Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1 | /*===------------- avx512bwintrin.h - AVX512BW intrinsics ------------------=== |
| 2 | * |
| 3 | * |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 4 | * 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 Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 7 | * |
| 8 | *===-----------------------------------------------------------------------=== |
| 9 | */ |
| 10 | #ifndef __IMMINTRIN_H |
| 11 | #error "Never use <avx512bwintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifndef __AVX512BWINTRIN_H |
| 15 | #define __AVX512BWINTRIN_H |
| 16 | |
| 17 | typedef unsigned int __mmask32; |
| 18 | typedef unsigned long long __mmask64; |
| 19 | |
| 20 | /* Define the default attributes for the functions in this file. */ |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 21 | #define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw"), __min_vector_width__(512))) |
| 22 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512bw"))) |
| 23 | |
| 24 | static __inline __mmask32 __DEFAULT_FN_ATTRS |
| 25 | _knot_mask32(__mmask32 __M) |
| 26 | { |
| 27 | return __builtin_ia32_knotsi(__M); |
| 28 | } |
| 29 | |
| 30 | static __inline __mmask64 __DEFAULT_FN_ATTRS |
| 31 | _knot_mask64(__mmask64 __M) |
| 32 | { |
| 33 | return __builtin_ia32_knotdi(__M); |
| 34 | } |
| 35 | |
| 36 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 37 | _kand_mask32(__mmask32 __A, __mmask32 __B) |
| 38 | { |
| 39 | return (__mmask32)__builtin_ia32_kandsi((__mmask32)__A, (__mmask32)__B); |
| 40 | } |
| 41 | |
| 42 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 43 | _kand_mask64(__mmask64 __A, __mmask64 __B) |
| 44 | { |
| 45 | return (__mmask64)__builtin_ia32_kanddi((__mmask64)__A, (__mmask64)__B); |
| 46 | } |
| 47 | |
| 48 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 49 | _kandn_mask32(__mmask32 __A, __mmask32 __B) |
| 50 | { |
| 51 | return (__mmask32)__builtin_ia32_kandnsi((__mmask32)__A, (__mmask32)__B); |
| 52 | } |
| 53 | |
| 54 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 55 | _kandn_mask64(__mmask64 __A, __mmask64 __B) |
| 56 | { |
| 57 | return (__mmask64)__builtin_ia32_kandndi((__mmask64)__A, (__mmask64)__B); |
| 58 | } |
| 59 | |
| 60 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 61 | _kor_mask32(__mmask32 __A, __mmask32 __B) |
| 62 | { |
| 63 | return (__mmask32)__builtin_ia32_korsi((__mmask32)__A, (__mmask32)__B); |
| 64 | } |
| 65 | |
| 66 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 67 | _kor_mask64(__mmask64 __A, __mmask64 __B) |
| 68 | { |
| 69 | return (__mmask64)__builtin_ia32_kordi((__mmask64)__A, (__mmask64)__B); |
| 70 | } |
| 71 | |
| 72 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 73 | _kxnor_mask32(__mmask32 __A, __mmask32 __B) |
| 74 | { |
| 75 | return (__mmask32)__builtin_ia32_kxnorsi((__mmask32)__A, (__mmask32)__B); |
| 76 | } |
| 77 | |
| 78 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 79 | _kxnor_mask64(__mmask64 __A, __mmask64 __B) |
| 80 | { |
| 81 | return (__mmask64)__builtin_ia32_kxnordi((__mmask64)__A, (__mmask64)__B); |
| 82 | } |
| 83 | |
| 84 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 85 | _kxor_mask32(__mmask32 __A, __mmask32 __B) |
| 86 | { |
| 87 | return (__mmask32)__builtin_ia32_kxorsi((__mmask32)__A, (__mmask32)__B); |
| 88 | } |
| 89 | |
| 90 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 91 | _kxor_mask64(__mmask64 __A, __mmask64 __B) |
| 92 | { |
| 93 | return (__mmask64)__builtin_ia32_kxordi((__mmask64)__A, (__mmask64)__B); |
| 94 | } |
| 95 | |
| 96 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 97 | _kortestc_mask32_u8(__mmask32 __A, __mmask32 __B) |
| 98 | { |
| 99 | return (unsigned char)__builtin_ia32_kortestcsi(__A, __B); |
| 100 | } |
| 101 | |
| 102 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 103 | _kortestz_mask32_u8(__mmask32 __A, __mmask32 __B) |
| 104 | { |
| 105 | return (unsigned char)__builtin_ia32_kortestzsi(__A, __B); |
| 106 | } |
| 107 | |
| 108 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 109 | _kortest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) { |
| 110 | *__C = (unsigned char)__builtin_ia32_kortestcsi(__A, __B); |
| 111 | return (unsigned char)__builtin_ia32_kortestzsi(__A, __B); |
| 112 | } |
| 113 | |
| 114 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 115 | _kortestc_mask64_u8(__mmask64 __A, __mmask64 __B) |
| 116 | { |
| 117 | return (unsigned char)__builtin_ia32_kortestcdi(__A, __B); |
| 118 | } |
| 119 | |
| 120 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 121 | _kortestz_mask64_u8(__mmask64 __A, __mmask64 __B) |
| 122 | { |
| 123 | return (unsigned char)__builtin_ia32_kortestzdi(__A, __B); |
| 124 | } |
| 125 | |
| 126 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 127 | _kortest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C) { |
| 128 | *__C = (unsigned char)__builtin_ia32_kortestcdi(__A, __B); |
| 129 | return (unsigned char)__builtin_ia32_kortestzdi(__A, __B); |
| 130 | } |
| 131 | |
| 132 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 133 | _ktestc_mask32_u8(__mmask32 __A, __mmask32 __B) |
| 134 | { |
| 135 | return (unsigned char)__builtin_ia32_ktestcsi(__A, __B); |
| 136 | } |
| 137 | |
| 138 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 139 | _ktestz_mask32_u8(__mmask32 __A, __mmask32 __B) |
| 140 | { |
| 141 | return (unsigned char)__builtin_ia32_ktestzsi(__A, __B); |
| 142 | } |
| 143 | |
| 144 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 145 | _ktest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C) { |
| 146 | *__C = (unsigned char)__builtin_ia32_ktestcsi(__A, __B); |
| 147 | return (unsigned char)__builtin_ia32_ktestzsi(__A, __B); |
| 148 | } |
| 149 | |
| 150 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 151 | _ktestc_mask64_u8(__mmask64 __A, __mmask64 __B) |
| 152 | { |
| 153 | return (unsigned char)__builtin_ia32_ktestcdi(__A, __B); |
| 154 | } |
| 155 | |
| 156 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 157 | _ktestz_mask64_u8(__mmask64 __A, __mmask64 __B) |
| 158 | { |
| 159 | return (unsigned char)__builtin_ia32_ktestzdi(__A, __B); |
| 160 | } |
| 161 | |
| 162 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 163 | _ktest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C) { |
| 164 | *__C = (unsigned char)__builtin_ia32_ktestcdi(__A, __B); |
| 165 | return (unsigned char)__builtin_ia32_ktestzdi(__A, __B); |
| 166 | } |
| 167 | |
| 168 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 169 | _kadd_mask32(__mmask32 __A, __mmask32 __B) |
| 170 | { |
| 171 | return (__mmask32)__builtin_ia32_kaddsi((__mmask32)__A, (__mmask32)__B); |
| 172 | } |
| 173 | |
| 174 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 175 | _kadd_mask64(__mmask64 __A, __mmask64 __B) |
| 176 | { |
| 177 | return (__mmask64)__builtin_ia32_kadddi((__mmask64)__A, (__mmask64)__B); |
| 178 | } |
| 179 | |
| 180 | #define _kshiftli_mask32(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 181 | ((__mmask32)__builtin_ia32_kshiftlisi((__mmask32)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 182 | |
| 183 | #define _kshiftri_mask32(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 184 | ((__mmask32)__builtin_ia32_kshiftrisi((__mmask32)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 185 | |
| 186 | #define _kshiftli_mask64(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 187 | ((__mmask64)__builtin_ia32_kshiftlidi((__mmask64)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 188 | |
| 189 | #define _kshiftri_mask64(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 190 | ((__mmask64)__builtin_ia32_kshiftridi((__mmask64)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 191 | |
| 192 | static __inline__ unsigned int __DEFAULT_FN_ATTRS |
| 193 | _cvtmask32_u32(__mmask32 __A) { |
| 194 | return (unsigned int)__builtin_ia32_kmovd((__mmask32)__A); |
| 195 | } |
| 196 | |
| 197 | static __inline__ unsigned long long __DEFAULT_FN_ATTRS |
| 198 | _cvtmask64_u64(__mmask64 __A) { |
| 199 | return (unsigned long long)__builtin_ia32_kmovq((__mmask64)__A); |
| 200 | } |
| 201 | |
| 202 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 203 | _cvtu32_mask32(unsigned int __A) { |
| 204 | return (__mmask32)__builtin_ia32_kmovd((__mmask32)__A); |
| 205 | } |
| 206 | |
| 207 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 208 | _cvtu64_mask64(unsigned long long __A) { |
| 209 | return (__mmask64)__builtin_ia32_kmovq((__mmask64)__A); |
| 210 | } |
| 211 | |
| 212 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
| 213 | _load_mask32(__mmask32 *__A) { |
| 214 | return (__mmask32)__builtin_ia32_kmovd(*(__mmask32 *)__A); |
| 215 | } |
| 216 | |
| 217 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
| 218 | _load_mask64(__mmask64 *__A) { |
| 219 | return (__mmask64)__builtin_ia32_kmovq(*(__mmask64 *)__A); |
| 220 | } |
| 221 | |
| 222 | static __inline__ void __DEFAULT_FN_ATTRS |
| 223 | _store_mask32(__mmask32 *__A, __mmask32 __B) { |
| 224 | *(__mmask32 *)__A = __builtin_ia32_kmovd((__mmask32)__B); |
| 225 | } |
| 226 | |
| 227 | static __inline__ void __DEFAULT_FN_ATTRS |
| 228 | _store_mask64(__mmask64 *__A, __mmask64 __B) { |
| 229 | *(__mmask64 *)__A = __builtin_ia32_kmovq((__mmask64)__B); |
| 230 | } |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 231 | |
| 232 | /* Integer compare */ |
| 233 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 234 | #define _mm512_cmp_epi8_mask(a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 235 | ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \ |
| 236 | (__v64qi)(__m512i)(b), (int)(p), \ |
| 237 | (__mmask64)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 238 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 239 | #define _mm512_mask_cmp_epi8_mask(m, a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 240 | ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \ |
| 241 | (__v64qi)(__m512i)(b), (int)(p), \ |
| 242 | (__mmask64)(m))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 243 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 244 | #define _mm512_cmp_epu8_mask(a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 245 | ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \ |
| 246 | (__v64qi)(__m512i)(b), (int)(p), \ |
| 247 | (__mmask64)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 248 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 249 | #define _mm512_mask_cmp_epu8_mask(m, a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 250 | ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \ |
| 251 | (__v64qi)(__m512i)(b), (int)(p), \ |
| 252 | (__mmask64)(m))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 253 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 254 | #define _mm512_cmp_epi16_mask(a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 255 | ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \ |
| 256 | (__v32hi)(__m512i)(b), (int)(p), \ |
| 257 | (__mmask32)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 258 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 259 | #define _mm512_mask_cmp_epi16_mask(m, a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 260 | ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \ |
| 261 | (__v32hi)(__m512i)(b), (int)(p), \ |
| 262 | (__mmask32)(m))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 263 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 264 | #define _mm512_cmp_epu16_mask(a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 265 | ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \ |
| 266 | (__v32hi)(__m512i)(b), (int)(p), \ |
| 267 | (__mmask32)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 268 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 269 | #define _mm512_mask_cmp_epu16_mask(m, a, b, p) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 270 | ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \ |
| 271 | (__v32hi)(__m512i)(b), (int)(p), \ |
| 272 | (__mmask32)(m))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 273 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 274 | #define _mm512_cmpeq_epi8_mask(A, B) \ |
| 275 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ) |
| 276 | #define _mm512_mask_cmpeq_epi8_mask(k, A, B) \ |
| 277 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 278 | #define _mm512_cmpge_epi8_mask(A, B) \ |
| 279 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GE) |
| 280 | #define _mm512_mask_cmpge_epi8_mask(k, A, B) \ |
| 281 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE) |
| 282 | #define _mm512_cmpgt_epi8_mask(A, B) \ |
| 283 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GT) |
| 284 | #define _mm512_mask_cmpgt_epi8_mask(k, A, B) \ |
| 285 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT) |
| 286 | #define _mm512_cmple_epi8_mask(A, B) \ |
| 287 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LE) |
| 288 | #define _mm512_mask_cmple_epi8_mask(k, A, B) \ |
| 289 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE) |
| 290 | #define _mm512_cmplt_epi8_mask(A, B) \ |
| 291 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LT) |
| 292 | #define _mm512_mask_cmplt_epi8_mask(k, A, B) \ |
