Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1 | /*===---- avx512dqintrin.h - AVX512DQ intrinsics ---------------------------=== |
| 2 | * |
Logan Chien | df4f766 | 2019-09-04 16:45:23 -0700 | [diff] [blame] | 3 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | * See https://llvm.org/LICENSE.txt for license information. |
| 5 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 6 | * |
| 7 | *===-----------------------------------------------------------------------=== |
| 8 | */ |
| 9 | |
| 10 | #ifndef __IMMINTRIN_H |
| 11 | #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifndef __AVX512DQINTRIN_H |
| 15 | #define __AVX512DQINTRIN_H |
| 16 | |
| 17 | /* Define the default attributes for the functions in this file. */ |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 18 | #define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), __min_vector_width__(512))) |
| 19 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 20 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 21 | static __inline __mmask8 __DEFAULT_FN_ATTRS |
| 22 | _knot_mask8(__mmask8 __M) |
| 23 | { |
| 24 | return __builtin_ia32_knotqi(__M); |
| 25 | } |
| 26 | |
| 27 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 28 | _kand_mask8(__mmask8 __A, __mmask8 __B) |
| 29 | { |
| 30 | return (__mmask8)__builtin_ia32_kandqi((__mmask8)__A, (__mmask8)__B); |
| 31 | } |
| 32 | |
| 33 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 34 | _kandn_mask8(__mmask8 __A, __mmask8 __B) |
| 35 | { |
| 36 | return (__mmask8)__builtin_ia32_kandnqi((__mmask8)__A, (__mmask8)__B); |
| 37 | } |
| 38 | |
| 39 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 40 | _kor_mask8(__mmask8 __A, __mmask8 __B) |
| 41 | { |
| 42 | return (__mmask8)__builtin_ia32_korqi((__mmask8)__A, (__mmask8)__B); |
| 43 | } |
| 44 | |
| 45 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 46 | _kxnor_mask8(__mmask8 __A, __mmask8 __B) |
| 47 | { |
| 48 | return (__mmask8)__builtin_ia32_kxnorqi((__mmask8)__A, (__mmask8)__B); |
| 49 | } |
| 50 | |
| 51 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 52 | _kxor_mask8(__mmask8 __A, __mmask8 __B) |
| 53 | { |
| 54 | return (__mmask8)__builtin_ia32_kxorqi((__mmask8)__A, (__mmask8)__B); |
| 55 | } |
| 56 | |
| 57 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 58 | _kortestc_mask8_u8(__mmask8 __A, __mmask8 __B) |
| 59 | { |
| 60 | return (unsigned char)__builtin_ia32_kortestcqi(__A, __B); |
| 61 | } |
| 62 | |
| 63 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 64 | _kortestz_mask8_u8(__mmask8 __A, __mmask8 __B) |
| 65 | { |
| 66 | return (unsigned char)__builtin_ia32_kortestzqi(__A, __B); |
| 67 | } |
| 68 | |
| 69 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 70 | _kortest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) { |
| 71 | *__C = (unsigned char)__builtin_ia32_kortestcqi(__A, __B); |
| 72 | return (unsigned char)__builtin_ia32_kortestzqi(__A, __B); |
| 73 | } |
| 74 | |
| 75 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 76 | _ktestc_mask8_u8(__mmask8 __A, __mmask8 __B) |
| 77 | { |
| 78 | return (unsigned char)__builtin_ia32_ktestcqi(__A, __B); |
| 79 | } |
| 80 | |
| 81 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 82 | _ktestz_mask8_u8(__mmask8 __A, __mmask8 __B) |
| 83 | { |
| 84 | return (unsigned char)__builtin_ia32_ktestzqi(__A, __B); |
| 85 | } |
| 86 | |
| 87 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 88 | _ktest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) { |
| 89 | *__C = (unsigned char)__builtin_ia32_ktestcqi(__A, __B); |
| 90 | return (unsigned char)__builtin_ia32_ktestzqi(__A, __B); |
| 91 | } |
| 92 | |
| 93 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 94 | _ktestc_mask16_u8(__mmask16 __A, __mmask16 __B) |
| 95 | { |
| 96 | return (unsigned char)__builtin_ia32_ktestchi(__A, __B); |
| 97 | } |
| 98 | |
| 99 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 100 | _ktestz_mask16_u8(__mmask16 __A, __mmask16 __B) |
| 101 | { |
| 102 | return (unsigned char)__builtin_ia32_ktestzhi(__A, __B); |
| 103 | } |
| 104 | |
| 105 | static __inline__ unsigned char __DEFAULT_FN_ATTRS |
| 106 | _ktest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C) { |
| 107 | *__C = (unsigned char)__builtin_ia32_ktestchi(__A, __B); |
| 108 | return (unsigned char)__builtin_ia32_ktestzhi(__A, __B); |
| 109 | } |
| 110 | |
| 111 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 112 | _kadd_mask8(__mmask8 __A, __mmask8 __B) |
| 113 | { |
| 114 | return (__mmask8)__builtin_ia32_kaddqi((__mmask8)__A, (__mmask8)__B); |
| 115 | } |
| 116 | |
| 117 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS |
| 118 | _kadd_mask16(__mmask16 __A, __mmask16 __B) |
| 119 | { |
| 120 | return (__mmask16)__builtin_ia32_kaddhi((__mmask16)__A, (__mmask16)__B); |
| 121 | } |
| 122 | |
| 123 | #define _kshiftli_mask8(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 124 | ((__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 125 | |
| 126 | #define _kshiftri_mask8(A, I) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 127 | ((__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I))) |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 128 | |
| 129 | static __inline__ unsigned int __DEFAULT_FN_ATTRS |
| 130 | _cvtmask8_u32(__mmask8 __A) { |
| 131 | return (unsigned int)__builtin_ia32_kmovb((__mmask8)__A); |
| 132 | } |
| 133 | |
| 134 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 135 | _cvtu32_mask8(unsigned int __A) { |
| 136 | return (__mmask8)__builtin_ia32_kmovb((__mmask8)__A); |
| 137 | } |
| 138 | |
| 139 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 140 | _load_mask8(__mmask8 *__A) { |
| 141 | return (__mmask8)__builtin_ia32_kmovb(*(__mmask8 *)__A); |
| 142 | } |
| 143 | |
| 144 | static __inline__ void __DEFAULT_FN_ATTRS |
| 145 | _store_mask8(__mmask8 *__A, __mmask8 __B) { |
| 146 | *(__mmask8 *)__A = __builtin_ia32_kmovb((__mmask8)__B); |
| 147 | } |
| 148 | |
| 149 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 150 | _mm512_mullo_epi64 (__m512i __A, __m512i __B) { |
| 151 | return (__m512i) ((__v8du) __A * (__v8du) __B); |
| 152 | } |
| 153 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 154 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 155 | _mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) { |
| 156 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, |
| 157 | (__v8di)_mm512_mullo_epi64(__A, __B), |
| 158 | (__v8di)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 159 | } |
| 160 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 161 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 162 | _mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B) { |
| 163 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, |
| 164 | (__v8di)_mm512_mullo_epi64(__A, __B), |
| 165 | (__v8di)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 166 | } |
| 167 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 168 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 169 | _mm512_xor_pd(__m512d __A, __m512d __B) { |
| 170 | return (__m512d)((__v8du)__A ^ (__v8du)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 171 | } |
| 172 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 173 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 174 | _mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 175 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 176 | (__v8df)_mm512_xor_pd(__A, __B), |
| 177 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 178 | } |
| 179 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 180 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 181 | _mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 182 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 183 | (__v8df)_mm512_xor_pd(__A, __B), |
| 184 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 185 | } |
| 186 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 187 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 188 | _mm512_xor_ps (__m512 __A, __m512 __B) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 189 | return (__m512)((__v16su)__A ^ (__v16su)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 190 | } |
