clang 20.0.0git
avx10_2convertintrin.h
Go to the documentation of this file.
1/*===--------------- avx10_2convertintrin.h - AVX10_2CONVERT ---------------===
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#ifndef __IMMINTRIN_H
10#error \
11 "Never use <avx10_2convertintrin.h> directly; include <immintrin.h> instead."
12#endif // __IMMINTRIN_H
13
14#ifdef __SSE2__
15
16#ifndef __AVX10_2CONVERTINTRIN_H
17#define __AVX10_2CONVERTINTRIN_H
18
19/* Define the default attributes for the functions in this file. */
20#define __DEFAULT_FN_ATTRS128 \
21 __attribute__((__always_inline__, __nodebug__, __target__("avx10.2-256"), \
22 __min_vector_width__(128)))
23#define __DEFAULT_FN_ATTRS256 \
24 __attribute__((__always_inline__, __nodebug__, __target__("avx10.2-256"), \
25 __min_vector_width__(256)))
26
27static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvtx2ps_ph(__m128 __A,
28 __m128 __B) {
29 return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
30 (__v4sf)__A, (__v4sf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)(-1));
31}
32
33static __inline__ __m128h __DEFAULT_FN_ATTRS128
34_mm_mask_cvtx2ps_ph(__m128h __W, __mmask8 __U, __m128 __A, __m128 __B) {
35 return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
36 (__v4sf)__A, (__v4sf)__B, (__v8hf)__W, (__mmask8)__U);
37}
38
39static __inline__ __m128h __DEFAULT_FN_ATTRS128
40_mm_maskz_cvtx2ps_ph(__mmask8 __U, __m128 __A, __m128 __B) {
41 return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
42 (__v4sf)__A, (__v4sf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U);
43}
44
45static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvtx2ps_ph(__m256 __A,
46 __m256 __B) {
47 return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
48 (__v8sf)__A, (__v8sf)__B, (__v16hf)_mm256_setzero_ph(), (__mmask16)(-1),
49 _MM_FROUND_CUR_DIRECTION);
50}
51
52static __inline__ __m256h __DEFAULT_FN_ATTRS256
53_mm256_mask_cvtx2ps_ph(__m256h __W, __mmask16 __U, __m256 __A, __m256 __B) {
54 return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
55 (__v8sf)__A, (__v8sf)__B, (__v16hf)__W, (__mmask16)__U,
56 _MM_FROUND_CUR_DIRECTION);
57}
58
59static __inline__ __m256h __DEFAULT_FN_ATTRS256
60_mm256_maskz_cvtx2ps_ph(__mmask16 __U, __m256 __A, __m256 __B) {
61 return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
62 (__v8sf)__A, (__v8sf)__B, (__v16hf)_mm256_setzero_ph(), (__mmask16)__U,
63 _MM_FROUND_CUR_DIRECTION);
64}
65
66#define _mm256_cvtx_round2ps_ph(A, B, R) \
67 ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask( \
68 (__v8sf)(A), (__v8sf)(B), (__v16hf)_mm256_undefined_ph(), \
69 (__mmask16)(-1), (const int)(R)))
70
71#define _mm256_mask_cvtx_round2ps_ph(W, U, A, B, R) \
72 ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask( \
73 (__v8sf)(A), (__v8sf)(B), (__v16hf)(W), (__mmask16)(U), (const int)(R)))
74
75#define _mm256_maskz_cvtx_round2ps_ph(U, A, B, R) \
76 ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask( \
77 (__v8sf)(A), (__v8sf)(B), (__v16hf)(_mm256_setzero_ph()), \
78 (__mmask16)(U), (const int)(R)))
79
80static __inline__ __m128i __DEFAULT_FN_ATTRS128
81_mm_cvtbiasph_pbf8(__m128i __A, __m128h __B) {
82 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
83 (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
84}
85
86static __inline__ __m128i __DEFAULT_FN_ATTRS128
87_mm_mask_cvtbiasph_pbf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
88 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
89 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
90}
91
92static __inline__ __m128i __DEFAULT_FN_ATTRS128