| 293 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT) |
| 294 | #define _mm512_cmpneq_epi8_mask(A, B) \ |
| 295 | _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_NE) |
| 296 | #define _mm512_mask_cmpneq_epi8_mask(k, A, B) \ |
| 297 | _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 298 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 299 | #define _mm512_cmpeq_epu8_mask(A, B) \ |
| 300 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ) |
| 301 | #define _mm512_mask_cmpeq_epu8_mask(k, A, B) \ |
| 302 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 303 | #define _mm512_cmpge_epu8_mask(A, B) \ |
| 304 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GE) |
| 305 | #define _mm512_mask_cmpge_epu8_mask(k, A, B) \ |
| 306 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE) |
| 307 | #define _mm512_cmpgt_epu8_mask(A, B) \ |
| 308 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GT) |
| 309 | #define _mm512_mask_cmpgt_epu8_mask(k, A, B) \ |
| 310 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT) |
| 311 | #define _mm512_cmple_epu8_mask(A, B) \ |
| 312 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LE) |
| 313 | #define _mm512_mask_cmple_epu8_mask(k, A, B) \ |
| 314 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE) |
| 315 | #define _mm512_cmplt_epu8_mask(A, B) \ |
| 316 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LT) |
| 317 | #define _mm512_mask_cmplt_epu8_mask(k, A, B) \ |
| 318 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT) |
| 319 | #define _mm512_cmpneq_epu8_mask(A, B) \ |
| 320 | _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_NE) |
| 321 | #define _mm512_mask_cmpneq_epu8_mask(k, A, B) \ |
| 322 | _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 323 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 324 | #define _mm512_cmpeq_epi16_mask(A, B) \ |
| 325 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ) |
| 326 | #define _mm512_mask_cmpeq_epi16_mask(k, A, B) \ |
| 327 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 328 | #define _mm512_cmpge_epi16_mask(A, B) \ |
| 329 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GE) |
| 330 | #define _mm512_mask_cmpge_epi16_mask(k, A, B) \ |
| 331 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE) |
| 332 | #define _mm512_cmpgt_epi16_mask(A, B) \ |
| 333 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GT) |
| 334 | #define _mm512_mask_cmpgt_epi16_mask(k, A, B) \ |
| 335 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT) |
| 336 | #define _mm512_cmple_epi16_mask(A, B) \ |
| 337 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LE) |
| 338 | #define _mm512_mask_cmple_epi16_mask(k, A, B) \ |
| 339 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE) |
| 340 | #define _mm512_cmplt_epi16_mask(A, B) \ |
| 341 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LT) |
| 342 | #define _mm512_mask_cmplt_epi16_mask(k, A, B) \ |
| 343 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT) |
| 344 | #define _mm512_cmpneq_epi16_mask(A, B) \ |
| 345 | _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_NE) |
| 346 | #define _mm512_mask_cmpneq_epi16_mask(k, A, B) \ |
| 347 | _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 348 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 349 | #define _mm512_cmpeq_epu16_mask(A, B) \ |
| 350 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ) |
| 351 | #define _mm512_mask_cmpeq_epu16_mask(k, A, B) \ |
| 352 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 353 | #define _mm512_cmpge_epu16_mask(A, B) \ |
| 354 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GE) |
| 355 | #define _mm512_mask_cmpge_epu16_mask(k, A, B) \ |
| 356 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE) |
| 357 | #define _mm512_cmpgt_epu16_mask(A, B) \ |
| 358 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GT) |
| 359 | #define _mm512_mask_cmpgt_epu16_mask(k, A, B) \ |
| 360 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT) |
| 361 | #define _mm512_cmple_epu16_mask(A, B) \ |
| 362 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LE) |
| 363 | #define _mm512_mask_cmple_epu16_mask(k, A, B) \ |
| 364 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE) |
| 365 | #define _mm512_cmplt_epu16_mask(A, B) \ |
| 366 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LT) |
| 367 | #define _mm512_mask_cmplt_epu16_mask(k, A, B) \ |
| 368 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT) |
| 369 | #define _mm512_cmpneq_epu16_mask(A, B) \ |
| 370 | _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_NE) |
| 371 | #define _mm512_mask_cmpneq_epu16_mask(k, A, B) \ |
| 372 | _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 373 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 374 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 375 | _mm512_add_epi8 (__m512i __A, __m512i __B) { |
| 376 | return (__m512i) ((__v64qu) __A + (__v64qu) __B); |
| 377 | } |
| 378 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 379 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 380 | _mm512_mask_add_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
| 381 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 382 | (__v64qi)_mm512_add_epi8(__A, __B), |
| 383 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 384 | } |
| 385 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 386 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 387 | _mm512_maskz_add_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
| 388 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 389 | (__v64qi)_mm512_add_epi8(__A, __B), |
| 390 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 391 | } |
| 392 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 393 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 394 | _mm512_sub_epi8 (__m512i __A, __m512i __B) { |
| 395 | return (__m512i) ((__v64qu) __A - (__v64qu) __B); |
| 396 | } |
| 397 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 398 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 399 | _mm512_mask_sub_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
| 400 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 401 | (__v64qi)_mm512_sub_epi8(__A, __B), |
| 402 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 403 | } |
| 404 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 405 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 406 | _mm512_maskz_sub_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
| 407 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 408 | (__v64qi)_mm512_sub_epi8(__A, __B), |
| 409 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 410 | } |
| 411 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 412 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 413 | _mm512_add_epi16 (__m512i __A, __m512i __B) { |
| 414 | return (__m512i) ((__v32hu) __A + (__v32hu) __B); |
| 415 | } |
| 416 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 417 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 418 | _mm512_mask_add_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
| 419 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 420 | (__v32hi)_mm512_add_epi16(__A, __B), |
| 421 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 422 | } |
| 423 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 424 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 425 | _mm512_maskz_add_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
| 426 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 427 | (__v32hi)_mm512_add_epi16(__A, __B), |
| 428 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 429 | } |
| 430 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 431 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 432 | _mm512_sub_epi16 (__m512i __A, __m512i __B) { |
| 433 | return (__m512i) ((__v32hu) __A - (__v32hu) __B); |
| 434 | } |
| 435 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 436 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 437 | _mm512_mask_sub_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
| 438 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 439 | (__v32hi)_mm512_sub_epi16(__A, __B), |
| 440 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 441 | } |
| 442 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 443 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 444 | _mm512_maskz_sub_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
| 445 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 446 | (__v32hi)_mm512_sub_epi16(__A, __B), |
| 447 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 448 | } |
| 449 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 450 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 451 | _mm512_mullo_epi16 (__m512i __A, __m512i __B) { |
| 452 | return (__m512i) ((__v32hu) __A * (__v32hu) __B); |
| 453 | } |
| 454 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 455 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 456 | _mm512_mask_mullo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
| 457 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 458 | (__v32hi)_mm512_mullo_epi16(__A, __B), |
| 459 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 460 | } |
| 461 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 462 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 463 | _mm512_maskz_mullo_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
| 464 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 465 | (__v32hi)_mm512_mullo_epi16(__A, __B), |
| 466 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 467 | } |
| 468 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 469 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 470 | _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W) |
| 471 | { |
| 472 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
| 473 | (__v64qi) __W, |
| 474 | (__v64qi) __A); |
| 475 | } |
| 476 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 477 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 478 | _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) |
| 479 | { |
| 480 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
| 481 | (__v32hi) __W, |
| 482 | (__v32hi) __A); |
| 483 | } |
| 484 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 485 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 486 | _mm512_abs_epi8 (__m512i __A) |
| 487 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 488 | return (__m512i)__builtin_ia32_pabsb512((__v64qi)__A); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 489 | } |
| 490 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 491 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 492 | _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) |
| 493 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 494 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 495 | (__v64qi)_mm512_abs_epi8(__A), |
| 496 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 497 | } |
| 498 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 499 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 500 | _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) |
| 501 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 502 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 503 | (__v64qi)_mm512_abs_epi8(__A), |
| 504 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 505 | } |
| 506 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 507 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 508 | _mm512_abs_epi16 (__m512i __A) |
| 509 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 510 | return (__m512i)__builtin_ia32_pabsw512((__v32hi)__A); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 511 | } |
| 512 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 513 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 514 | _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) |
| 515 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 516 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 517 | (__v32hi)_mm512_abs_epi16(__A), |
| 518 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 519 | } |
| 520 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 521 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 522 | _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) |
| 523 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 524 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 525 | (__v32hi)_mm512_abs_epi16(__A), |
| 526 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 527 | } |
| 528 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 529 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 530 | _mm512_packs_epi32(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 531 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 532 | return (__m512i)__builtin_ia32_packssdw512((__v16si)__A, (__v16si)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 533 | } |
| 534 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 535 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 536 | _mm512_maskz_packs_epi32(__mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 537 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 538 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 539 | (__v32hi)_mm512_packs_epi32(__A, __B), |
