blob: c5b96ae8ada70a922088ce03460720ebb30604d1 [file] [log] [blame]
Logan Chien2833ffb2018-10-09 10:03:24 +08001/*===------------- avx512vbmivlintrin.h - VBMI intrinsics ------------------===
2 *
3 *
Logan Chiendf4f7662019-09-04 16:45:23 -07004 * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5 * See https://llvm.org/LICENSE.txt for license information.
6 * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
Logan Chien2833ffb2018-10-09 10:03:24 +08007 *
8 *===-----------------------------------------------------------------------===
9 */
10#ifndef __IMMINTRIN_H
11#error "Never use <avx512vbmivlintrin.h> directly; include <immintrin.h> instead."
12#endif
13
14#ifndef __VBMIVLINTRIN_H
15#define __VBMIVLINTRIN_H
16
17/* Define the default attributes for the functions in this file. */
Logan Chien55afb0a2018-10-15 10:42:14 +080018#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi,avx512vl"), __min_vector_width__(128)))
19#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi,avx512vl"), __min_vector_width__(256)))
Logan Chien2833ffb2018-10-09 10:03:24 +080020
21
Logan Chien55afb0a2018-10-15 10:42:14 +080022static __inline__ __m128i __DEFAULT_FN_ATTRS128
23_mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080024{
Logan Chien55afb0a2018-10-15 10:42:14 +080025 return (__m128i)__builtin_ia32_vpermi2varqi128((__v16qi)__A,
26 (__v16qi)__I,
27 (__v16qi)__B);
Logan Chien2833ffb2018-10-09 10:03:24 +080028}
29
Logan Chien55afb0a2018-10-15 10:42:14 +080030static __inline__ __m128i __DEFAULT_FN_ATTRS128
31_mm_mask_permutex2var_epi8(__m128i __A, __mmask16 __U, __m128i __I,
32 __m128i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080033{
Logan Chien55afb0a2018-10-15 10:42:14 +080034 return (__m128i)__builtin_ia32_selectb_128(__U,
35 (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
36 (__v16qi)__A);
Logan Chien2833ffb2018-10-09 10:03:24 +080037}
38
Logan Chien55afb0a2018-10-15 10:42:14 +080039static __inline__ __m128i __DEFAULT_FN_ATTRS128
40_mm_mask2_permutex2var_epi8(__m128i __A, __m128i __I, __mmask16 __U,
41 __m128i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080042{
Logan Chien55afb0a2018-10-15 10:42:14 +080043 return (__m128i)__builtin_ia32_selectb_128(__U,
44 (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
45 (__v16qi)__I);
Logan Chien2833ffb2018-10-09 10:03:24 +080046}
47
Logan Chien55afb0a2018-10-15 10:42:14 +080048static __inline__ __m128i __DEFAULT_FN_ATTRS128
49_mm_maskz_permutex2var_epi8(__mmask16 __U, __m128i __A, __m128i __I,
50 __m128i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080051{
Logan Chien55afb0a2018-10-15 10:42:14 +080052 return (__m128i)__builtin_ia32_selectb_128(__U,
53 (__v16qi)_mm_permutex2var_epi8(__A, __I, __B),
54 (__v16qi)_mm_setzero_si128());
Logan Chien2833ffb2018-10-09 10:03:24 +080055}
56
Logan Chien55afb0a2018-10-15 10:42:14 +080057static __inline__ __m256i __DEFAULT_FN_ATTRS256
58_mm256_permutex2var_epi8(__m256i __A, __m256i __I, __m256i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080059{
Logan Chien55afb0a2018-10-15 10:42:14 +080060 return (__m256i)__builtin_ia32_vpermi2varqi256((__v32qi)__A, (__v32qi)__I,
61 (__v32qi)__B);
Logan Chien2833ffb2018-10-09 10:03:24 +080062}
63
Logan Chien55afb0a2018-10-15 10:42:14 +080064static __inline__ __m256i __DEFAULT_FN_ATTRS256
65_mm256_mask_permutex2var_epi8(__m256i __A, __mmask32 __U, __m256i __I,