93_mm_maskz_cvtbiasph_pbf8(__mmask8 __U, __m128i __A, __m128h __B) {
94 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
95 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
96 (__mmask8)__U);
97}
98
99static __inline__ __m128i __DEFAULT_FN_ATTRS256
100_mm256_cvtbiasph_pbf8(__m256i __A, __m256h __B) {
101 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
102 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_undefined_si128(),
103 (__mmask16)-1);
104}
105
106static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiasph_pbf8(
107 __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
108 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
109 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
110}
111
112static __inline__ __m128i __DEFAULT_FN_ATTRS256
113_mm256_maskz_cvtbiasph_pbf8(__mmask16 __U, __m256i __A, __m256h __B) {
114 return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
115 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
116 (__mmask16)__U);
117}
118
119static __inline__ __m128i __DEFAULT_FN_ATTRS128
120_mm_cvtbiassph_pbf8(__m128i __A, __m128h __B) {
121 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
122 (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
123}
124
125static __inline__ __m128i __DEFAULT_FN_ATTRS128
126_mm_mask_cvtbiassph_pbf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
127 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
128 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
129}
130
131static __inline__ __m128i __DEFAULT_FN_ATTRS128
132_mm_maskz_cvtbiassph_pbf8(__mmask8 __U, __m128i __A, __m128h __B) {
133 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
134 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
135 (__mmask8)__U);
136}
137
138static __inline__ __m128i __DEFAULT_FN_ATTRS256
139_mm256_cvtbiassph_pbf8(__m256i __A, __m256h __B) {
140 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
141 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_undefined_si128(),
142 (__mmask16)-1);
143}
144
145static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiassph_pbf8(
146 __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
147 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
148 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
149}
150
151static __inline__ __m128i __DEFAULT_FN_ATTRS256
152_mm256_maskz_cvtbiassph_pbf8(__mmask16 __U, __m256i __A, __m256h __B) {
153 return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
154 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
155 (__mmask16)__U);
156}
157
158static __inline__ __m128i __DEFAULT_FN_ATTRS128
159_mm_cvtbiasph_phf8(__m128i __A, __m128h __B) {
160 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
161 (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
162}
163
164static __inline__ __m128i __DEFAULT_FN_ATTRS128
165_mm_mask_cvtbiasph_phf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
166 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
167 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
168}
169
170static __inline__ __m128i __DEFAULT_FN_ATTRS128
171_mm_maskz_cvtbiasph_phf8(__mmask8 __U, __m128i __A, __m128h __B) {
172 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
173 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
174 (__mmask8)__U);
175}
176
177static __inline__ __m128i __DEFAULT_FN_ATTRS256
178_mm256_cvtbiasph_phf8(__m256i __A, __m256h __B) {
179 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
180 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_undefined_si128(),
181 (__mmask16)-1);