| 540 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 541 | } |
| 542 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 543 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 544 | _mm512_mask_packs_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 545 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 546 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 547 | (__v32hi)_mm512_packs_epi32(__A, __B), |
| 548 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 549 | } |
| 550 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 551 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 552 | _mm512_packs_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 553 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 554 | return (__m512i)__builtin_ia32_packsswb512((__v32hi)__A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 555 | } |
| 556 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 557 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 558 | _mm512_mask_packs_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 559 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 560 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 561 | (__v64qi)_mm512_packs_epi16(__A, __B), |
| 562 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 563 | } |
| 564 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 565 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 566 | _mm512_maskz_packs_epi16(__mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 567 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 568 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 569 | (__v64qi)_mm512_packs_epi16(__A, __B), |
| 570 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 571 | } |
| 572 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 573 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 574 | _mm512_packus_epi32(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 575 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 576 | return (__m512i)__builtin_ia32_packusdw512((__v16si) __A, (__v16si) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 577 | } |
| 578 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 579 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 580 | _mm512_maskz_packus_epi32(__mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 581 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 582 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 583 | (__v32hi)_mm512_packus_epi32(__A, __B), |
| 584 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 585 | } |
| 586 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 587 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 588 | _mm512_mask_packus_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 589 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 590 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 591 | (__v32hi)_mm512_packus_epi32(__A, __B), |
| 592 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 593 | } |
| 594 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 595 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 596 | _mm512_packus_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 597 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 598 | return (__m512i)__builtin_ia32_packuswb512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 599 | } |
| 600 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 601 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 602 | _mm512_mask_packus_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 603 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 604 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 605 | (__v64qi)_mm512_packus_epi16(__A, __B), |
| 606 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 607 | } |
| 608 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 609 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 610 | _mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 611 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 612 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 613 | (__v64qi)_mm512_packus_epi16(__A, __B), |
| 614 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 615 | } |
| 616 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 617 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 618 | _mm512_adds_epi8 (__m512i __A, __m512i __B) |
| 619 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 620 | return (__m512i)__builtin_ia32_paddsb512((__v64qi)__A, (__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 621 | } |
| 622 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 623 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 624 | _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 625 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 626 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 627 | (__v64qi)_mm512_adds_epi8(__A, __B), |
| 628 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 629 | } |
| 630 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 631 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 632 | _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) |
| 633 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 634 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 635 | (__v64qi)_mm512_adds_epi8(__A, __B), |
| 636 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 637 | } |
| 638 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 639 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 640 | _mm512_adds_epi16 (__m512i __A, __m512i __B) |
| 641 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 642 | return (__m512i)__builtin_ia32_paddsw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 643 | } |
| 644 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 645 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 646 | _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 647 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 648 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 649 | (__v32hi)_mm512_adds_epi16(__A, __B), |
| 650 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 651 | } |
| 652 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 653 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 654 | _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 655 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 656 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 657 | (__v32hi)_mm512_adds_epi16(__A, __B), |
| 658 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 659 | } |
| 660 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 661 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 662 | _mm512_adds_epu8 (__m512i __A, __m512i __B) |
| 663 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 664 | return (__m512i)__builtin_ia32_paddusb512((__v64qi) __A, (__v64qi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 665 | } |
| 666 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 667 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 668 | _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 669 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 670 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 671 | (__v64qi)_mm512_adds_epu8(__A, __B), |
| 672 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 673 | } |
| 674 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 675 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 676 | _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
| 677 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 678 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 679 | (__v64qi)_mm512_adds_epu8(__A, __B), |
| 680 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 681 | } |
| 682 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 683 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 684 | _mm512_adds_epu16 (__m512i __A, __m512i __B) |
| 685 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 686 | return (__m512i)__builtin_ia32_paddusw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 687 | } |
| 688 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 689 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 690 | _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 691 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 692 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 693 | (__v32hi)_mm512_adds_epu16(__A, __B), |
| 694 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 695 | } |
| 696 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 697 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 698 | _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 699 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 700 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 701 | (__v32hi)_mm512_adds_epu16(__A, __B), |
| 702 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 703 | } |
| 704 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 705 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 706 | _mm512_avg_epu8 (__m512i __A, __m512i __B) |
| 707 | { |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 708 | return (__m512i)__builtin_ia32_pavgb512((__v64qi)__A, (__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 709 | } |
| 710 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 711 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 712 | _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A, |
| 713 | __m512i __B) |
| 714 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 715 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 716 | (__v64qi)_mm512_avg_epu8(__A, __B), |
| 717 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 718 | } |
| 719 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 720 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 721 | _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
| 722 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 723 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 724 | (__v64qi)_mm512_avg_epu8(__A, __B), |
| 725 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 726 | } |
| 727 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 728 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 729 | _mm512_avg_epu16 (__m512i __A, __m512i __B) |
| 730 | { |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 731 | return (__m512i)__builtin_ia32_pavgw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 732 | } |
| 733 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 734 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 735 | _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A, |
| 736 | __m512i __B) |
| 737 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 738 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 739 | (__v32hi)_mm512_avg_epu16(__A, __B), |
| 740 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 741 | } |
| 742 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 743 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 744 | _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 745 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 746 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 747 | (__v32hi)_mm512_avg_epu16(__A, __B), |
| 748 | (__v32hi) _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 749 | } |
| 750 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 751 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 752 | _mm512_max_epi8 (__m512i __A, __m512i __B) |
| 753 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 754 | return (__m512i)__builtin_ia32_pmaxsb512((__v64qi) __A, (__v64qi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 755 | } |
| 756 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 757 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 758 | _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) |
| 759 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 760 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 761 | (__v64qi)_mm512_max_epi8(__A, __B), |
| 762 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 763 | } |
| 764 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 765 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 766 | _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 767 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 768 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 769 | (__v64qi)_mm512_max_epi8(__A, __B), |
| 770 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 771 | } |
| 772 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 773 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 774 | _mm512_max_epi16 (__m512i __A, __m512i __B) |
| 775 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 776 | return (__m512i)__builtin_ia32_pmaxsw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 777 | } |
| 778 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 779 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 780 | _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) |
| 781 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 782 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 783 | (__v32hi)_mm512_max_epi16(__A, __B), |
| 784 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 785 | } |