66 __m256i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080067{
Logan Chien55afb0a2018-10-15 10:42:14 +080068 return (__m256i)__builtin_ia32_selectb_256(__U,
69 (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
70 (__v32qi)__A);
Logan Chien2833ffb2018-10-09 10:03:24 +080071}
72
Logan Chien55afb0a2018-10-15 10:42:14 +080073static __inline__ __m256i __DEFAULT_FN_ATTRS256
74_mm256_mask2_permutex2var_epi8(__m256i __A, __m256i __I, __mmask32 __U,
75 __m256i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080076{
Logan Chien55afb0a2018-10-15 10:42:14 +080077 return (__m256i)__builtin_ia32_selectb_256(__U,
78 (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
79 (__v32qi)__I);
Logan Chien2833ffb2018-10-09 10:03:24 +080080}
81
Logan Chien55afb0a2018-10-15 10:42:14 +080082static __inline__ __m256i __DEFAULT_FN_ATTRS256
83_mm256_maskz_permutex2var_epi8(__mmask32 __U, __m256i __A, __m256i __I,
84 __m256i __B)
Logan Chien2833ffb2018-10-09 10:03:24 +080085{
Logan Chien55afb0a2018-10-15 10:42:14 +080086 return (__m256i)__builtin_ia32_selectb_256(__U,
87 (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B),
88 (__v32qi)_mm256_setzero_si256());
Logan Chien2833ffb2018-10-09 10:03:24 +080089}
90
Logan Chien55afb0a2018-10-15 10:42:14 +080091static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080092_mm_permutexvar_epi8 (__m128i __A, __m128i __B)
93{
Logan Chien55afb0a2018-10-15 10:42:14 +080094 return (__m128i)__builtin_ia32_permvarqi128((__v16qi)__B, (__v16qi)__A);
Logan Chien2833ffb2018-10-09 10:03:24 +080095}
96
Logan Chien55afb0a2018-10-15 10:42:14 +080097static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080098_mm_maskz_permutexvar_epi8 (__mmask16 __M, __m128i __A, __m128i __B)
99{
Logan Chien55afb0a2018-10-15 10:42:14 +0800100 return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
101 (__v16qi)_mm_permutexvar_epi8(__A, __B),
102 (__v16qi)_mm_setzero_si128());
Logan Chien2833ffb2018-10-09 10:03:24 +0800103}
104
Logan Chien55afb0a2018-10-15 10:42:14 +0800105static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800106_mm_mask_permutexvar_epi8 (__m128i __W, __mmask16 __M, __m128i __A,
107 __m128i __B)
108{
Logan Chien55afb0a2018-10-15 10:42:14 +0800109 return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
110 (__v16qi)_mm_permutexvar_epi8(__A, __B),
111 (__v16qi)__W);
Logan Chien2833ffb2018-10-09 10:03:24 +0800112}
113
Logan Chien55afb0a2018-10-15 10:42:14 +0800114static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800115_mm256_permutexvar_epi8 (__m256i __A, __m256i __B)
116{
Logan Chien55afb0a2018-10-15 10:42:14 +0800117 return (__m256i)__builtin_ia32_permvarqi256((__v32qi) __B, (__v32qi) __A);
Logan Chien2833ffb2018-10-09 10:03:24 +0800118}
119
Logan Chien55afb0a2018-10-15 10:42:14 +0800120static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800121_mm256_maskz_permutexvar_epi8 (__mmask32 __M, __m256i __A,
122 __m256i __B)
123{
Logan Chien55afb0a2018-10-15 10:42:14 +0800124 return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
125 (__v32qi)_mm256_permutexvar_epi8(__A, __B),
126 (__v32qi)_mm256_setzero_si256());
Logan Chien2833ffb2018-10-09 10:03:24 +0800127}
128
Logan Chien55afb0a2018-10-15 10:42:14 +0800129static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800130_mm256_mask_permutexvar_epi8 (__m256i __W, __mmask32 __M, __m256i __A,
