_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
2127 {
2128  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2129  (__v16hi)_mm256_srai_epi16(__A, (int)__B),
2130  (__v16hi)_mm256_setzero_si256());
2131 }
2132 
2133 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2134 _mm_mask_srl_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
2135 {
2136  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2137  (__v8hi)_mm_srl_epi16(__A, __B),
2138  (__v8hi)__W);
2139 }
2140 
2141 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2142 _mm_maskz_srl_epi16 (__mmask8 __U, __m128i __A, __m128i __B)
2143 {
2144  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2145  (__v8hi)_mm_srl_epi16(__A, __B),
2146  (__v8hi)_mm_setzero_si128());
2147 }
2148 
2149 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2150 _mm256_mask_srl_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
2151 {
2152  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2153  (__v16hi)_mm256_srl_epi16(__A, __B),
2154  (__v16hi)__W);
2155 }
2156 
2157 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2158 _mm256_maskz_srl_epi16(__mmask16 __U, __m256i __A, __m128i __B)
2159 {
2160  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2161  (__v16hi)_mm256_srl_epi16(__A, __B),
2162  (__v16hi)_mm256_setzero_si256());
2163 }
2164 
2165 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2166 _mm_mask_srli_epi16(__m128i __W, __mmask8 __U, __m128i __A, int __B)
2167 {
2168  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2169  (__v8hi)_mm_srli_epi16(__A, __B),
2170  (__v8hi)__W);
2171 }
2172 
2173 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2174 _mm_maskz_srli_epi16 (__mmask8 __U, __m128i __A, int __B)
2175 {
2176  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U,
2177  (__v8hi)_mm_srli_epi16(__A, __B),
2178  (__v8hi)_mm_setzero_si128());
2179 }
2180 
2181 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2182 _mm256_mask_srli_epi16(__m256i __W, __mmask16 __U, __m256i __A, int __B)
2183 {
2184  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2185  (__v16hi)_mm256_srli_epi16(__A, __B),
2186  (__v16hi)__W);
2187 }
2188 
2189 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2190 _mm256_maskz_srli_epi16(__mmask16 __U, __m256i __A, int __B)
2191 {
2192  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U,
2193  (__v16hi)_mm256_srli_epi16(__A, __B),
2194  (__v16hi)_mm256_setzero_si256());
2195 }
2196 
2197 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2198 _mm_mask_mov_epi16 (__m128i __W, __mmask8 __U, __m128i __A)
2199 {
2200  return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
2201  (__v8hi) __A,
2202  (__v8hi) __W);
2203 }
2204 
2205 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2206 _mm_maskz_mov_epi16 (__mmask8 __U, __m128i __A)
2207 {
2208  return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U,
2209  (__v8hi) __A,
2210  (__v8hi) _mm_setzero_si128 ());
2211 }
2212 
2213 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2214 _mm256_mask_mov_epi16 (__m256i __W, __mmask16 __U, __m256i __A)
2215 {
2216  return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
2217  (__v16hi) __A,
2218  (__v16hi) __W);
2219 }
2220 
2221 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2222 _mm256_maskz_mov_epi16 (__mmask16 __U, __m256i __A)
2223 {
2224  return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U,
2225  (__v16hi) __A,
2226  (__v16hi) _mm256_setzero_si256 ());
2227 }
2228 
2229 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2230 _mm_mask_mov_epi8 (__m128i __W, __mmask16 __U, __m128i __A)
2231 {
2232  return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
2233  (__v16qi) __A,
2234  (__v16qi) __W);
2235 }
2236 
2237 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2238 _mm_maskz_mov_epi8 (__mmask16 __U, __m128i __A)
2239 {
2240  return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U,
2241  (__v16qi) __A,
2242  (__v16qi) _mm_setzero_si128 ());
2243 }
2244 
2245 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2246 _mm256_mask_mov_epi8 (__m256i __W, __mmask32 __U, __m256i __A)
2247 {
2248  return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
2249  (__v32qi) __A,
2250  (__v32qi) __W);
2251 }
2252 
2253 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2254 _mm256_maskz_mov_epi8 (__mmask32 __U, __m256i __A)
2255 {
2256  return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U,
2257  (__v32qi) __A,
2258  (__v32qi) _mm256_setzero_si256 ());
2259 }
2260 
2261 
2262 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2263 _mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A)
2264 {
2265  return (__m128i) __builtin_ia32_selectb_128(__M,
2266  (__v16qi) _mm_set1_epi8(__A),
2267  (__v16qi) __O);
2268 }
2269 
2270 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2271 _mm_maskz_set1_epi8 (__mmask16 __M, char __A)
2272 {
2273  return (__m128i) __builtin_ia32_selectb_128(__M,
2274  (__v16qi) _mm_set1_epi8(__A),
2275  (__v16qi) _mm_setzero_si128());
2276 }
2277 
2278 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2279 _mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A)
2280 {
2281  return (__m256i) __builtin_ia32_selectb_256(__M,
2282  (__v32qi) _mm256_set1_epi8(__A),
2283  (__v32qi) __O);
2284 }
2285 
2286 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2287 _mm256_maskz_set1_epi8 (__mmask32 __M, char __A)
2288 {
2289  return (__m256i) __builtin_ia32_selectb_256(__M,
2290  (__v32qi) _mm256_set1_epi8(__A),
2291  (__v32qi) _mm256_setzero_si256());
2292 }
2293 
2294 static __inline __m128i __DEFAULT_FN_ATTRS128
2295 _mm_loadu_epi16 (void const *__P)
2296 {
2297  struct __loadu_epi16 {
2298  __m128i_u __v;
2299  } __attribute__((__packed__, __may_alias__));
2300  return ((const struct __loadu_epi16*)__P)->__v;
2301 }
2302 
2303 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2304 _mm_mask_loadu_epi16 (__m128i __W, __mmask8 __U, void const *__P)
2305 {
2306  return (__m128i) __builtin_ia32_loaddquhi128_mask ((const __v8hi *) __P,
2307  (__v8hi) __W,
2308  (__mmask8) __U);
2309 }
2310 
2311 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2312 _mm_maskz_loadu_epi16 (__mmask8 __U, void const *__P)
2313 {
2314  return (__m128i) __builtin_ia32_loaddquhi128_mask ((const __v8hi *) __P,
2315  (__v8hi)
2316  _mm_setzero_si128 (),
2317  (__mmask8) __U);
2318 }
2319 
2320 static __inline __m256i __DEFAULT_FN_ATTRS256
2321 _mm256_loadu_epi16 (void const *__P)
2322 {
2323  struct __loadu_epi16 {
2324  __m256i_u __v;
2325  } __attribute__((__packed__, __may_alias__));
2326  return ((const struct __loadu_epi16*)__P)->__v;
2327 }
2328 
2329 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2330 _mm256_mask_loadu_epi16 (__m256i __W, __mmask16 __U, void const *__P)
2331 {
2332  return (__m256i) __builtin_ia32_loaddquhi256_mask ((const __v16hi *) __P,
2333  (__v16hi) __W,
2334  (__mmask16) __U);
2335 }
2336 
2337 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2338 _mm256_maskz_loadu_epi16 (__mmask16 __U, void const *__P)
2339 {
2340  return (__m256i) __builtin_ia32_loaddquhi256_mask ((const __v16hi *) __P,
2341  (__v16hi)
2342  _mm256_setzero_si256 (),
2343  (__mmask16) __U);
2344 }
2345 
2346 static __inline __m128i __DEFAULT_FN_ATTRS128
2347 _mm_loadu_epi8 (void const *__P)
2348 {
2349  struct __loadu_epi8 {
2350  __m128i_u __v;
2351  } __attribute__((__packed__, __may_alias__));
2352  return ((const struct __loadu_epi8*)__P)->__v;
2353 }
2354 
2355 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2356 _mm_mask_loadu_epi8 (__m128i __W, __mmask16 __U, void const *__P)
2357 {
2358  return (__m128i) __builtin_ia32_loaddquqi128_mask ((const __v16qi *) __P,
2359  (__v16qi) __W,
2360  (__mmask16) __U);
2361 }
2362 
2363 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2364 _mm_maskz_loadu_epi8 (__mmask16 __U, void const *__P)
2365 {
2366  return (__m128i) __builtin_ia32_loaddquqi128_mask ((const __v16qi *) __P,
2367  (__v16qi)
2368  _mm_setzero_si128 (),
2369  (__mmask16) __U);
2370 }
2371 
2372 static __inline __m256i __DEFAULT_FN_ATTRS256
2373 _mm256_loadu_epi8 (void const *__P)
2374 {
2375  struct __loadu_epi8 {
2376  __m256i_u __v;
2377  } __attribute__((__packed__, __may_alias__));
2378  return ((const struct __loadu_epi8*)__P)->__v;
2379 }
2380 
2381 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2382 _mm256_mask_loadu_epi8 (__m256i __W, __mmask32 __U, void const *__P)
2383 {
2384  return (__m256i) __builtin_ia32_loaddquqi256_mask ((const __v32qi *) __P,
2385  (__v32qi) __W,
2386  (__mmask32) __U);
2387 }
2388 
2389 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2390 _mm256_maskz_loadu_epi8 (__mmask32 __U, void const *__P)
2391 {
2392  return (__m256i) __builtin_ia32_loaddquqi256_mask ((const __v32qi *) __P,
2393  (__v32qi)
2394  _mm256_setzero_si256 (),
2395  (__mmask32) __U);
2396 }
2397 
2398 static __inline void __DEFAULT_FN_ATTRS128
2399 _mm_storeu_epi16 (void *__P, __m128i __A)
2400 {
2401  struct __storeu_epi16 {
2402  __m128i_u __v;
2403  } __attribute__((__packed__, __may_alias__));
2404  ((struct __storeu_epi16*)__P)->__v = __A;
2405 }
2406 
2407 static __inline__ void __DEFAULT_FN_ATTRS128
2408 _mm_mask_storeu_epi16 (void *__P, __mmask8 __U, __m128i __A)
2409 {
2410  __builtin_ia32_storedquhi128_mask ((__v8hi *) __P,
2411  (__v8hi) __A,
2412  (__mmask8) __U);
2413 }
2414 
2415 static __inline void __DEFAULT_FN_ATTRS256
2416 _mm256_storeu_epi16 (void *__P, __m256i __A)
2417 {
2418  struct __storeu_epi16 {
2419  __m256i_u __v;
2420  } __attribute__((__packed__, __may_alias__));
2421  ((struct __storeu_epi16*)__P)->__v = __A;
2422 }
2423 
2424 static __inline__ void __DEFAULT_FN_ATTRS256
2425 _mm256_mask_storeu_epi16 (void *__P, __mmask16 __U, __m256i __A)
2426 {
2427  __builtin_ia32_storedquhi256_mask ((__v16hi *) __P,
2428  (__v16hi) __A,
2429  (__mmask16) __U);
2430 }
2431 
2432 static __inline void __DEFAULT_FN_ATTRS128
2433 _mm_storeu_epi8 (void *__P, __m128i __A)
2434 {
2435  struct __storeu_epi8 {
2436  __m128i_u __v;
2437  } __attribute__((__packed__, __may_alias__));
2438  ((struct __storeu_epi8*)__P)->__v = __A;
2439 }
2440 
2441 static __inline__ void __DEFAULT_FN_ATTRS128
2442 _mm_mask_storeu_epi8 (void *__P, __mmask16 __U, __m128i __A)
2443 {
2444  __builtin_ia32_storedquqi128_mask ((__v16qi *) __P,
2445  (__v16qi) __A,
2446  (__mmask16) __U);
2447 }
2448 
2449 static __inline void __DEFAULT_FN_ATTRS256
2450 _mm256_storeu_epi8 (void *__P, __m256i __A)
2451 {
2452  struct __storeu_epi8 {
2453  __m256i_u __v;
2454  } __attribute__((__packed__, __may_alias__));
2455  ((struct __storeu_epi8*)__P)->__v = __A;
2456 }
2457 
2458 static __inline__ void __DEFAULT_FN_ATTRS256
2459 _mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A)
2460 {
2461  __builtin_ia32_storedquqi256_mask ((__v32qi *) __P,
2462  (__v32qi) __A,
2463  (__mmask32) __U);
2464 }
2465 
2466 static __inline__ __mmask16 __DEFAULT_FN_ATTRS128
2467 _mm_test_epi8_mask (__m128i __A, __m128i __B)
2468 {
2469  return _mm_cmpneq_epi8_mask (_mm_and_si128(__A, __B), _mm_setzero_si128());
2470 }