| 191 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 192 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 193 | _mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 194 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 195 | (__v16sf)_mm512_xor_ps(__A, __B), |
| 196 | (__v16sf)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 197 | } |
| 198 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 199 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 200 | _mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 201 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 202 | (__v16sf)_mm512_xor_ps(__A, __B), |
| 203 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 204 | } |
| 205 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 206 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 207 | _mm512_or_pd(__m512d __A, __m512d __B) { |
| 208 | return (__m512d)((__v8du)__A | (__v8du)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 209 | } |
| 210 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 211 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 212 | _mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 213 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 214 | (__v8df)_mm512_or_pd(__A, __B), |
| 215 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 216 | } |
| 217 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 218 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 219 | _mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 220 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 221 | (__v8df)_mm512_or_pd(__A, __B), |
| 222 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 223 | } |
| 224 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 225 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 226 | _mm512_or_ps(__m512 __A, __m512 __B) { |
| 227 | return (__m512)((__v16su)__A | (__v16su)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 228 | } |
| 229 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 230 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 231 | _mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 232 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 233 | (__v16sf)_mm512_or_ps(__A, __B), |
| 234 | (__v16sf)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 235 | } |
| 236 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 237 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 238 | _mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 239 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 240 | (__v16sf)_mm512_or_ps(__A, __B), |
| 241 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 242 | } |
| 243 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 244 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 245 | _mm512_and_pd(__m512d __A, __m512d __B) { |
| 246 | return (__m512d)((__v8du)__A & (__v8du)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 247 | } |
| 248 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 249 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 250 | _mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 251 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 252 | (__v8df)_mm512_and_pd(__A, __B), |
| 253 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 254 | } |
| 255 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 256 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 257 | _mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 258 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 259 | (__v8df)_mm512_and_pd(__A, __B), |
| 260 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 261 | } |
| 262 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 263 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 264 | _mm512_and_ps(__m512 __A, __m512 __B) { |
| 265 | return (__m512)((__v16su)__A & (__v16su)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 266 | } |
| 267 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 268 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 269 | _mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 270 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 271 | (__v16sf)_mm512_and_ps(__A, __B), |
| 272 | (__v16sf)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 273 | } |
| 274 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 275 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 276 | _mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 277 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 278 | (__v16sf)_mm512_and_ps(__A, __B), |
| 279 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 280 | } |
| 281 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 282 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 283 | _mm512_andnot_pd(__m512d __A, __m512d __B) { |
| 284 | return (__m512d)(~(__v8du)__A & (__v8du)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 285 | } |
| 286 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 287 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 288 | _mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 289 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 290 | (__v8df)_mm512_andnot_pd(__A, __B), |
| 291 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 292 | } |
| 293 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 294 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 295 | _mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 296 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 297 | (__v8df)_mm512_andnot_pd(__A, __B), |
| 298 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 299 | } |
| 300 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 301 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 302 | _mm512_andnot_ps(__m512 __A, __m512 __B) { |
| 303 | return (__m512)(~(__v16su)__A & (__v16su)__B); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 304 | } |
| 305 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 306 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 307 | _mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 308 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 309 | (__v16sf)_mm512_andnot_ps(__A, __B), |
| 310 | (__v16sf)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 311 | } |
| 312 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 313 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 314 | _mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 315 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 316 | (__v16sf)_mm512_andnot_ps(__A, __B), |
| 317 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 318 | } |
| 319 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 320 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 321 | _mm512_cvtpd_epi64 (__m512d __A) { |
| 322 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 323 | (__v8di) _mm512_setzero_si512(), |
| 324 | (__mmask8) -1, |
| 325 | _MM_FROUND_CUR_DIRECTION); |
| 326 | } |
| 327 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 328 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 329 | _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 330 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 331 | (__v8di) __W, |
| 332 | (__mmask8) __U, |
| 333 | _MM_FROUND_CUR_DIRECTION); |
| 334 | } |
| 335 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 336 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 337 | _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) { |
| 338 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 339 | (__v8di) _mm512_setzero_si512(), |
| 340 | (__mmask8) __U, |
| 341 | _MM_FROUND_CUR_DIRECTION); |
| 342 | } |
| 343 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 344 | #define _mm512_cvt_roundpd_epi64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 345 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 346 | (__v8di)_mm512_setzero_si512(), \ |
| 347 | (__mmask8)-1, (int)(R))) |
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_mask_cvt_roundpd_epi64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 350 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 351 | (__v8di)(__m512i)(W), \ |
| 352 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 353 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 354 | #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 355 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 356 | (__v8di)_mm512_setzero_si512(), \ |