131 __m256i __B)
132{
Logan Chien55afb0a2018-10-15 10:42:14 +0800133 return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
134 (__v32qi)_mm256_permutexvar_epi8(__A, __B),
135 (__v32qi)__W);
Logan Chien2833ffb2018-10-09 10:03:24 +0800136}
137
Logan Chien55afb0a2018-10-15 10:42:14 +0800138static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chiendbcf4122019-03-21 10:50:25 +0800139_mm_multishift_epi64_epi8(__m128i __X, __m128i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800140{
Logan Chiendbcf4122019-03-21 10:50:25 +0800141 return (__m128i)__builtin_ia32_vpmultishiftqb128((__v16qi)__X, (__v16qi)__Y);
Logan Chien2833ffb2018-10-09 10:03:24 +0800142}
143
Logan Chien55afb0a2018-10-15 10:42:14 +0800144static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chiendbcf4122019-03-21 10:50:25 +0800145_mm_mask_multishift_epi64_epi8(__m128i __W, __mmask16 __M, __m128i __X,
146 __m128i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800147{
Logan Chiendbcf4122019-03-21 10:50:25 +0800148 return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
149 (__v16qi)_mm_multishift_epi64_epi8(__X, __Y),
150 (__v16qi)__W);
Logan Chien2833ffb2018-10-09 10:03:24 +0800151}
152
Logan Chien55afb0a2018-10-15 10:42:14 +0800153static __inline__ __m128i __DEFAULT_FN_ATTRS128
Logan Chiendbcf4122019-03-21 10:50:25 +0800154_mm_maskz_multishift_epi64_epi8(__mmask16 __M, __m128i __X, __m128i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800155{
Logan Chiendbcf4122019-03-21 10:50:25 +0800156 return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M,
157 (__v16qi)_mm_multishift_epi64_epi8(__X, __Y),
158 (__v16qi)_mm_setzero_si128());
Logan Chien2833ffb2018-10-09 10:03:24 +0800159}
160
Logan Chien55afb0a2018-10-15 10:42:14 +0800161static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chiendbcf4122019-03-21 10:50:25 +0800162_mm256_multishift_epi64_epi8(__m256i __X, __m256i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800163{
Logan Chiendbcf4122019-03-21 10:50:25 +0800164 return (__m256i)__builtin_ia32_vpmultishiftqb256((__v32qi)__X, (__v32qi)__Y);
Logan Chien2833ffb2018-10-09 10:03:24 +0800165}
166
Logan Chien55afb0a2018-10-15 10:42:14 +0800167static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chiendbcf4122019-03-21 10:50:25 +0800168_mm256_mask_multishift_epi64_epi8(__m256i __W, __mmask32 __M, __m256i __X,
169 __m256i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800170{
Logan Chiendbcf4122019-03-21 10:50:25 +0800171 return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
172 (__v32qi)_mm256_multishift_epi64_epi8(__X, __Y),
173 (__v32qi)__W);
Logan Chien2833ffb2018-10-09 10:03:24 +0800174}
175
Logan Chien55afb0a2018-10-15 10:42:14 +0800176static __inline__ __m256i __DEFAULT_FN_ATTRS256
Logan Chiendbcf4122019-03-21 10:50:25 +0800177_mm256_maskz_multishift_epi64_epi8(__mmask32 __M, __m256i __X, __m256i __Y)
Logan Chien2833ffb2018-10-09 10:03:24 +0800178{
Logan Chiendbcf4122019-03-21 10:50:25 +0800179 return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M,
180 (__v32qi)_mm256_multishift_epi64_epi8(__X, __Y),
181 (__v32qi)_mm256_setzero_si256());
Logan Chien2833ffb2018-10-09 10:03:24 +0800182}
183
184
Logan Chien55afb0a2018-10-15 10:42:14 +0800185#undef __DEFAULT_FN_ATTRS128
186#undef __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800187
188#endif