2471 
2472 static __inline__ __mmask16 __DEFAULT_FN_ATTRS128
2473 _mm_mask_test_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B)
2474 {
2475  return _mm_mask_cmpneq_epi8_mask (__U, _mm_and_si128 (__A, __B),
2476  _mm_setzero_si128());
2477 }
2478 
2479 static __inline__ __mmask32 __DEFAULT_FN_ATTRS256
2480 _mm256_test_epi8_mask (__m256i __A, __m256i __B)
2481 {
2482  return _mm256_cmpneq_epi8_mask (_mm256_and_si256(__A, __B),
2483  _mm256_setzero_si256());
2484 }
2485 
2486 static __inline__ __mmask32 __DEFAULT_FN_ATTRS256
2487 _mm256_mask_test_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B)
2488 {
2489  return _mm256_mask_cmpneq_epi8_mask (__U, _mm256_and_si256(__A, __B),
2490  _mm256_setzero_si256());
2491 }
2492 
2493 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
2494 _mm_test_epi16_mask (__m128i __A, __m128i __B)
2495 {
2496  return _mm_cmpneq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
2497 }
2498 
2499 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
2500 _mm_mask_test_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B)
2501 {
2502  return _mm_mask_cmpneq_epi16_mask (__U, _mm_and_si128 (__A, __B),
2503  _mm_setzero_si128());
2504 }
2505 
2506 static __inline__ __mmask16 __DEFAULT_FN_ATTRS256
2507 _mm256_test_epi16_mask (__m256i __A, __m256i __B)
2508 {
2509  return _mm256_cmpneq_epi16_mask (_mm256_and_si256 (__A, __B),
2510  _mm256_setzero_si256 ());
2511 }
2512 
2513 static __inline__ __mmask16 __DEFAULT_FN_ATTRS256
2514 _mm256_mask_test_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B)
2515 {
2516  return _mm256_mask_cmpneq_epi16_mask (__U, _mm256_and_si256(__A, __B),
2517  _mm256_setzero_si256());
2518 }
2519 
2520 static __inline__ __mmask16 __DEFAULT_FN_ATTRS128
2521 _mm_testn_epi8_mask (__m128i __A, __m128i __B)
2522 {
2523  return _mm_cmpeq_epi8_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
2524 }
2525 
2526 static __inline__ __mmask16 __DEFAULT_FN_ATTRS128
2527 _mm_mask_testn_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B)
2528 {
2529  return _mm_mask_cmpeq_epi8_mask (__U, _mm_and_si128 (__A, __B),
2530  _mm_setzero_si128());
2531 }
2532 
2533 static __inline__ __mmask32 __DEFAULT_FN_ATTRS256
2534 _mm256_testn_epi8_mask (__m256i __A, __m256i __B)
2535 {
2536  return _mm256_cmpeq_epi8_mask (_mm256_and_si256 (__A, __B),
2537  _mm256_setzero_si256());
2538 }
2539 
2540 static __inline__ __mmask32 __DEFAULT_FN_ATTRS256
2541 _mm256_mask_testn_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B)
2542 {
2543  return _mm256_mask_cmpeq_epi8_mask (__U, _mm256_and_si256 (__A, __B),
2544  _mm256_setzero_si256());
2545 }
2546 
2547 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
2548 _mm_testn_epi16_mask (__m128i __A, __m128i __B)
2549 {
2550  return _mm_cmpeq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128());
2551 }
2552 
2553 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
2554 _mm_mask_testn_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B)
2555 {
2556  return _mm_mask_cmpeq_epi16_mask (__U, _mm_and_si128(__A, __B), _mm_setzero_si128());
2557 }
2558 
2559 static __inline__ __mmask16 __DEFAULT_FN_ATTRS256
2560 _mm256_testn_epi16_mask (__m256i __A, __m256i __B)
2561 {
2562  return _mm256_cmpeq_epi16_mask (_mm256_and_si256(__A, __B),
2563  _mm256_setzero_si256());
2564 }
2565 
2566 static __inline__ __mmask16 __DEFAULT_FN_ATTRS256
2567 _mm256_mask_testn_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B)
2568 {
2569  return _mm256_mask_cmpeq_epi16_mask (__U, _mm256_and_si256 (__A, __B),
2570  _mm256_setzero_si256());
2571 }
2572 
2573 static __inline__ __mmask16 __DEFAULT_FN_ATTRS128
2574 _mm_movepi8_mask (__m128i __A)
2575 {
2576  return (__mmask16) __builtin_ia32_cvtb2mask128 ((__v16qi) __A);
2577 }
2578 
2579 static __inline__ __mmask32 __DEFAULT_FN_ATTRS256
2580 _mm256_movepi8_mask (__m256i __A)
2581 {
2582  return (__mmask32) __builtin_ia32_cvtb2mask256 ((__v32qi) __A);
2583 }
2584 
2585 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
2586 _mm_movepi16_mask (__m128i __A)
2587 {
2588  return (__mmask8) __builtin_ia32_cvtw2mask128 ((__v8hi) __A);
2589 }
2590 
2591 static __inline__ __mmask16 __DEFAULT_FN_ATTRS256
2592 _mm256_movepi16_mask (__m256i __A)
2593 {
2594  return (__mmask16) __builtin_ia32_cvtw2mask256 ((__v16hi) __A);
2595 }
2596 
2597 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2598 _mm_movm_epi8 (__mmask16 __A)
2599 {
2600  return (__m128i) __builtin_ia32_cvtmask2b128 (__A);
2601 }
2602 
2603 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2604 _mm256_movm_epi8 (__mmask32 __A)
2605 {
2606  return (__m256i) __builtin_ia32_cvtmask2b256 (__A);
2607 }
2608 
2609 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2610 _mm_movm_epi16 (__mmask8 __A)
2611 {
2612  return (__m128i) __builtin_ia32_cvtmask2w128 (__A);
2613 }
2614 
2615 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2616 _mm256_movm_epi16 (__mmask16 __A)
2617 {
2618  return (__m256i) __builtin_ia32_cvtmask2w256 (__A);
2619 }
2620 
2621 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2622 _mm_mask_broadcastb_epi8 (__m128i __O, __mmask16 __M, __m128i __A)
2623 {
2624  return (__m128i)__builtin_ia32_selectb_128(__M,
2625  (__v16qi) _mm_broadcastb_epi8(__A),
2626  (__v16qi) __O);
2627 }
2628 
2629 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2630 _mm_maskz_broadcastb_epi8 (__mmask16 __M, __m128i __A)
2631 {
2632  return (__m128i)__builtin_ia32_selectb_128(__M,
2633  (__v16qi) _mm_broadcastb_epi8(__A),
2634  (__v16qi) _mm_setzero_si128());
2635 }
2636 
2637 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2638 _mm256_mask_broadcastb_epi8 (__m256i __O, __mmask32 __M, __m128i __A)
2639 {
2640  return (__m256i)__builtin_ia32_selectb_256(__M,
2641  (__v32qi) _mm256_broadcastb_epi8(__A),
2642  (__v32qi) __O);
2643 }
2644 
2645 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2646 _mm256_maskz_broadcastb_epi8 (__mmask32 __M, __m128i __A)
2647 {
2648  return (__m256i)__builtin_ia32_selectb_256(__M,
2649  (__v32qi) _mm256_broadcastb_epi8(__A),
2650  (__v32qi) _mm256_setzero_si256());
2651 }
2652 
2653 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2654 _mm_mask_broadcastw_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
2655 {
2656  return (__m128i)__builtin_ia32_selectw_128(__M,
2657  (__v8hi) _mm_broadcastw_epi16(__A),
2658  (__v8hi) __O);
2659 }
2660 
2661 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2662 _mm_maskz_broadcastw_epi16 (__mmask8 __M, __m128i __A)
2663 {
2664  return (__m128i)__builtin_ia32_selectw_128(__M,
2665  (__v8hi) _mm_broadcastw_epi16(__A),
2666  (__v8hi) _mm_setzero_si128());
2667 }
2668 
2669 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2670 _mm256_mask_broadcastw_epi16 (__m256i __O, __mmask16 __M, __m128i __A)
2671 {
2672  return (__m256i)__builtin_ia32_selectw_256(__M,
2673  (__v16hi) _mm256_broadcastw_epi16(__A),
2674  (__v16hi) __O);
2675 }
2676 
2677 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2678 _mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A)
2679 {
2680  return (__m256i)__builtin_ia32_selectw_256(__M,
2681  (__v16hi) _mm256_broadcastw_epi16(__A),
2682  (__v16hi) _mm256_setzero_si256());
2683 }
2684 
2685 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2686 _mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A)
2687 {
2688  return (__m256i) __builtin_ia32_selectw_256 (__M,
2689  (__v16hi) _mm256_set1_epi16(__A),
2690  (__v16hi) __O);
2691 }
2692 
2693 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2694 _mm256_maskz_set1_epi16 (__mmask16 __M, short __A)
2695 {
2696  return (__m256i) __builtin_ia32_selectw_256(__M,
2697  (__v16hi)_mm256_set1_epi16(__A),
2698  (__v16hi) _mm256_setzero_si256());
2699 }
2700 
2701 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2702 _mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A)
2703 {
2704  return (__m128i) __builtin_ia32_selectw_128(__M,
2705  (__v8hi) _mm_set1_epi16(__A),
2706  (__v8hi) __O);
2707 }
2708 
2709 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2710 _mm_maskz_set1_epi16 (__mmask8 __M, short __A)
2711 {
2712  return (__m128i) __builtin_ia32_selectw_128(__M,
2713  (__v8hi) _mm_set1_epi16(__A),
2714  (__v8hi) _mm_setzero_si128());
2715 }
2716 
2717 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2718 _mm_permutexvar_epi16 (__m128i __A, __m128i __B)
2719 {
2720  return (__m128i)__builtin_ia32_permvarhi128((__v8hi) __B, (__v8hi) __A);
2721 }
2722 
2723 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2724 _mm_maskz_permutexvar_epi16 (__mmask8 __M, __m128i __A, __m128i __B)
2725 {
2726  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
2727  (__v8hi)_mm_permutexvar_epi16(__A, __B),
2728  (__v8hi) _mm_setzero_si128());
2729 }
2730 
2731 static __inline__ __m128i __DEFAULT_FN_ATTRS128
2732 _mm_mask_permutexvar_epi16 (__m128i __W, __mmask8 __M, __m128i __A,
2733  __m128i __B)
2734 {
2735  return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M,
2736  (__v8hi)_mm_permutexvar_epi16(__A, __B),
2737  (__v8hi)__W);
2738 }
2739 
2740 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2741 _mm256_permutexvar_epi16 (__m256i __A, __m256i __B)
2742 {
2743  return (__m256i)__builtin_ia32_permvarhi256((__v16hi) __B, (__v16hi) __A);
2744 }
2745 
2746 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2747 _mm256_maskz_permutexvar_epi16 (__mmask16 __M, __m256i __A,
2748  __m256i __B)
2749 {
2750  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
2751  (__v16hi)_mm256_permutexvar_epi16(__A, __B),
2752  (__v16hi)_mm256_setzero_si256());
2753 }
2754 
2755 static __inline__ __m256i __DEFAULT_FN_ATTRS256
2756 _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A,
2757  __m256i __B)
2758 {
2759  return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M,
2760  (__v16hi)_mm256_permutexvar_epi16(__A, __B),
2761  (__v16hi)__W);
2762 }
2763 
2764 #define _mm_mask_alignr_epi8(W, U, A, B, N) \
2765  ((__m128i)__builtin_ia32_selectb_128((__mmask16)(U), \
2766  (__v16qi)_mm_alignr_epi8((A), (B), (int)(N)), \
2767  (__v16qi)(__m128i)(W)))
2768 
2769 #define _mm_maskz_alignr_epi8(U, A, B, N) \
2770  ((__m128i)__builtin_ia32_selectb_128((__mmask16)(U), \
2771  (__v16qi)_mm_alignr_epi8((A), (B), (int)(N)), \
2772  (__v16qi)_mm_setzero_si128()))
2773 
2774 #define _mm256_mask_alignr_epi8(W, U, A, B, N) \
2775  ((__m256i)__builtin_ia32_selectb_256((__mmask32)(U), \
2776  (__v32qi)_mm256_alignr_epi8((A), (B), (int)(N)), \
2777  (__v32qi)(__m256i)(W)))
2778 
2779 #define _mm256_maskz_alignr_epi8(U, A, B, N) \
2780  ((__m256i)__builtin_ia32_selectb_256((__mmask32)(U), \
2781  (__v32qi)_mm256_alignr_epi8((A), (B), (int)(N)), \
2782  (__v32qi)_mm256_setzero_si256()))
2783 
2784 #define _mm_dbsad_epu8(A, B, imm) \
2785  ((__m128i)__builtin_ia32_dbpsadbw128((__v16qi)(__m128i)(A), \
2786  (__v16qi)(__m128i)(B), (int)(imm)))
2787 
2788 #define _mm_mask_dbsad_epu8(W, U, A, B, imm) \
2789  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
2790  (__v8hi)_mm_dbsad_epu8((A), (B), (imm)), \
2791  (__v8hi)(__m128i)(W)))
2792 
