native 0.0.1
Vectors, masks and wide register packs for C++26
Loading...
Searching...
No Matches
native.x86.vbmi.ccm
1// SPDX-License-Identifier: BSD-2-Clause OR Apache-2.0
2module;
3#include "native/isa_import.h"
4#include "native/x86/vbmi.h"
5export module native.x86.vbmi;
6export import native.x86.features;
7export import native.simd;
8// SPDX-License-Identifier: BSD-2-Clause OR Apache-2.0
9
10#if NATIVE_HOST_X86 || defined(NATIVE_DOXYGEN)
11export namespace native {
23
25 template<isa<x86> Arch>
26 requires(Arch.has(x86_feature::avx512f) &&
27 Arch.has(x86_feature::avx512bw) &&
28 Arch.has(x86_feature::avx512vbmi) &&
29 Arch.has(x86_feature::avx512vl))
30 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
31 constexpr simd<std::uint8_t, 16, Arch> vpermb(
32 simd<std::uint8_t, 16, Arch> indices,
33 simd<std::uint8_t, 16, Arch> value) noexcept {
34 if consteval {
35 return detail::x86_vbmi_constant::permute<false>(
36 indices, value, simd<std::uint8_t, 16, Arch>{}, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
37 } else {
38 return simd<std::uint8_t, 16, Arch>::from_native(
39 detail::x86_vbmi::vpermb<Arch>(
40 indices.to_native(), value.to_native()));
41 }
42 }
43
45 template<isa<x86> Arch>
46 requires(Arch.has(x86_feature::avx512f) &&
47 Arch.has(x86_feature::avx512bw) &&
48 Arch.has(x86_feature::avx512vbmi) &&
49 Arch.has(x86_feature::avx512vl))
50 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
51 constexpr simd<std::uint8_t, 16, Arch> mask_vpermb(
52 simd<std::uint8_t, 16, Arch> source,
53 predicate<16, Arch> mask,
54 simd<std::uint8_t, 16, Arch> indices,
55 simd<std::uint8_t, 16, Arch> value) noexcept {
56 if consteval {
57 return detail::x86_vbmi_constant::permute<false>(
58 indices, value, simd<std::uint8_t, 16, Arch>{}, source, mask.to_bitset());
59 } else {
60 return simd<std::uint8_t, 16, Arch>::from_native(
61 detail::x86_vbmi::mask_vpermb<Arch>(
62 source.to_native(), mask.to_bitset(), indices.to_native(), value.to_native()));
63 }
64 }
65
67 template<isa<x86> Arch>
68 requires(Arch.has(x86_feature::avx512f) &&
69 Arch.has(x86_feature::avx512bw) &&
70 Arch.has(x86_feature::avx512vbmi) &&
71 Arch.has(x86_feature::avx512vl))
72 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
73 constexpr simd<std::uint8_t, 16, Arch> maskz_vpermb(
74 predicate<16, Arch> mask,
75 simd<std::uint8_t, 16, Arch> indices,
76 simd<std::uint8_t, 16, Arch> value) noexcept {
77 if consteval {
78 return detail::x86_vbmi_constant::permute<false>(
79 indices, value, simd<std::uint8_t, 16, Arch>{}, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
80 } else {
81 return simd<std::uint8_t, 16, Arch>::from_native(
82 detail::x86_vbmi::maskz_vpermb<Arch>(
83 mask.to_bitset(), indices.to_native(), value.to_native()));
84 }
85 }
86
88 template<isa<x86> Arch>
89 requires(Arch.has(x86_feature::avx512f) &&
90 Arch.has(x86_feature::avx512bw) &&
91 Arch.has(x86_feature::avx512vbmi) &&
92 Arch.has(x86_feature::avx512vl))
93 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
94 constexpr simd<std::uint8_t, 16, Arch> vpermt2b(
95 simd<std::uint8_t, 16, Arch> a,
96 simd<std::uint8_t, 16, Arch> indices,
97 simd<std::uint8_t, 16, Arch> b) noexcept {
98 if consteval {
99 return detail::x86_vbmi_constant::permute<true>(
100 indices, a, b, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
101 } else {
102 return simd<std::uint8_t, 16, Arch>::from_native(
103 detail::x86_vbmi::vpermt2b<Arch>(
104 a.to_native(), indices.to_native(), b.to_native()));
105 }
106 }
107
109 template<isa<x86> Arch>
110 requires(Arch.has(x86_feature::avx512f) &&
111 Arch.has(x86_feature::avx512bw) &&
112 Arch.has(x86_feature::avx512vbmi) &&
113 Arch.has(x86_feature::avx512vl))
114 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
115 constexpr simd<std::uint8_t, 16, Arch> mask_vpermt2b(
116 simd<std::uint8_t, 16, Arch> a,
117 predicate<16, Arch> mask,
118 simd<std::uint8_t, 16, Arch> indices,
119 simd<std::uint8_t, 16, Arch> b) noexcept {
120 if consteval {
121 return detail::x86_vbmi_constant::permute<true>(
122 indices, a, b, a, mask.to_bitset());
123 } else {
124 return simd<std::uint8_t, 16, Arch>::from_native(
125 detail::x86_vbmi::mask_vpermt2b<Arch>(
126 a.to_native(), mask.to_bitset(), indices.to_native(), b.to_native()));
127 }
128 }
129
131 template<isa<x86> Arch>
132 requires(Arch.has(x86_feature::avx512f) &&
133 Arch.has(x86_feature::avx512bw) &&
134 Arch.has(x86_feature::avx512vbmi) &&
135 Arch.has(x86_feature::avx512vl))
136 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
137 constexpr simd<std::uint8_t, 16, Arch> maskz_vpermt2b(
138 predicate<16, Arch> mask,
139 simd<std::uint8_t, 16, Arch> a,
140 simd<std::uint8_t, 16, Arch> indices,
141 simd<std::uint8_t, 16, Arch> b) noexcept {
142 if consteval {
143 return detail::x86_vbmi_constant::permute<true>(
144 indices, a, b, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
145 } else {
146 return simd<std::uint8_t, 16, Arch>::from_native(
147 detail::x86_vbmi::maskz_vpermt2b<Arch>(
148 mask.to_bitset(), a.to_native(), indices.to_native(), b.to_native()));
149 }
150 }
151
153 template<isa<x86> Arch>
154 requires(Arch.has(x86_feature::avx512f) &&
155 Arch.has(x86_feature::avx512bw) &&
156 Arch.has(x86_feature::avx512vbmi) &&
157 Arch.has(x86_feature::avx512vl))
158 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
159 constexpr simd<std::uint8_t, 16, Arch> vpermi2b(
160 simd<std::uint8_t, 16, Arch> indices,
161 simd<std::uint8_t, 16, Arch> a,
162 simd<std::uint8_t, 16, Arch> b) noexcept {
163 if consteval {
164 return detail::x86_vbmi_constant::permute<true>(
165 indices, a, b, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
166 } else {
167 return simd<std::uint8_t, 16, Arch>::from_native(
168 detail::x86_vbmi::vpermi2b<Arch>(
169 indices.to_native(), a.to_native(), b.to_native()));
170 }
171 }
172
174 template<isa<x86> Arch>
175 requires(Arch.has(x86_feature::avx512f) &&
176 Arch.has(x86_feature::avx512bw) &&
177 Arch.has(x86_feature::avx512vbmi) &&
178 Arch.has(x86_feature::avx512vl))
179 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
180 constexpr simd<std::uint8_t, 16, Arch> mask_vpermi2b(
181 simd<std::uint8_t, 16, Arch> indices,
182 predicate<16, Arch> mask,
183 simd<std::uint8_t, 16, Arch> a,
184 simd<std::uint8_t, 16, Arch> b) noexcept {
185 if consteval {
186 return detail::x86_vbmi_constant::permute<true>(
187 indices, a, b, indices, mask.to_bitset());
188 } else {
189 return simd<std::uint8_t, 16, Arch>::from_native(
190 detail::x86_vbmi::mask_vpermi2b<Arch>(
191 indices.to_native(), mask.to_bitset(), a.to_native(), b.to_native()));
192 }
193 }
194
196 template<isa<x86> Arch>
197 requires(Arch.has(x86_feature::avx512f) &&
198 Arch.has(x86_feature::avx512bw) &&
199 Arch.has(x86_feature::avx512vbmi) &&
200 Arch.has(x86_feature::avx512vl))
201 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