| 357 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 358 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 359 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 360 | _mm512_cvtpd_epu64 (__m512d __A) { |
| 361 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 362 | (__v8di) _mm512_setzero_si512(), |
| 363 | (__mmask8) -1, |
| 364 | _MM_FROUND_CUR_DIRECTION); |
| 365 | } |
| 366 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 367 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 368 | _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 369 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 370 | (__v8di) __W, |
| 371 | (__mmask8) __U, |
| 372 | _MM_FROUND_CUR_DIRECTION); |
| 373 | } |
| 374 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 375 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 376 | _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) { |
| 377 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 378 | (__v8di) _mm512_setzero_si512(), |
| 379 | (__mmask8) __U, |
| 380 | _MM_FROUND_CUR_DIRECTION); |
| 381 | } |
| 382 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 383 | #define _mm512_cvt_roundpd_epu64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 384 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 385 | (__v8di)_mm512_setzero_si512(), \ |
| 386 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 387 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 388 | #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 389 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 390 | (__v8di)(__m512i)(W), \ |
| 391 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 392 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 393 | #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 394 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 395 | (__v8di)_mm512_setzero_si512(), \ |
| 396 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 397 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 398 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 399 | _mm512_cvtps_epi64 (__m256 __A) { |
| 400 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 401 | (__v8di) _mm512_setzero_si512(), |
| 402 | (__mmask8) -1, |
| 403 | _MM_FROUND_CUR_DIRECTION); |
| 404 | } |
| 405 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 406 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 407 | _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 408 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 409 | (__v8di) __W, |
| 410 | (__mmask8) __U, |
| 411 | _MM_FROUND_CUR_DIRECTION); |
| 412 | } |
| 413 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 414 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 415 | _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) { |
| 416 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 417 | (__v8di) _mm512_setzero_si512(), |
| 418 | (__mmask8) __U, |
| 419 | _MM_FROUND_CUR_DIRECTION); |
| 420 | } |
| 421 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 422 | #define _mm512_cvt_roundps_epi64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 423 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 424 | (__v8di)_mm512_setzero_si512(), \ |
| 425 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 426 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 427 | #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 428 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 429 | (__v8di)(__m512i)(W), \ |
| 430 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 431 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 432 | #define _mm512_maskz_cvt_roundps_epi64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 433 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 434 | (__v8di)_mm512_setzero_si512(), \ |
| 435 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 436 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 437 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 438 | _mm512_cvtps_epu64 (__m256 __A) { |
| 439 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 440 | (__v8di) _mm512_setzero_si512(), |
| 441 | (__mmask8) -1, |
| 442 | _MM_FROUND_CUR_DIRECTION); |
| 443 | } |
| 444 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 445 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 446 | _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 447 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 448 | (__v8di) __W, |
| 449 | (__mmask8) __U, |
| 450 | _MM_FROUND_CUR_DIRECTION); |
| 451 | } |
| 452 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 453 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 454 | _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) { |
| 455 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 456 | (__v8di) _mm512_setzero_si512(), |
| 457 | (__mmask8) __U, |
| 458 | _MM_FROUND_CUR_DIRECTION); |
| 459 | } |
| 460 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 461 | #define _mm512_cvt_roundps_epu64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 462 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 463 | (__v8di)_mm512_setzero_si512(), \ |
| 464 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 465 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 466 | #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 467 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 468 | (__v8di)(__m512i)(W), \ |
| 469 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 470 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 471 | #define _mm512_maskz_cvt_roundps_epu64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 472 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 473 | (__v8di)_mm512_setzero_si512(), \ |
| 474 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 475 | |
| 476 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 477 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 478 | _mm512_cvtepi64_pd (__m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 479 | return (__m512d)__builtin_convertvector((__v8di)__A, __v8df); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 480 | } |
| 481 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 482 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 483 | _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 484 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 485 | (__v8df)_mm512_cvtepi64_pd(__A), |
| 486 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 487 | } |
| 488 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 489 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 490 | _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 491 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 492 | (__v8df)_mm512_cvtepi64_pd(__A), |
| 493 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 494 | } |
| 495 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 496 | #define _mm512_cvt_roundepi64_pd(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 497 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 498 | (__v8df)_mm512_setzero_pd(), \ |
| 499 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 500 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 501 | #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 502 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 503 | (__v8df)(__m512d)(W), \ |
| 504 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 505 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 506 | #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 507 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 508 | (__v8df)_mm512_setzero_pd(), \ |
| 509 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 510 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 511 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 512 | _mm512_cvtepi64_ps (__m512i __A) { |
| 513 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 514 | (__v8sf) _mm256_setzero_ps(), |
| 515 | (__mmask8) -1, |
| 516 | _MM_FROUND_CUR_DIRECTION); |
| 517 | } |
| 518 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 519 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 520 | _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) { |
| 521 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 522 | (__v8sf) __W, |
| 523 | (__mmask8) __U, |
| 524 | _MM_FROUND_CUR_DIRECTION); |
