avx512vbmi2intrin.h revision 344779
1142425Snectar/*===------------- avx512vbmi2intrin.h - VBMI2 intrinsics ------------------===
2160814Ssimon *
3142425Snectar *
4142425Snectar * Permission is hereby granted, free of charge, to any person obtaining a copy
5142425Snectar * of this software and associated documentation files (the "Software"), to deal
6142425Snectar * in the Software without restriction, including without limitation the rights
7142425Snectar * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8142425Snectar * copies of the Software, and to permit persons to whom the Software is
9142425Snectar * furnished to do so, subject to the following conditions:
10142425Snectar *
11142425Snectar * The above copyright notice and this permission notice shall be included in
12142425Snectar * all copies or substantial portions of the Software.
13142425Snectar *
14142425Snectar * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15142425Snectar * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16142425Snectar * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17142425Snectar * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18142425Snectar * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19142425Snectar * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20160814Ssimon * THE SOFTWARE.
21142425Snectar *
22142425Snectar *===-----------------------------------------------------------------------===
23142425Snectar */
24142425Snectar#ifndef __IMMINTRIN_H
25142425Snectar#error "Never use <avx512vbmi2intrin.h> directly; include <immintrin.h> instead."
26142425Snectar#endif
27142425Snectar
28142425Snectar#ifndef __AVX512VBMI2INTRIN_H
29142425Snectar#define __AVX512VBMI2INTRIN_H
30142425Snectar
31142425Snectar/* Define the default attributes for the functions in this file. */
32142425Snectar#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi2"), __min_vector_width__(512)))
33142425Snectar
34142425Snectar
35142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
36142425Snectar_mm512_mask_compress_epi16(__m512i __S, __mmask32 __U, __m512i __D)
37142425Snectar{
38142425Snectar  return (__m512i) __builtin_ia32_compresshi512_mask ((__v32hi) __D,
39142425Snectar              (__v32hi) __S,
40142425Snectar              __U);
41142425Snectar}
42142425Snectar
43142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
44142425Snectar_mm512_maskz_compress_epi16(__mmask32 __U, __m512i __D)
45142425Snectar{
46142425Snectar  return (__m512i) __builtin_ia32_compresshi512_mask ((__v32hi) __D,
47160814Ssimon              (__v32hi) _mm512_setzero_si512(),
48160814Ssimon              __U);
49160814Ssimon}
50160814Ssimon
51160814Ssimonstatic __inline__ __m512i __DEFAULT_FN_ATTRS
52160814Ssimon_mm512_mask_compress_epi8(__m512i __S, __mmask64 __U, __m512i __D)
53142425Snectar{
54160814Ssimon  return (__m512i) __builtin_ia32_compressqi512_mask ((__v64qi) __D,
55160814Ssimon              (__v64qi) __S,
56142425Snectar              __U);
57142425Snectar}
58142425Snectar
59142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
60142425Snectar_mm512_maskz_compress_epi8(__mmask64 __U, __m512i __D)
61142425Snectar{
62142425Snectar  return (__m512i) __builtin_ia32_compressqi512_mask ((__v64qi) __D,
63142425Snectar              (__v64qi) _mm512_setzero_si512(),
64142425Snectar              __U);
65160814Ssimon}
66160814Ssimon
67160814Ssimonstatic __inline__ void __DEFAULT_FN_ATTRS
68160814Ssimon_mm512_mask_compressstoreu_epi16(void *__P, __mmask32 __U, __m512i __D)
69160814Ssimon{
70160814Ssimon  __builtin_ia32_compressstorehi512_mask ((__v32hi *) __P, (__v32hi) __D,
71142425Snectar              __U);
72142425Snectar}
73142425Snectar
74142425Snectarstatic __inline__ void __DEFAULT_FN_ATTRS
75142425Snectar_mm512_mask_compressstoreu_epi8(void *__P, __mmask64 __U, __m512i __D)
76142425Snectar{
77142425Snectar  __builtin_ia32_compressstoreqi512_mask ((__v64qi *) __P, (__v64qi) __D,
78142425Snectar              __U);
79142425Snectar}
80142425Snectar
81142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
82142425Snectar_mm512_mask_expand_epi16(__m512i __S, __mmask32 __U, __m512i __D)