182}
183
184static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiasph_phf8(
185 __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
186 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
187 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
188}
189
190static __inline__ __m128i __DEFAULT_FN_ATTRS256
191_mm256_maskz_cvtbiasph_phf8(__mmask16 __U, __m256i __A, __m256h __B) {
192 return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
193 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
194 (__mmask16)__U);
195}
196
197static __inline__ __m128i __DEFAULT_FN_ATTRS128
198_mm_cvtbiassph_phf8(__m128i __A, __m128h __B) {
199 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
200 (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
201}
202
203static __inline__ __m128i __DEFAULT_FN_ATTRS128
204_mm_mask_cvtbiassph_phf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
205 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
206 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
207}
208
209static __inline__ __m128i __DEFAULT_FN_ATTRS128
210_mm_maskz_cvtbiassph_phf8(__mmask8 __U, __m128i __A, __m128h __B) {
211 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
212 (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
213 (__mmask8)__U);
214}
215
216static __inline__ __m128i __DEFAULT_FN_ATTRS256
217_mm256_cvtbiassph_phf8(__m256i __A, __m256h __B) {
218 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
219 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_undefined_si128(),
220 (__mmask16)-1);
221}
222
223static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiassph_phf8(
224 __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
225 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
226 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
227}
228
229static __inline__ __m128i __DEFAULT_FN_ATTRS256
230_mm256_maskz_cvtbiassph_phf8(__mmask16 __U, __m256i __A, __m256h __B) {
231 return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
232 (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)_mm_setzero_si128(),
233 (__mmask16)__U);
234}
235
236static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtne2ph_pbf8(__m128h __A,
237 __m128h __B) {
238 return (__m128i)__builtin_ia32_vcvtne2ph2bf8_128((__v8hf)(__A),
239 (__v8hf)(__B));
240}
241
242static __inline__ __m128i __DEFAULT_FN_ATTRS128
243_mm_mask_cvtne2ph_pbf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
244 return (__m128i)__builtin_ia32_selectb_128(
245 (__mmask16)__U, (__v16qi)_mm_cvtne2ph_pbf8(__A, __B), (__v16qi)__W);
246}
247
248static __inline__ __m128i __DEFAULT_FN_ATTRS128
249_mm_maskz_cvtne2ph_pbf8(__mmask16 __U, __m128h __A, __m128h __B) {
250 return (__m128i)__builtin_ia32_selectb_128(
251 (__mmask16)__U, (__v16qi)_mm_cvtne2ph_pbf8(__A, __B),
252 (__v16qi)(__m128i)_mm_setzero_si128());
253}
254
255static __inline__ __m256i __DEFAULT_FN_ATTRS256
256_mm256_cvtne2ph_pbf8(__m256h __A, __m256h __B) {
257 return (__m256i)__builtin_ia32_vcvtne2ph2bf8_256((__v16hf)(__A),
258 (__v16hf)(__B));
259}
260
261static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtne2ph_pbf8(
262 __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
263 return (__m256i)__builtin_ia32_selectb_256(
264 (__mmask16)__U, (__v32qi)_mm256_cvtne2ph_pbf8(__A, __B), (__v32qi)__W);
265}
266
267static __inline__ __m256i __DEFAULT_FN_ATTRS256
268_mm256_maskz_cvtne2ph_pbf8(__mmask32 __U, __m256h __A, __m256h __B) {
269 return (__m256i)__builtin_ia32_selectb_256(
270 (__mmask16)__U, (__v32qi)_mm256_cvtne2ph_pbf8(__A, __B),