| 525 | } |
| 526 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 527 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 528 | _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) { |
| 529 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 530 | (__v8sf) _mm256_setzero_ps(), |
| 531 | (__mmask8) __U, |
| 532 | _MM_FROUND_CUR_DIRECTION); |
| 533 | } |
| 534 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 535 | #define _mm512_cvt_roundepi64_ps(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 536 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 537 | (__v8sf)_mm256_setzero_ps(), \ |
| 538 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 539 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 540 | #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 541 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 542 | (__v8sf)(__m256)(W), (__mmask8)(U), \ |
| 543 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 544 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 545 | #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 546 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 547 | (__v8sf)_mm256_setzero_ps(), \ |
| 548 | (__mmask8)(U), (int)(R))) |
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 | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 552 | _mm512_cvttpd_epi64 (__m512d __A) { |
| 553 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 554 | (__v8di) _mm512_setzero_si512(), |
| 555 | (__mmask8) -1, |
| 556 | _MM_FROUND_CUR_DIRECTION); |
| 557 | } |
| 558 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 559 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 560 | _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 561 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 562 | (__v8di) __W, |
| 563 | (__mmask8) __U, |
| 564 | _MM_FROUND_CUR_DIRECTION); |
| 565 | } |
| 566 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 567 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 568 | _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) { |
| 569 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 570 | (__v8di) _mm512_setzero_si512(), |
| 571 | (__mmask8) __U, |
| 572 | _MM_FROUND_CUR_DIRECTION); |
| 573 | } |
| 574 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 575 | #define _mm512_cvtt_roundpd_epi64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 576 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 577 | (__v8di)_mm512_setzero_si512(), \ |
| 578 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 579 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 580 | #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 581 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 582 | (__v8di)(__m512i)(W), \ |
| 583 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 584 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 585 | #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 586 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 587 | (__v8di)_mm512_setzero_si512(), \ |
| 588 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 589 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 590 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 591 | _mm512_cvttpd_epu64 (__m512d __A) { |
| 592 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 593 | (__v8di) _mm512_setzero_si512(), |
| 594 | (__mmask8) -1, |
| 595 | _MM_FROUND_CUR_DIRECTION); |
| 596 | } |
| 597 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 598 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 599 | _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 600 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 601 | (__v8di) __W, |
| 602 | (__mmask8) __U, |
| 603 | _MM_FROUND_CUR_DIRECTION); |
| 604 | } |
| 605 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 606 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 607 | _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) { |
| 608 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 609 | (__v8di) _mm512_setzero_si512(), |
| 610 | (__mmask8) __U, |
| 611 | _MM_FROUND_CUR_DIRECTION); |
| 612 | } |
| 613 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 614 | #define _mm512_cvtt_roundpd_epu64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 615 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 616 | (__v8di)_mm512_setzero_si512(), \ |
| 617 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 618 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 619 | #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 620 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 621 | (__v8di)(__m512i)(W), \ |
| 622 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 623 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 624 | #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 625 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 626 | (__v8di)_mm512_setzero_si512(), \ |
| 627 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 628 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 629 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 630 | _mm512_cvttps_epi64 (__m256 __A) { |
| 631 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 632 | (__v8di) _mm512_setzero_si512(), |
| 633 | (__mmask8) -1, |
| 634 | _MM_FROUND_CUR_DIRECTION); |
| 635 | } |
| 636 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 637 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 638 | _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 639 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 640 | (__v8di) __W, |
| 641 | (__mmask8) __U, |
| 642 | _MM_FROUND_CUR_DIRECTION); |
| 643 | } |
| 644 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 645 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 646 | _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) { |
| 647 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 648 | (__v8di) _mm512_setzero_si512(), |
| 649 | (__mmask8) __U, |
| 650 | _MM_FROUND_CUR_DIRECTION); |
| 651 | } |
| 652 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 653 | #define _mm512_cvtt_roundps_epi64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 654 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 655 | (__v8di)_mm512_setzero_si512(), \ |
| 656 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 657 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 658 | #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 659 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 660 | (__v8di)(__m512i)(W), \ |
| 661 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 662 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 663 | #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 664 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 665 | (__v8di)_mm512_setzero_si512(), \ |
| 666 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 667 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 668 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 669 | _mm512_cvttps_epu64 (__m256 __A) { |
| 670 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 671 | (__v8di) _mm512_setzero_si512(), |
| 672 | (__mmask8) -1, |
| 673 | _MM_FROUND_CUR_DIRECTION); |
| 674 | } |
| 675 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 676 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 677 | _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 678 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 679 | (__v8di) __W, |
| 680 | (__mmask8) __U, |
| 681 | _MM_FROUND_CUR_DIRECTION); |
| 682 | } |
| 683 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 684 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 685 | _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) { |
| 686 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 687 | (__v8di) _mm512_setzero_si512(), |
| 688 | (__mmask8) __U, |
| 689 | _MM_FROUND_CUR_DIRECTION); |
| 690 | } |
| 691 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 692 | #define _mm512_cvtt_roundps_epu64(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 693 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 694 | (__v8di)_mm512_setzero_si512(), \ |
| 695 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 696 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 697 | #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 698 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 699 | (__v8di)(__m512i)(W), \ |