2793 #define _mm_maskz_dbsad_epu8(U, A, B, imm) \
2794  ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \
2795  (__v8hi)_mm_dbsad_epu8((A), (B), (imm)), \
2796  (__v8hi)_mm_setzero_si128()))
2797 
2798 #define _mm256_dbsad_epu8(A, B, imm) \
2799  ((__m256i)__builtin_ia32_dbpsadbw256((__v32qi)(__m256i)(A), \
2800  (__v32qi)(__m256i)(B), (int)(imm)))
2801 
2802 #define _mm256_mask_dbsad_epu8(W, U, A, B, imm) \
2803  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
2804  (__v16hi)_mm256_dbsad_epu8((A), (B), (imm)), \
2805  (__v16hi)(__m256i)(W)))
2806 
2807 #define _mm256_maskz_dbsad_epu8(U, A, B, imm) \
2808  ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \
2809  (__v16hi)_mm256_dbsad_epu8((A), (B), (imm)), \
2810  (__v16hi)_mm256_setzero_si256()))
2811 
2812 static __inline__ short __DEFAULT_FN_ATTRS128
2813 _mm_reduce_add_epi16(__m128i __W) {
2814  return __builtin_reduce_add((__v8hi)__W);
2815 }
2816 
2817 static __inline__ short __DEFAULT_FN_ATTRS128
2818 _mm_reduce_mul_epi16(__m128i __W) {
2819  return __builtin_reduce_mul((__v8hi)__W);
2820 }
2821 
2822 static __inline__ short __DEFAULT_FN_ATTRS128
2823 _mm_reduce_and_epi16(__m128i __W) {
2824  return __builtin_reduce_and((__v8hi)__W);
2825 }
2826 
2827 static __inline__ short __DEFAULT_FN_ATTRS128
2828 _mm_reduce_or_epi16(__m128i __W) {
2829  return __builtin_reduce_or((__v8hi)__W);
2830 }
2831 
2832 static __inline__ short __DEFAULT_FN_ATTRS128
2833 _mm_mask_reduce_add_epi16( __mmask8 __M, __m128i __W) {
2834  __W = _mm_maskz_mov_epi16(__M, __W);
2835  return __builtin_reduce_add((__v8hi)__W);
2836 }
2837 
2838 static __inline__ short __DEFAULT_FN_ATTRS128
2839 _mm_mask_reduce_mul_epi16( __mmask8 __M, __m128i __W) {
2840  __W = _mm_mask_mov_epi16(_mm_set1_epi16(1), __M, __W);
2841  return __builtin_reduce_mul((__v8hi)__W);
2842 }
2843 
2844 static __inline__ short __DEFAULT_FN_ATTRS128
2845 _mm_mask_reduce_and_epi16( __mmask8 __M, __m128i __W) {
2846  __W = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __W);
2847  return __builtin_reduce_and((__v8hi)__W);
2848 }
2849 
2850 static __inline__ short __DEFAULT_FN_ATTRS128
2851 _mm_mask_reduce_or_epi16(__mmask8 __M, __m128i __W) {
2852  __W = _mm_maskz_mov_epi16(__M, __W);
2853  return __builtin_reduce_or((__v8hi)__W);
2854 }
2855 
2856 static __inline__ short __DEFAULT_FN_ATTRS128
2857 _mm_reduce_max_epi16(__m128i __V) {
2858  return __builtin_reduce_max((__v8hi)__V);
2859 }
2860 
2861 static __inline__ unsigned short __DEFAULT_FN_ATTRS128
2862 _mm_reduce_max_epu16(__m128i __V) {
2863  return __builtin_reduce_max((__v8hu)__V);
2864 }
2865 
2866 static __inline__ short __DEFAULT_FN_ATTRS128
2867 _mm_reduce_min_epi16(__m128i __V) {
2868  return __builtin_reduce_min((__v8hi)__V);
2869 }
2870 
2871 static __inline__ unsigned short __DEFAULT_FN_ATTRS128
2872 _mm_reduce_min_epu16(__m128i __V) {
2873  return __builtin_reduce_min((__v8hu)__V);
2874 }
2875 
2876 static __inline__ short __DEFAULT_FN_ATTRS128
2877 _mm_mask_reduce_max_epi16(__mmask16 __M, __m128i __V) {
2878  __V = _mm_mask_mov_epi16(_mm_set1_epi16(-32767-1), __M, __V);
2879  return __builtin_reduce_max((__v8hi)__V);
2880 }
2881 
2882 static __inline__ unsigned short __DEFAULT_FN_ATTRS128
2883 _mm_mask_reduce_max_epu16(__mmask16 __M, __m128i __V) {
2884  __V = _mm_maskz_mov_epi16(__M, __V);
2885  return __builtin_reduce_max((__v8hu)__V);
2886 }
2887 
2888 static __inline__ short __DEFAULT_FN_ATTRS128
2889 _mm_mask_reduce_min_epi16(__mmask16 __M, __m128i __V) {
2890  __V = _mm_mask_mov_epi16(_mm_set1_epi16(32767), __M, __V);
2891  return __builtin_reduce_min((__v8hi)__V);
2892 }
2893 
2894 static __inline__ unsigned short __DEFAULT_FN_ATTRS128
2895 _mm_mask_reduce_min_epu16(__mmask16 __M, __m128i __V) {
2896  __V = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __V);
2897  return __builtin_reduce_min((__v8hu)__V);
2898 }
2899 
2900 static __inline__ short __DEFAULT_FN_ATTRS256
2901 _mm256_reduce_add_epi16(__m256i __W) {
2902  return __builtin_reduce_add((__v16hi)__W);
2903 }
2904 
2905 static __inline__ short __DEFAULT_FN_ATTRS256
2906 _mm256_reduce_mul_epi16(__m256i __W) {
2907  return __builtin_reduce_mul((__v16hi)__W);
2908 }
2909 
2910 static __inline__ short __DEFAULT_FN_ATTRS256
2911 _mm256_reduce_and_epi16(__m256i __W) {
2912  return __builtin_reduce_and((__v16hi)__W);
2913 }
2914 
2915 static __inline__ short __DEFAULT_FN_ATTRS256
2916 _mm256_reduce_or_epi16(__m256i __W) {
2917  return __builtin_reduce_or((__v16hi)__W);
2918 }
2919 
2920 static __inline__ short __DEFAULT_FN_ATTRS256
2921 _mm256_mask_reduce_add_epi16( __mmask16 __M, __m256i __W) {
2922  __W = _mm256_maskz_mov_epi16(__M, __W);
2923  return __builtin_reduce_add((__v16hi)__W);
2924 }
2925 
2926 static __inline__ short __DEFAULT_FN_ATTRS256
2927 _mm256_mask_reduce_mul_epi16( __mmask16 __M, __m256i __W) {
2928  __W = _mm256_mask_mov_epi16(_mm256_set1_epi16(1), __M, __W);
2929  return __builtin_reduce_mul((__v16hi)__W);
2930 }
2931 
2932 static __inline__ short __DEFAULT_FN_ATTRS256
2933 _mm256_mask_reduce_and_epi16( __mmask16 __M, __m256i __W) {
2934  __W = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __W);
2935  return __builtin_reduce_and((__v16hi)__W);
2936 }
2937 
2938 static __inline__ short __DEFAULT_FN_ATTRS256
2939 _mm256_mask_reduce_or_epi16(__mmask16 __M, __m256i __W) {
2940  __W = _mm256_maskz_mov_epi16(__M, __W);
2941  return __builtin_reduce_or((__v16hi)__W);
2942 }
2943 
2944 static __inline__ short __DEFAULT_FN_ATTRS256
2945 _mm256_reduce_max_epi16(__m256i __V) {
2946  return __builtin_reduce_max((__v16hi)__V);
2947 }
2948 
2949 static __inline__ unsigned short __DEFAULT_FN_ATTRS256
2950 _mm256_reduce_max_epu16(__m256i __V) {
2951  return __builtin_reduce_max((__v16hu)__V);
2952 }
2953 
2954 static __inline__ short __DEFAULT_FN_ATTRS256
2955 _mm256_reduce_min_epi16(__m256i __V) {
2956  return __builtin_reduce_min((__v16hi)__V);
2957 }
2958 
2959 static __inline__ unsigned short __DEFAULT_FN_ATTRS256
2960 _mm256_reduce_min_epu16(__m256i __V) {
2961  return __builtin_reduce_min((__v16hu)__V);
2962 }
2963 
2964 static __inline__ short __DEFAULT_FN_ATTRS256
2965 _mm256_mask_reduce_max_epi16(__mmask16 __M, __m256i __V) {
2966  __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-32767-1), __M, __V);
2967  return __builtin_reduce_max((__v16hi)__V);
2968 }
2969 
2970 static __inline__ unsigned short __DEFAULT_FN_ATTRS256
2971 _mm256_mask_reduce_max_epu16(__mmask16 __M, __m256i __V) {
2972  __V = _mm256_maskz_mov_epi16(__M, __V);
2973  return __builtin_reduce_max((__v16hu)__V);
2974 }
2975 
2976 static __inline__ short __DEFAULT_FN_ATTRS256
2977 _mm256_mask_reduce_min_epi16(__mmask16 __M, __m256i __V) {
2978  __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(32767), __M, __V);
2979  return __builtin_reduce_min((__v16hi)__V);
2980 }
2981 
2982 static __inline__ unsigned short __DEFAULT_FN_ATTRS256
2983 _mm256_mask_reduce_min_epu16(__mmask16 __M, __m256i __V) {
2984  __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __V);
2985  return __builtin_reduce_min((__v16hu)__V);
2986 }
2987 
2988 static __inline__ signed char __DEFAULT_FN_ATTRS128
2989 _mm_reduce_add_epi8(__m128i __W) {
2990  return __builtin_reduce_add((__v16qs)__W);
2991 }
2992 
2993 static __inline__ signed char __DEFAULT_FN_ATTRS128
2994 _mm_reduce_mul_epi8(__m128i __W) {
2995  return __builtin_reduce_mul((__v16qs)__W);
2996 }
2997 
2998 static __inline__ signed char __DEFAULT_FN_ATTRS128
2999 _mm_reduce_and_epi8(__m128i __W) {
3000  return __builtin_reduce_and((__v16qs)__W);
3001 }
3002 
3003 static __inline__ signed char __DEFAULT_FN_ATTRS128
3004 _mm_reduce_or_epi8(__m128i __W) {
3005  return __builtin_reduce_or((__v16qs)__W);
3006 }
3007 
3008 static __inline__ signed char __DEFAULT_FN_ATTRS128
3009 _mm_mask_reduce_add_epi8(__mmask16 __M, __m128i __W) {
3010  __W = _mm_maskz_mov_epi8(__M, __W);
3011  return __builtin_reduce_add((__v16qs)__W);
3012 }
3013 
3014 static __inline__ signed char __DEFAULT_FN_ATTRS128
3015 _mm_mask_reduce_mul_epi8(__mmask16 __M, __m128i __W) {
3016  __W = _mm_mask_mov_epi8(_mm_set1_epi8(1), __M, __W);
3017  return __builtin_reduce_mul((__v16qs)__W);
3018 }
3019 
3020 static __inline__ signed char __DEFAULT_FN_ATTRS128
3021 _mm_mask_reduce_and_epi8(__mmask16 __M, __m128i __W) {
3022  __W = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __W);
3023  return __builtin_reduce_and((__v16qs)__W);
3024 }
3025 
3026 static __inline__ signed char __DEFAULT_FN_ATTRS128
3027 _mm_mask_reduce_or_epi8(__mmask16 __M, __m128i __W) {
3028  __W = _mm_maskz_mov_epi8(__M, __W);
3029  return __builtin_reduce_or((__v16qs)__W);
3030 }
3031 
3032 static __inline__ signed char __DEFAULT_FN_ATTRS128
3033 _mm_reduce_max_epi8(__m128i __V) {
3034  return __builtin_reduce_max((__v16qs)__V);
3035 }
3036 
3037 static __inline__ unsigned char __DEFAULT_FN_ATTRS128
3038 _mm_reduce_max_epu8(__m128i __V) {
3039  return __builtin_reduce_max((__v16qu)__V);
3040 }
3041 
3042 static __inline__ signed char __DEFAULT_FN_ATTRS128
3043 _mm_reduce_min_epi8(__m128i __V) {
3044  return __builtin_reduce_min((__v16qs)__V);
3045 }
3046 
3047 static __inline__ unsigned char __DEFAULT_FN_ATTRS128
3048 _mm_reduce_min_epu8(__m128i __V) {
3049  return __builtin_reduce_min((__v16qu)__V);
3050 }
3051 
3052 static __inline__ signed char __DEFAULT_FN_ATTRS128
3053 _mm_mask_reduce_max_epi8(__mmask16 __M, __m128i __V) {
3054  __V = _mm_mask_mov_epi8(_mm_set1_epi8(-127-1), __M, __V);
3055  return __builtin_reduce_max((__v16qs)__V);
3056 }
3057 
3058 static __inline__ unsigned char __DEFAULT_FN_ATTRS128
3059 _mm_mask_reduce_max_epu8(__mmask16 __M, __m128i __V) {
3060  __V = _mm_maskz_mov_epi8(__M, __V);
3061  return __builtin_reduce_max((__v16qu)__V);
3062 }
3063 
3064 static __inline__ signed char __DEFAULT_FN_ATTRS128
3065 _mm_mask_reduce_min_epi8(__mmask16 __M, __m128i __V) {
3066  __V = _mm_mask_mov_epi8(_mm_set1_epi8(127), __M, __V);
3067  return __builtin_reduce_min((__v16qs)__V);
3068 }
3069 
3070 static __inline__ unsigned char __DEFAULT_FN_ATTRS128
3071 _mm_mask_reduce_min_epu8(__mmask16 __M, __m128i __V) {
3072  __V = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __V);
3073  return __builtin_reduce_min((__v16qu)__V);
3074 }
3075 
3076 static __inline__ signed char __DEFAULT_FN_ATTRS256
3077 _mm256_reduce_add_epi8(__m256i __W) {
3078  return __builtin_reduce_add((__v32qs)__W);
3079 }
3080 
3081 static __inline__ signed char __DEFAULT_FN_ATTRS256
3082 _mm256_reduce_mul_epi8(__m256i __W) {
3083  return __builtin_reduce_mul((__v32qs)__W);
3084 }
3085 
3086 static __inline__ signed char __DEFAULT_FN_ATTRS256
3087 _mm256_reduce_and_epi8(__m256i __W) {
3088  return __builtin_reduce_and((__v32qs)__W);
3089 }
3090 
3091 static __inline__ signed char __DEFAULT_FN_ATTRS256
3092 _mm256_reduce_or_epi8(__m256i __W) {
3093  return __builtin_reduce_or((__v32qs)__W);
3094 }
3095 
3096 static __inline__ signed char __DEFAULT_FN_ATTRS256
3097 _mm256_mask_reduce_add_epi8(__mmask32 __M, __m256i __W) {
3098  __W = _mm256_maskz_mov_epi8(__M, __W);
3099  return __builtin_reduce_add((__v32qs)__W);