271 (__v32qi)(__m256i)_mm256_setzero_si256());
272}
273
274static __inline__ __m128i __DEFAULT_FN_ATTRS128
275_mm_cvtnes2ph_pbf8(__m128h __A, __m128h __B) {
276 return (__m128i)__builtin_ia32_vcvtne2ph2bf8s_128((__v8hf)(__A),
277 (__v8hf)(__B));
278}
279
280static __inline__ __m128i __DEFAULT_FN_ATTRS128
281_mm_mask_cvtnes2ph_pbf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
282 return (__m128i)__builtin_ia32_selectb_128(
283 (__mmask16)__U, (__v16qi)_mm_cvtnes2ph_pbf8(__A, __B), (__v16qi)__W);
284}
285
286static __inline__ __m128i __DEFAULT_FN_ATTRS128
287_mm_maskz_cvtnes2ph_pbf8(__mmask16 __U, __m128h __A, __m128h __B) {
288 return (__m128i)__builtin_ia32_selectb_128(
289 (__mmask16)__U, (__v16qi)_mm_cvtnes2ph_pbf8(__A, __B),
290 (__v16qi)(__m128i)_mm_setzero_si128());
291}
292
293static __inline__ __m256i __DEFAULT_FN_ATTRS256
294_mm256_cvtnes2ph_pbf8(__m256h __A, __m256h __B) {
295 return (__m256i)__builtin_ia32_vcvtne2ph2bf8s_256((__v16hf)(__A),
296 (__v16hf)(__B));
297}
298
299static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtnes2ph_pbf8(
300 __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
301 return (__m256i)__builtin_ia32_selectb_256(
302 (__mmask16)__U, (__v32qi)_mm256_cvtnes2ph_pbf8(__A, __B), (__v32qi)__W);
303}
304
305static __inline__ __m256i __DEFAULT_FN_ATTRS256
306_mm256_maskz_cvtnes2ph_pbf8(__mmask32 __U, __m256h __A, __m256h __B) {
307 return (__m256i)__builtin_ia32_selectb_256(
308 (__mmask16)__U, (__v32qi)_mm256_cvtnes2ph_pbf8(__A, __B),
309 (__v32qi)(__m256i)_mm256_setzero_si256());
310}
311
312static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtne2ph_phf8(__m128h __A,
313 __m128h __B) {
314 return (__m128i)__builtin_ia32_vcvtne2ph2hf8_128((__v8hf)(__A),
315 (__v8hf)(__B));
316}
317
318static __inline__ __m128i __DEFAULT_FN_ATTRS128
319_mm_mask_cvtne2ph_phf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
320 return (__m128i)__builtin_ia32_selectb_128(
321 (__mmask16)__U, (__v16qi)_mm_cvtne2ph_phf8(__A, __B), (__v16qi)__W);
322}
323
324static __inline__ __m128i __DEFAULT_FN_ATTRS128
325_mm_maskz_cvtne2ph_phf8(__mmask16 __U, __m128h __A, __m128h __B) {
326 return (__m128i)__builtin_ia32_selectb_128(
327 (__mmask16)__U, (__v16qi)_mm_cvtne2ph_phf8(__A, __B),
328 (__v16qi)(__m128i)_mm_setzero_si128());
329}
330
331static __inline__ __m256i __DEFAULT_FN_ATTRS256
332_mm256_cvtne2ph_phf8(__m256h __A, __m256h __B) {
333 return (__m256i)__builtin_ia32_vcvtne2ph2hf8_256((__v16hf)(__A),
334 (__v16hf)(__B));
335}
336
337static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtne2ph_phf8(
338 __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
339 return (__m256i)__builtin_ia32_selectb_256(
340 (__mmask16)__U, (__v32qi)_mm256_cvtne2ph_phf8(__A, __B), (__v32qi)__W);
341}
342
343static __inline__ __m256i __DEFAULT_FN_ATTRS256
344_mm256_maskz_cvtne2ph_phf8(__mmask32 __U, __m256h __A, __m256h __B) {
345 return (__m256i)__builtin_ia32_selectb_256(
346 (__mmask16)__U, (__v32qi)_mm256_cvtne2ph_phf8(__A, __B),
347 (__v32qi)(__m256i)_mm256_setzero_si256());
348}
349
350static __inline__ __m128i __DEFAULT_FN_ATTRS128
351_mm_cvtnes2ph_phf8(__m128h __A, __m128h __B) {
352 return (__m128i)__builtin_ia32_vcvtne2ph2hf8s_128((__v8hf)(__A),
353 (__v8hf)(__B));
354}
355
356static __inline__ __m128i __DEFAULT_FN_ATTRS128
357_mm_mask_cvtnes2ph_phf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
358 return (__m128i)__builtin_ia32_selectb_128(
359 (__mmask16)__U, (__v16qi)_mm_cvtnes2ph_phf8(__A, __B), (__v16qi)__W);
360}
361
362static __inline__ __m128i __DEFAULT_FN_ATTRS128