| 786 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 787 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 788 | _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, |
| 789 | __m512i __B) |
| 790 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 791 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 792 | (__v32hi)_mm512_max_epi16(__A, __B), |
| 793 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 794 | } |
| 795 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 796 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 797 | _mm512_max_epu8 (__m512i __A, __m512i __B) |
| 798 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 799 | return (__m512i)__builtin_ia32_pmaxub512((__v64qi)__A, (__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 800 | } |
| 801 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 802 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 803 | _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) |
| 804 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 805 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 806 | (__v64qi)_mm512_max_epu8(__A, __B), |
| 807 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 808 | } |
| 809 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 810 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 811 | _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 812 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 813 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 814 | (__v64qi)_mm512_max_epu8(__A, __B), |
| 815 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 816 | } |
| 817 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 818 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 819 | _mm512_max_epu16 (__m512i __A, __m512i __B) |
| 820 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 821 | return (__m512i)__builtin_ia32_pmaxuw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 822 | } |
| 823 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 824 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 825 | _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) |
| 826 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 827 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 828 | (__v32hi)_mm512_max_epu16(__A, __B), |
| 829 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 830 | } |
| 831 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 832 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 833 | _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 834 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 835 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 836 | (__v32hi)_mm512_max_epu16(__A, __B), |
| 837 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 838 | } |
| 839 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 840 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 841 | _mm512_min_epi8 (__m512i __A, __m512i __B) |
| 842 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 843 | return (__m512i)__builtin_ia32_pminsb512((__v64qi) __A, (__v64qi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 844 | } |
| 845 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 846 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 847 | _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) |
| 848 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 849 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 850 | (__v64qi)_mm512_min_epi8(__A, __B), |
| 851 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 852 | } |
| 853 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 854 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 855 | _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 856 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 857 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 858 | (__v64qi)_mm512_min_epi8(__A, __B), |
| 859 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 860 | } |
| 861 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 862 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 863 | _mm512_min_epi16 (__m512i __A, __m512i __B) |
| 864 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 865 | return (__m512i)__builtin_ia32_pminsw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 866 | } |
| 867 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 868 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 869 | _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) |
| 870 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 871 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 872 | (__v32hi)_mm512_min_epi16(__A, __B), |
| 873 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 874 | } |
| 875 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 876 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 877 | _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 878 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 879 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 880 | (__v32hi)_mm512_min_epi16(__A, __B), |
| 881 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 882 | } |
| 883 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 884 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 885 | _mm512_min_epu8 (__m512i __A, __m512i __B) |
| 886 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 887 | return (__m512i)__builtin_ia32_pminub512((__v64qi)__A, (__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 888 | } |
| 889 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 890 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 891 | _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) |
| 892 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 893 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 894 | (__v64qi)_mm512_min_epu8(__A, __B), |
| 895 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 896 | } |
| 897 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 898 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 899 | _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 900 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 901 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, |
| 902 | (__v64qi)_mm512_min_epu8(__A, __B), |
| 903 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 904 | } |
| 905 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 906 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 907 | _mm512_min_epu16 (__m512i __A, __m512i __B) |
| 908 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 909 | return (__m512i)__builtin_ia32_pminuw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 910 | } |
| 911 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 912 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 913 | _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) |
| 914 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 915 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 916 | (__v32hi)_mm512_min_epu16(__A, __B), |
| 917 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 918 | } |
| 919 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 920 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 921 | _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 922 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 923 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 924 | (__v32hi)_mm512_min_epu16(__A, __B), |
| 925 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 926 | } |
| 927 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 928 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 929 | _mm512_shuffle_epi8(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 930 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 931 | return (__m512i)__builtin_ia32_pshufb512((__v64qi)__A,(__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 932 | } |
| 933 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 934 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 935 | _mm512_mask_shuffle_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 936 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 937 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 938 | (__v64qi)_mm512_shuffle_epi8(__A, __B), |
| 939 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 940 | } |
| 941 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 942 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 943 | _mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 944 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 945 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 946 | (__v64qi)_mm512_shuffle_epi8(__A, __B), |
| 947 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 948 | } |
| 949 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 950 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 951 | _mm512_subs_epi8 (__m512i __A, __m512i __B) |
| 952 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 953 | return (__m512i)__builtin_ia32_psubsb512((__v64qi)__A, (__v64qi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 954 | } |
| 955 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 956 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 957 | _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 958 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 959 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 960 | (__v64qi)_mm512_subs_epi8(__A, __B), |
| 961 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 962 | } |
| 963 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 964 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 965 | _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) |
| 966 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 967 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 968 | (__v64qi)_mm512_subs_epi8(__A, __B), |
| 969 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 970 | } |
| 971 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 972 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 973 | _mm512_subs_epi16 (__m512i __A, __m512i __B) |
| 974 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 975 | return (__m512i)__builtin_ia32_psubsw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 976 | } |
| 977 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 978 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 979 | _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 980 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 981 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 982 | (__v32hi)_mm512_subs_epi16(__A, __B), |
| 983 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 984 | } |
| 985 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 986 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 987 | _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 988 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 989 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 990 | (__v32hi)_mm512_subs_epi16(__A, __B), |
| 991 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 992 | } |
| 993 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 994 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 995 | _mm512_subs_epu8 (__m512i __A, __m512i __B) |
| 996 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 997 | return (__m512i)__builtin_ia32_psubusb512((__v64qi) __A, (__v64qi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 998 | } |
| 999 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1000 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 1001 | _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1002 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1003 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1004 | (__v64qi)_mm512_subs_epu8(__A, __B), |
| 1005 | (__v64qi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1006 | } |
| 1007 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1008 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1009 | _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B) |
| 1010 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1011 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1012 | (__v64qi)_mm512_subs_epu8(__A, __B), |
| 1013 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1014 | } |
| 1015 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1016 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1017 | _mm512_subs_epu16 (__m512i __A, __m512i __B) |
| 1018 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1019 | return (__m512i)__builtin_ia32_psubusw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1020 | } |
| 1021 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1022 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 1023 | _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1024 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1025 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1026 | (__v32hi)_mm512_subs_epu16(__A, __B), |
| 1027 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1028 | } |
| 1029 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1030 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1031 | _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 1032 | { |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1033 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1034 | (__v32hi)_mm512_subs_epu16(__A, __B), |