3100 }
3101 
3102 static __inline__ signed char __DEFAULT_FN_ATTRS256
3103 _mm256_mask_reduce_mul_epi8(__mmask32 __M, __m256i __W) {
3104  __W = _mm256_mask_mov_epi8(_mm256_set1_epi8(1), __M, __W);
3105  return __builtin_reduce_mul((__v32qs)__W);
3106 }
3107 
3108 static __inline__ signed char __DEFAULT_FN_ATTRS256
3109 _mm256_mask_reduce_and_epi8(__mmask32 __M, __m256i __W) {
3110  __W = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __W);
3111  return __builtin_reduce_and((__v32qs)__W);
3112 }
3113 
3114 static __inline__ signed char __DEFAULT_FN_ATTRS256
3115 _mm256_mask_reduce_or_epi8(__mmask32 __M, __m256i __W) {
3116  __W = _mm256_maskz_mov_epi8(__M, __W);
3117  return __builtin_reduce_or((__v32qs)__W);
3118 }
3119 
3120 static __inline__ signed char __DEFAULT_FN_ATTRS256
3121 _mm256_reduce_max_epi8(__m256i __V) {
3122  return __builtin_reduce_max((__v32qs)__V);
3123 }
3124 
3125 static __inline__ unsigned char __DEFAULT_FN_ATTRS256
3126 _mm256_reduce_max_epu8(__m256i __V) {
3127  return __builtin_reduce_max((__v32qu)__V);
3128 }
3129 
3130 static __inline__ signed char __DEFAULT_FN_ATTRS256
3131 _mm256_reduce_min_epi8(__m256i __V) {
3132  return __builtin_reduce_min((__v32qs)__V);
3133 }
3134 
3135 static __inline__ unsigned char __DEFAULT_FN_ATTRS256
3136 _mm256_reduce_min_epu8(__m256i __V) {
3137  return __builtin_reduce_min((__v32qu)__V);
3138 }
3139 
3140 static __inline__ signed char __DEFAULT_FN_ATTRS256
3141 _mm256_mask_reduce_max_epi8(__mmask32 __M, __m256i __V) {
3142  __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-127-1), __M, __V);
3143  return __builtin_reduce_max((__v32qs)__V);
3144 }
3145 
3146 static __inline__ unsigned char __DEFAULT_FN_ATTRS256
3147 _mm256_mask_reduce_max_epu8(__mmask32 __M, __m256i __V) {
3148  __V = _mm256_maskz_mov_epi8(__M, __V);
3149  return __builtin_reduce_max((__v32qu)__V);
3150 }
3151 
3152 static __inline__ signed char __DEFAULT_FN_ATTRS256
3153 _mm256_mask_reduce_min_epi8(__mmask32 __M, __m256i __V) {
3154  __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(127), __M, __V);
3155  return __builtin_reduce_min((__v32qs)__V);
3156 }
3157 
3158 static __inline__ unsigned char __DEFAULT_FN_ATTRS256
3159 _mm256_mask_reduce_min_epu8(__mmask32 __M, __m256i __V) {
3160  __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __V);
3161  return __builtin_reduce_min((__v32qu)__V);
3162 }
3163 
3164 #undef __DEFAULT_FN_ATTRS128
3165 #undef __DEFAULT_FN_ATTRS256
3166 
3167 #endif /* __AVX512VLBWINTRIN_H */
__attribute__
_Float16 __2f16 __attribute__((ext_vector_type(2)))
Zeroes the upper 128 bits (bits 255:128) of all YMM registers.
Definition: __clang_hip_libdevice_declares.h:293
_mm256_abs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_abs_epi8(__m256i __a)
Computes the absolute value of each signed byte in the 256-bit integer vector __a and returns each va...
Definition: avx2intrin.h:99
_mm256_srl_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_srl_epi16(__m256i __a, __m128i __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a right by the number of bits giv...
Definition: avx2intrin.h:2394
_mm256_and_si256
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_and_si256(__m256i __a, __m256i __b)
Computes the bitwise AND of the 256-bit integer vectors in __a and __b.
Definition: avx2intrin.h:455
_mm256_max_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_max_epu16(__m256i __a, __m256i __b)
Compares the corresponding unsigned 16-bit integers in the two 256-bit vectors of [16 x i16] in __a a...
Definition: avx2intrin.h:1176
_mm256_min_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_min_epu16(__m256i __a, __m256i __b)
Compares the corresponding unsigned 16-bit integers in the two 256-bit vectors of [16 x i16] in __a a...
Definition: avx2intrin.h:1290
_mm256_packs_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packs_epi32(__m256i __a, __m256i __b)
Converts the elements of two 256-bit vectors of [8 x i32] to 16-bit integers using signed saturation,...
Definition: avx2intrin.h:196
_mm256_cvtepi8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_cvtepi8_epi16(__m128i __V)
Sign-extends bytes from the 128-bit integer vector in __V and returns the 16-bit values in the corres...
Definition: avx2intrin.h:1358
_mm256_shuffle_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_shuffle_epi8(__m256i __a, __m256i __b)
Shuffles 8-bit integers in the 256-bit integer vector __a according to control information in the 256...
Definition: avx2intrin.h:1901
_mm_broadcastb_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastb_epi8(__m128i __X)
Broadcasts the low byte from the 128-bit integer vector in __X to all bytes of the 128-bit result.
Definition: avx2intrin.h:3227
_mm256_cvtepu8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_cvtepu8_epi16(__m128i __V)
Zero-extends bytes from the 128-bit integer vector in __V and returns the 16-bit values in the corres...
Definition: avx2intrin.h:1517
_mm256_subs_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_subs_epu8(__m256i __a, __m256i __b)
Subtracts 8-bit integers from corresponding bytes of two 256-bit integer vectors using unsigned satur...
Definition: avx2intrin.h:2659
_mm256_min_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_min_epi8(__m256i __a, __m256i __b)
Compares the corresponding signed bytes in the two 256-bit integer vectors in __a and __b and returns...
Definition: avx2intrin.h:1214
_mm256_sub_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_sub_epi16(__m256i __a, __m256i __b)
Subtracts 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16].
Definition: avx2intrin.h:2528
_mm256_max_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_max_epi8(__m256i __a, __m256i __b)
Compares the corresponding signed bytes in the two 256-bit integer vectors in __a and __b and returns...
Definition: avx2intrin.h:1100
_mm256_maddubs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maddubs_epi16(__m256i __a, __m256i __b)
Multiplies each unsigned byte from the 256-bit integer vector in __a with the corresponding signed by...
Definition: avx2intrin.h:1049
_mm256_unpackhi_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Unpacks and interleaves 8-bit integers from parts of the 256-bit integer vectors in __a and __b to fo...
Definition: avx2intrin.h:2719
_mm256_adds_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_adds_epu16(__m256i __a, __m256i __b)
Adds 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16] using unsigned ...
Definition: avx2intrin.h:409
_mm256_mulhrs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mulhrs_epi16(__m256i __a, __m256i __b)
Multiplies signed 16-bit integer elements of two 256-bit vectors of [16 x i16], truncates the 32-bit ...
Definition: avx2intrin.h:1697
_mm256_sll_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_sll_epi16(__m256i __a, __m128i __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a left by the number of bits spec...
Definition: avx2intrin.h:2150
_mm256_add_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_add_epi8(__m256i __a, __m256i __b)
Adds 8-bit integers from corresponding bytes of two 256-bit integer vectors and returns the lower 8 b...
Definition: avx2intrin.h:278
_mm256_add_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_add_epi16(__m256i __a, __m256i __b)
Adds 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16] and returns the...
Definition: avx2intrin.h:297
_mm256_subs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_subs_epi8(__m256i __a, __m256i __b)
Subtracts 8-bit integers from corresponding bytes of two 256-bit integer vectors using signed saturat...
Definition: avx2intrin.h:2606
_mm256_max_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_max_epu8(__m256i __a, __m256i __b)
Compares the corresponding unsigned bytes in the two 256-bit integer vectors in __a and __b and retur...
Definition: avx2intrin.h:1157
_mm256_min_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_min_epi16(__m256i __a, __m256i __b)
Compares the corresponding signed 16-bit integers in the two 256-bit vectors of [16 x i16] in __a and...
Definition: avx2intrin.h:1233
_mm256_max_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_max_epi16(__m256i __a, __m256i __b)
Compares the corresponding signed 16-bit integers in the two 256-bit vectors of [16 x i16] in __a and...
Definition: avx2intrin.h:1119
_mm256_broadcastb_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastb_epi8(__m128i __X)
Broadcasts the low byte from the 128-bit integer vector in __X to all bytes of the 256-bit result.
Definition: avx2intrin.h:3163
_mm256_unpackhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Unpacks and interleaves 16-bit integers from parts of the 256-bit vectors of [16 x i16] in __a and __...
Definition: avx2intrin.h:2754
_mm256_mulhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mulhi_epi16(__m256i __a, __m256i __b)
Multiplies signed 16-bit integer elements of two 256-bit vectors of [16 x i16], and returns the upper...
Definition: avx2intrin.h:1735
_mm256_packs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packs_epi16(__m256i __a, __m256i __b)
Converts the elements of two 256-bit vectors of [16 x i16] to 8-bit integers using signed saturation,...
Definition: avx2intrin.h:164
_mm_broadcastw_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_broadcastw_epi16(__m128i __X)
Broadcasts the low element from the 128-bit vector of [8 x i16] in __X to all elements of the result'...
Definition: avx2intrin.h:3243
_mm256_sub_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_sub_epi8(__m256i __a, __m256i __b)
Subtracts 8-bit integers from corresponding bytes of two 256-bit integer vectors.
Definition: avx2intrin.h:2501
_mm256_mulhi_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mulhi_epu16(__m256i __a, __m256i __b)
Multiplies unsigned 16-bit integer elements of two 256-bit vectors of [16 x i16], and returns the upp...
Definition: avx2intrin.h:1716
_mm256_slli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_slli_epi16(__m256i __a, int __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a left by __count bits,...
Definition: avx2intrin.h:2129
_mm256_madd_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_madd_epi16(__m256i __a, __m256i __b)
Multiplies corresponding 16-bit elements of two 256-bit vectors of [16 x i16], forming 32-bit interme...
Definition: avx2intrin.h:1081
_mm256_avg_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_avg_epu16(__m256i __a, __m256i __b)
Computes the averages of the corresponding unsigned 16-bit integers in the two 256-bit vectors of [16...
Definition: avx2intrin.h:525
_mm256_min_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_min_epu8(__m256i __a, __m256i __b)
Compares the corresponding unsigned bytes in the two 256-bit integer vectors in __a and __b and retur...
Definition: avx2intrin.h:1271
_mm256_packus_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packus_epi32(__m256i __V1, __m256i __V2)
Converts elements from two 256-bit vectors of [8 x i32] to 16-bit integers using unsigned saturation,...