363_mm_maskz_cvtnes2ph_phf8(__mmask16 __U, __m128h __A, __m128h __B) {
364 return (__m128i)__builtin_ia32_selectb_128(
365 (__mmask16)__U, (__v16qi)_mm_cvtnes2ph_phf8(__A, __B),
366 (__v16qi)(__m128i)_mm_setzero_si128());
367}
368
369static __inline__ __m256i __DEFAULT_FN_ATTRS256
370_mm256_cvtnes2ph_phf8(__m256h __A, __m256h __B) {
371 return (__m256i)__builtin_ia32_vcvtne2ph2hf8s_256((__v16hf)(__A),
372 (__v16hf)(__B));
373}
374
375static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtnes2ph_phf8(
376 __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
377 return (__m256i)__builtin_ia32_selectb_256(
378 (__mmask16)__U, (__v32qi)_mm256_cvtnes2ph_phf8(__A, __B), (__v32qi)__W);
379}
380
381static __inline__ __m256i __DEFAULT_FN_ATTRS256
382_mm256_maskz_cvtnes2ph_phf8(__mmask32 __U, __m256h __A, __m256h __B) {
383 return (__m256i)__builtin_ia32_selectb_256(
384 (__mmask16)__U, (__v32qi)_mm256_cvtnes2ph_phf8(__A, __B),
385 (__v32qi)(__m256i)_mm256_setzero_si256());
386}
387
388static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvtnehf8_ph(__m128i __A) {
389 return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
390 (__v16qi)__A, (__v8hf)(__m128h)_mm_undefined_ph(), (__mmask8)-1);
391}
392
393static __inline__ __m128h __DEFAULT_FN_ATTRS128
394_mm_mask_cvtnehf8_ph(__m128h __W, __mmask8 __U, __m128i __A) {
395 return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
396 (__v16qi)__A, (__v8hf)(__m128h)__W, (__mmask8)__U);
397}
398
399static __inline__ __m128h __DEFAULT_FN_ATTRS128
400_mm_maskz_cvtnehf8_ph(__mmask8 __U, __m128i __A) {
401 return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
402 (__v16qi)__A, (__v8hf)(__m128h)_mm_setzero_ph(), (__mmask8)__U);
403}
404
405static __inline__ __m256h __DEFAULT_FN_ATTRS256
406_mm256_cvtnehf8_ph(__m128i __A) {
407 return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
408 (__v16qi)__A, (__v16hf)(__m256h)_mm256_undefined_ph(), (__mmask16)-1);
409}
410
411static __inline__ __m256h __DEFAULT_FN_ATTRS256
412_mm256_mask_cvtnehf8_ph(__m256h __W, __mmask16 __U, __m128i __A) {
413 return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
414 (__v16qi)__A, (__v16hf)(__m256h)__W, (__mmask16)__U);
415}
416
417static __inline__ __m256h __DEFAULT_FN_ATTRS256
418_mm256_maskz_cvtnehf8_ph(__mmask16 __U, __m128i __A) {
419 return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
420 (__v16qi)__A, (__v16hf)(__m256h)_mm256_setzero_ph(), (__mmask16)__U);
421}
422
423static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtneph_pbf8(__m128h __A) {
424 return (__m128i)__builtin_ia32_vcvtneph2bf8_128_mask(
425 (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
426}
427
428static __inline__ __m128i __DEFAULT_FN_ATTRS128
429_mm_mask_cvtneph_pbf8(__m128i __W, __mmask8 __U, __m128h __A) {
430 return (__m128i)__builtin_ia32_vcvtneph2bf8_128_mask(
431 (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
432}
433
434static __inline__ __m128i __DEFAULT_FN_ATTRS128
435_mm_maskz_cvtneph_pbf8(__mmask8 __U, __m128h __A) {
436 return (__m128i)__builtin_ia32_vcvtneph2bf8_128_mask(
437 (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
438}
439
440static __inline__ __m128i __DEFAULT_FN_ATTRS256
441_mm256_cvtneph_pbf8(__m256h __A) {
442 return (__m128i)__builtin_ia32_vcvtneph2bf8_256_mask(
443 (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
444}
445
446static __inline__ __m128i __DEFAULT_FN_ATTRS256
447_mm256_mask_cvtneph_pbf8(__m128i __W, __mmask16 __U, __m256h __A) {
448 return (__m128i)__builtin_ia32_vcvtneph2bf8_256_mask(
449 (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