202 constexpr simd<std::uint8_t, 16, Arch> maskz_vpermi2b(
203 predicate<16, Arch> mask,
204 simd<std::uint8_t, 16, Arch> indices,
205 simd<std::uint8_t, 16, Arch> a,
206 simd<std::uint8_t, 16, Arch> b) noexcept {
207 if consteval {
208 return detail::x86_vbmi_constant::permute<true>(
209 indices, a, b, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
210 } else {
211 return simd<std::uint8_t, 16, Arch>::from_native(
212 detail::x86_vbmi::maskz_vpermi2b<Arch>(
213 mask.to_bitset(), indices.to_native(), a.to_native(), b.to_native()));
214 }
215 }
216
218 template<isa<x86> Arch>
219 requires(Arch.has(x86_feature::avx512f) &&
220 Arch.has(x86_feature::avx512bw) &&
221 Arch.has(x86_feature::avx512vbmi) &&
222 Arch.has(x86_feature::avx512vl))
223 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
224 constexpr simd<std::uint8_t, 16, Arch> vpmultishiftqb(
225 simd<std::uint8_t, 16, Arch> control,
226 simd<std::uint64_t, 2, Arch> value) noexcept {
227 if consteval {
228 return detail::x86_vbmi_constant::multishift(
229 control, value, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
230 } else {
231 return simd<std::uint8_t, 16, Arch>::from_native(
232 detail::x86_vbmi::vpmultishiftqb<Arch>(
233 control.to_native(), value.to_native()));
234 }
235 }
236
238 template<isa<x86> Arch>
239 requires(Arch.has(x86_feature::avx512f) &&
240 Arch.has(x86_feature::avx512bw) &&
241 Arch.has(x86_feature::avx512vbmi) &&
242 Arch.has(x86_feature::avx512vl))
243 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
244 constexpr simd<std::uint8_t, 16, Arch> mask_vpmultishiftqb(
245 simd<std::uint8_t, 16, Arch> source,
246 predicate<16, Arch> mask,
247 simd<std::uint8_t, 16, Arch> control,
248 simd<std::uint64_t, 2, Arch> value) noexcept {
249 if consteval {
250 return detail::x86_vbmi_constant::multishift(
251 control, value, source, mask.to_bitset());
252 } else {
253 return simd<std::uint8_t, 16, Arch>::from_native(
254 detail::x86_vbmi::mask_vpmultishiftqb<Arch>(
255 source.to_native(), mask.to_bitset(), control.to_native(), value.to_native()));
256 }
257 }
258
260 template<isa<x86> Arch>
261 requires(Arch.has(x86_feature::avx512f) &&
262 Arch.has(x86_feature::avx512bw) &&
263 Arch.has(x86_feature::avx512vbmi) &&
264 Arch.has(x86_feature::avx512vl))
265 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
266 constexpr simd<std::uint8_t, 16, Arch> maskz_vpmultishiftqb(
267 predicate<16, Arch> mask,
268 simd<std::uint8_t, 16, Arch> control,
269 simd<std::uint64_t, 2, Arch> value) noexcept {
270 if consteval {
271 return detail::x86_vbmi_constant::multishift(
272 control, value, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
273 } else {
274 return simd<std::uint8_t, 16, Arch>::from_native(
275 detail::x86_vbmi::maskz_vpmultishiftqb<Arch>(
276 mask.to_bitset(), control.to_native(), value.to_native()));
277 }
278 }
279
281 template<isa<x86> Arch>
282 requires(Arch.has(x86_feature::avx512f) &&
283 Arch.has(x86_feature::avx512bw) &&
284 Arch.has(x86_feature::avx512vbmi) &&
285 Arch.has(x86_feature::avx512vl))
286 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
287 constexpr simd<std::uint8_t, 32, Arch> vpermb(
288 simd<std::uint8_t, 32, Arch> indices,
289 simd<std::uint8_t, 32, Arch> value) noexcept {
290 if consteval {
291 return detail::x86_vbmi_constant::permute<false>(
292 indices, value, simd<std::uint8_t, 32, Arch>{}, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
293 } else {
294 return simd<std::uint8_t, 32, Arch>::from_native(
295 detail::x86_vbmi::vpermb<Arch>(
296 indices.to_native(), value.to_native()));
297 }
298 }
299
301 template<isa<x86> Arch>
302 requires(Arch.has(x86_feature::avx512f) &&
303 Arch.has(x86_feature::avx512bw) &&
304 Arch.has(x86_feature::avx512vbmi) &&
305 Arch.has(x86_feature::avx512vl))
306 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
307 constexpr simd<std::uint8_t, 32, Arch> mask_vpermb(
308 simd<std::uint8_t, 32, Arch> source,
309 predicate<32, Arch> mask,
310 simd<std::uint8_t, 32, Arch> indices,
311 simd<std::uint8_t, 32, Arch> value) noexcept {
312 if consteval {
313 return detail::x86_vbmi_constant::permute<false>(
314 indices, value, simd<std::uint8_t, 32, Arch>{}, source, mask.to_bitset());
315 } else {
316 return simd<std::uint8_t, 32, Arch>::from_native(
317 detail::x86_vbmi::mask_vpermb<Arch>(
318 source.to_native(), mask.to_bitset(), indices.to_native(), value.to_native()));
319 }
320 }
321
323 template<isa<x86> Arch>
324 requires(Arch.has(x86_feature::avx512f) &&
325 Arch.has(x86_feature::avx512bw) &&
326 Arch.has(x86_feature::avx512vbmi) &&
327 Arch.has(x86_feature::avx512vl))
328 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
329 constexpr simd<std::uint8_t, 32, Arch> maskz_vpermb(
330 predicate<32, Arch> mask,
331 simd<std::uint8_t, 32, Arch> indices,
332 simd<std::uint8_t, 32, Arch> value) noexcept {
333 if consteval {
334 return detail::x86_vbmi_constant::permute<false>(
335 indices, value, simd<std::uint8_t, 32, Arch>{}, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
336 } else {
337 return simd<std::uint8_t, 32, Arch>::from_native(
338 detail::x86_vbmi::maskz_vpermb<Arch>(
339 mask.to_bitset(), indices.to_native(), value.to_native()));
340 }
341 }
342
344 template<isa<x86> Arch>
345 requires(Arch.has(x86_feature::avx512f) &&
346 Arch.has(x86_feature::avx512bw) &&
347 Arch.has(x86_feature::avx512vbmi) &&
348 Arch.has(x86_feature::avx512vl))
349 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
350 constexpr simd<std::uint8_t, 32, Arch> vpermt2b(
351 simd<std::uint8_t, 32, Arch> a,
352 simd<std::uint8_t, 32, Arch> indices,
353 simd<std::uint8_t, 32, Arch> b) noexcept {
354 if consteval {
355 return detail::x86_vbmi_constant::permute<true>(
356 indices, a, b, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
357 } else {
358 return simd<std::uint8_t, 32, Arch>::from_native(
359 detail::x86_vbmi::vpermt2b<Arch>(
360 a.to_native(), indices.to_native(), b.to_native()));
361 }
362 }
363
365 template<isa<x86> Arch>
366 requires(Arch.has(x86_feature::avx512f) &&
367 Arch.has(x86_feature::avx512bw) &&
368 Arch.has(x86_feature::avx512vbmi) &&
369 Arch.has(x86_feature::avx512vl))
370 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
371 constexpr simd<std::uint8_t, 32, Arch> mask_vpermt2b(
372 simd<std::uint8_t, 32, Arch> a,
373 predicate<32, Arch> mask,
374 simd<std::uint8_t, 32, Arch> indices,
375 simd<std::uint8_t, 32, Arch> b) noexcept {
376 if consteval {
377 return detail::x86_vbmi_constant::permute<true>(
378 indices, a, b, a, mask.to_bitset());
379 } else {
380 return simd<std::uint8_t, 32, Arch>::from_native(
381 detail::x86_vbmi::mask_vpermt2b<Arch>(
382 a.to_native(), mask.to_bitset(), indices.to_native(), b.to_native()));
383 }
384 }
385
387 template<isa<x86> Arch>
388 requires(Arch.has(x86_feature::avx512f) &&
389 Arch.has(x86_feature::avx512bw) &&