Definition: avx2intrin.h:259
_mm256_unpacklo_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Unpacks and interleaves 8-bit integers from parts of the 256-bit integer vectors in __a and __b to fo...
Definition: avx2intrin.h:2852
_mm256_subs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_subs_epi16(__m256i __a, __m256i __b)
Subtracts 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16] using sign...
Definition: avx2intrin.h:2632
_mm256_broadcastw_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_broadcastw_epi16(__m128i __X)
Broadcasts the low element from the 128-bit vector of [8 x i16] in __X to all elements of the result'...
Definition: avx2intrin.h:3179
_mm256_mullo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mullo_epi16(__m256i __a, __m256i __b)
Multiplies signed 16-bit integer elements of two 256-bit vectors of [16 x i16], and returns the lower...
Definition: avx2intrin.h:1754
_mm256_srai_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_srai_epi16(__m256i __a, int __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a right by __count bits,...
Definition: avx2intrin.h:2250
_mm256_unpacklo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Unpacks and interleaves 16-bit integers from parts of the 256-bit vectors of [16 x i16] in __a and __...
Definition: avx2intrin.h:2887
_mm256_adds_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_adds_epu8(__m256i __a, __m256i __b)
Adds 8-bit integers from corresponding bytes of two 256-bit integer vectors using unsigned saturation...
Definition: avx2intrin.h:391
_mm256_packus_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_packus_epi16(__m256i __a, __m256i __b)
Converts elements from two 256-bit vectors of [16 x i16] to 8-bit integers using unsigned saturation,...
Definition: avx2intrin.h:227
_mm256_adds_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_adds_epi8(__m256i __a, __m256i __b)
Adds 8-bit integers from corresponding bytes of two 256-bit integer vectors using signed saturation,...
Definition: avx2intrin.h:354
_mm256_sra_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_sra_epi16(__m256i __a, __m128i __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a right by the number of bits giv...
Definition: avx2intrin.h:2272
_mm256_abs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_abs_epi16(__m256i __a)
Computes the absolute value of each signed 16-bit element in the 256-bit vector of [16 x i16] in __a ...
Definition: avx2intrin.h:116
_mm256_adds_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_adds_epi16(__m256i __a, __m256i __b)
Adds 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16] using signed sa...
Definition: avx2intrin.h:372
_mm256_avg_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_avg_epu8(__m256i __a, __m256i __b)
Computes the averages of the corresponding unsigned bytes in the two 256-bit integer vectors in __a a...
Definition: avx2intrin.h:499
_mm256_subs_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_subs_epu16(__m256i __a, __m256i __b)
Subtracts 16-bit integers from corresponding elements of two 256-bit vectors of [16 x i16] using unsi...
Definition: avx2intrin.h:2685
_mm256_srli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_srli_epi16(__m256i __a, int __count)
Shifts each 16-bit element of the 256-bit vector of [16 x i16] in __a right by __count bits,...
Definition: avx2intrin.h:2373
__mmask32
unsigned int __mmask32
Definition: avx512bwintrin.h:17
__mmask8
unsigned char __mmask8
Definition: avx512fintrin.h:41
__mmask16
unsigned short __mmask16
Definition: avx512fintrin.h:42
_mm256_mask_sra_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_sra_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2085
_mm256_mask_madd_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_madd_epi16(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1392
_mm_maskz_srlv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_srlv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2017
_mm256_mask_blend_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_blend_epi8(__mmask32 __U, __m256i __A, __m256i __W)
Definition: avx512vlbwintrin.h:458
_mm256_mask_cmpeq_epi16_mask
#define _mm256_mask_cmpeq_epi16_mask(k, A, B)
Definition: avx512vlbwintrin.h:236
_mm_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtepi16_epi8(__m128i __A)
Definition: avx512vlbwintrin.h:1490
_mm_mask_slli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_slli_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:1948
_mm256_mask_max_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_max_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:953
_mm256_mask_test_epi8_mask
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 _mm256_mask_test_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2487
_mm256_maskz_srlv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_srlv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1995
_mm_storeu_epi16
static __inline void __DEFAULT_FN_ATTRS128 _mm_storeu_epi16(void *__P, __m128i __A)
Definition: avx512vlbwintrin.h:2399
_mm256_srlv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_srlv_epi16(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1981
_mm_mask_cvtepu8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_cvtepu8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1797
_mm_mask_broadcastb_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_broadcastb_epi8(__m128i __O, __mmask16 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2622
_mm256_mask_subs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_subs_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1201
_mm_maskz_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtsepi16_epi8(__mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1420
_mm_movepi8_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 _mm_movepi8_mask(__m128i __A)
Definition: avx512vlbwintrin.h:2574
_mm_mask_loadu_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_loadu_epi8(__m128i __W, __mmask16 __U, void const *__P)
Definition: avx512vlbwintrin.h:2356
_mm_mask_reduce_max_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_max_epi8(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:3053
_mm_maskz_mulhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mulhi_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1631
_mm_maskz_mullo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mullo_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:443
_mm_maskz_cvtepi8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1772
_mm_mask_avg_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_avg_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:801
_mm_loadu_epi8
static __inline __m128i __DEFAULT_FN_ATTRS128 _mm_loadu_epi8(void const *__P)
Definition: avx512vlbwintrin.h:2347
_mm_srlv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_srlv_epi16(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2003
_mm256_mask_broadcastb_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_broadcastb_epi8(__m256i __O, __mmask32 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2638
_mm256_mask_reduce_add_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_add_epi16(__mmask16 __M, __m256i __W)
Definition: avx512vlbwintrin.h:2921
_mm_maskz_abs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_abs_epi8(__mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:490
_mm256_cmpeq_epi8_mask
#define _mm256_cmpeq_epi8_mask(A, B)
Definition: avx512vlbwintrin.h:134
_mm_reduce_max_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_max_epi8(__m128i __V)
Definition: avx512vlbwintrin.h:3033
_mm256_reduce_min_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS256 _mm256_reduce_min_epu8(__m256i __V)
Definition: avx512vlbwintrin.h:3136
_mm256_maskz_slli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_slli_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:1973
_mm_movepi16_mask
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 _mm_movepi16_mask(__m128i __A)
Definition: avx512vlbwintrin.h:2586
_mm256_reduce_min_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS256 _mm256_reduce_min_epu16(__m256i __V)
Definition: avx512vlbwintrin.h:2960
_mm_mask_subs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_subs_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1153
_mm_maskz_avg_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_avg_epu16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:841
_mm_mask2_permutex2var_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask2_permutex2var_epi16(__m128i __A, __m128i __I, __mmask8 __U, __m128i __B)
Definition: avx512vlbwintrin.h:1297
_mm_mask_adds_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_adds_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:769
_mm_mask_reduce_min_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_min_epi8(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:3065
_mm_maskz_cvtepu8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtepu8_epi16(__mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1805
_mm256_mask_reduce_max_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_max_epu8(__mmask32 __M, __m256i __V)
Definition: avx512vlbwintrin.h:3147
_mm_cmpneq_epi16_mask
#define _mm_cmpneq_epi16_mask(A, B)
Definition: avx512vlbwintrin.h:229
_mm_mask_maddubs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_maddubs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
Definition: avx512vlbwintrin.h:1349
_mm256_mask_unpackhi_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_unpackhi_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1666
_mm_maskz_mulhi_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mulhi_epu16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1603
_mm256_mask_abs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_abs_epi16(__m256i __W, __mmask16 __U, __m256i __A)
Definition: avx512vlbwintrin.h:530
_mm256_mask_avg_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_avg_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:849
_mm256_reduce_and_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_and_epi8(__m256i __W)
Definition: avx512vlbwintrin.h:3087
_mm256_mask_unpackhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_unpackhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1694
_mm_maskz_add_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_add_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:373
_mm_mask_cmpeq_epi8_mask
#define _mm_mask_cmpeq_epi8_mask(k, A, B)
Definition: avx512vlbwintrin.h:111
_mm_mask_adds_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_adds_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:705
_mm_maskz_sub_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_sub_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:415
_mm256_mask_sub_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_sub_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:352
_mm256_mask_reduce_and_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_and_epi16(__mmask16 __M, __m256i __W)
Definition: avx512vlbwintrin.h:2933
_mm256_mask_srai_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_srai_epi16(__m256i __W, __mmask16 __U, __m256i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:2117
_mm_maskz_unpackhi_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_unpackhi_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1659
_mm256_maskz_unpackhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_unpackhi_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1701
_mm_reduce_or_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_or_epi16(__m128i __W)
Definition: avx512vlbwintrin.h:2828
_mm256_reduce_and_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_and_epi16(__m256i __W)
Definition: avx512vlbwintrin.h:2911
_mm_mask_reduce_max_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_max_epi16(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:2877
_mm_testn_epi8_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 _mm_testn_epi8_mask(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2521
_mm256_maskz_mullo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mullo_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:429
_mm_mask_testn_epi16_mask
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 _mm_mask_testn_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2554
_mm256_maskz_broadcastb_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_broadcastb_epi8(__mmask32 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2646
_mm256_maskz_avg_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_avg_epu16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:857
_mm_cmpneq_epi8_mask
#define _mm_cmpneq_epi8_mask(A, B)
Definition: avx512vlbwintrin.h:129
_mm256_mask_cvtusepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS256 _mm256_mask_cvtusepi16_storeu_epi8(void *__P, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1562
_mm_maskz_add_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_add_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:387
_mm_testn_epi16_mask
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 _mm_testn_epi16_mask(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2548
_mm256_mask_cmpneq_epi8_mask
#define _mm256_mask_cmpneq_epi8_mask(k, A, B)
Definition: avx512vlbwintrin.h:156
_mm256_mask_sll_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_sll_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1932
_mm_mask_mov_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mov_epi8(__m128i __W, __mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2230
_mm256_maskz_permutexvar_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutexvar_epi16(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2747
_mm_mask_reduce_add_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_add_epi16(__mmask8 __M, __m128i __W)
Definition: avx512vlbwintrin.h:2833
_mm_mask_test_epi16_mask
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 _mm_mask_test_epi16_mask(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2500
_mm256_mask_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtepi16_epi8(__m128i __O, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1536
_mm_maskz_subs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_subs_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1161
_mm256_mask_packs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:601
_mm256_maskz_min_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_min_epi8(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1009
_mm_mask_packus_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:649
_mm_maskz_packus_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:609
_mm_maskz_abs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_abs_epi16(__mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:522
_mm_mask_subs_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_subs_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1249
_mm256_loadu_epi16
static __inline __m256i __DEFAULT_FN_ATTRS256 _mm256_loadu_epi16(void const *__P)
Definition: avx512vlbwintrin.h:2321
_mm256_mask_subs_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_subs_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1233
_mm256_maskz_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtepi16_epi8(__mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1543
_mm256_mask_storeu_epi16
static __inline__ void __DEFAULT_FN_ATTRS256 _mm256_mask_storeu_epi16(void *__P, __mmask16 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2425
_mm_mask_avg_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_avg_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:833
_mm_maskz_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtusepi16_epi8(__mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1462
_mm256_maskz_add_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_add_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:331
_mm256_mask_min_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_min_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1081
_mm256_maskz_packus_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:625
_mm_mask_sub_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_sub_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:408
_mm_maskz_max_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_max_epu8(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:929
_mm_maskz_srli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_srli_epi16(__mmask8 __U, __m128i __A, int __B)
Definition: avx512vlbwintrin.h:2174
_mm256_cmpneq_epi8_mask
#define _mm256_cmpneq_epi8_mask(A, B)
Definition: avx512vlbwintrin.h:154
_mm256_mask_permutexvar_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutexvar_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2756
_mm_mask_min_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_min_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1065
_mm256_mask_adds_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_adds_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:689
_mm_maskz_shuffle_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1129
_mm256_mask_srav_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_srav_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2031
_mm_mask_testn_epi8_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 _mm_mask_testn_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2527
_mm256_mask_min_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_min_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1017
_mm_maskz_unpackhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_unpackhi_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1687
_mm256_mask_test_epi16_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 _mm256_mask_test_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2514
_mm_reduce_min_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_min_epi8(__m128i __V)
Definition: avx512vlbwintrin.h:3043
_mm256_mask_add_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_add_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:324
_mm_mask_packs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:585
_mm256_maskz_packs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:593
_mm_maskz_set1_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_set1_epi16(__mmask8 __M, short __A)
Definition: avx512vlbwintrin.h:2710
_mm256_mask_reduce_min_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_min_epi8(__mmask32 __M, __m256i __V)
Definition: avx512vlbwintrin.h:3153
_mm256_mask_min_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_min_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1049
_mm256_mask_set1_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_set1_epi8(__m256i __O, __mmask32 __M, char __A)
Definition: avx512vlbwintrin.h:2279
_mm256_mask_reduce_and_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_and_epi8(__mmask32 __M, __m256i __W)
Definition: avx512vlbwintrin.h:3109
_mm_maskz_loadu_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_loadu_epi8(__mmask16 __U, void const *__P)
Definition: avx512vlbwintrin.h:2364
_mm256_maskz_min_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_min_epu8(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1073
_mm256_reduce_min_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_min_epi8(__m256i __V)
Definition: avx512vlbwintrin.h:3131