| 1035 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1036 | } |
| 1037 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1038 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1039 | _mm512_permutex2var_epi16(__m512i __A, __m512i __I, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1040 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1041 | return (__m512i)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I, |
| 1042 | (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1043 | } |
| 1044 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1045 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1046 | _mm512_mask_permutex2var_epi16(__m512i __A, __mmask32 __U, __m512i __I, |
| 1047 | __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1048 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1049 | return (__m512i)__builtin_ia32_selectw_512(__U, |
| 1050 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
| 1051 | (__v32hi)__A); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1052 | } |
| 1053 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1054 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1055 | _mm512_mask2_permutex2var_epi16(__m512i __A, __m512i __I, __mmask32 __U, |
| 1056 | __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1057 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1058 | return (__m512i)__builtin_ia32_selectw_512(__U, |
| 1059 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
| 1060 | (__v32hi)__I); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1061 | } |
| 1062 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1063 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1064 | _mm512_maskz_permutex2var_epi16(__mmask32 __U, __m512i __A, __m512i __I, |
| 1065 | __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1066 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1067 | return (__m512i)__builtin_ia32_selectw_512(__U, |
| 1068 | (__v32hi)_mm512_permutex2var_epi16(__A, __I, __B), |
| 1069 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1070 | } |
| 1071 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1072 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1073 | _mm512_mulhrs_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1074 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1075 | return (__m512i)__builtin_ia32_pmulhrsw512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1076 | } |
| 1077 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1078 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1079 | _mm512_mask_mulhrs_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1080 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1081 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1082 | (__v32hi)_mm512_mulhrs_epi16(__A, __B), |
| 1083 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1084 | } |
| 1085 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1086 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1087 | _mm512_maskz_mulhrs_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1088 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1089 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1090 | (__v32hi)_mm512_mulhrs_epi16(__A, __B), |
| 1091 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1092 | } |
| 1093 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1094 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1095 | _mm512_mulhi_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1096 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1097 | return (__m512i)__builtin_ia32_pmulhw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1098 | } |
| 1099 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1100 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1101 | _mm512_mask_mulhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1102 | __m512i __B) |
| 1103 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1104 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1105 | (__v32hi)_mm512_mulhi_epi16(__A, __B), |
| 1106 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1107 | } |
| 1108 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1109 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1110 | _mm512_maskz_mulhi_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1111 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1112 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1113 | (__v32hi)_mm512_mulhi_epi16(__A, __B), |
| 1114 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1115 | } |
| 1116 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1117 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1118 | _mm512_mulhi_epu16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1119 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1120 | return (__m512i)__builtin_ia32_pmulhuw512((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1121 | } |
| 1122 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1123 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1124 | _mm512_mask_mulhi_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1125 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1126 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1127 | (__v32hi)_mm512_mulhi_epu16(__A, __B), |
| 1128 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1129 | } |
| 1130 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1131 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1132 | _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B) |
| 1133 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1134 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1135 | (__v32hi)_mm512_mulhi_epu16(__A, __B), |
| 1136 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1137 | } |
| 1138 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1139 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1140 | _mm512_maddubs_epi16(__m512i __X, __m512i __Y) { |
| 1141 | return (__m512i)__builtin_ia32_pmaddubsw512((__v64qi)__X, (__v64qi)__Y); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1142 | } |
| 1143 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1144 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1145 | _mm512_mask_maddubs_epi16(__m512i __W, __mmask32 __U, __m512i __X, |
| 1146 | __m512i __Y) { |
| 1147 | return (__m512i)__builtin_ia32_selectw_512((__mmask32) __U, |
| 1148 | (__v32hi)_mm512_maddubs_epi16(__X, __Y), |
| 1149 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1150 | } |
| 1151 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1152 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1153 | _mm512_maskz_maddubs_epi16(__mmask32 __U, __m512i __X, __m512i __Y) { |
| 1154 | return (__m512i)__builtin_ia32_selectw_512((__mmask32) __U, |
| 1155 | (__v32hi)_mm512_maddubs_epi16(__X, __Y), |
| 1156 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1157 | } |
| 1158 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1159 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1160 | _mm512_madd_epi16(__m512i __A, __m512i __B) { |
| 1161 | return (__m512i)__builtin_ia32_pmaddwd512((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1162 | } |
| 1163 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1164 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1165 | _mm512_mask_madd_epi16(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) { |
| 1166 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__U, |
| 1167 | (__v16si)_mm512_madd_epi16(__A, __B), |
| 1168 | (__v16si)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1169 | } |
| 1170 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1171 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1172 | _mm512_maskz_madd_epi16(__mmask16 __U, __m512i __A, __m512i __B) { |
| 1173 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__U, |
| 1174 | (__v16si)_mm512_madd_epi16(__A, __B), |
| 1175 | (__v16si)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1176 | } |
| 1177 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1178 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1179 | _mm512_cvtsepi16_epi8 (__m512i __A) { |
| 1180 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
| 1181 | (__v32qi)_mm256_setzero_si256(), |
| 1182 | (__mmask32) -1); |
| 1183 | } |
| 1184 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1185 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1186 | _mm512_mask_cvtsepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
| 1187 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
| 1188 | (__v32qi)__O, |
| 1189 | __M); |
| 1190 | } |
| 1191 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1192 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1193 | _mm512_maskz_cvtsepi16_epi8 (__mmask32 __M, __m512i __A) { |
| 1194 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, |
| 1195 | (__v32qi) _mm256_setzero_si256(), |
| 1196 | __M); |
| 1197 | } |
| 1198 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1199 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1200 | _mm512_cvtusepi16_epi8 (__m512i __A) { |
| 1201 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
| 1202 | (__v32qi) _mm256_setzero_si256(), |
| 1203 | (__mmask32) -1); |
| 1204 | } |
| 1205 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1206 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1207 | _mm512_mask_cvtusepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
| 1208 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
| 1209 | (__v32qi) __O, |
| 1210 | __M); |
| 1211 | } |
| 1212 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1213 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1214 | _mm512_maskz_cvtusepi16_epi8 (__mmask32 __M, __m512i __A) { |
| 1215 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, |
| 1216 | (__v32qi) _mm256_setzero_si256(), |
| 1217 | __M); |
| 1218 | } |
| 1219 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1220 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1221 | _mm512_cvtepi16_epi8 (__m512i __A) { |
| 1222 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1223 | (__v32qi) _mm256_undefined_si256(), |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1224 | (__mmask32) -1); |
| 1225 | } |
| 1226 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1227 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1228 | _mm512_mask_cvtepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) { |
| 1229 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
| 1230 | (__v32qi) __O, |
| 1231 | __M); |
| 1232 | } |
| 1233 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1234 | static __inline__ __m256i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1235 | _mm512_maskz_cvtepi16_epi8 (__mmask32 __M, __m512i __A) { |
| 1236 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, |
| 1237 | (__v32qi) _mm256_setzero_si256(), |
| 1238 | __M); |
| 1239 | } |
| 1240 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1241 | static __inline__ void __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1242 | _mm512_mask_cvtepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
| 1243 | { |
| 1244 | __builtin_ia32_pmovwb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
| 1245 | } |
| 1246 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1247 | static __inline__ void __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1248 | _mm512_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
| 1249 | { |
| 1250 | __builtin_ia32_pmovswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
| 1251 | } |
| 1252 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1253 | static __inline__ void __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1254 | _mm512_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) |
| 1255 | { |
| 1256 | __builtin_ia32_pmovuswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); |
| 1257 | } |
| 1258 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1259 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1260 | _mm512_unpackhi_epi8(__m512i __A, __m512i __B) { |
| 1261 | return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B, |
| 1262 | 8, 64+8, 9, 64+9, |
| 1263 | 10, 64+10, 11, 64+11, |
| 1264 | 12, 64+12, 13, 64+13, |
| 1265 | 14, 64+14, 15, 64+15, |
| 1266 | 24, 64+24, 25, 64+25, |
| 1267 | 26, 64+26, 27, 64+27, |
| 1268 | 28, 64+28, 29, 64+29, |
| 1269 | 30, 64+30, 31, 64+31, |
| 1270 | 40, 64+40, 41, 64+41, |
| 1271 | 42, 64+42, 43, 64+43, |
| 1272 | 44, 64+44, 45, 64+45, |
| 1273 | 46, 64+46, 47, 64+47, |
| 1274 | 56, 64+56, 57, 64+57, |
| 1275 | 58, 64+58, 59, 64+59, |
| 1276 | 60, 64+60, 61, 64+61, |
| 1277 | 62, 64+62, 63, 64+63); |
| 1278 | } |
| 1279 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1280 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1281 | _mm512_mask_unpackhi_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
| 1282 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1283 | (__v64qi)_mm512_unpackhi_epi8(__A, __B), |
| 1284 | (__v64qi)__W); |
| 1285 | } |
| 1286 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1287 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1288 | _mm512_maskz_unpackhi_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