390 Arch.has(x86_feature::avx512vbmi) &&
391 Arch.has(x86_feature::avx512vl))
392 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
393 constexpr simd<std::uint8_t, 32, Arch> maskz_vpermt2b(
394 predicate<32, Arch> mask,
395 simd<std::uint8_t, 32, Arch> a,
396 simd<std::uint8_t, 32, Arch> indices,
397 simd<std::uint8_t, 32, Arch> b) noexcept {
398 if consteval {
399 return detail::x86_vbmi_constant::permute<true>(
400 indices, a, b, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
401 } else {
402 return simd<std::uint8_t, 32, Arch>::from_native(
403 detail::x86_vbmi::maskz_vpermt2b<Arch>(
404 mask.to_bitset(), a.to_native(), indices.to_native(), b.to_native()));
405 }
406 }
407
409 template<isa<x86> Arch>
410 requires(Arch.has(x86_feature::avx512f) &&
411 Arch.has(x86_feature::avx512bw) &&
412 Arch.has(x86_feature::avx512vbmi) &&
413 Arch.has(x86_feature::avx512vl))
414 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
415 constexpr simd<std::uint8_t, 32, Arch> vpermi2b(
416 simd<std::uint8_t, 32, Arch> indices,
417 simd<std::uint8_t, 32, Arch> a,
418 simd<std::uint8_t, 32, Arch> b) noexcept {
419 if consteval {
420 return detail::x86_vbmi_constant::permute<true>(
421 indices, a, b, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
422 } else {
423 return simd<std::uint8_t, 32, Arch>::from_native(
424 detail::x86_vbmi::vpermi2b<Arch>(
425 indices.to_native(), a.to_native(), b.to_native()));
426 }
427 }
428
430 template<isa<x86> Arch>
431 requires(Arch.has(x86_feature::avx512f) &&
432 Arch.has(x86_feature::avx512bw) &&
433 Arch.has(x86_feature::avx512vbmi) &&
434 Arch.has(x86_feature::avx512vl))
435 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
436 constexpr simd<std::uint8_t, 32, Arch> mask_vpermi2b(
437 simd<std::uint8_t, 32, Arch> indices,
438 predicate<32, Arch> mask,
439 simd<std::uint8_t, 32, Arch> a,
440 simd<std::uint8_t, 32, Arch> b) noexcept {
441 if consteval {
442 return detail::x86_vbmi_constant::permute<true>(
443 indices, a, b, indices, mask.to_bitset());
444 } else {
445 return simd<std::uint8_t, 32, Arch>::from_native(
446 detail::x86_vbmi::mask_vpermi2b<Arch>(
447 indices.to_native(), mask.to_bitset(), a.to_native(), b.to_native()));
448 }
449 }
450
452 template<isa<x86> Arch>
453 requires(Arch.has(x86_feature::avx512f) &&
454 Arch.has(x86_feature::avx512bw) &&
455 Arch.has(x86_feature::avx512vbmi) &&
456 Arch.has(x86_feature::avx512vl))
457 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
458 constexpr simd<std::uint8_t, 32, Arch> maskz_vpermi2b(
459 predicate<32, Arch> mask,
460 simd<std::uint8_t, 32, Arch> indices,
461 simd<std::uint8_t, 32, Arch> a,
462 simd<std::uint8_t, 32, Arch> b) noexcept {
463 if consteval {
464 return detail::x86_vbmi_constant::permute<true>(
465 indices, a, b, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
466 } else {
467 return simd<std::uint8_t, 32, Arch>::from_native(
468 detail::x86_vbmi::maskz_vpermi2b<Arch>(
469 mask.to_bitset(), indices.to_native(), a.to_native(), b.to_native()));
470 }
471 }
472
474 template<isa<x86> Arch>
475 requires(Arch.has(x86_feature::avx512f) &&
476 Arch.has(x86_feature::avx512bw) &&
477 Arch.has(x86_feature::avx512vbmi) &&
478 Arch.has(x86_feature::avx512vl))
479 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
480 constexpr simd<std::uint8_t, 32, Arch> vpmultishiftqb(
481 simd<std::uint8_t, 32, Arch> control,
482 simd<std::uint64_t, 4, Arch> value) noexcept {
483 if consteval {
484 return detail::x86_vbmi_constant::multishift(
485 control, value, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
486 } else {
487 return simd<std::uint8_t, 32, Arch>::from_native(
488 detail::x86_vbmi::vpmultishiftqb<Arch>(
489 control.to_native(), value.to_native()));
490 }
491 }
492
494 template<isa<x86> Arch>
495 requires(Arch.has(x86_feature::avx512f) &&
496 Arch.has(x86_feature::avx512bw) &&
497 Arch.has(x86_feature::avx512vbmi) &&
498 Arch.has(x86_feature::avx512vl))
499 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
500 constexpr simd<std::uint8_t, 32, Arch> mask_vpmultishiftqb(
501 simd<std::uint8_t, 32, Arch> source,
502 predicate<32, Arch> mask,
503 simd<std::uint8_t, 32, Arch> control,
504 simd<std::uint64_t, 4, Arch> value) noexcept {
505 if consteval {
506 return detail::x86_vbmi_constant::multishift(
507 control, value, source, mask.to_bitset());
508 } else {
509 return simd<std::uint8_t, 32, Arch>::from_native(
510 detail::x86_vbmi::mask_vpmultishiftqb<Arch>(
511 source.to_native(), mask.to_bitset(), control.to_native(), value.to_native()));
512 }
513 }
514
516 template<isa<x86> Arch>
517 requires(Arch.has(x86_feature::avx512f) &&
518 Arch.has(x86_feature::avx512bw) &&
519 Arch.has(x86_feature::avx512vbmi) &&
520 Arch.has(x86_feature::avx512vl))
521 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi,avx512vl")
522 constexpr simd<std::uint8_t, 32, Arch> maskz_vpmultishiftqb(
523 predicate<32, Arch> mask,
524 simd<std::uint8_t, 32, Arch> control,
525 simd<std::uint64_t, 4, Arch> value) noexcept {
526 if consteval {
527 return detail::x86_vbmi_constant::multishift(
528 control, value, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
529 } else {
530 return simd<std::uint8_t, 32, Arch>::from_native(
531 detail::x86_vbmi::maskz_vpmultishiftqb<Arch>(
532 mask.to_bitset(), control.to_native(), value.to_native()));
533 }
534 }
535
537 template<isa<x86> Arch>
538 requires(Arch.has(x86_feature::avx512f) &&
539 Arch.has(x86_feature::avx512bw) &&
540 Arch.has(x86_feature::avx512vbmi))
541 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
542 constexpr simd<std::uint8_t, 64, Arch> vpermb(
543 simd<std::uint8_t, 64, Arch> indices,
544 simd<std::uint8_t, 64, Arch> value) noexcept {
545 if consteval {
546 return detail::x86_vbmi_constant::permute<false>(
547 indices, value, simd<std::uint8_t, 64, Arch>{}, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
548 } else {
549 return simd<std::uint8_t, 64, Arch>::from_native(
550 detail::x86_vbmi::vpermb<Arch>(
551 indices.to_native(), value.to_native()));
552 }
553 }
554
556 template<isa<x86> Arch>
557 requires(Arch.has(x86_feature::avx512f) &&
558 Arch.has(x86_feature::avx512bw) &&
559 Arch.has(x86_feature::avx512vbmi))
560 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
561 constexpr simd<std::uint8_t, 64, Arch> mask_vpermb(
562 simd<std::uint8_t, 64, Arch> source,
563 predicate<64, Arch> mask,
564 simd<std::uint8_t, 64, Arch> indices,
565 simd<std::uint8_t, 64, Arch> value) noexcept {
566 if consteval {
567 return detail::x86_vbmi_constant::permute<false>(
568 indices, value, simd<std::uint8_t, 64, Arch>{}, source, mask.to_bitset());
569 } else {
570 return simd<std::uint8_t, 64, Arch>::from_native(
571 detail::x86_vbmi::mask_vpermb<Arch>(
572 source.to_native(), mask.to_bitset(), indices.to_native(), value.to_native()));
573 }
574 }
575
577 template<isa<x86> Arch>