| 700 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 701 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 702 | #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 703 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 704 | (__v8di)_mm512_setzero_si512(), \ |
| 705 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 706 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 707 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 708 | _mm512_cvtepu64_pd (__m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 709 | return (__m512d)__builtin_convertvector((__v8du)__A, __v8df); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 710 | } |
| 711 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 712 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 713 | _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 714 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 715 | (__v8df)_mm512_cvtepu64_pd(__A), |
| 716 | (__v8df)__W); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 717 | } |
| 718 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 719 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 720 | _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 721 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 722 | (__v8df)_mm512_cvtepu64_pd(__A), |
| 723 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 724 | } |
| 725 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 726 | #define _mm512_cvt_roundepu64_pd(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 727 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 728 | (__v8df)_mm512_setzero_pd(), \ |
| 729 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 730 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 731 | #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 732 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 733 | (__v8df)(__m512d)(W), \ |
| 734 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 735 | |
| 736 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 737 | #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 738 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 739 | (__v8df)_mm512_setzero_pd(), \ |
| 740 | (__mmask8)(U), (int)(R))) |
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__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 744 | _mm512_cvtepu64_ps (__m512i __A) { |
| 745 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 746 | (__v8sf) _mm256_setzero_ps(), |
| 747 | (__mmask8) -1, |
| 748 | _MM_FROUND_CUR_DIRECTION); |
| 749 | } |
| 750 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 751 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 752 | _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) { |
| 753 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 754 | (__v8sf) __W, |
| 755 | (__mmask8) __U, |
| 756 | _MM_FROUND_CUR_DIRECTION); |
| 757 | } |
| 758 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 759 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 760 | _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) { |
| 761 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 762 | (__v8sf) _mm256_setzero_ps(), |
| 763 | (__mmask8) __U, |
| 764 | _MM_FROUND_CUR_DIRECTION); |
| 765 | } |
| 766 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 767 | #define _mm512_cvt_roundepu64_ps(A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 768 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 769 | (__v8sf)_mm256_setzero_ps(), \ |
| 770 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 771 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 772 | #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 773 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 774 | (__v8sf)(__m256)(W), (__mmask8)(U), \ |
| 775 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 776 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 777 | #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 778 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 779 | (__v8sf)_mm256_setzero_ps(), \ |
| 780 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 781 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 782 | #define _mm512_range_pd(A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 783 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 784 | (__v8df)(__m512d)(B), (int)(C), \ |
| 785 | (__v8df)_mm512_setzero_pd(), \ |
| 786 | (__mmask8)-1, \ |
| 787 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 788 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 789 | #define _mm512_mask_range_pd(W, U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 790 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 791 | (__v8df)(__m512d)(B), (int)(C), \ |
| 792 | (__v8df)(__m512d)(W), (__mmask8)(U), \ |
| 793 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 794 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 795 | #define _mm512_maskz_range_pd(U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 796 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 797 | (__v8df)(__m512d)(B), (int)(C), \ |
| 798 | (__v8df)_mm512_setzero_pd(), \ |
| 799 | (__mmask8)(U), \ |
| 800 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 801 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 802 | #define _mm512_range_round_pd(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 803 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 804 | (__v8df)(__m512d)(B), (int)(C), \ |
| 805 | (__v8df)_mm512_setzero_pd(), \ |
| 806 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 807 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 808 | #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 809 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 810 | (__v8df)(__m512d)(B), (int)(C), \ |
| 811 | (__v8df)(__m512d)(W), (__mmask8)(U), \ |
| 812 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 813 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 814 | #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 815 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 816 | (__v8df)(__m512d)(B), (int)(C), \ |
| 817 | (__v8df)_mm512_setzero_pd(), \ |
| 818 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 819 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 820 | #define _mm512_range_ps(A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 821 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 822 | (__v16sf)(__m512)(B), (int)(C), \ |
| 823 | (__v16sf)_mm512_setzero_ps(), \ |
| 824 | (__mmask16)-1, \ |
| 825 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 826 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 827 | #define _mm512_mask_range_ps(W, U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 828 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 829 | (__v16sf)(__m512)(B), (int)(C), \ |
| 830 | (__v16sf)(__m512)(W), (__mmask16)(U), \ |
| 831 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 832 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 833 | #define _mm512_maskz_range_ps(U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 834 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 835 | (__v16sf)(__m512)(B), (int)(C), \ |
| 836 | (__v16sf)_mm512_setzero_ps(), \ |
| 837 | (__mmask16)(U), \ |
| 838 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 839 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 840 | #define _mm512_range_round_ps(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 841 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 842 | (__v16sf)(__m512)(B), (int)(C), \ |
| 843 | (__v16sf)_mm512_setzero_ps(), \ |
| 844 | (__mmask16)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 845 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 846 | #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 847 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 848 | (__v16sf)(__m512)(B), (int)(C), \ |
| 849 | (__v16sf)(__m512)(W), (__mmask16)(U), \ |
| 850 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 851 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 852 | #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 853 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 854 | (__v16sf)(__m512)(B), (int)(C), \ |