83142425Snectar{
84142425Snectar  return (__m512i) __builtin_ia32_expandhi512_mask ((__v32hi) __D,
85160814Ssimon              (__v32hi) __S,
86142425Snectar              __U);
87142425Snectar}
88142425Snectar
89142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
90142425Snectar_mm512_maskz_expand_epi16(__mmask32 __U, __m512i __D)
91142425Snectar{
92142425Snectar  return (__m512i) __builtin_ia32_expandhi512_mask ((__v32hi) __D,
93160814Ssimon              (__v32hi) _mm512_setzero_si512(),
94142425Snectar              __U);
95142425Snectar}
96142425Snectar
97142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
98142425Snectar_mm512_mask_expand_epi8(__m512i __S, __mmask64 __U, __m512i __D)
99142425Snectar{
100142425Snectar  return (__m512i) __builtin_ia32_expandqi512_mask ((__v64qi) __D,
101142425Snectar              (__v64qi) __S,
102142425Snectar              __U);
103142425Snectar}
104142425Snectar
105142425Snectarstatic __inline__ __m512i __DEFAULT_FN_ATTRS
106142425Snectar_mm512_maskz_expand_epi8(__mmask64 __U, __m512i __D)
107142425Snectar{
108  return (__m512i) __builtin_ia32_expandqi512_mask ((__v64qi) __D,
109              (__v64qi) _mm512_setzero_si512(),
110              __U);
111}
112
113static __inline__ __m512i __DEFAULT_FN_ATTRS
114_mm512_mask_expandloadu_epi16(__m512i __S, __mmask32 __U, void const *__P)
115{
116  return (__m512i) __builtin_ia32_expandloadhi512_mask ((const __v32hi *)__P,
117              (__v32hi) __S,
118              __U);
119}
120
121static __inline__ __m512i __DEFAULT_FN_ATTRS
122_mm512_maskz_expandloadu_epi16(__mmask32 __U, void const *__P)
123{
124  return (__m512i) __builtin_ia32_expandloadhi512_mask ((const __v32hi *)__P,
125              (__v32hi) _mm512_setzero_si512(),
126              __U);
127}
128
129static __inline__ __m512i __DEFAULT_FN_ATTRS
130_mm512_mask_expandloadu_epi8(__m512i __S, __mmask64 __U, void const *__P)
131{
132  return (__m512i) __builtin_ia32_expandloadqi512_mask ((const __v64qi *)__P,
133              (__v64qi) __S,
134              __U);
135}
136
137static __inline__ __m512i __DEFAULT_FN_ATTRS
138_mm512_maskz_expandloadu_epi8(__mmask64 __U, void const *__P)
139{
140  return (__m512i) __builtin_ia32_expandloadqi512_mask ((const __v64qi *)__P,
141              (__v64qi) _mm512_setzero_si512(),
142              __U);
143}
144
145#define _mm512_shldi_epi64(A, B, I) \
146  (__m512i)__builtin_ia32_vpshldq512((__v8di)(__m512i)(A), \
147                                     (__v8di)(__m512i)(B), (int)(I))
148
149#define _mm512_mask_shldi_epi64(S, U, A, B, I) \
150  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
151                                    (__v8di)_mm512_shldi_epi64((A), (B), (I)), \
152                                    (__v8di)(__m512i)(S))
153
154#define _mm512_maskz_shldi_epi64(U, A, B, I) \
155  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
156                                    (__v8di)_mm512_shldi_epi64((A), (B), (I)), \
157                                    (__v8di)_mm512_setzero_si512())
158
159#define _mm512_shldi_epi32(A, B, I) \
160  (__m512i)__builtin_ia32_vpshldd512((__v16si)(__m512i)(A), \
161                                     (__v16si)(__m512i)(B), (int)(I))
162
163#define _mm512_mask_shldi_epi32(S, U, A, B, I) \
164  (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
165                                   (__v16si)_mm512_shldi_epi32((A), (B), (I)), \
166                                   (__v16si)(__m512i)(S))
167
168#define _mm512_maskz_shldi_epi32(U, A, B, I) \
169  (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
170                                   (__v16si)_mm512_shldi_epi32((A), (B), (I)), \
171                                   (__v16si)_mm512_setzero_si512())
172
173#define _mm512_shldi_epi16(A, B, I) \
174  (__m512i)__builtin_ia32_vpshldw512((__v32hi)(__m512i)(A), \
175                                     (__v32hi)(__m512i)(B), (int)(I))
176
177#define _mm512_mask_shldi_epi16(S, U, A, B, I) \
178  (__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
179                                   (__v32hi)_mm512_shldi_epi16((A), (B), (I)), \