578 requires(Arch.has(x86_feature::avx512f) &&
579 Arch.has(x86_feature::avx512bw) &&
580 Arch.has(x86_feature::avx512vbmi))
581 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
582 constexpr simd<std::uint8_t, 64, Arch> maskz_vpermb(
583 predicate<64, Arch> mask,
584 simd<std::uint8_t, 64, Arch> indices,
585 simd<std::uint8_t, 64, Arch> value) noexcept {
586 if consteval {
587 return detail::x86_vbmi_constant::permute<false>(
588 indices, value, simd<std::uint8_t, 64, Arch>{}, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
589 } else {
590 return simd<std::uint8_t, 64, Arch>::from_native(
591 detail::x86_vbmi::maskz_vpermb<Arch>(
592 mask.to_bitset(), indices.to_native(), value.to_native()));
593 }
594 }
595
597 template<isa<x86> Arch>
598 requires(Arch.has(x86_feature::avx512f) &&
599 Arch.has(x86_feature::avx512bw) &&
600 Arch.has(x86_feature::avx512vbmi))
601 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
602 constexpr simd<std::uint8_t, 64, Arch> vpermt2b(
603 simd<std::uint8_t, 64, Arch> a,
604 simd<std::uint8_t, 64, Arch> indices,
605 simd<std::uint8_t, 64, Arch> b) noexcept {
606 if consteval {
607 return detail::x86_vbmi_constant::permute<true>(
608 indices, a, b, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
609 } else {
610 return simd<std::uint8_t, 64, Arch>::from_native(
611 detail::x86_vbmi::vpermt2b<Arch>(
612 a.to_native(), indices.to_native(), b.to_native()));
613 }
614 }
615
617 template<isa<x86> Arch>
618 requires(Arch.has(x86_feature::avx512f) &&
619 Arch.has(x86_feature::avx512bw) &&
620 Arch.has(x86_feature::avx512vbmi))
621 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
622 constexpr simd<std::uint8_t, 64, Arch> mask_vpermt2b(
623 simd<std::uint8_t, 64, Arch> a,
624 predicate<64, Arch> mask,
625 simd<std::uint8_t, 64, Arch> indices,
626 simd<std::uint8_t, 64, Arch> b) noexcept {
627 if consteval {
628 return detail::x86_vbmi_constant::permute<true>(
629 indices, a, b, a, mask.to_bitset());
630 } else {
631 return simd<std::uint8_t, 64, Arch>::from_native(
632 detail::x86_vbmi::mask_vpermt2b<Arch>(
633 a.to_native(), mask.to_bitset(), indices.to_native(), b.to_native()));
634 }
635 }
636
638 template<isa<x86> Arch>
639 requires(Arch.has(x86_feature::avx512f) &&
640 Arch.has(x86_feature::avx512bw) &&
641 Arch.has(x86_feature::avx512vbmi))
642 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
643 constexpr simd<std::uint8_t, 64, Arch> maskz_vpermt2b(
644 predicate<64, Arch> mask,
645 simd<std::uint8_t, 64, Arch> a,
646 simd<std::uint8_t, 64, Arch> indices,
647 simd<std::uint8_t, 64, Arch> b) noexcept {
648 if consteval {
649 return detail::x86_vbmi_constant::permute<true>(
650 indices, a, b, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
651 } else {
652 return simd<std::uint8_t, 64, Arch>::from_native(
653 detail::x86_vbmi::maskz_vpermt2b<Arch>(
654 mask.to_bitset(), a.to_native(), indices.to_native(), b.to_native()));
655 }
656 }
657
659 template<isa<x86> Arch>
660 requires(Arch.has(x86_feature::avx512f) &&
661 Arch.has(x86_feature::avx512bw) &&
662 Arch.has(x86_feature::avx512vbmi))
663 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
664 constexpr simd<std::uint8_t, 64, Arch> vpermi2b(
665 simd<std::uint8_t, 64, Arch> indices,
666 simd<std::uint8_t, 64, Arch> a,
667 simd<std::uint8_t, 64, Arch> b) noexcept {
668 if consteval {
669 return detail::x86_vbmi_constant::permute<true>(
670 indices, a, b, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
671 } else {
672 return simd<std::uint8_t, 64, Arch>::from_native(
673 detail::x86_vbmi::vpermi2b<Arch>(
674 indices.to_native(), a.to_native(), b.to_native()));
675 }
676 }
677
679 template<isa<x86> Arch>
680 requires(Arch.has(x86_feature::avx512f) &&
681 Arch.has(x86_feature::avx512bw) &&
682 Arch.has(x86_feature::avx512vbmi))
683 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
684 constexpr simd<std::uint8_t, 64, Arch> mask_vpermi2b(
685 simd<std::uint8_t, 64, Arch> indices,
686 predicate<64, Arch> mask,
687 simd<std::uint8_t, 64, Arch> a,
688 simd<std::uint8_t, 64, Arch> b) noexcept {
689 if consteval {
690 return detail::x86_vbmi_constant::permute<true>(
691 indices, a, b, indices, mask.to_bitset());
692 } else {
693 return simd<std::uint8_t, 64, Arch>::from_native(
694 detail::x86_vbmi::mask_vpermi2b<Arch>(
695 indices.to_native(), mask.to_bitset(), a.to_native(), b.to_native()));
696 }
697 }
698
700 template<isa<x86> Arch>
701 requires(Arch.has(x86_feature::avx512f) &&
702 Arch.has(x86_feature::avx512bw) &&
703 Arch.has(x86_feature::avx512vbmi))
704 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
705 constexpr simd<std::uint8_t, 64, Arch> maskz_vpermi2b(
706 predicate<64, Arch> mask,
707 simd<std::uint8_t, 64, Arch> indices,
708 simd<std::uint8_t, 64, Arch> a,
709 simd<std::uint8_t, 64, Arch> b) noexcept {
710 if consteval {
711 return detail::x86_vbmi_constant::permute<true>(
712 indices, a, b, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
713 } else {
714 return simd<std::uint8_t, 64, Arch>::from_native(
715 detail::x86_vbmi::maskz_vpermi2b<Arch>(
716 mask.to_bitset(), indices.to_native(), a.to_native(), b.to_native()));
717 }
718 }
719
721 template<isa<x86> Arch>
722 requires(Arch.has(x86_feature::avx512f) &&
723 Arch.has(x86_feature::avx512bw) &&
724 Arch.has(x86_feature::avx512vbmi))
725 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
726 constexpr simd<std::uint8_t, 64, Arch> vpmultishiftqb(
727 simd<std::uint8_t, 64, Arch> control,
728 simd<std::uint64_t, 8, Arch> value) noexcept {
729 if consteval {
730 return detail::x86_vbmi_constant::multishift(
731 control, value, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
732 } else {
733 return simd<std::uint8_t, 64, Arch>::from_native(
734 detail::x86_vbmi::vpmultishiftqb<Arch>(
735 control.to_native(), value.to_native()));
736 }
737 }
738
740 template<isa<x86> Arch>
741 requires(Arch.has(x86_feature::avx512f) &&
742 Arch.has(x86_feature::avx512bw) &&
743 Arch.has(x86_feature::avx512vbmi))
744 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
745 constexpr simd<std::uint8_t, 64, Arch> mask_vpmultishiftqb(
746 simd<std::uint8_t, 64, Arch> source,
747 predicate<64, Arch> mask,
748 simd<std::uint8_t, 64, Arch> control,
749 simd<std::uint64_t, 8, Arch> value) noexcept {
750 if consteval {
751 return detail::x86_vbmi_constant::multishift(
752 control, value, source, mask.to_bitset());
753 } else {
754 return simd<std::uint8_t, 64, Arch>::from_native(
755 detail::x86_vbmi::mask_vpmultishiftqb<Arch>(
756 source.to_native(), mask.to_bitset(), control.to_native(), value.to_native()));
757 }
758 }
759
761 template<isa<x86> Arch>
762 requires(Arch.has(x86_feature::avx512f) &&
763 Arch.has(x86_feature::avx512bw) &&
764 Arch.has(x86_feature::avx512vbmi))
765 native_nodiscard native_inline native_const native_target("avx512f,avx512bw,avx512vbmi")