450}
451
452static __inline__ __m128i __DEFAULT_FN_ATTRS256
453_mm256_maskz_cvtneph_pbf8(__mmask16 __U, __m256h __A) {
454 return (__m128i)__builtin_ia32_vcvtneph2bf8_256_mask(
455 (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
456}
457
458static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtnesph_pbf8(__m128h __A) {
459 return (__m128i)__builtin_ia32_vcvtneph2bf8s_128_mask(
460 (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
461}
462
463static __inline__ __m128i __DEFAULT_FN_ATTRS128
464_mm_mask_cvtnesph_pbf8(__m128i __W, __mmask8 __U, __m128h __A) {
465 return (__m128i)__builtin_ia32_vcvtneph2bf8s_128_mask(
466 (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
467}
468
469static __inline__ __m128i __DEFAULT_FN_ATTRS128
470_mm_maskz_cvtnesph_pbf8(__mmask8 __U, __m128h __A) {
471 return (__m128i)__builtin_ia32_vcvtneph2bf8s_128_mask(
472 (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
473}
474
475static __inline__ __m128i __DEFAULT_FN_ATTRS256
476_mm256_cvtnesph_pbf8(__m256h __A) {
477 return (__m128i)__builtin_ia32_vcvtneph2bf8s_256_mask(
478 (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
479}
480
481static __inline__ __m128i __DEFAULT_FN_ATTRS256
482_mm256_mask_cvtnesph_pbf8(__m128i __W, __mmask16 __U, __m256h __A) {
483 return (__m128i)__builtin_ia32_vcvtneph2bf8s_256_mask(
484 (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
485}
486
487static __inline__ __m128i __DEFAULT_FN_ATTRS256
488_mm256_maskz_cvtnesph_pbf8(__mmask16 __U, __m256h __A) {
489 return (__m128i)__builtin_ia32_vcvtneph2bf8s_256_mask(
490 (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
491}
492
493static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtneph_phf8(__m128h __A) {
494 return (__m128i)__builtin_ia32_vcvtneph2hf8_128_mask(
495 (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
496}
497
498static __inline__ __m128i __DEFAULT_FN_ATTRS128
499_mm_mask_cvtneph_phf8(__m128i __W, __mmask8 __U, __m128h __A) {
500 return (__m128i)__builtin_ia32_vcvtneph2hf8_128_mask(
501 (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
502}
503
504static __inline__ __m128i __DEFAULT_FN_ATTRS128
505_mm_maskz_cvtneph_phf8(__mmask8 __U, __m128h __A) {
506 return (__m128i)__builtin_ia32_vcvtneph2hf8_128_mask(
507 (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
508}
509
510static __inline__ __m128i __DEFAULT_FN_ATTRS256
511_mm256_cvtneph_phf8(__m256h __A) {
512 return (__m128i)__builtin_ia32_vcvtneph2hf8_256_mask(
513 (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
514}
515
516static __inline__ __m128i __DEFAULT_FN_ATTRS256
517_mm256_mask_cvtneph_phf8(__m128i __W, __mmask16 __U, __m256h __A) {
518 return (__m128i)__builtin_ia32_vcvtneph2hf8_256_mask(
519 (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
520}
521
522static __inline__ __m128i __DEFAULT_FN_ATTRS256
523_mm256_maskz_cvtneph_phf8(__mmask16 __U, __m256h __A) {
524 return (__m128i)__builtin_ia32_vcvtneph2hf8_256_mask(
525 (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
526}
527
528static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtnesph_phf8(__m128h __A) {
529 return (__m128i)__builtin_ia32_vcvtneph2hf8s_128_mask(
530 (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
531}
532
533static __inline__ __m128i __DEFAULT_FN_ATTRS128
534_mm_mask_cvtnesph_phf8(__m128i __W, __mmask8 __U, __m128h __A) {
535 return (__m128i)__builtin_ia32_vcvtneph2hf8s_128_mask(
536 (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
537}
538
539static __inline__ __m128i __DEFAULT_FN_ATTRS128