180                                   (__v32hi)(__m512i)(S))
181
182#define _mm512_maskz_shldi_epi16(U, A, B, I) \
183  (__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
184                                   (__v32hi)_mm512_shldi_epi16((A), (B), (I)), \
185                                   (__v32hi)_mm512_setzero_si512())
186
187#define _mm512_shrdi_epi64(A, B, I) \
188  (__m512i)__builtin_ia32_vpshrdq512((__v8di)(__m512i)(A), \
189                                     (__v8di)(__m512i)(B), (int)(I))
190
191#define _mm512_mask_shrdi_epi64(S, U, A, B, I) \
192  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
193                                    (__v8di)_mm512_shrdi_epi64((A), (B), (I)), \
194                                    (__v8di)(__m512i)(S))
195
196#define _mm512_maskz_shrdi_epi64(U, A, B, I) \
197  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
198                                    (__v8di)_mm512_shrdi_epi64((A), (B), (I)), \
199                                    (__v8di)_mm512_setzero_si512())
200
201#define _mm512_shrdi_epi32(A, B, I) \
202  (__m512i)__builtin_ia32_vpshrdd512((__v16si)(__m512i)(A), \
203                                     (__v16si)(__m512i)(B), (int)(I))
204
205#define _mm512_mask_shrdi_epi32(S, U, A, B, I) \
206  (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
207                                   (__v16si)_mm512_shrdi_epi32((A), (B), (I)), \
208                                   (__v16si)(__m512i)(S))
209
210#define _mm512_maskz_shrdi_epi32(U, A, B, I) \
211  (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
212                                   (__v16si)_mm512_shrdi_epi32((A), (B), (I)), \
213                                   (__v16si)_mm512_setzero_si512())
214
215#define _mm512_shrdi_epi16(A, B, I) \
216  (__m512i)__builtin_ia32_vpshrdw512((__v32hi)(__m512i)(A), \
217                                     (__v32hi)(__m512i)(B), (int)(I))
218
219#define _mm512_mask_shrdi_epi16(S, U, A, B, I) \
220  (__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
221                                   (__v32hi)_mm512_shrdi_epi16((A), (B), (I)), \
222                                   (__v32hi)(__m512i)(S))
223
224#define _mm512_maskz_shrdi_epi16(U, A, B, I) \
225  (__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
226                                   (__v32hi)_mm512_shrdi_epi16((A), (B), (I)), \
227                                   (__v32hi)_mm512_setzero_si512())
228
229static __inline__ __m512i __DEFAULT_FN_ATTRS
230_mm512_shldv_epi64(__m512i __A, __m512i __B, __m512i __C)
231{
232  return (__m512i)__builtin_ia32_vpshldvq512((__v8di)__A, (__v8di)__B,
233                                             (__v8di)__C);
234}
235
236static __inline__ __m512i __DEFAULT_FN_ATTRS
237_mm512_mask_shldv_epi64(__m512i __A, __mmask8 __U, __m512i __B, __m512i __C)
238{
239  return (__m512i)__builtin_ia32_selectq_512(__U,
240                                      (__v8di)_mm512_shldv_epi64(__A, __B, __C),
241                                      (__v8di)__A);
242}
243
244static __inline__ __m512i __DEFAULT_FN_ATTRS
245_mm512_maskz_shldv_epi64(__mmask8 __U, __m512i __A, __m512i __B, __m512i __C)
246{
247  return (__m512i)__builtin_ia32_selectq_512(__U,
248                                      (__v8di)_mm512_shldv_epi64(__A, __B, __C),
249                                      (__v8di)_mm512_setzero_si512());
250}
251
252static __inline__ __m512i __DEFAULT_FN_ATTRS
253_mm512_shldv_epi32(__m512i __A, __m512i __B, __m512i __C)
254{
255  return (__m512i)__builtin_ia32_vpshldvd512((__v16si)__A, (__v16si)__B,
256                                             (__v16si)__C);
257}
258
259static __inline__ __m512i __DEFAULT_FN_ATTRS
260_mm512_mask_shldv_epi32(__m512i __A, __mmask16 __U, __m512i __B, __m512i __C)
261{
262  return (__m512i)__builtin_ia32_selectd_512(__U,
263                                     (__v16si)_mm512_shldv_epi32(__A, __B, __C),
264                                     (__v16si)__A);
265}
266
267static __inline__ __m512i __DEFAULT_FN_ATTRS