766 constexpr simd<std::uint8_t, 64, Arch> maskz_vpmultishiftqb(
767 predicate<64, Arch> mask,
768 simd<std::uint8_t, 64, Arch> control,
769 simd<std::uint64_t, 8, Arch> value) noexcept {
770 if consteval {
771 return detail::x86_vbmi_constant::multishift(
772 control, value, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
773 } else {
774 return simd<std::uint8_t, 64, Arch>::from_native(
775 detail::x86_vbmi::maskz_vpmultishiftqb<Arch>(
776 mask.to_bitset(), control.to_native(), value.to_native()));
777 }
778 }
779
781 template<isa<x86> Arch>
782 requires(!(Arch.has(x86_feature::avx512f) &&
783 Arch.has(x86_feature::avx512bw) &&
784 Arch.has(x86_feature::avx512vbmi) &&
785 Arch.has(x86_feature::avx512vl)) &&
786 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
787 native_nodiscard consteval simd<std::uint8_t, 16, Arch> vpermb(
788 simd<std::uint8_t, 16, Arch> indices,
789 simd<std::uint8_t, 16, Arch> value) noexcept {
790 return detail::x86_vbmi_constant::permute<false>(
791 indices, value, simd<std::uint8_t, 16, Arch>{}, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
792 }
793
795 template<isa<x86> Arch>
796 requires(!(Arch.has(x86_feature::avx512f) &&
797 Arch.has(x86_feature::avx512bw) &&
798 Arch.has(x86_feature::avx512vbmi) &&
799 Arch.has(x86_feature::avx512vl)) &&
800 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
801 native_nodiscard consteval simd<std::uint8_t, 16, Arch> mask_vpermb(
802 simd<std::uint8_t, 16, Arch> source,
803 predicate<16, Arch> mask,
804 simd<std::uint8_t, 16, Arch> indices,
805 simd<std::uint8_t, 16, Arch> value) noexcept {
806 return detail::x86_vbmi_constant::permute<false>(
807 indices, value, simd<std::uint8_t, 16, Arch>{}, source, mask.to_bitset());
808 }
809
811 template<isa<x86> Arch>
812 requires(!(Arch.has(x86_feature::avx512f) &&
813 Arch.has(x86_feature::avx512bw) &&
814 Arch.has(x86_feature::avx512vbmi) &&
815 Arch.has(x86_feature::avx512vl)) &&
816 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
817 native_nodiscard consteval simd<std::uint8_t, 16, Arch> maskz_vpermb(
818 predicate<16, Arch> mask,
819 simd<std::uint8_t, 16, Arch> indices,
820 simd<std::uint8_t, 16, Arch> value) noexcept {
821 return detail::x86_vbmi_constant::permute<false>(
822 indices, value, simd<std::uint8_t, 16, Arch>{}, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
823 }
824
826 template<isa<x86> Arch>
827 requires(!(Arch.has(x86_feature::avx512f) &&
828 Arch.has(x86_feature::avx512bw) &&
829 Arch.has(x86_feature::avx512vbmi) &&
830 Arch.has(x86_feature::avx512vl)) &&
831 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
832 native_nodiscard consteval simd<std::uint8_t, 16, Arch> vpermt2b(
833 simd<std::uint8_t, 16, Arch> a,
834 simd<std::uint8_t, 16, Arch> indices,
835 simd<std::uint8_t, 16, Arch> b) noexcept {
836 return detail::x86_vbmi_constant::permute<true>(
837 indices, a, b, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
838 }
839
841 template<isa<x86> Arch>
842 requires(!(Arch.has(x86_feature::avx512f) &&
843 Arch.has(x86_feature::avx512bw) &&
844 Arch.has(x86_feature::avx512vbmi) &&
845 Arch.has(x86_feature::avx512vl)) &&
846 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
847 native_nodiscard consteval simd<std::uint8_t, 16, Arch> mask_vpermt2b(
848 simd<std::uint8_t, 16, Arch> a,
849 predicate<16, Arch> mask,
850 simd<std::uint8_t, 16, Arch> indices,
851 simd<std::uint8_t, 16, Arch> b) noexcept {
852 return detail::x86_vbmi_constant::permute<true>(
853 indices, a, b, a, mask.to_bitset());
854 }
855
857 template<isa<x86> Arch>
858 requires(!(Arch.has(x86_feature::avx512f) &&
859 Arch.has(x86_feature::avx512bw) &&
860 Arch.has(x86_feature::avx512vbmi) &&
861 Arch.has(x86_feature::avx512vl)) &&
862 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
863 native_nodiscard consteval simd<std::uint8_t, 16, Arch> maskz_vpermt2b(
864 predicate<16, Arch> mask,
865 simd<std::uint8_t, 16, Arch> a,
866 simd<std::uint8_t, 16, Arch> indices,
867 simd<std::uint8_t, 16, Arch> b) noexcept {
868 return detail::x86_vbmi_constant::permute<true>(
869 indices, a, b, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
870 }
871
873 template<isa<x86> Arch>
874 requires(!(Arch.has(x86_feature::avx512f) &&
875 Arch.has(x86_feature::avx512bw) &&
876 Arch.has(x86_feature::avx512vbmi) &&
877 Arch.has(x86_feature::avx512vl)) &&
878 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
879 native_nodiscard consteval simd<std::uint8_t, 16, Arch> vpermi2b(
880 simd<std::uint8_t, 16, Arch> indices,
881 simd<std::uint8_t, 16, Arch> a,
882 simd<std::uint8_t, 16, Arch> b) noexcept {
883 return detail::x86_vbmi_constant::permute<true>(
884 indices, a, b, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
885 }
886
888 template<isa<x86> Arch>
889 requires(!(Arch.has(x86_feature::avx512f) &&
890 Arch.has(x86_feature::avx512bw) &&
891 Arch.has(x86_feature::avx512vbmi) &&
892 Arch.has(x86_feature::avx512vl)) &&
893 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
894 native_nodiscard consteval simd<std::uint8_t, 16, Arch> mask_vpermi2b(
895 simd<std::uint8_t, 16, Arch> indices,
896 predicate<16, Arch> mask,
897 simd<std::uint8_t, 16, Arch> a,
898 simd<std::uint8_t, 16, Arch> b) noexcept {
899 return detail::x86_vbmi_constant::permute<true>(
900 indices, a, b, indices, mask.to_bitset());
901 }
902
904 template<isa<x86> Arch>
905 requires(!(Arch.has(x86_feature::avx512f) &&
906 Arch.has(x86_feature::avx512bw) &&
907 Arch.has(x86_feature::avx512vbmi) &&
908 Arch.has(x86_feature::avx512vl)) &&
909 requires { sizeof(simd<std::uint8_t, 16, Arch>); })
910 native_nodiscard consteval simd<std::uint8_t, 16, Arch> maskz_vpermi2b(
911 predicate<16, Arch> mask,
912 simd<std::uint8_t, 16, Arch> indices,
913 simd<std::uint8_t, 16, Arch> a,
914 simd<std::uint8_t, 16, Arch> b) noexcept {
915 return detail::x86_vbmi_constant::permute<true>(
916 indices, a, b, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
917 }
918
920 template<isa<x86> Arch>
921 requires(!(Arch.has(x86_feature::avx512f) &&
922 Arch.has(x86_feature::avx512bw) &&
923 Arch.has(x86_feature::avx512vbmi) &&
924 Arch.has(x86_feature::avx512vl)) &&
925 requires { sizeof(simd<std::uint8_t, 16, Arch>); sizeof(simd<std::uint64_t, 2, Arch>); })
926 native_nodiscard consteval simd<std::uint8_t, 16, Arch> vpmultishiftqb(
927 simd<std::uint8_t, 16, Arch> control,
928 simd<std::uint64_t, 2, Arch> value) noexcept {
929 return detail::x86_vbmi_constant::multishift(
930 control, value, simd<std::uint8_t, 16, Arch>{}, ~std::uint64_t{0});
931 }
932
934 template<isa<x86> Arch>
935 requires(!(Arch.has(x86_feature::avx512f) &&
936 Arch.has(x86_feature::avx512bw) &&
937 Arch.has(x86_feature::avx512vbmi) &&
938 Arch.has(x86_feature::avx512vl)) &&
939 requires { sizeof(simd<std::uint8_t, 16, Arch>); sizeof(simd<std::uint64_t, 2, Arch>); })
940 native_nodiscard consteval simd<std::uint8_t, 16, Arch> mask_vpmultishiftqb(