| 1289 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1290 | (__v64qi)_mm512_unpackhi_epi8(__A, __B), |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1291 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1292 | } |
| 1293 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1294 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1295 | _mm512_unpackhi_epi16(__m512i __A, __m512i __B) { |
| 1296 | return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B, |
| 1297 | 4, 32+4, 5, 32+5, |
| 1298 | 6, 32+6, 7, 32+7, |
| 1299 | 12, 32+12, 13, 32+13, |
| 1300 | 14, 32+14, 15, 32+15, |
| 1301 | 20, 32+20, 21, 32+21, |
| 1302 | 22, 32+22, 23, 32+23, |
| 1303 | 28, 32+28, 29, 32+29, |
| 1304 | 30, 32+30, 31, 32+31); |
| 1305 | } |
| 1306 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1307 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1308 | _mm512_mask_unpackhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
| 1309 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1310 | (__v32hi)_mm512_unpackhi_epi16(__A, __B), |
| 1311 | (__v32hi)__W); |
| 1312 | } |
| 1313 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1314 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1315 | _mm512_maskz_unpackhi_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
| 1316 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1317 | (__v32hi)_mm512_unpackhi_epi16(__A, __B), |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1318 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1319 | } |
| 1320 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1321 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1322 | _mm512_unpacklo_epi8(__m512i __A, __m512i __B) { |
| 1323 | return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B, |
| 1324 | 0, 64+0, 1, 64+1, |
| 1325 | 2, 64+2, 3, 64+3, |
| 1326 | 4, 64+4, 5, 64+5, |
| 1327 | 6, 64+6, 7, 64+7, |
| 1328 | 16, 64+16, 17, 64+17, |
| 1329 | 18, 64+18, 19, 64+19, |
| 1330 | 20, 64+20, 21, 64+21, |
| 1331 | 22, 64+22, 23, 64+23, |
| 1332 | 32, 64+32, 33, 64+33, |
| 1333 | 34, 64+34, 35, 64+35, |
| 1334 | 36, 64+36, 37, 64+37, |
| 1335 | 38, 64+38, 39, 64+39, |
| 1336 | 48, 64+48, 49, 64+49, |
| 1337 | 50, 64+50, 51, 64+51, |
| 1338 | 52, 64+52, 53, 64+53, |
| 1339 | 54, 64+54, 55, 64+55); |
| 1340 | } |
| 1341 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1342 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1343 | _mm512_mask_unpacklo_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { |
| 1344 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1345 | (__v64qi)_mm512_unpacklo_epi8(__A, __B), |
| 1346 | (__v64qi)__W); |
| 1347 | } |
| 1348 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1349 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1350 | _mm512_maskz_unpacklo_epi8(__mmask64 __U, __m512i __A, __m512i __B) { |
| 1351 | return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, |
| 1352 | (__v64qi)_mm512_unpacklo_epi8(__A, __B), |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1353 | (__v64qi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1354 | } |
| 1355 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1356 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1357 | _mm512_unpacklo_epi16(__m512i __A, __m512i __B) { |
| 1358 | return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B, |
| 1359 | 0, 32+0, 1, 32+1, |
| 1360 | 2, 32+2, 3, 32+3, |
| 1361 | 8, 32+8, 9, 32+9, |
| 1362 | 10, 32+10, 11, 32+11, |
| 1363 | 16, 32+16, 17, 32+17, |
| 1364 | 18, 32+18, 19, 32+19, |
| 1365 | 24, 32+24, 25, 32+25, |
| 1366 | 26, 32+26, 27, 32+27); |
| 1367 | } |
| 1368 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1369 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1370 | _mm512_mask_unpacklo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { |
| 1371 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1372 | (__v32hi)_mm512_unpacklo_epi16(__A, __B), |
| 1373 | (__v32hi)__W); |
| 1374 | } |
| 1375 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1376 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1377 | _mm512_maskz_unpacklo_epi16(__mmask32 __U, __m512i __A, __m512i __B) { |
| 1378 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1379 | (__v32hi)_mm512_unpacklo_epi16(__A, __B), |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1380 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1381 | } |
| 1382 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1383 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1384 | _mm512_cvtepi8_epi16(__m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1385 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1386 | /* This function always performs a signed extension, but __v32qi is a char |
| 1387 | which may be signed or unsigned, so use __v32qs. */ |
| 1388 | return (__m512i)__builtin_convertvector((__v32qs)__A, __v32hi); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1389 | } |
| 1390 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1391 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1392 | _mm512_mask_cvtepi8_epi16(__m512i __W, __mmask32 __U, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1393 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1394 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1395 | (__v32hi)_mm512_cvtepi8_epi16(__A), |
| 1396 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1397 | } |
| 1398 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1399 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1400 | _mm512_maskz_cvtepi8_epi16(__mmask32 __U, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1401 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1402 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1403 | (__v32hi)_mm512_cvtepi8_epi16(__A), |
| 1404 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1405 | } |
| 1406 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1407 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1408 | _mm512_cvtepu8_epi16(__m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1409 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1410 | return (__m512i)__builtin_convertvector((__v32qu)__A, __v32hi); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1411 | } |
| 1412 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1413 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1414 | _mm512_mask_cvtepu8_epi16(__m512i __W, __mmask32 __U, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1415 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1416 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1417 | (__v32hi)_mm512_cvtepu8_epi16(__A), |
| 1418 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1419 | } |
| 1420 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1421 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1422 | _mm512_maskz_cvtepu8_epi16(__mmask32 __U, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1423 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1424 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1425 | (__v32hi)_mm512_cvtepu8_epi16(__A), |
| 1426 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1427 | } |
| 1428 | |
| 1429 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1430 | #define _mm512_shufflehi_epi16(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1431 | ((__m512i)__builtin_ia32_pshufhw512((__v32hi)(__m512i)(A), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1432 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1433 | #define _mm512_mask_shufflehi_epi16(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1434 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 1435 | (__v32hi)_mm512_shufflehi_epi16((A), \ |
| 1436 | (imm)), \ |
| 1437 | (__v32hi)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1438 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1439 | #define _mm512_maskz_shufflehi_epi16(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1440 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 1441 | (__v32hi)_mm512_shufflehi_epi16((A), \ |
| 1442 | (imm)), \ |
| 1443 | (__v32hi)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1444 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1445 | #define _mm512_shufflelo_epi16(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1446 | ((__m512i)__builtin_ia32_pshuflw512((__v32hi)(__m512i)(A), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1447 | |
| 1448 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1449 | #define _mm512_mask_shufflelo_epi16(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1450 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 1451 | (__v32hi)_mm512_shufflelo_epi16((A), \ |
| 1452 | (imm)), \ |
| 1453 | (__v32hi)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1454 | |
| 1455 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1456 | #define _mm512_maskz_shufflelo_epi16(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1457 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 1458 | (__v32hi)_mm512_shufflelo_epi16((A), \ |
| 1459 | (imm)), \ |
| 1460 | (__v32hi)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1461 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1462 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1463 | _mm512_sllv_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1464 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1465 | return (__m512i)__builtin_ia32_psllv32hi((__v32hi) __A, (__v32hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1466 | } |
| 1467 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1468 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1469 | _mm512_mask_sllv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1470 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1471 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1472 | (__v32hi)_mm512_sllv_epi16(__A, __B), |
| 1473 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1474 | } |
| 1475 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1476 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1477 | _mm512_maskz_sllv_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1478 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1479 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1480 | (__v32hi)_mm512_sllv_epi16(__A, __B), |
| 1481 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1482 | } |
| 1483 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1484 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1485 | _mm512_sll_epi16(__m512i __A, __m128i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1486 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1487 | return (__m512i)__builtin_ia32_psllw512((__v32hi) __A, (__v8hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1488 | } |
| 1489 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1490 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1491 | _mm512_mask_sll_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1492 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1493 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1494 | (__v32hi)_mm512_sll_epi16(__A, __B), |
| 1495 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1496 | } |
| 1497 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1498 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1499 | _mm512_maskz_sll_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1500 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1501 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1502 | (__v32hi)_mm512_sll_epi16(__A, __B), |
| 1503 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1504 | } |
| 1505 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1506 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1507 | _mm512_slli_epi16(__m512i __A, unsigned int __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1508 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1509 | return (__m512i)__builtin_ia32_psllwi512((__v32hi)__A, __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1510 | } |
| 1511 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1512 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1513 | _mm512_mask_slli_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
| 1514 | unsigned int __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1515 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1516 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1517 | (__v32hi)_mm512_slli_epi16(__A, __B), |
| 1518 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1519 | } |
| 1520 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1521 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1522 | _mm512_maskz_slli_epi16(__mmask32 __U, __m512i __A, unsigned int __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1523 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1524 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1525 | (__v32hi)_mm512_slli_epi16(__A, __B), |
| 1526 | (__v32hi)_mm512_setzero_si512()); |
| 1527 | } |
| 1528 | |
| 1529 | #define _mm512_bslli_epi128(a, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1530 | ((__m512i)__builtin_ia32_pslldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm))) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1531 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1532 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1533 | _mm512_srlv_epi16(__m512i __A, __m512i __B) |