| 855 | (__v16sf)_mm512_setzero_ps(), \ |
| 856 | (__mmask16)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 857 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 858 | #define _mm_range_round_ss(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 859 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 860 | (__v4sf)(__m128)(B), \ |
| 861 | (__v4sf)_mm_setzero_ps(), \ |
| 862 | (__mmask8) -1, (int)(C),\ |
| 863 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 864 | |
| 865 | #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 866 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 867 | #define _mm_mask_range_round_ss(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 868 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 869 | (__v4sf)(__m128)(B), \ |
| 870 | (__v4sf)(__m128)(W),\ |
| 871 | (__mmask8)(U), (int)(C),\ |
| 872 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 873 | |
| 874 | #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION) |
| 875 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 876 | #define _mm_maskz_range_round_ss(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 877 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 878 | (__v4sf)(__m128)(B), \ |
| 879 | (__v4sf)_mm_setzero_ps(), \ |
| 880 | (__mmask8)(U), (int)(C),\ |
| 881 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 882 | |
| 883 | #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 884 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 885 | #define _mm_range_round_sd(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 886 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 887 | (__v2df)(__m128d)(B), \ |
| 888 | (__v2df)_mm_setzero_pd(), \ |
| 889 | (__mmask8) -1, (int)(C),\ |
| 890 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 891 | |
| 892 | #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 893 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 894 | #define _mm_mask_range_round_sd(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 895 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 896 | (__v2df)(__m128d)(B), \ |
| 897 | (__v2df)(__m128d)(W),\ |
| 898 | (__mmask8)(U), (int)(C),\ |
| 899 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 900 | |
| 901 | #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 902 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 903 | #define _mm_maskz_range_round_sd(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 904 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 905 | (__v2df)(__m128d)(B), \ |
| 906 | (__v2df)_mm_setzero_pd(), \ |
| 907 | (__mmask8)(U), (int)(C),\ |
| 908 | (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 909 | |
| 910 | #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 911 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 912 | #define _mm512_reduce_pd(A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 913 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 914 | (__v8df)_mm512_setzero_pd(), \ |
| 915 | (__mmask8)-1, \ |
| 916 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 917 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 918 | #define _mm512_mask_reduce_pd(W, U, A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 919 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 920 | (__v8df)(__m512d)(W), \ |
| 921 | (__mmask8)(U), \ |
| 922 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 923 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 924 | #define _mm512_maskz_reduce_pd(U, A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 925 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 926 | (__v8df)_mm512_setzero_pd(), \ |
| 927 | (__mmask8)(U), \ |
| 928 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 929 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 930 | #define _mm512_reduce_ps(A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 931 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 932 | (__v16sf)_mm512_setzero_ps(), \ |
| 933 | (__mmask16)-1, \ |
| 934 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 935 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 936 | #define _mm512_mask_reduce_ps(W, U, A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 937 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 938 | (__v16sf)(__m512)(W), \ |
| 939 | (__mmask16)(U), \ |
| 940 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 941 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 942 | #define _mm512_maskz_reduce_ps(U, A, B) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 943 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 944 | (__v16sf)_mm512_setzero_ps(), \ |
| 945 | (__mmask16)(U), \ |
| 946 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 947 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 948 | #define _mm512_reduce_round_pd(A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 949 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 950 | (__v8df)_mm512_setzero_pd(), \ |
| 951 | (__mmask8)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 952 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 953 | #define _mm512_mask_reduce_round_pd(W, U, A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 954 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 955 | (__v8df)(__m512d)(W), \ |
| 956 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 957 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 958 | #define _mm512_maskz_reduce_round_pd(U, A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 959 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 960 | (__v8df)_mm512_setzero_pd(), \ |
| 961 | (__mmask8)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 962 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 963 | #define _mm512_reduce_round_ps(A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 964 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 965 | (__v16sf)_mm512_setzero_ps(), \ |
| 966 | (__mmask16)-1, (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 967 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 968 | #define _mm512_mask_reduce_round_ps(W, U, A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 969 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 970 | (__v16sf)(__m512)(W), \ |
| 971 | (__mmask16)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 972 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 973 | #define _mm512_maskz_reduce_round_ps(U, A, B, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 974 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 975 | (__v16sf)_mm512_setzero_ps(), \ |
| 976 | (__mmask16)(U), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 977 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 978 | #define _mm_reduce_ss(A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 979 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 980 | (__v4sf)(__m128)(B), \ |
| 981 | (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ |
| 982 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 983 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 984 | #define _mm_mask_reduce_ss(W, U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 985 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 986 | (__v4sf)(__m128)(B), \ |
| 987 | (__v4sf)(__m128)(W), (__mmask8)(U), \ |
| 988 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 989 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 990 | #define _mm_maskz_reduce_ss(U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 991 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 992 | (__v4sf)(__m128)(B), \ |
| 993 | (__v4sf)_mm_setzero_ps(), \ |
| 994 | (__mmask8)(U), (int)(C), \ |
| 995 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 996 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 997 | #define _mm_reduce_round_ss(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 998 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 999 | (__v4sf)(__m128)(B), \ |
| 1000 | (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ |
| 1001 | (int)(C), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1002 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1003 | #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1004 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 1005 | (__v4sf)(__m128)(B), \ |