_mm256_mask_mulhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mulhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1638
_mm_sllv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_sllv_epi16(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1894
_mm_mask_adds_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_adds_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:673
_mm256_maskz_mulhi_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mulhi_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1645
_mm_test_epi16_mask
static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 _mm_test_epi16_mask(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2494
_mm256_maskz_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtsepi16_epi8(__mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1441
_mm_mask_srai_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_srai_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:2101
_mm_mask_set1_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_set1_epi8(__m128i __O, __mmask16 __M, char __A)
Definition: avx512vlbwintrin.h:2263
_mm_mask_max_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_max_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:969
_mm_mask_reduce_min_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_min_epi16(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:2889
_mm256_maskz_adds_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_adds_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:697
_mm256_mask_subs_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_subs_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1265
_mm_maskz_subs_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_subs_epu8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1225
_mm_mask_subs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_subs_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1185
_mm_maskz_adds_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_adds_epu8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:745
_mm256_mask_testn_epi8_mask
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 _mm256_mask_testn_epi8_mask(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2541
_mm_maskz_mulhrs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mulhrs_epi16(__mmask8 __U, __m128i __X, __m128i __Y)
Definition: avx512vlbwintrin.h:1575
_mm_maskz_sub_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_sub_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:401
_mm256_movepi8_mask
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 _mm256_movepi8_mask(__m256i __A)
Definition: avx512vlbwintrin.h:2580
_mm256_maskz_srav_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_srav_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2039
_mm_cmpeq_epi8_mask
#define _mm_cmpeq_epi8_mask(A, B)
Definition: avx512vlbwintrin.h:109
_mm256_mask_max_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_max_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:985
_mm_mask_madd_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_madd_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1378
_mm_maskz_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtepi16_epi8(__mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1505
_mm_mask_packus_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_packus_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:617
_mm256_maskz_add_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_add_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:317
_mm256_storeu_epi16
static __inline void __DEFAULT_FN_ATTRS256 _mm256_storeu_epi16(void *__P, __m256i __A)
Definition: avx512vlbwintrin.h:2416
_mm_maskz_sll_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_sll_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1924
_mm256_maskz_sub_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_sub_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:359
_mm_mask_reduce_max_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_max_epu16(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:2883
_mm256_maskz_set1_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_set1_epi8(__mmask32 __M, char __A)
Definition: avx512vlbwintrin.h:2287
_mm256_mask_reduce_or_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_or_epi16(__mmask16 __M, __m256i __W)
Definition: avx512vlbwintrin.h:2939
_mm_mask_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_cvtsepi16_epi8(__m128i __O, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1413
_mm256_mask_min_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1113
_mm256_testn_epi8_mask
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 _mm256_testn_epi8_mask(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2534
_mm256_mask_reduce_add_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_add_epi8(__mmask32 __M, __m256i __W)
Definition: avx512vlbwintrin.h:3097
_mm256_maskz_sub_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_sub_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:345
_mm256_maskz_cvtepu8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtepu8_epi16(__mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1821
_mm_mask_adds_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_adds_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:737
_mm256_mask_loadu_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_loadu_epi8(__m256i __W, __mmask32 __U, void const *__P)
Definition: avx512vlbwintrin.h:2382
_mm_mask_mulhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mulhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1624
_mm256_mask_reduce_max_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_max_epu16(__mmask16 __M, __m256i __V)
Definition: avx512vlbwintrin.h:2971
_mm256_permutex2var_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B)
Definition: avx512vlbwintrin.h:1315
_mm_maskz_sllv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_sllv_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1908
_mm256_reduce_mul_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_mul_epi16(__m256i __W)
Definition: avx512vlbwintrin.h:2906
_mm256_maskz_mov_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mov_epi8(__mmask32 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2254
_mm_mask_sub_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_sub_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:394
_mm_mask_reduce_min_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_min_epu8(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:3071
_mm_maskz_srav_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_srav_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2061
_mm256_mask_adds_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_adds_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:721
_mm_maskz_permutex2var_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutex2var_epi16(__mmask8 __U, __m128i __A, __m128i __I, __m128i __B)
Definition: avx512vlbwintrin.h:1306
_mm256_mask_sub_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_sub_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:338
_mm256_maskz_max_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_max_epu8(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:945
_mm_srav_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_srav_epi16(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2047
_mm_reduce_and_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_and_epi8(__m128i __W)
Definition: avx512vlbwintrin.h:2999
_mm256_reduce_or_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_or_epi16(__m256i __W)
Definition: avx512vlbwintrin.h:2916
_mm_maskz_adds_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_adds_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:681
_mm_maskz_subs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_subs_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1193
_mm256_mask_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS256 _mm256_mask_storeu_epi8(void *__P, __mmask32 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2459
_mm_maskz_min_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_min_epi16(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1025
_mm256_mask_srlv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_srlv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1987
_mm256_mask_mulhi_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mulhi_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1610
_mm_maskz_max_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_max_epi8(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:865
_mm256_sllv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_sllv_epi16(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1872
_mm256_test_epi8_mask
static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 _mm256_test_epi8_mask(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2480
_mm256_maskz_set1_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_set1_epi16(__mmask16 __M, short __A)
Definition: avx512vlbwintrin.h:2694
_mm_maskz_packs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:577
_mm256_maskz_packs_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:561
_mm_reduce_add_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_add_epi8(__m128i __W)
Definition: avx512vlbwintrin.h:2989
_mm_mask_storeu_epi16
static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_storeu_epi16(void *__P, __mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2408
_mm256_maskz_abs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A)
Definition: avx512vlbwintrin.h:538
_mm_mask_srlv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_srlv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2009
_mm_movm_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_movm_epi8(__mmask16 __A)
Definition: avx512vlbwintrin.h:2598
_mm256_maskz_unpacklo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_unpacklo_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1757
_mm256_maskz_packus_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:657
_mm256_mask_mullo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mullo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:422
_mm_mask_blend_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_blend_epi8(__mmask16 __U, __m128i __A, __m128i __W)
Definition: avx512vlbwintrin.h:450
_mm256_cmpeq_epi16_mask
#define _mm256_cmpeq_epi16_mask(A, B)
Definition: avx512vlbwintrin.h:234
_mm256_maskz_min_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1105
_mm_storeu_epi8
static __inline void __DEFAULT_FN_ATTRS128 _mm_storeu_epi8(void *__P, __m128i __A)
Definition: avx512vlbwintrin.h:2433
_mm_mask_permutex2var_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutex2var_epi16(__m128i __A, __mmask8 __U, __m128i __I, __m128i __B)
Definition: avx512vlbwintrin.h:1288
_mm256_reduce_add_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_add_epi16(__m256i __W)
Definition: avx512vlbwintrin.h:2901
_mm_maskz_max_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_max_epi16(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:897
_mm_mask_reduce_mul_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_mul_epi16(__mmask8 __M, __m128i __W)
Definition: avx512vlbwintrin.h:2839
_mm256_maskz_srai_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_srai_epi16(__mmask16 __U, __m256i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:2126
_mm_maskz_mov_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mov_epi16(__mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2206
_mm_maskz_packus_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:641
_mm256_maskz_sra_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_sra_epi16(__mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2093
_mm256_maskz_abs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_abs_epi8(__mmask32 __U, __m256i __A)
Definition: avx512vlbwintrin.h:506
_mm_mask_reduce_add_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_add_epi8(__mmask16 __M, __m128i __W)
Definition: avx512vlbwintrin.h:3009
_mm_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtusepi16_epi8(__m128i __A)
Definition: avx512vlbwintrin.h:1448
_mm256_mask_reduce_or_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_or_epi8(__mmask32 __M, __m256i __W)
Definition: avx512vlbwintrin.h:3115
_mm_maskz_adds_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_adds_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:713
_mm256_movepi16_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 _mm256_movepi16_mask(__m256i __A)
Definition: avx512vlbwintrin.h:2592
_mm256_mask_cvtsepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS256 _mm256_mask_cvtsepi16_storeu_epi8(void *__P, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1556
_mm256_maskz_srli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_srli_epi16(__mmask16 __U, __m256i __A, int __B)
Definition: avx512vlbwintrin.h:2190
_mm_mask_set1_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_set1_epi16(__m128i __O, __mmask8 __M, short __A)
Definition: avx512vlbwintrin.h:2702
_mm256_mask_loadu_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_loadu_epi16(__m256i __W, __mmask16 __U, void const *__P)
Definition: avx512vlbwintrin.h:2330
_mm_maskz_permutexvar_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutexvar_epi16(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2724
_mm256_maskz_subs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_subs_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1177
_mm256_maskz_srl_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_srl_epi16(__mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2158
_mm_mask_cvtepi8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_cvtepi8_epi16(__m128i __W, __mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1764
_mm_mask_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_cvtepi16_epi8(__m128i __O, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1498
_mm256_reduce_mul_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_mul_epi8(__m256i __W)
Definition: avx512vlbwintrin.h:3082
_mm_maskz_broadcastw_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_broadcastw_epi16(__mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2662
_mm256_srav_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_srav_epi16(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2025
_mm_mask_cvtsepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_cvtsepi16_storeu_epi8(void *__P, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1519
_mm256_maskz_adds_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_adds_epu16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:793
_mm256_mask_abs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_abs_epi8(__m256i __W, __mmask32 __U, __m256i __A)
Definition: avx512vlbwintrin.h:498
_mm_mask_min_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_min_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1033
_mm_mask_mulhrs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mulhrs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
Definition: avx512vlbwintrin.h:1568
_mm_maskz_set1_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_set1_epi8(__mmask16 __M, char __A)
Definition: avx512vlbwintrin.h:2271
_mm256_maskz_unpackhi_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_unpackhi_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1673
_mm256_maskz_loadu_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_loadu_epi8(__mmask32 __U, void const *__P)
Definition: avx512vlbwintrin.h:2390
_mm256_mask_mov_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mov_epi16(__m256i __W, __mmask16 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2214
_mm256_mask_permutex2var_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutex2var_epi16(__m256i __A, __mmask16 __U, __m256i __I, __m256i __B)
Definition: avx512vlbwintrin.h:1322
_mm_mask_cvtepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_cvtepi16_storeu_epi8(void *__P, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1512
_mm_test_epi8_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 _mm_test_epi8_mask(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2467
_mm_mask_cvtusepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_cvtusepi16_storeu_epi8(void *__P, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1525
_mm256_maskz_adds_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_adds_epu8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:761
_mm_maskz_subs_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_subs_epu16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1257
_mm256_mask_subs_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_subs_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1169
_mm256_mask_reduce_min_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_min_epi16(__mmask16 __M, __m256i __V)
Definition: avx512vlbwintrin.h:2977
_mm_reduce_or_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_or_epi8(__m128i __W)
Definition: avx512vlbwintrin.h:3004
_mm256_mask_cvtepi16_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS256 _mm256_mask_cvtepi16_storeu_epi8(void *__P, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1550
_mm_maskz_avg_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_avg_epu8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:809
_mm256_maskz_permutex2var_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutex2var_epi16(__mmask16 __U, __m256i __A, __m256i __I, __m256i __B)
Definition: avx512vlbwintrin.h:1340
_mm_mask_packs_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_packs_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:553
__DEFAULT_FN_ATTRS256
#define __DEFAULT_FN_ATTRS256
Definition: avx512vlbwintrin.h:22
_mm_mask_cmpneq_epi16_mask
#define _mm_mask_cmpneq_epi16_mask(k, A, B)
Definition: avx512vlbwintrin.h:231
_mm_mask_test_epi8_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 _mm_mask_test_epi8_mask(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2473
_mm256_test_epi16_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 _mm256_test_epi16_mask(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2507
_mm256_maskz_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtusepi16_epi8(__mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1483
_mm256_maskz_subs_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_subs_epu8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1241
_mm256_mask_reduce_max_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_max_epi16(__mmask16 __M, __m256i __V)
Definition: avx512vlbwintrin.h:2965
_mm256_maskz_maddubs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_maddubs_epi16(__mmask16 __U, __m256i __X, __m256i __Y)
Definition: avx512vlbwintrin.h:1371
_mm256_maskz_avg_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_avg_epu8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:825
_mm256_maskz_shuffle_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_shuffle_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1145
_mm256_mask_add_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_add_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:310
_mm_maskz_min_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_min_epu8(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1057
_mm256_movm_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_movm_epi8(__mmask32 __A)
Definition: avx512vlbwintrin.h:2604
_mm_cmpeq_epi16_mask