941 simd<std::uint8_t, 16, Arch> source,
942 predicate<16, Arch> mask,
943 simd<std::uint8_t, 16, Arch> control,
944 simd<std::uint64_t, 2, Arch> value) noexcept {
945 return detail::x86_vbmi_constant::multishift(
946 control, value, source, mask.to_bitset());
947 }
948
950 template<isa<x86> Arch>
951 requires(!(Arch.has(x86_feature::avx512f) &&
952 Arch.has(x86_feature::avx512bw) &&
953 Arch.has(x86_feature::avx512vbmi) &&
954 Arch.has(x86_feature::avx512vl)) &&
955 requires { sizeof(simd<std::uint8_t, 16, Arch>); sizeof(simd<std::uint64_t, 2, Arch>); })
956 native_nodiscard consteval simd<std::uint8_t, 16, Arch> maskz_vpmultishiftqb(
957 predicate<16, Arch> mask,
958 simd<std::uint8_t, 16, Arch> control,
959 simd<std::uint64_t, 2, Arch> value) noexcept {
960 return detail::x86_vbmi_constant::multishift(
961 control, value, simd<std::uint8_t, 16, Arch>{}, mask.to_bitset());
962 }
963
965 template<isa<x86> Arch>
966 requires(!(Arch.has(x86_feature::avx512f) &&
967 Arch.has(x86_feature::avx512bw) &&
968 Arch.has(x86_feature::avx512vbmi) &&
969 Arch.has(x86_feature::avx512vl)) &&
970 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
971 native_nodiscard consteval simd<std::uint8_t, 32, Arch> vpermb(
972 simd<std::uint8_t, 32, Arch> indices,
973 simd<std::uint8_t, 32, Arch> value) noexcept {
974 return detail::x86_vbmi_constant::permute<false>(
975 indices, value, simd<std::uint8_t, 32, Arch>{}, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
976 }
977
979 template<isa<x86> Arch>
980 requires(!(Arch.has(x86_feature::avx512f) &&
981 Arch.has(x86_feature::avx512bw) &&
982 Arch.has(x86_feature::avx512vbmi) &&
983 Arch.has(x86_feature::avx512vl)) &&
984 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
985 native_nodiscard consteval simd<std::uint8_t, 32, Arch> mask_vpermb(
986 simd<std::uint8_t, 32, Arch> source,
987 predicate<32, Arch> mask,
988 simd<std::uint8_t, 32, Arch> indices,
989 simd<std::uint8_t, 32, Arch> value) noexcept {
990 return detail::x86_vbmi_constant::permute<false>(
991 indices, value, simd<std::uint8_t, 32, Arch>{}, source, mask.to_bitset());
992 }
993
995 template<isa<x86> Arch>
996 requires(!(Arch.has(x86_feature::avx512f) &&
997 Arch.has(x86_feature::avx512bw) &&
998 Arch.has(x86_feature::avx512vbmi) &&
999 Arch.has(x86_feature::avx512vl)) &&
1000 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1001 native_nodiscard consteval simd<std::uint8_t, 32, Arch> maskz_vpermb(
1002 predicate<32, Arch> mask,
1003 simd<std::uint8_t, 32, Arch> indices,
1004 simd<std::uint8_t, 32, Arch> value) noexcept {
1005 return detail::x86_vbmi_constant::permute<false>(
1006 indices, value, simd<std::uint8_t, 32, Arch>{}, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
1007 }
1008
1010 template<isa<x86> Arch>
1011 requires(!(Arch.has(x86_feature::avx512f) &&
1012 Arch.has(x86_feature::avx512bw) &&
1013 Arch.has(x86_feature::avx512vbmi) &&
1014 Arch.has(x86_feature::avx512vl)) &&
1015 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1016 native_nodiscard consteval simd<std::uint8_t, 32, Arch> vpermt2b(
1017 simd<std::uint8_t, 32, Arch> a,
1018 simd<std::uint8_t, 32, Arch> indices,
1019 simd<std::uint8_t, 32, Arch> b) noexcept {
1020 return detail::x86_vbmi_constant::permute<true>(
1021 indices, a, b, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
1022 }
1023
1025 template<isa<x86> Arch>
1026 requires(!(Arch.has(x86_feature::avx512f) &&
1027 Arch.has(x86_feature::avx512bw) &&
1028 Arch.has(x86_feature::avx512vbmi) &&
1029 Arch.has(x86_feature::avx512vl)) &&
1030 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1031 native_nodiscard consteval simd<std::uint8_t, 32, Arch> mask_vpermt2b(
1032 simd<std::uint8_t, 32, Arch> a,
1033 predicate<32, Arch> mask,
1034 simd<std::uint8_t, 32, Arch> indices,
1035 simd<std::uint8_t, 32, Arch> b) noexcept {
1036 return detail::x86_vbmi_constant::permute<true>(
1037 indices, a, b, a, mask.to_bitset());
1038 }
1039
1041 template<isa<x86> Arch>
1042 requires(!(Arch.has(x86_feature::avx512f) &&
1043 Arch.has(x86_feature::avx512bw) &&
1044 Arch.has(x86_feature::avx512vbmi) &&
1045 Arch.has(x86_feature::avx512vl)) &&
1046 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1047 native_nodiscard consteval simd<std::uint8_t, 32, Arch> maskz_vpermt2b(
1048 predicate<32, Arch> mask,
1049 simd<std::uint8_t, 32, Arch> a,
1050 simd<std::uint8_t, 32, Arch> indices,
1051 simd<std::uint8_t, 32, Arch> b) noexcept {
1052 return detail::x86_vbmi_constant::permute<true>(
1053 indices, a, b, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
1054 }
1055
1057 template<isa<x86> Arch>
1058 requires(!(Arch.has(x86_feature::avx512f) &&
1059 Arch.has(x86_feature::avx512bw) &&
1060 Arch.has(x86_feature::avx512vbmi) &&
1061 Arch.has(x86_feature::avx512vl)) &&
1062 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1063 native_nodiscard consteval simd<std::uint8_t, 32, Arch> vpermi2b(
1064 simd<std::uint8_t, 32, Arch> indices,
1065 simd<std::uint8_t, 32, Arch> a,
1066 simd<std::uint8_t, 32, Arch> b) noexcept {
1067 return detail::x86_vbmi_constant::permute<true>(
1068 indices, a, b, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
1069 }
1070
1072 template<isa<x86> Arch>
1073 requires(!(Arch.has(x86_feature::avx512f) &&
1074 Arch.has(x86_feature::avx512bw) &&
1075 Arch.has(x86_feature::avx512vbmi) &&
1076 Arch.has(x86_feature::avx512vl)) &&
1077 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1078 native_nodiscard consteval simd<std::uint8_t, 32, Arch> mask_vpermi2b(
1079 simd<std::uint8_t, 32, Arch> indices,
1080 predicate<32, Arch> mask,
1081 simd<std::uint8_t, 32, Arch> a,
1082 simd<std::uint8_t, 32, Arch> b) noexcept {
1083 return detail::x86_vbmi_constant::permute<true>(
1084 indices, a, b, indices, mask.to_bitset());
1085 }
1086
1088 template<isa<x86> Arch>
1089 requires(!(Arch.has(x86_feature::avx512f) &&
1090 Arch.has(x86_feature::avx512bw) &&
1091 Arch.has(x86_feature::avx512vbmi) &&
1092 Arch.has(x86_feature::avx512vl)) &&
1093 requires { sizeof(simd<std::uint8_t, 32, Arch>); })
1094 native_nodiscard consteval simd<std::uint8_t, 32, Arch> maskz_vpermi2b(
1095 predicate<32, Arch> mask,
1096 simd<std::uint8_t, 32, Arch> indices,
1097 simd<std::uint8_t, 32, Arch> a,
1098 simd<std::uint8_t, 32, Arch> b) noexcept {
1099 return detail::x86_vbmi_constant::permute<true>(
1100 indices, a, b, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
1101 }
1102
1104 template<isa<x86> Arch>
1105 requires(!(Arch.has(x86_feature::avx512f) &&
1106 Arch.has(x86_feature::avx512bw) &&
1107 Arch.has(x86_feature::avx512vbmi) &&
1108 Arch.has(x86_feature::avx512vl)) &&
1109 requires { sizeof(simd<std::uint8_t, 32, Arch>); sizeof(simd<std::uint64_t, 4, Arch>); })
1110 native_nodiscard consteval simd<std::uint8_t, 32, Arch> vpmultishiftqb(
1111 simd<std::uint8_t, 32, Arch> control,
1112 simd<std::uint64_t, 4, Arch> value) noexcept {