| 1006 | (__v4sf)(__m128)(W), (__mmask8)(U), \ |
| 1007 | (int)(C), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1008 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1009 | #define _mm_maskz_reduce_round_ss(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1010 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 1011 | (__v4sf)(__m128)(B), \ |
| 1012 | (__v4sf)_mm_setzero_ps(), \ |
| 1013 | (__mmask8)(U), (int)(C), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1014 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1015 | #define _mm_reduce_sd(A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1016 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1017 | (__v2df)(__m128d)(B), \ |
| 1018 | (__v2df)_mm_setzero_pd(), \ |
| 1019 | (__mmask8)-1, (int)(C), \ |
| 1020 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1021 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1022 | #define _mm_mask_reduce_sd(W, U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1023 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1024 | (__v2df)(__m128d)(B), \ |
| 1025 | (__v2df)(__m128d)(W), (__mmask8)(U), \ |
| 1026 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1027 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1028 | #define _mm_maskz_reduce_sd(U, A, B, C) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1029 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1030 | (__v2df)(__m128d)(B), \ |
| 1031 | (__v2df)_mm_setzero_pd(), \ |
| 1032 | (__mmask8)(U), (int)(C), \ |
| 1033 | _MM_FROUND_CUR_DIRECTION)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1034 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1035 | #define _mm_reduce_round_sd(A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1036 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1037 | (__v2df)(__m128d)(B), \ |
| 1038 | (__v2df)_mm_setzero_pd(), \ |
| 1039 | (__mmask8)-1, (int)(C), (int)(R))) |
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 | #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1042 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1043 | (__v2df)(__m128d)(B), \ |
| 1044 | (__v2df)(__m128d)(W), (__mmask8)(U), \ |
| 1045 | (int)(C), (int)(R))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1046 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1047 | #define _mm_maskz_reduce_round_sd(U, A, B, C, R) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1048 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1049 | (__v2df)(__m128d)(B), \ |
| 1050 | (__v2df)_mm_setzero_pd(), \ |
| 1051 | (__mmask8)(U), (int)(C), (int)(R))) |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1052 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1053 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1054 | _mm512_movepi32_mask (__m512i __A) |
| 1055 | { |
| 1056 | return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A); |
| 1057 | } |
| 1058 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1059 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1060 | _mm512_movm_epi32 (__mmask16 __A) |
| 1061 | { |
| 1062 | return (__m512i) __builtin_ia32_cvtmask2d512 (__A); |
| 1063 | } |
| 1064 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1065 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1066 | _mm512_movm_epi64 (__mmask8 __A) |
| 1067 | { |
| 1068 | return (__m512i) __builtin_ia32_cvtmask2q512 (__A); |
| 1069 | } |
| 1070 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1071 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1072 | _mm512_movepi64_mask (__m512i __A) |
| 1073 | { |
| 1074 | return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A); |
| 1075 | } |
| 1076 | |
| 1077 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1078 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1079 | _mm512_broadcast_f32x2 (__m128 __A) |
| 1080 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1081 | return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A, |
| 1082 | 0, 1, 0, 1, 0, 1, 0, 1, |
| 1083 | 0, 1, 0, 1, 0, 1, 0, 1); |
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__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1087 | _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A) |
| 1088 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1089 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1090 | (__v16sf)_mm512_broadcast_f32x2(__A), |
| 1091 | (__v16sf)__O); |
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__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1095 | _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A) |
| 1096 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1097 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1098 | (__v16sf)_mm512_broadcast_f32x2(__A), |
| 1099 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1100 | } |
| 1101 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1102 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1103 | _mm512_broadcast_f32x8(__m256 __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1104 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1105 | return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A, |
| 1106 | 0, 1, 2, 3, 4, 5, 6, 7, |
| 1107 | 0, 1, 2, 3, 4, 5, 6, 7); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1108 | } |
| 1109 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1110 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1111 | _mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1112 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1113 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1114 | (__v16sf)_mm512_broadcast_f32x8(__A), |
| 1115 | (__v16sf)__O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1116 | } |
| 1117 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1118 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1119 | _mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1120 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1121 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1122 | (__v16sf)_mm512_broadcast_f32x8(__A), |
| 1123 | (__v16sf)_mm512_setzero_ps()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1124 | } |
| 1125 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1126 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1127 | _mm512_broadcast_f64x2(__m128d __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1128 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1129 | return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A, |
| 1130 | 0, 1, 0, 1, 0, 1, 0, 1); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1131 | } |
| 1132 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1133 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1134 | _mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1135 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1136 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M, |
| 1137 | (__v8df)_mm512_broadcast_f64x2(__A), |
| 1138 | (__v8df)__O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1139 | } |
| 1140 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1141 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1142 | _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1143 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1144 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M, |
| 1145 | (__v8df)_mm512_broadcast_f64x2(__A), |
| 1146 | (__v8df)_mm512_setzero_pd()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1147 | } |
| 1148 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1149 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1150 | _mm512_broadcast_i32x2 (__m128i __A) |
| 1151 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1152 | return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
| 1153 | 0, 1, 0, 1, 0, 1, 0, 1, |
| 1154 | 0, 1, 0, 1, 0, 1, 0, 1); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1155 | } |
| 1156 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1157 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1158 | _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A) |
| 1159 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1160 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1161 | (__v16si)_mm512_broadcast_i32x2(__A), |