#define _mm_cmpeq_epi16_mask(A, B)
Definition: avx512vlbwintrin.h:209
_mm_mask_cmpeq_epi16_mask
#define _mm_mask_cmpeq_epi16_mask(k, A, B)
Definition: avx512vlbwintrin.h:211
_mm_mask_max_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_max_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:937
__DEFAULT_FN_ATTRS128
#define __DEFAULT_FN_ATTRS128
Definition: avx512vlbwintrin.h:18
_mm256_maskz_mulhi_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mulhi_epu16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1617
_mm256_mask_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtsepi16_epi8(__m128i __O, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1434
_mm256_maskz_adds_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_adds_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:729
_mm256_storeu_epi8
static __inline void __DEFAULT_FN_ATTRS256 _mm256_storeu_epi8(void *__P, __m256i __A)
Definition: avx512vlbwintrin.h:2450
_mm256_movm_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_movm_epi16(__mmask16 __A)
Definition: avx512vlbwintrin.h:2616
_mm_maskz_mov_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_mov_epi8(__mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2238
_mm_permutexvar_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutexvar_epi16(__m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2718
_mm_mask_storeu_epi8
static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_storeu_epi8(void *__P, __mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2442
_mm256_maskz_min_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_min_epi16(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1041
_mm256_reduce_max_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS256 _mm256_reduce_max_epu8(__m256i __V)
Definition: avx512vlbwintrin.h:3126
_mm_mask_permutexvar_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutexvar_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2732
_mm_mask_mov_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mov_epi16(__m128i __W, __mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:2198
_mm_mask_shuffle_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1121
_mm_mask_sra_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_sra_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2069
_mm_maskz_unpacklo_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_unpacklo_epi8(__mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1715
_mm_mask_mullo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mullo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:436
_mm256_mask_adds_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_adds_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:753
_mm256_mask_max_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_max_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:889
_mm_mask_abs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_abs_epi8(__m128i __W, __mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:482
_mm_maskz_loadu_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_loadu_epi16(__mmask8 __U, void const *__P)
Definition: avx512vlbwintrin.h:2312
_mm256_mask_reduce_max_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_max_epi8(__mmask32 __M, __m256i __V)
Definition: avx512vlbwintrin.h:3141
_mm_reduce_min_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS128 _mm_reduce_min_epu16(__m128i __V)
Definition: avx512vlbwintrin.h:2872
_mm256_mask_reduce_min_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_min_epu16(__mmask16 __M, __m256i __V)
Definition: avx512vlbwintrin.h:2983
_mm_reduce_mul_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_mul_epi16(__m128i __W)
Definition: avx512vlbwintrin.h:2818
_mm256_maskz_sllv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_sllv_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1886
_mm256_mask_packs_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:569
_mm_maskz_max_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_max_epu16(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:961
_mm_maskz_maddubs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_maddubs_epi16(__mmask8 __U, __m128i __X, __m128i __Y)
Definition: avx512vlbwintrin.h:1356
_mm256_mask_slli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_slli_epi16(__m256i __W, __mmask16 __U, __m256i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:1964
_mm_permutex2var_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B)
Definition: avx512vlbwintrin.h:1281
_mm256_mask_broadcastw_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_broadcastw_epi16(__m256i __O, __mmask16 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2670
_mm_mask_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_cvtusepi16_epi8(__m128i __O, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:1455
_mm256_testn_epi16_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 _mm256_testn_epi16_mask(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2560
_mm256_mask_testn_epi16_mask
static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 _mm256_mask_testn_epi16_mask(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2567
_mm256_maskz_unpacklo_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_unpacklo_epi8(__mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1729
_mm256_maskz_mov_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mov_epi16(__mmask16 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2222
_mm_mask_unpackhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_unpackhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1680
_mm256_mask_sllv_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_sllv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1878
_mm_maskz_slli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_slli_epi16(__mmask8 __U, __m128i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:1956
_mm_reduce_max_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS128 _mm_reduce_max_epu16(__m128i __V)
Definition: avx512vlbwintrin.h:2862
_mm256_mask_adds_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_adds_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:785
_mm256_reduce_max_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_max_epi16(__m256i __V)
Definition: avx512vlbwintrin.h:2945
_mm256_reduce_max_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_max_epi8(__m256i __V)
Definition: avx512vlbwintrin.h:3121
_mm256_loadu_epi8
static __inline __m256i __DEFAULT_FN_ATTRS256 _mm256_loadu_epi8(void const *__P)
Definition: avx512vlbwintrin.h:2373
_mm256_mask_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtusepi16_epi8(__m128i __O, __mmask16 __M, __m256i __A)
Definition: avx512vlbwintrin.h:1476
_mm_mask_reduce_and_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_and_epi16(__mmask8 __M, __m128i __W)
Definition: avx512vlbwintrin.h:2845
_mm_maskz_srai_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_srai_epi16(__mmask8 __U, __m128i __A, unsigned int __B)
Definition: avx512vlbwintrin.h:2109
_mm_mask_abs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_abs_epi16(__m128i __W, __mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:514
_mm_maskz_min_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_min_epu16(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1089
_mm256_maskz_mulhrs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_mulhrs_epi16(__mmask16 __U, __m256i __X, __m256i __Y)
Definition: avx512vlbwintrin.h:1589
_mm_maskz_packs_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:546
_mm256_reduce_add_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_add_epi8(__m256i __W)
Definition: avx512vlbwintrin.h:3077
_mm_mask_add_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_add_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:380
_mm_mask_unpacklo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_unpacklo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1736
_mm_mask_min_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_min_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1097
_mm256_maskz_subs_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1273
_mm_mask_sllv_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_sllv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1900
_mm256_mask_cmpeq_epi8_mask
#define _mm256_mask_cmpeq_epi8_mask(k, A, B)
Definition: avx512vlbwintrin.h:136
_mm_mask_add_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_add_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:366
_mm_mask_subs_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_subs_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1217
_mm_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtsepi16_epi8(__m128i __A)
Definition: avx512vlbwintrin.h:1406
_mm_mask_unpacklo_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_unpacklo_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1708
_mm256_cvtsepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtsepi16_epi8(__m256i __A)
Definition: avx512vlbwintrin.h:1427
_mm256_mask_avg_epu8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_avg_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:817
_mm_mask_reduce_or_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_or_epi8(__mmask16 __M, __m128i __W)
Definition: avx512vlbwintrin.h:3027
_mm_mask_broadcastw_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_broadcastw_epi16(__m128i __O, __mmask8 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2654
_mm_mask_reduce_and_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_and_epi8(__mmask16 __M, __m128i __W)
Definition: avx512vlbwintrin.h:3021
_mm256_mask_mulhrs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mulhrs_epi16(__m256i __W, __mmask16 __U, __m256i __X, __m256i __Y)
Definition: avx512vlbwintrin.h:1582
_mm_reduce_max_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_max_epi16(__m128i __V)
Definition: avx512vlbwintrin.h:2857
_mm256_mask_reduce_min_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_min_epu8(__mmask32 __M, __m256i __V)
Definition: avx512vlbwintrin.h:3159
_mm_maskz_broadcastb_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_broadcastb_epi8(__mmask16 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2630
_mm256_mask_cmpneq_epi16_mask
#define _mm256_mask_cmpneq_epi16_mask(k, A, B)
Definition: avx512vlbwintrin.h:256
_mm256_maskz_madd_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_madd_epi16(__mmask8 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1399
_mm256_mask_srli_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_srli_epi16(__m256i __W, __mmask16 __U, __m256i __A, int __B)
Definition: avx512vlbwintrin.h:2182
_mm_mask_sll_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_sll_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1916
_mm_maskz_min_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_min_epi8(__mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:993
_mm256_mask_reduce_mul_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_mul_epi16(__mmask16 __M, __m256i __W)
Definition: avx512vlbwintrin.h:2927
_mm256_mask_unpacklo_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_unpacklo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1750
_mm_reduce_min_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_min_epi16(__m128i __V)
Definition: avx512vlbwintrin.h:2867
_mm_mask_blend_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_blend_epi16(__mmask8 __U, __m128i __A, __m128i __W)
Definition: avx512vlbwintrin.h:466
_mm256_maskz_max_epu16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_max_epu16(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:977
_mm256_reduce_or_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_reduce_or_epi8(__m256i __W)
Definition: avx512vlbwintrin.h:3092
_mm256_mask_set1_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_set1_epi16(__m256i __O, __mmask16 __M, short __A)
Definition: avx512vlbwintrin.h:2686
_mm_maskz_sra_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_sra_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2077
_mm256_maskz_sll_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_sll_epi16(__mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1940
_mm256_maskz_cvtepi8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1788
_mm_mask_reduce_max_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_max_epu8(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:3059
_mm_maskz_unpacklo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_unpacklo_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1743
_mm256_mask_max_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_max_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:921
_mm256_maskz_subs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_subs_epi16(__mmask16 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1209
_mm_mask_loadu_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_loadu_epi16(__m128i __W, __mmask8 __U, void const *__P)
Definition: avx512vlbwintrin.h:2304
_mm_movm_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_movm_epi16(__mmask8 __A)
Definition: avx512vlbwintrin.h:2610
_mm_maskz_madd_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_madd_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1385
_mm256_reduce_max_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS256 _mm256_reduce_max_epu16(__m256i __V)
Definition: avx512vlbwintrin.h:2950
_mm_mask_srav_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_srav_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2053
_mm_mask_max_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_max_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:905
_mm_reduce_add_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_add_epi16(__m128i __W)
Definition: avx512vlbwintrin.h:2813
_mm256_cvtusepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtusepi16_epi8(__m256i __A)
Definition: avx512vlbwintrin.h:1469
_mm256_maskz_max_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_max_epi16(__mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:913
_mm_mask_unpackhi_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_unpackhi_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1652
_mm_mask_srl_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_srl_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2134
_mm256_cvtepi16_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtepi16_epi8(__m256i __A)
Definition: avx512vlbwintrin.h:1531
_mm256_maskz_max_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_max_epi8(__mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:881
_mm_maskz_srl_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_srl_epi16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2142
_mm_mask_cmpneq_epi8_mask
#define _mm_mask_cmpneq_epi8_mask(k, A, B)
Definition: avx512vlbwintrin.h:131
_mm256_mask_maddubs_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_maddubs_epi16(__m256i __W, __mmask16 __U, __m256i __X, __m256i __Y)
Definition: avx512vlbwintrin.h:1363
_mm256_maskz_broadcastw_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_broadcastw_epi16(__mmask16 __M, __m128i __A)
Definition: avx512vlbwintrin.h:2678
_mm_reduce_and_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_reduce_and_epi16(__m128i __W)
Definition: avx512vlbwintrin.h:2823
_mm256_mask2_permutex2var_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask2_permutex2var_epi16(__m256i __A, __m256i __I, __mmask16 __U, __m256i __B)
Definition: avx512vlbwintrin.h:1331
_mm256_mask_shuffle_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_shuffle_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1137
_mm256_mask_cvtepu8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtepu8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1813
_mm_mask_reduce_mul_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_mask_reduce_mul_epi8(__mmask16 __M, __m128i __W)
Definition: avx512vlbwintrin.h:3015
_mm256_mask_packus_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:665
_mm256_cmpneq_epi16_mask
#define _mm256_cmpneq_epi16_mask(A, B)
Definition: avx512vlbwintrin.h:254
_mm_mask_mulhi_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_mulhi_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1596
_mm256_maskz_loadu_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_loadu_epi16(__mmask16 __U, void const *__P)
Definition: avx512vlbwintrin.h:2338
_mm_mask_srli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_srli_epi16(__m128i __W, __mmask8 __U, __m128i __A, int __B)
Definition: avx512vlbwintrin.h:2166
_mm256_mask_mov_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_mov_epi8(__m256i __W, __mmask32 __U, __m256i __A)
Definition: avx512vlbwintrin.h:2246
_mm_mask_reduce_or_epi16
static __inline__ short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_or_epi16(__mmask8 __M, __m128i __W)
Definition: avx512vlbwintrin.h:2851
_mm_reduce_max_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS128 _mm_reduce_max_epu8(__m128i __V)
Definition: avx512vlbwintrin.h:3038
_mm_reduce_mul_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS128 _mm_reduce_mul_epi8(__m128i __W)
Definition: avx512vlbwintrin.h:2994
_mm_mask_min_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_min_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:1001
_mm256_permutexvar_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutexvar_epi16(__m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:2741
_mm256_mask_unpacklo_epi8
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_unpacklo_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:1722
_mm256_mask_srl_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_srl_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B)
Definition: avx512vlbwintrin.h:2150
_mm_mask_max_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_max_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:873
_mm256_mask_cvtepi8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtepi8_epi16(__m256i __W, __mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1780
_mm_reduce_min_epu8
static __inline__ unsigned char __DEFAULT_FN_ATTRS128 _mm_reduce_min_epu8(__m128i __V)
Definition: avx512vlbwintrin.h:3048
_mm_mask_reduce_min_epu16
static __inline__ unsigned short __DEFAULT_FN_ATTRS128 _mm_mask_reduce_min_epu16(__mmask16 __M, __m128i __V)
Definition: avx512vlbwintrin.h:2895
_mm256_mask_reduce_mul_epi8
static __inline__ signed char __DEFAULT_FN_ATTRS256 _mm256_mask_reduce_mul_epi8(__mmask32 __M, __m256i __W)
Definition: avx512vlbwintrin.h:3103
_mm_maskz_adds_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_adds_epu16(__mmask8 __U, __m128i __A, __m128i __B)
Definition: avx512vlbwintrin.h:777
_mm_loadu_epi16
static __inline __m128i __DEFAULT_FN_ATTRS128 _mm_loadu_epi16(void const *__P)
Definition: avx512vlbwintrin.h:2295
_mm256_reduce_min_epi16
static __inline__ short __DEFAULT_FN_ATTRS256 _mm256_reduce_min_epi16(__m256i __V)
Definition: avx512vlbwintrin.h:2955
_mm256_mask_packus_epi32
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B)
Definition: avx512vlbwintrin.h:633
_mm256_mask_blend_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_blend_epi16(__mmask16 __U, __m256i __A, __m256i __W)
Definition: avx512vlbwintrin.h:474
_mm256_set1_epi8
static __inline __m256i __DEFAULT_FN_ATTRS _mm256_set1_epi8(char __b)
Constructs a 256-bit integer vector of [32 x i8], with each of the 8-bit integral vector elements set...