540_mm_maskz_cvtnesph_phf8(__mmask8 __U, __m128h __A) {
541 return (__m128i)__builtin_ia32_vcvtneph2hf8s_128_mask(
542 (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
543}
544
545static __inline__ __m128i __DEFAULT_FN_ATTRS256
546_mm256_cvtnesph_phf8(__m256h __A) {
547 return (__m128i)__builtin_ia32_vcvtneph2hf8s_256_mask(
548 (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
549}
550
551static __inline__ __m128i __DEFAULT_FN_ATTRS256
552_mm256_mask_cvtnesph_phf8(__m128i __W, __mmask16 __U, __m256h __A) {
553 return (__m128i)__builtin_ia32_vcvtneph2hf8s_256_mask(
554 (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
555}
556
557static __inline__ __m128i __DEFAULT_FN_ATTRS256
558_mm256_maskz_cvtnesph_phf8(__mmask16 __U, __m256h __A) {
559 return (__m128i)__builtin_ia32_vcvtneph2hf8s_256_mask(
560 (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
561}
562
563static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvtpbf8_ph(__m128i __A) {
564 return _mm_castsi128_ph(_mm_slli_epi16(_mm_cvtepi8_epi16(__A), 8));
565}
566
567static __inline__ __m128h __DEFAULT_FN_ATTRS128
568_mm_mask_cvtpbf8_ph(__m128h __S, __mmask8 __U, __m128i __A) {
569 return _mm_castsi128_ph(
570 _mm_mask_slli_epi16((__m128i)__S, __U, _mm_cvtepi8_epi16(__A), 8));
571}
572
573static __inline__ __m128h __DEFAULT_FN_ATTRS128
574_mm_maskz_cvtpbf8_ph(__mmask8 __U, __m128i __A) {
575 return _mm_castsi128_ph(_mm_slli_epi16(_mm_maskz_cvtepi8_epi16(__U, __A), 8));
576}
577
578static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvtpbf8_ph(__m128i __A) {
579 return _mm256_castsi256_ph(_mm256_slli_epi16(_mm256_cvtepi8_epi16(__A), 8));
580}
581
582static __inline__ __m256h __DEFAULT_FN_ATTRS256
583_mm256_mask_cvtpbf8_ph(__m256h __S, __mmask8 __U, __m128i __A) {
584 return _mm256_castsi256_ph(
585 _mm256_mask_slli_epi16((__m256i)__S, __U, _mm256_cvtepi8_epi16(__A), 8));
586}
587
588static __inline__ __m256h __DEFAULT_FN_ATTRS256
589_mm256_maskz_cvtpbf8_ph(__mmask8 __U, __m128i __A) {
590 return _mm256_castsi256_ph(
591 _mm256_slli_epi16(_mm256_maskz_cvtepi8_epi16(__U, __A), 8));
592}
593
594#undef __DEFAULT_FN_ATTRS128
595#undef __DEFAULT_FN_ATTRS256
596
597#endif // __AVX10_2CONVERTINTRIN_H
598#endif // __SSE2__
_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:1367
_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:2138
__DEFAULT_FN_ATTRS256
#define __DEFAULT_FN_ATTRS256
Definition: avx2intrin.h:26
__DEFAULT_FN_ATTRS128
#define __DEFAULT_FN_ATTRS128
Definition: avx2intrin.h:29
__mmask32
unsigned int __mmask32
Definition: avx512bwintrin.h:17
__mmask8
unsigned char __mmask8
Definition: avx512fintrin.h:41
__mmask16
unsigned short __mmask16
Definition: avx512fintrin.h:42
_MM_FROUND_CUR_DIRECTION
#define _MM_FROUND_CUR_DIRECTION
Definition: avx512fintrin.h:49
_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
_mm_maskz_cvtepi8_epi16
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1772
_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
_mm256_maskz_cvtepi8_epi16
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A)
Definition: avx512vlbwintrin.h:1788
_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:4362
_mm_undefined_si128
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_undefined_si128(void)
Generates a 128-bit vector of [4 x i32] with unspecified content.
Definition: emmintrin.h:3484
_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:2764
_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:3865
_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:1227
Generated on Fri Nov 22 2024 09:49:18 for clang by doxygen 1.9.6