| 1534 | { |
| 1535 | return (__m512i)__builtin_ia32_psrlv32hi((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1536 | } |
| 1537 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1538 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1539 | _mm512_mask_srlv_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1540 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1541 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1542 | (__v32hi)_mm512_srlv_epi16(__A, __B), |
| 1543 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1544 | } |
| 1545 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1546 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1547 | _mm512_maskz_srlv_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1548 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1549 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1550 | (__v32hi)_mm512_srlv_epi16(__A, __B), |
| 1551 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1552 | } |
| 1553 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1554 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1555 | _mm512_srav_epi16(__m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1556 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1557 | return (__m512i)__builtin_ia32_psrav32hi((__v32hi)__A, (__v32hi)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1558 | } |
| 1559 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1560 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1561 | _mm512_mask_srav_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1562 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1563 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1564 | (__v32hi)_mm512_srav_epi16(__A, __B), |
| 1565 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1566 | } |
| 1567 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1568 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1569 | _mm512_maskz_srav_epi16(__mmask32 __U, __m512i __A, __m512i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1570 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1571 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1572 | (__v32hi)_mm512_srav_epi16(__A, __B), |
| 1573 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1574 | } |
| 1575 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1576 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1577 | _mm512_sra_epi16(__m512i __A, __m128i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1578 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1579 | return (__m512i)__builtin_ia32_psraw512((__v32hi) __A, (__v8hi) __B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1580 | } |
| 1581 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1582 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1583 | _mm512_mask_sra_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1584 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1585 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1586 | (__v32hi)_mm512_sra_epi16(__A, __B), |
| 1587 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1588 | } |
| 1589 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1590 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1591 | _mm512_maskz_sra_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
| 1592 | { |
| 1593 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1594 | (__v32hi)_mm512_sra_epi16(__A, __B), |
| 1595 | (__v32hi)_mm512_setzero_si512()); |
| 1596 | } |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1597 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1598 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1599 | _mm512_srai_epi16(__m512i __A, unsigned int __B) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1600 | { |
| 1601 | return (__m512i)__builtin_ia32_psrawi512((__v32hi)__A, __B); |
| 1602 | } |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1603 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1604 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1605 | _mm512_mask_srai_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
| 1606 | unsigned int __B) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1607 | { |
| 1608 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1609 | (__v32hi)_mm512_srai_epi16(__A, __B), |
| 1610 | (__v32hi)__W); |
| 1611 | } |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1612 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1613 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1614 | _mm512_maskz_srai_epi16(__mmask32 __U, __m512i __A, unsigned int __B) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1615 | { |
| 1616 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1617 | (__v32hi)_mm512_srai_epi16(__A, __B), |
| 1618 | (__v32hi)_mm512_setzero_si512()); |
| 1619 | } |
| 1620 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1621 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1622 | _mm512_srl_epi16(__m512i __A, __m128i __B) |
| 1623 | { |
| 1624 | return (__m512i)__builtin_ia32_psrlw512((__v32hi) __A, (__v8hi) __B); |
| 1625 | } |
| 1626 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1627 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1628 | _mm512_mask_srl_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B) |
| 1629 | { |
| 1630 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1631 | (__v32hi)_mm512_srl_epi16(__A, __B), |
| 1632 | (__v32hi)__W); |
| 1633 | } |
| 1634 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1635 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1636 | _mm512_maskz_srl_epi16(__mmask32 __U, __m512i __A, __m128i __B) |
| 1637 | { |
| 1638 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1639 | (__v32hi)_mm512_srl_epi16(__A, __B), |
| 1640 | (__v32hi)_mm512_setzero_si512()); |
| 1641 | } |
| 1642 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1643 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1644 | _mm512_srli_epi16(__m512i __A, unsigned int __B) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1645 | { |
| 1646 | return (__m512i)__builtin_ia32_psrlwi512((__v32hi)__A, __B); |
| 1647 | } |
| 1648 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1649 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Sasha Smundak | 0fc590b | 2020-10-07 08:11:59 -0700 | [diff] [blame] | 1650 | _mm512_mask_srli_epi16(__m512i __W, __mmask32 __U, __m512i __A, |
| 1651 | unsigned int __B) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1652 | { |
| 1653 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1654 | (__v32hi)_mm512_srli_epi16(__A, __B), |
| 1655 | (__v32hi)__W); |
| 1656 | } |
| 1657 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1658 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1659 | _mm512_maskz_srli_epi16(__mmask32 __U, __m512i __A, int __B) |
| 1660 | { |
| 1661 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, |
| 1662 | (__v32hi)_mm512_srli_epi16(__A, __B), |
| 1663 | (__v32hi)_mm512_setzero_si512()); |
| 1664 | } |
| 1665 | |
| 1666 | #define _mm512_bsrli_epi128(a, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1667 | ((__m512i)__builtin_ia32_psrldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1668 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1669 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1670 | _mm512_mask_mov_epi16 (__m512i __W, __mmask32 __U, __m512i __A) |
| 1671 | { |
| 1672 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
| 1673 | (__v32hi) __A, |
| 1674 | (__v32hi) __W); |
| 1675 | } |
| 1676 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1677 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1678 | _mm512_maskz_mov_epi16 (__mmask32 __U, __m512i __A) |
| 1679 | { |
| 1680 | return (__m512i) __builtin_ia32_selectw_512 ((__mmask32) __U, |
| 1681 | (__v32hi) __A, |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1682 | (__v32hi) _mm512_setzero_si512 ()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1683 | } |
| 1684 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1685 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1686 | _mm512_mask_mov_epi8 (__m512i __W, __mmask64 __U, __m512i __A) |
| 1687 | { |
| 1688 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
| 1689 | (__v64qi) __A, |
| 1690 | (__v64qi) __W); |
| 1691 | } |
| 1692 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1693 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1694 | _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A) |
| 1695 | { |
| 1696 | return (__m512i) __builtin_ia32_selectb_512 ((__mmask64) __U, |
| 1697 | (__v64qi) __A, |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1698 | (__v64qi) _mm512_setzero_si512 ()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1699 | } |
| 1700 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1701 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1702 | _mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A) |
| 1703 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1704 | return (__m512i) __builtin_ia32_selectb_512(__M, |
| 1705 | (__v64qi)_mm512_set1_epi8(__A), |
| 1706 | (__v64qi) __O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1707 | } |
| 1708 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1709 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1710 | _mm512_maskz_set1_epi8 (__mmask64 __M, char __A) |
| 1711 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1712 | return (__m512i) __builtin_ia32_selectb_512(__M, |
| 1713 | (__v64qi) _mm512_set1_epi8(__A), |
| 1714 | (__v64qi) _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1715 | } |
| 1716 | |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 1717 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1718 | _mm512_kunpackd (__mmask64 __A, __mmask64 __B) |
| 1719 | { |
| 1720 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, |
| 1721 | (__mmask64) __B); |
| 1722 | } |
| 1723 | |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 1724 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1725 | _mm512_kunpackw (__mmask32 __A, __mmask32 __B) |
| 1726 | { |
| 1727 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, |
| 1728 | (__mmask32) __B); |
| 1729 | } |
| 1730 | |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1731 | static __inline __m512i __DEFAULT_FN_ATTRS512 |
| 1732 | _mm512_loadu_epi16 (void const *__P) |
| 1733 | { |
| 1734 | struct __loadu_epi16 { |
Logan Chien | dbcf412 | 2019-03-21 10:50:25 +0800 | [diff] [blame] | 1735 | __m512i_u __v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1736 | } __attribute__((__packed__, __may_alias__)); |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1737 | return ((const struct __loadu_epi16*)__P)->__v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1738 | } |
| 1739 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1740 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1741 | _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P) |
| 1742 | { |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1743 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const __v32hi *) __P, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1744 | (__v32hi) __W, |
| 1745 | (__mmask32) __U); |
| 1746 | } |
| 1747 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1748 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1749 | _mm512_maskz_loadu_epi16 (__mmask32 __U, void const *__P) |
| 1750 | { |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1751 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const __v32hi *) __P, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1752 | (__v32hi) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1753 | _mm512_setzero_si512 (), |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1754 | (__mmask32) __U); |
| 1755 | } |
| 1756 | |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1757 | static __inline __m512i __DEFAULT_FN_ATTRS512 |
| 1758 | _mm512_loadu_epi8 (void const *__P) |
| 1759 | { |
| 1760 | struct __loadu_epi8 { |
Logan Chien | dbcf412 | 2019-03-21 10:50:25 +0800 | [diff] [blame] | 1761 | __m512i_u __v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1762 | } __attribute__((__packed__, __may_alias__)); |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1763 | return ((const struct __loadu_epi8*)__P)->__v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1764 | } |
| 1765 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1766 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1767 | _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P) |
| 1768 | { |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1769 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const __v64qi *) __P, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1770 | (__v64qi) __W, |
| 1771 | (__mmask64) __U); |
| 1772 | } |
| 1773 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1774 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1775 | _mm512_maskz_loadu_epi8 (__mmask64 __U, void const *__P) |
| 1776 | { |