268_mm512_maskz_shldv_epi32(__mmask16 __U, __m512i __A, __m512i __B, __m512i __C)
269{
270  return (__m512i)__builtin_ia32_selectd_512(__U,
271                                     (__v16si)_mm512_shldv_epi32(__A, __B, __C),
272                                     (__v16si)_mm512_setzero_si512());
273}
274
275static __inline__ __m512i __DEFAULT_FN_ATTRS
276_mm512_shldv_epi16(__m512i __A, __m512i __B, __m512i __C)
277{
278  return (__m512i)__builtin_ia32_vpshldvw512((__v32hi)__A, (__v32hi)__B,
279                                             (__v32hi)__C);
280}
281
282static __inline__ __m512i __DEFAULT_FN_ATTRS
283_mm512_mask_shldv_epi16(__m512i __A, __mmask32 __U, __m512i __B, __m512i __C)
284{
285  return (__m512i)__builtin_ia32_selectw_512(__U,
286                                     (__v32hi)_mm512_shldv_epi16(__A, __B, __C),
287                                     (__v32hi)__A);
288}
289
290static __inline__ __m512i __DEFAULT_FN_ATTRS
291_mm512_maskz_shldv_epi16(__mmask32 __U, __m512i __A, __m512i __B, __m512i __C)
292{
293  return (__m512i)__builtin_ia32_selectw_512(__U,
294                                     (__v32hi)_mm512_shldv_epi16(__A, __B, __C),
295                                     (__v32hi)_mm512_setzero_si512());
296}
297
298static __inline__ __m512i __DEFAULT_FN_ATTRS
299_mm512_shrdv_epi64(__m512i __A, __m512i __B, __m512i __C)
300{
301  return (__m512i)__builtin_ia32_vpshrdvq512((__v8di)__A, (__v8di)__B,
302                                             (__v8di)__C);
303}
304
305static __inline__ __m512i __DEFAULT_FN_ATTRS
306_mm512_mask_shrdv_epi64(__m512i __A, __mmask8 __U, __m512i __B, __m512i __C)
307{
308  return (__m512i)__builtin_ia32_selectq_512(__U,
309                                      (__v8di)_mm512_shrdv_epi64(__A, __B, __C),
310                                      (__v8di)__A);
311}
312
313static __inline__ __m512i __DEFAULT_FN_ATTRS
314_mm512_maskz_shrdv_epi64(__mmask8 __U, __m512i __A, __m512i __B, __m512i __C)
315{
316  return (__m512i)__builtin_ia32_selectq_512(__U,
317                                      (__v8di)_mm512_shrdv_epi64(__A, __B, __C),
318                                      (__v8di)_mm512_setzero_si512());
319}
320
321static __inline__ __m512i __DEFAULT_FN_ATTRS
322_mm512_shrdv_epi32(__m512i __A, __m512i __B, __m512i __C)
323{
324  return (__m512i)__builtin_ia32_vpshrdvd512((__v16si)__A, (__v16si)__B,
325                                             (__v16si)__C);
326}
327
328static __inline__ __m512i __DEFAULT_FN_ATTRS
329_mm512_mask_shrdv_epi32(__m512i __A, __mmask16 __U, __m512i __B, __m512i __C)
330{
331  return (__m512i) __builtin_ia32_selectd_512(__U,
332                                     (__v16si)_mm512_shrdv_epi32(__A, __B, __C),
333                                     (__v16si)__A);
334}
335
336static __inline__ __m512i __DEFAULT_FN_ATTRS
337_mm512_maskz_shrdv_epi32(__mmask16 __U, __m512i __A, __m512i __B, __m512i __C)
338{
339  return (__m512i) __builtin_ia32_selectd_512(__U,
340                                     (__v16si)_mm512_shrdv_epi32(__A, __B, __C),
341                                     (__v16si)_mm512_setzero_si512());
342}
343
344static __inline__ __m512i __DEFAULT_FN_ATTRS
345_mm512_shrdv_epi16(__m512i __A, __m512i __B, __m512i __C)
346{
347  return (__m512i)__builtin_ia32_vpshrdvw512((__v32hi)__A, (__v32hi)__B,
348                                             (__v32hi)__C);
349}
350
351static __inline__ __m512i __DEFAULT_FN_ATTRS
352_mm512_mask_shrdv_epi16(__m512i __A, __mmask32 __U, __m512i __B, __m512i __C)
353{
354  return (__m512i)__builtin_ia32_selectw_512(__U,
355                                     (__v32hi)_mm512_shrdv_epi16(__A, __B, __C),
356                                     (__v32hi)__A);
357}
358
359static __inline__ __m512i __DEFAULT_FN_ATTRS
360_mm512_maskz_shrdv_epi16(__mmask32 __U, __m512i __A, __m512i __B, __m512i __C)
361{
362  return (__m512i)__builtin_ia32_selectw_512(__U,
363                                     (__v32hi)_mm512_shrdv_epi16(__A, __B, __C),
364                                     (__v32hi)_mm512_setzero_si512());
365}
366
367
368#undef __DEFAULT_FN_ATTRS
369
370#endif
371
372