1113 return detail::x86_vbmi_constant::multishift(
1114 control, value, simd<std::uint8_t, 32, Arch>{}, ~std::uint64_t{0});
1115 }
1116
1118 template<isa<x86> Arch>
1119 requires(!(Arch.has(x86_feature::avx512f) &&
1120 Arch.has(x86_feature::avx512bw) &&
1121 Arch.has(x86_feature::avx512vbmi) &&
1122 Arch.has(x86_feature::avx512vl)) &&
1123 requires { sizeof(simd<std::uint8_t, 32, Arch>); sizeof(simd<std::uint64_t, 4, Arch>); })
1124 native_nodiscard consteval simd<std::uint8_t, 32, Arch> mask_vpmultishiftqb(
1125 simd<std::uint8_t, 32, Arch> source,
1126 predicate<32, Arch> mask,
1127 simd<std::uint8_t, 32, Arch> control,
1128 simd<std::uint64_t, 4, Arch> value) noexcept {
1129 return detail::x86_vbmi_constant::multishift(
1130 control, value, source, mask.to_bitset());
1131 }
1132
1134 template<isa<x86> Arch>
1135 requires(!(Arch.has(x86_feature::avx512f) &&
1136 Arch.has(x86_feature::avx512bw) &&
1137 Arch.has(x86_feature::avx512vbmi) &&
1138 Arch.has(x86_feature::avx512vl)) &&
1139 requires { sizeof(simd<std::uint8_t, 32, Arch>); sizeof(simd<std::uint64_t, 4, Arch>); })
1140 native_nodiscard consteval simd<std::uint8_t, 32, Arch> maskz_vpmultishiftqb(
1141 predicate<32, Arch> mask,
1142 simd<std::uint8_t, 32, Arch> control,
1143 simd<std::uint64_t, 4, Arch> value) noexcept {
1144 return detail::x86_vbmi_constant::multishift(
1145 control, value, simd<std::uint8_t, 32, Arch>{}, mask.to_bitset());
1146 }
1147
1149 template<isa<x86> Arch>
1150 requires(!(Arch.has(x86_feature::avx512f) &&
1151 Arch.has(x86_feature::avx512bw) &&
1152 Arch.has(x86_feature::avx512vbmi)) &&
1153 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1154 native_nodiscard consteval simd<std::uint8_t, 64, Arch> vpermb(
1155 simd<std::uint8_t, 64, Arch> indices,
1156 simd<std::uint8_t, 64, Arch> value) noexcept {
1157 return detail::x86_vbmi_constant::permute<false>(
1158 indices, value, simd<std::uint8_t, 64, Arch>{}, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
1159 }
1160
1162 template<isa<x86> Arch>
1163 requires(!(Arch.has(x86_feature::avx512f) &&
1164 Arch.has(x86_feature::avx512bw) &&
1165 Arch.has(x86_feature::avx512vbmi)) &&
1166 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1167 native_nodiscard consteval simd<std::uint8_t, 64, Arch> mask_vpermb(
1168 simd<std::uint8_t, 64, Arch> source,
1169 predicate<64, Arch> mask,
1170 simd<std::uint8_t, 64, Arch> indices,
1171 simd<std::uint8_t, 64, Arch> value) noexcept {
1172 return detail::x86_vbmi_constant::permute<false>(
1173 indices, value, simd<std::uint8_t, 64, Arch>{}, source, mask.to_bitset());
1174 }
1175
1177 template<isa<x86> Arch>
1178 requires(!(Arch.has(x86_feature::avx512f) &&
1179 Arch.has(x86_feature::avx512bw) &&
1180 Arch.has(x86_feature::avx512vbmi)) &&
1181 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1182 native_nodiscard consteval simd<std::uint8_t, 64, Arch> maskz_vpermb(
1183 predicate<64, Arch> mask,
1184 simd<std::uint8_t, 64, Arch> indices,
1185 simd<std::uint8_t, 64, Arch> value) noexcept {
1186 return detail::x86_vbmi_constant::permute<false>(
1187 indices, value, simd<std::uint8_t, 64, Arch>{}, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
1188 }
1189
1191 template<isa<x86> Arch>
1192 requires(!(Arch.has(x86_feature::avx512f) &&
1193 Arch.has(x86_feature::avx512bw) &&
1194 Arch.has(x86_feature::avx512vbmi)) &&
1195 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1196 native_nodiscard consteval simd<std::uint8_t, 64, Arch> vpermt2b(
1197 simd<std::uint8_t, 64, Arch> a,
1198 simd<std::uint8_t, 64, Arch> indices,
1199 simd<std::uint8_t, 64, Arch> b) noexcept {
1200 return detail::x86_vbmi_constant::permute<true>(
1201 indices, a, b, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
1202 }
1203
1205 template<isa<x86> Arch>
1206 requires(!(Arch.has(x86_feature::avx512f) &&
1207 Arch.has(x86_feature::avx512bw) &&
1208 Arch.has(x86_feature::avx512vbmi)) &&
1209 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1210 native_nodiscard consteval simd<std::uint8_t, 64, Arch> mask_vpermt2b(
1211 simd<std::uint8_t, 64, Arch> a,
1212 predicate<64, Arch> mask,
1213 simd<std::uint8_t, 64, Arch> indices,
1214 simd<std::uint8_t, 64, Arch> b) noexcept {
1215 return detail::x86_vbmi_constant::permute<true>(
1216 indices, a, b, a, mask.to_bitset());
1217 }
1218
1220 template<isa<x86> Arch>
1221 requires(!(Arch.has(x86_feature::avx512f) &&
1222 Arch.has(x86_feature::avx512bw) &&
1223 Arch.has(x86_feature::avx512vbmi)) &&
1224 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1225 native_nodiscard consteval simd<std::uint8_t, 64, Arch> maskz_vpermt2b(
1226 predicate<64, Arch> mask,
1227 simd<std::uint8_t, 64, Arch> a,
1228 simd<std::uint8_t, 64, Arch> indices,
1229 simd<std::uint8_t, 64, Arch> b) noexcept {
1230 return detail::x86_vbmi_constant::permute<true>(
1231 indices, a, b, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
1232 }
1233
1235 template<isa<x86> Arch>
1236 requires(!(Arch.has(x86_feature::avx512f) &&
1237 Arch.has(x86_feature::avx512bw) &&
1238 Arch.has(x86_feature::avx512vbmi)) &&
1239 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1240 native_nodiscard consteval simd<std::uint8_t, 64, Arch> vpermi2b(
1241 simd<std::uint8_t, 64, Arch> indices,
1242 simd<std::uint8_t, 64, Arch> a,
1243 simd<std::uint8_t, 64, Arch> b) noexcept {
1244 return detail::x86_vbmi_constant::permute<true>(
1245 indices, a, b, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
1246 }
1247
1249 template<isa<x86> Arch>
1250 requires(!(Arch.has(x86_feature::avx512f) &&
1251 Arch.has(x86_feature::avx512bw) &&
1252 Arch.has(x86_feature::avx512vbmi)) &&
1253 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1254 native_nodiscard consteval simd<std::uint8_t, 64, Arch> mask_vpermi2b(
1255 simd<std::uint8_t, 64, Arch> indices,
1256 predicate<64, Arch> mask,
1257 simd<std::uint8_t, 64, Arch> a,
1258 simd<std::uint8_t, 64, Arch> b) noexcept {
1259 return detail::x86_vbmi_constant::permute<true>(
1260 indices, a, b, indices, mask.to_bitset());
1261 }
1262
1264 template<isa<x86> Arch>
1265 requires(!(Arch.has(x86_feature::avx512f) &&
1266 Arch.has(x86_feature::avx512bw) &&
1267 Arch.has(x86_feature::avx512vbmi)) &&
1268 requires { sizeof(simd<std::uint8_t, 64, Arch>); })
1269 native_nodiscard consteval simd<std::uint8_t, 64, Arch> maskz_vpermi2b(
1270 predicate<64, Arch> mask,
1271 simd<std::uint8_t, 64, Arch> indices,
1272 simd<std::uint8_t, 64, Arch> a,
1273 simd<std::uint8_t, 64, Arch> b) noexcept {
1274 return detail::x86_vbmi_constant::permute<true>(
1275 indices, a, b, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
1276 }
1277
1279 template<isa<x86> Arch>
1280 requires(!(Arch.has(x86_feature::avx512f) &&
1281 Arch.has(x86_feature::avx512bw) &&
1282 Arch.has(x86_feature::avx512vbmi)) &&
1283 requires { sizeof(simd<std::uint8_t, 64, Arch>); sizeof(simd<std::uint64_t, 8, Arch>); })