Sasha Smundak | 33d5ddd | 2020-05-04 13:37:26 -0700 | [diff] [blame] | 1777 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const __v64qi *) __P, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1778 | (__v64qi) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1779 | _mm512_setzero_si512 (), |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1780 | (__mmask64) __U); |
| 1781 | } |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1782 | |
| 1783 | static __inline void __DEFAULT_FN_ATTRS512 |
| 1784 | _mm512_storeu_epi16 (void *__P, __m512i __A) |
| 1785 | { |
| 1786 | struct __storeu_epi16 { |
Logan Chien | dbcf412 | 2019-03-21 10:50:25 +0800 | [diff] [blame] | 1787 | __m512i_u __v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1788 | } __attribute__((__packed__, __may_alias__)); |
| 1789 | ((struct __storeu_epi16*)__P)->__v = __A; |
| 1790 | } |
| 1791 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1792 | static __inline__ void __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1793 | _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A) |
| 1794 | { |
| 1795 | __builtin_ia32_storedquhi512_mask ((__v32hi *) __P, |
| 1796 | (__v32hi) __A, |
| 1797 | (__mmask32) __U); |
| 1798 | } |
| 1799 | |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1800 | static __inline void __DEFAULT_FN_ATTRS512 |
| 1801 | _mm512_storeu_epi8 (void *__P, __m512i __A) |
| 1802 | { |
| 1803 | struct __storeu_epi8 { |
Logan Chien | dbcf412 | 2019-03-21 10:50:25 +0800 | [diff] [blame] | 1804 | __m512i_u __v; |
Logan Chien | 969aea6 | 2018-12-05 18:40:57 +0800 | [diff] [blame] | 1805 | } __attribute__((__packed__, __may_alias__)); |
| 1806 | ((struct __storeu_epi8*)__P)->__v = __A; |
| 1807 | } |
| 1808 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1809 | static __inline__ void __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1810 | _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A) |
| 1811 | { |
| 1812 | __builtin_ia32_storedquqi512_mask ((__v64qi *) __P, |
| 1813 | (__v64qi) __A, |
| 1814 | (__mmask64) __U); |
| 1815 | } |
| 1816 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1817 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1818 | _mm512_test_epi8_mask (__m512i __A, __m512i __B) |
| 1819 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1820 | return _mm512_cmpneq_epi8_mask (_mm512_and_epi32 (__A, __B), |
| 1821 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1822 | } |
| 1823 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1824 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1825 | _mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) |
| 1826 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1827 | return _mm512_mask_cmpneq_epi8_mask (__U, _mm512_and_epi32 (__A, __B), |
| 1828 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1829 | } |
| 1830 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1831 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1832 | _mm512_test_epi16_mask (__m512i __A, __m512i __B) |
| 1833 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1834 | return _mm512_cmpneq_epi16_mask (_mm512_and_epi32 (__A, __B), |
| 1835 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1836 | } |
| 1837 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1838 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1839 | _mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) |
| 1840 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1841 | return _mm512_mask_cmpneq_epi16_mask (__U, _mm512_and_epi32 (__A, __B), |
| 1842 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1843 | } |
| 1844 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1845 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1846 | _mm512_testn_epi8_mask (__m512i __A, __m512i __B) |
| 1847 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1848 | return _mm512_cmpeq_epi8_mask (_mm512_and_epi32 (__A, __B), _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1849 | } |
| 1850 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1851 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1852 | _mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) |
| 1853 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1854 | return _mm512_mask_cmpeq_epi8_mask (__U, _mm512_and_epi32 (__A, __B), |
| 1855 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1856 | } |
| 1857 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1858 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1859 | _mm512_testn_epi16_mask (__m512i __A, __m512i __B) |
| 1860 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1861 | return _mm512_cmpeq_epi16_mask (_mm512_and_epi32 (__A, __B), |
| 1862 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1863 | } |
| 1864 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1865 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1866 | _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) |
| 1867 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1868 | return _mm512_mask_cmpeq_epi16_mask (__U, _mm512_and_epi32 (__A, __B), |
| 1869 | _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1870 | } |
| 1871 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1872 | static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1873 | _mm512_movepi8_mask (__m512i __A) |
| 1874 | { |
| 1875 | return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A); |
| 1876 | } |
| 1877 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1878 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1879 | _mm512_movepi16_mask (__m512i __A) |
| 1880 | { |
| 1881 | return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); |
| 1882 | } |
| 1883 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1884 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1885 | _mm512_movm_epi8 (__mmask64 __A) |
| 1886 | { |
| 1887 | return (__m512i) __builtin_ia32_cvtmask2b512 (__A); |
| 1888 | } |
| 1889 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1890 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1891 | _mm512_movm_epi16 (__mmask32 __A) |
| 1892 | { |
| 1893 | return (__m512i) __builtin_ia32_cvtmask2w512 (__A); |
| 1894 | } |
| 1895 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1896 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1897 | _mm512_broadcastb_epi8 (__m128i __A) |
| 1898 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1899 | return (__m512i)__builtin_shufflevector((__v16qi) __A, (__v16qi) __A, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1900 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1901 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1902 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1903 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); |
| 1904 | } |
| 1905 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1906 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1907 | _mm512_mask_broadcastb_epi8 (__m512i __O, __mmask64 __M, __m128i __A) |
| 1908 | { |
| 1909 | return (__m512i)__builtin_ia32_selectb_512(__M, |
| 1910 | (__v64qi) _mm512_broadcastb_epi8(__A), |
| 1911 | (__v64qi) __O); |
| 1912 | } |
| 1913 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1914 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1915 | _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A) |
| 1916 | { |
| 1917 | return (__m512i)__builtin_ia32_selectb_512(__M, |
| 1918 | (__v64qi) _mm512_broadcastb_epi8(__A), |
| 1919 | (__v64qi) _mm512_setzero_si512()); |
| 1920 | } |
| 1921 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1922 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1923 | _mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A) |
| 1924 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1925 | return (__m512i) __builtin_ia32_selectw_512(__M, |
| 1926 | (__v32hi) _mm512_set1_epi16(__A), |
| 1927 | (__v32hi) __O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1928 | } |
| 1929 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1930 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1931 | _mm512_maskz_set1_epi16 (__mmask32 __M, short __A) |
| 1932 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1933 | return (__m512i) __builtin_ia32_selectw_512(__M, |
| 1934 | (__v32hi) _mm512_set1_epi16(__A), |
| 1935 | (__v32hi) _mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1936 | } |
| 1937 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1938 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1939 | _mm512_broadcastw_epi16 (__m128i __A) |
| 1940 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1941 | return (__m512i)__builtin_shufflevector((__v8hi) __A, (__v8hi) __A, |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1942 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1943 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); |
| 1944 | } |
| 1945 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1946 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1947 | _mm512_mask_broadcastw_epi16 (__m512i __O, __mmask32 __M, __m128i __A) |
| 1948 | { |
| 1949 | return (__m512i)__builtin_ia32_selectw_512(__M, |
| 1950 | (__v32hi) _mm512_broadcastw_epi16(__A), |
| 1951 | (__v32hi) __O); |
| 1952 | } |
| 1953 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1954 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1955 | _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A) |
| 1956 | { |
| 1957 | return (__m512i)__builtin_ia32_selectw_512(__M, |
| 1958 | (__v32hi) _mm512_broadcastw_epi16(__A), |
| 1959 | (__v32hi) _mm512_setzero_si512()); |
| 1960 | } |
| 1961 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1962 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1963 | _mm512_permutexvar_epi16 (__m512i __A, __m512i __B) |
| 1964 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1965 | return (__m512i)__builtin_ia32_permvarhi512((__v32hi)__B, (__v32hi)__A); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1966 | } |
| 1967 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1968 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1969 | _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A, |
| 1970 | __m512i __B) |
| 1971 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1972 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 1973 | (__v32hi)_mm512_permutexvar_epi16(__A, __B), |
| 1974 | (__v32hi)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1975 | } |
| 1976 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1977 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1978 | _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A, |
| 1979 | __m512i __B) |
| 1980 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1981 | return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M, |
| 1982 | (__v32hi)_mm512_permutexvar_epi16(__A, __B), |
| 1983 | (__v32hi)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1984 | } |
| 1985 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1986 | #define _mm512_alignr_epi8(A, B, N) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1987 | ((__m512i)__builtin_ia32_palignr512((__v64qi)(__m512i)(A), \ |
| 1988 | (__v64qi)(__m512i)(B), (int)(N))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1989 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1990 | #define _mm512_mask_alignr_epi8(W, U, A, B, N) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1991 | ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \ |
| 1992 | (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \ |
| 1993 | (__v64qi)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1994 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1995 | #define _mm512_maskz_alignr_epi8(U, A, B, N) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1996 | ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1997 | (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1998 | (__v64qi)(__m512i)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1999 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 2000 | #define _mm512_dbsad_epu8(A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 2001 | ((__m512i)__builtin_ia32_dbpsadbw512((__v64qi)(__m512i)(A), \ |
| 2002 | (__v64qi)(__m512i)(B), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 2003 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 2004 | #define _mm512_mask_dbsad_epu8(W, U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 2005 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 2006 | (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 2007 | (__v32hi)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 2008 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 2009 | #define _mm512_maskz_dbsad_epu8(U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 2010 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 2011 | (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 2012 | (__v32hi)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 2013 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 2014 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 2015 | _mm512_sad_epu8 (__m512i __A, __m512i __B) |
| 2016 | { |
| 2017 | return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A, |
| 2018 | (__v64qi) __B); |
| 2019 | } |
| 2020 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 2021 | #undef __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 2022 | #undef __DEFAULT_FN_ATTRS |
| 2023 | |
| 2024 | #endif |