Definition: avxintrin.h:4290
_mm256_setzero_si256
static __inline __m256i __DEFAULT_FN_ATTRS _mm256_setzero_si256(void)
Constructs a 256-bit integer vector initialized to zero.
Definition: avxintrin.h:4353
_mm256_set1_epi16
static __inline __m256i __DEFAULT_FN_ATTRS _mm256_set1_epi16(short __w)
Constructs a 256-bit integer vector of [16 x i16], with each of the 16-bit integral vector elements s...
Definition: avxintrin.h:4272
_mm_unpacklo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_unpacklo_epi16(__m128i __a, __m128i __b)
Unpacks the low-order (index 0-3) values from each of the two 128-bit vectors of [8 x i16] and interl...
Definition: emmintrin.h:4531
_mm_packus_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_packus_epi16(__m128i __a, __m128i __b)
Converts, with saturation, 16-bit signed integers from both 128-bit integer vector operands into 8-bi...
Definition: emmintrin.h:4188
_mm_min_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epu8(__m128i __a, __m128i __b)
Compares corresponding elements of two 128-bit unsigned [16 x i8] vectors, saving the smaller value f...
Definition: emmintrin.h:2359
_mm_and_si128
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_and_si128(__m128i __a, __m128i __b)
Performs a bitwise AND of two 128-bit integer vectors.
Definition: emmintrin.h:2662
_mm_adds_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_adds_epi8(__m128i __a, __m128i __b)
Adds, with saturation, the corresponding elements of two 128-bit signed [16 x i8] vectors,...
Definition: emmintrin.h:2154
_mm_sub_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_sub_epi16(__m128i __a, __m128i __b)
Subtracts the corresponding 16-bit integer values in the operands.
Definition: emmintrin.h:2507
_mm_mullo_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mullo_epi16(__m128i __a, __m128i __b)
Multiplies the corresponding elements of two signed [8 x i16] vectors, saving the lower 16 bits of ea...
Definition: emmintrin.h:2416
_mm_srai_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_srai_epi16(__m128i __a, int __count)
Right-shifts each 16-bit value in the 128-bit integer vector operand by the specified number of bits.
Definition: emmintrin.h:2866
_mm_max_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_max_epu8(__m128i __a, __m128i __b)
Compares corresponding elements of two 128-bit unsigned [16 x i8] vectors, saving the greater value f...
Definition: emmintrin.h:2321
_mm_avg_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_avg_epu16(__m128i __a, __m128i __b)
Computes the rounded averages of corresponding elements of two 128-bit unsigned [8 x i16] vectors,...
Definition: emmintrin.h:2258
_mm_srl_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_srl_epi16(__m128i __a, __m128i __count)
Right-shifts each of 16-bit values in the 128-bit integer vector operand by the specified number of b...
Definition: emmintrin.h:2984
_mm_unpacklo_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_unpacklo_epi8(__m128i __a, __m128i __b)
Unpacks the low-order (index 0-7) values from two 128-bit vectors of [16 x i8] and interleaves them i...
Definition: emmintrin.h:4502
_mm_subs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_subs_epi8(__m128i __a, __m128i __b)
Subtracts, with saturation, corresponding 8-bit signed integer values in the input and returns the di...
Definition: emmintrin.h:2581
_mm_subs_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_subs_epu16(__m128i __a, __m128i __b)
Subtracts, with saturation, corresponding 16-bit unsigned integer values in the input and returns the...
Definition: emmintrin.h:2645
_mm_unpackhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_unpackhi_epi16(__m128i __a, __m128i __b)
Unpacks the high-order (index 4-7) values from two 128-bit vectors of [8 x i16] and interleaves them ...
Definition: emmintrin.h:4424
_mm_madd_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_madd_epi16(__m128i __a, __m128i __b)
Multiplies the corresponding elements of two 128-bit signed [8 x i16] vectors, producing eight interm...
Definition: emmintrin.h:2283
_mm_adds_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_adds_epu16(__m128i __a, __m128i __b)
Adds, with saturation, the corresponding elements of two 128-bit unsigned [8 x i16] vectors,...
Definition: emmintrin.h:2220
_mm_slli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_slli_epi16(__m128i __a, int __count)
Left-shifts each 16-bit value in the 128-bit integer vector operand by the specified number of bits.
Definition: emmintrin.h:2757
_mm_adds_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_adds_epu8(__m128i __a, __m128i __b)
Adds, with saturation, the corresponding elements of two 128-bit unsigned [16 x i8] vectors,...
Definition: emmintrin.h:2198
_mm_adds_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_adds_epi16(__m128i __a, __m128i __b)
Adds, with saturation, the corresponding elements of two 128-bit signed [8 x i16] vectors,...
Definition: emmintrin.h:2176
_mm_subs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_subs_epi16(__m128i __a, __m128i __b)
Subtracts, with saturation, corresponding 16-bit signed integer values in the input and returns the d...
Definition: emmintrin.h:2603
_mm_sub_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_sub_epi8(__m128i __a, __m128i __b)
Subtracts the corresponding 8-bit integer values in the operands.
Definition: emmintrin.h:2490
_mm_mulhi_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mulhi_epi16(__m128i __a, __m128i __b)
Multiplies the corresponding elements of two signed [8 x i16] vectors, saving the upper 16 bits of ea...
Definition: emmintrin.h:2378
_mm_mulhi_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mulhi_epu16(__m128i __a, __m128i __b)
Multiplies the corresponding elements of two unsigned [8 x i16] vectors, saving the upper 16 bits of ...
Definition: emmintrin.h:2397
_mm_max_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_max_epi16(__m128i __a, __m128i __b)
Compares corresponding elements of two 128-bit signed [8 x i16] vectors, saving the greater value fro...
Definition: emmintrin.h:2302
_mm_packs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_packs_epi16(__m128i __a, __m128i __b)
Converts, with saturation, 16-bit signed integers from both 128-bit integer vector operands into 8-bi...
Definition: emmintrin.h:4142
_mm_sra_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_sra_epi16(__m128i __a, __m128i __count)
Right-shifts each 16-bit value in the 128-bit integer vector operand by the specified number of bits.
Definition: emmintrin.h:2885
_mm_set1_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_set1_epi8(char __b)
Initializes all values in a 128-bit vector of [16 x i8] with the specified 8-bit value.
Definition: emmintrin.h:3723
_mm_packs_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_packs_epi32(__m128i __a, __m128i __b)
Converts, with saturation, 32-bit signed integers from both 128-bit integer vector operands into 16-b...
Definition: emmintrin.h:4165
_mm_sll_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_sll_epi16(__m128i __a, __m128i __count)
Left-shifts each 16-bit value in the 128-bit integer vector operand by the specified number of bits.
Definition: emmintrin.h:2775
_mm_add_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_add_epi8(__m128i __a, __m128i __b)
Adds the corresponding elements of two 128-bit vectors of [16 x i8], saving the lower 8 bits of each ...
Definition: emmintrin.h:2052
_mm_min_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epi16(__m128i __a, __m128i __b)
Compares corresponding elements of two 128-bit signed [8 x i16] vectors, saving the smaller value fro...
Definition: emmintrin.h:2340
_mm_set1_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_set1_epi16(short __w)
Initializes all values in a 128-bit vector of [8 x i16] with the specified 16-bit value.
Definition: emmintrin.h:3706
_mm_avg_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_avg_epu8(__m128i __a, __m128i __b)
Computes the rounded averages of corresponding elements of two 128-bit unsigned [16 x i8] vectors,...
Definition: emmintrin.h:2239
_mm_unpackhi_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_unpackhi_epi8(__m128i __a, __m128i __b)
Unpacks the high-order (index 8-15) values from two 128-bit vectors of [16 x i8] and interleaves them...
Definition: emmintrin.h:4396
_mm_srli_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_srli_epi16(__m128i __a, int __count)
Right-shifts each of 16-bit values in the 128-bit integer vector operand by the specified number of b...
Definition: emmintrin.h:2966
_mm_add_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_add_epi16(__m128i __a, __m128i __b)
Adds the corresponding elements of two 128-bit vectors of [8 x i16], saving the lower 16 bits of each...
Definition: emmintrin.h:2073
_mm_subs_epu8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_subs_epu8(__m128i __a, __m128i __b)
Subtracts, with saturation, corresponding 8-bit unsigned integer values in the input and returns the ...
Definition: emmintrin.h:2624
_mm_setzero_si128
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_setzero_si128(void)
Creates a 128-bit integer vector initialized to zero.
Definition: emmintrin.h:3858
__v
struct __storeu_i16 *__P __v
Definition: immintrin.h:472
__P
__inline unsigned int unsigned int unsigned int * __P
Definition: bmi2intrin.h:25
__Y
__inline unsigned int unsigned int __Y
Definition: bmi2intrin.h:19
_mm_min_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epu16(__m128i __V1, __m128i __V2)
Compares the corresponding elements of two 128-bit vectors of [8 x u16] and returns a 128-bit vector ...
Definition: smmintrin.h:702
_mm_cvtepu8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvtepu8_epi16(__m128i __V)
Zero-extends each of the lower eight 8-bit integer elements of a 128-bit vector of [16 x i8] to 16-bi...
Definition: smmintrin.h:1338
_mm_min_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_min_epi8(__m128i __V1, __m128i __V2)
Compares the corresponding elements of two 128-bit vectors of [16 x i8] and returns a 128-bit vector ...
Definition: smmintrin.h:666
_mm_packus_epi32
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_packus_epi32(__m128i __V1, __m128i __V2)
Converts, with saturation, 32-bit signed integers from both 128-bit integer vector operands into 16-b...
Definition: smmintrin.h:1454
_mm_max_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_max_epi8(__m128i __V1, __m128i __V2)
Compares the corresponding elements of two 128-bit vectors of [16 x i8] and returns a 128-bit vector ...
Definition: smmintrin.h:684
_mm_max_epu16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_max_epu16(__m128i __V1, __m128i __V2)
Compares the corresponding elements of two 128-bit vectors of [8 x u16] and returns a 128-bit vector ...
Definition: smmintrin.h:720
_mm_cvtepi8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvtepi8_epi16(__m128i __V)
Sign-extends each of the lower eight 8-bit integer elements of a 128-bit vector of [16 x i8] to 16-bi...
Definition: smmintrin.h:1221
_mm_abs_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_abs_epi8(__m128i __a)
Computes the absolute value of each of the packed 8-bit signed integers in the source operand and sto...
Definition: tmmintrin.h:59
_mm_mulhrs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_mulhrs_epi16(__m128i __a, __m128i __b)
Multiplies packed 16-bit signed integer values, truncates the 32-bit products to the 18 most signific...
Definition: tmmintrin.h:549
_mm_abs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_abs_epi16(__m128i __a)
Computes the absolute value of each of the packed 16-bit signed integers in the source operand and st...
Definition: tmmintrin.h:95
_mm_maddubs_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_maddubs_epi16(__m128i __a, __m128i __b)
Multiplies corresponding pairs of packed 8-bit unsigned integer values contained in the first source ...
Definition: tmmintrin.h:499
_mm_shuffle_epi8
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_shuffle_epi8(__m128i __a, __m128i __b)
Copies the 8-bit integers from a 128-bit integer vector to the destination or clears 8-bit values in ...
Definition: tmmintrin.h:595
Generated on Tue Jul 2 2024 01:37:34 for clang by doxygen 1.9.1