1284 native_nodiscard consteval simd<std::uint8_t, 64, Arch> vpmultishiftqb(
1285 simd<std::uint8_t, 64, Arch> control,
1286 simd<std::uint64_t, 8, Arch> value) noexcept {
1287 return detail::x86_vbmi_constant::multishift(
1288 control, value, simd<std::uint8_t, 64, Arch>{}, ~std::uint64_t{0});
1289 }
1290
1292 template<isa<x86> Arch>
1293 requires(!(Arch.has(x86_feature::avx512f) &&
1294 Arch.has(x86_feature::avx512bw) &&
1295 Arch.has(x86_feature::avx512vbmi)) &&
1296 requires { sizeof(simd<std::uint8_t, 64, Arch>); sizeof(simd<std::uint64_t, 8, Arch>); })
1297 native_nodiscard consteval simd<std::uint8_t, 64, Arch> mask_vpmultishiftqb(
1298 simd<std::uint8_t, 64, Arch> source,
1299 predicate<64, Arch> mask,
1300 simd<std::uint8_t, 64, Arch> control,
1301 simd<std::uint64_t, 8, Arch> value) noexcept {
1302 return detail::x86_vbmi_constant::multishift(
1303 control, value, source, mask.to_bitset());
1304 }
1305
1307 template<isa<x86> Arch>
1308 requires(!(Arch.has(x86_feature::avx512f) &&
1309 Arch.has(x86_feature::avx512bw) &&
1310 Arch.has(x86_feature::avx512vbmi)) &&
1311 requires { sizeof(simd<std::uint8_t, 64, Arch>); sizeof(simd<std::uint64_t, 8, Arch>); })
1312 native_nodiscard consteval simd<std::uint8_t, 64, Arch> maskz_vpmultishiftqb(
1313 predicate<64, Arch> mask,
1314 simd<std::uint8_t, 64, Arch> control,
1315 simd<std::uint64_t, 8, Arch> value) noexcept {
1316 return detail::x86_vbmi_constant::multishift(
1317 control, value, simd<std::uint8_t, 64, Arch>{}, mask.to_bitset());
1318 }
1319
1321 template<isa<x86> Arch, class... Args>
1322 void vpermb(Args...) = delete;
1323
1325 template<isa<x86> Arch, class... Args>
1326 void mask_vpermb(Args...) = delete;
1327
1329 template<isa<x86> Arch, class... Args>
1330 void maskz_vpermb(Args...) = delete;
1331
1333 template<isa<x86> Arch, class... Args>
1334 void vpermt2b(Args...) = delete;
1335
1337 template<isa<x86> Arch, class... Args>
1338 void mask_vpermt2b(Args...) = delete;
1339
1341 template<isa<x86> Arch, class... Args>
1342 void maskz_vpermt2b(Args...) = delete;
1343
1345 template<isa<x86> Arch, class... Args>
1346 void vpermi2b(Args...) = delete;
1347
1349 template<isa<x86> Arch, class... Args>
1350 void mask_vpermi2b(Args...) = delete;
1351
1353 template<isa<x86> Arch, class... Args>
1354 void maskz_vpermi2b(Args...) = delete;
1355
1357 template<isa<x86> Arch, class... Args>
1358 void vpmultishiftqb(Args...) = delete;
1359
1361 template<isa<x86> Arch, class... Args>
1362 void mask_vpmultishiftqb(Args...) = delete;
1363
1365 template<isa<x86> Arch, class... Args>
1366 void maskz_vpmultishiftqb(Args...) = delete;
1367
1369}
1370#endif
native_inline
#define native_inline
inline [[always_inline]]
Definition attributes.h:212
native_nodiscard
#define native_nodiscard
C++17 [[nodiscard]].
Definition attributes.h:189
native_const
#define native_const
[[const]] is not const
Definition attributes.h:108
native_target
#define native_target(x)
this indicates a required feature set for the current multiversioned function.
Definition attributes.h:476
native::mask
typename mask_traits< std::remove_cvref_t< T > >::type mask
Definition mask_traits.h:22
native::mask_vpmultishiftqb
constexpr simd< std::uint8_t, 16, Arch > mask_vpmultishiftqb(simd< std::uint8_t, 16, Arch > source, predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > control, simd< std::uint64_t, 2, Arch > value) noexcept
Inactive destination bytes retain source.
Definition native.x86.vbmi.ccm:244
native::vpmultishiftqb
constexpr simd< std::uint8_t, 16, Arch > vpmultishiftqb(simd< std::uint8_t, 16, Arch > control, simd< std::uint64_t, 2, Arch > value) noexcept
Select eight wrapping bits per control byte from its corresponding source qword.
Definition native.x86.vbmi.ccm:224
native::vpermb
constexpr simd< std::uint8_t, 16, Arch > vpermb(simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > value) noexcept
Permute bytes across the whole vector using the low index bits.
Definition native.x86.vbmi.ccm:31
native::mask_vpermi2b
constexpr simd< std::uint8_t, 16, Arch > mask_vpermi2b(simd< std::uint8_t, 16, Arch > indices, predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > a, simd< std::uint8_t, 16, Arch > b) noexcept
Inactive destination bytes retain the index vector.
Definition native.x86.vbmi.ccm:180
native::maskz_vpermt2b
constexpr simd< std::uint8_t, 16, Arch > maskz_vpermt2b(predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > a, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > b) noexcept
Inactive destination bytes become zero.
Definition native.x86.vbmi.ccm:137
native::mask_vpermb
constexpr simd< std::uint8_t, 16, Arch > mask_vpermb(simd< std::uint8_t, 16, Arch > source, predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > value) noexcept
Inactive destination bytes retain source.
Definition native.x86.vbmi.ccm:51
native::vpermi2b
constexpr simd< std::uint8_t, 16, Arch > vpermi2b(simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > a, simd< std::uint8_t, 16, Arch > b) noexcept
Permute bytes from two tables with the index vector as the tied destination.
Definition native.x86.vbmi.ccm:159
native::vpermt2b
constexpr simd< std::uint8_t, 16, Arch > vpermt2b(simd< std::uint8_t, 16, Arch > a, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > b) noexcept
Permute bytes from two tables; the table index spans both complete vectors.
Definition native.x86.vbmi.ccm:94
native::maskz_vpmultishiftqb
constexpr simd< std::uint8_t, 16, Arch > maskz_vpmultishiftqb(predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > control, simd< std::uint64_t, 2, Arch > value) noexcept
Inactive destination bytes become zero.
Definition native.x86.vbmi.ccm:266
native::mask_vpermt2b
constexpr simd< std::uint8_t, 16, Arch > mask_vpermt2b(simd< std::uint8_t, 16, Arch > a, predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > b) noexcept
Inactive destination bytes retain the first table a.
Definition native.x86.vbmi.ccm:115
native::maskz_vpermi2b
constexpr simd< std::uint8_t, 16, Arch > maskz_vpermi2b(predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > a, simd< std::uint8_t, 16, Arch > b) noexcept
Inactive destination bytes become zero.
Definition native.x86.vbmi.ccm:202
native::maskz_vpermb
constexpr simd< std::uint8_t, 16, Arch > maskz_vpermb(predicate< 16, Arch > mask, simd< std::uint8_t, 16, Arch > indices, simd< std::uint8_t, 16, Arch > value) noexcept
Inactive destination bytes become zero.
Definition native.x86.vbmi.ccm:73
native.simd
native.x86.features
native.x86.vbmi
native
Architecture-tagged vectors, register packs and supporting value types. Native arithmetic follows its...
Definition native.arm.aes.ccm:10
std
Standard-library adaptations documented here for SIMD value types.
native::simd
Omitted architecture arguments use the native.simd provider's baseline.
Definition native.simd.ccm:19