| 1162 | (__v16si)__O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1163 | } |
| 1164 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1165 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1166 | _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A) |
| 1167 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1168 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1169 | (__v16si)_mm512_broadcast_i32x2(__A), |
| 1170 | (__v16si)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1171 | } |
| 1172 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1173 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1174 | _mm512_broadcast_i32x8(__m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1175 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1176 | return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A, |
| 1177 | 0, 1, 2, 3, 4, 5, 6, 7, |
| 1178 | 0, 1, 2, 3, 4, 5, 6, 7); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1179 | } |
| 1180 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1181 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1182 | _mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1183 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1184 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1185 | (__v16si)_mm512_broadcast_i32x8(__A), |
| 1186 | (__v16si)__O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1187 | } |
| 1188 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1189 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1190 | _mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1191 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1192 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1193 | (__v16si)_mm512_broadcast_i32x8(__A), |
| 1194 | (__v16si)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1195 | } |
| 1196 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1197 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1198 | _mm512_broadcast_i64x2(__m128i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1199 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1200 | return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A, |
| 1201 | 0, 1, 0, 1, 0, 1, 0, 1); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1202 | } |
| 1203 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1204 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1205 | _mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1206 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1207 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, |
| 1208 | (__v8di)_mm512_broadcast_i64x2(__A), |
| 1209 | (__v8di)__O); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1210 | } |
| 1211 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1212 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1213 | _mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1214 | { |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1215 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, |
| 1216 | (__v8di)_mm512_broadcast_i64x2(__A), |
| 1217 | (__v8di)_mm512_setzero_si512()); |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1218 | } |
| 1219 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1220 | #define _mm512_extractf32x8_ps(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1221 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1222 | (__v8sf)_mm256_undefined_ps(), \ |
| 1223 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1224 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1225 | #define _mm512_mask_extractf32x8_ps(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1226 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1227 | (__v8sf)(__m256)(W), \ |
| 1228 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1229 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1230 | #define _mm512_maskz_extractf32x8_ps(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1231 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1232 | (__v8sf)_mm256_setzero_ps(), \ |
| 1233 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1234 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1235 | #define _mm512_extractf64x2_pd(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1236 | ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ |
| 1237 | (int)(imm), \ |
| 1238 | (__v2df)_mm_undefined_pd(), \ |
| 1239 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1240 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1241 | #define _mm512_mask_extractf64x2_pd(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1242 | ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ |
| 1243 | (int)(imm), \ |
| 1244 | (__v2df)(__m128d)(W), \ |
| 1245 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1246 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1247 | #define _mm512_maskz_extractf64x2_pd(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1248 | ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ |
| 1249 | (int)(imm), \ |
| 1250 | (__v2df)_mm_setzero_pd(), \ |
| 1251 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1252 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1253 | #define _mm512_extracti32x8_epi32(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1254 | ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ |
| 1255 | (__v8si)_mm256_undefined_si256(), \ |
| 1256 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1257 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1258 | #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1259 | ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ |
| 1260 | (__v8si)(__m256i)(W), \ |
| 1261 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1262 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1263 | #define _mm512_maskz_extracti32x8_epi32(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1264 | ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ |
| 1265 | (__v8si)_mm256_setzero_si256(), \ |
| 1266 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1267 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1268 | #define _mm512_extracti64x2_epi64(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1269 | ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1270 | (int)(imm), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1271 | (__v2di)_mm_undefined_si128(), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1272 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1273 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1274 | #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1275 | ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ |
| 1276 | (int)(imm), \ |
| 1277 | (__v2di)(__m128i)(W), \ |
| 1278 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1279 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1280 | #define _mm512_maskz_extracti64x2_epi64(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1281 | ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ |
| 1282 | (int)(imm), \ |
| 1283 | (__v2di)_mm_setzero_si128(), \ |
| 1284 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1285 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1286 | #define _mm512_insertf32x8(A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1287 | ((__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \ |
| 1288 | (__v8sf)(__m256)(B), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1289 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1290 | #define _mm512_mask_insertf32x8(W, U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1291 | ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1292 | (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1293 | (__v16sf)(__m512)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1294 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1295 | #define _mm512_maskz_insertf32x8(U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1296 | ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1297 | (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1298 | (__v16sf)_mm512_setzero_ps())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1299 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1300 | #define _mm512_insertf64x2(A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1301 | ((__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \ |
| 1302 | (__v2df)(__m128d)(B), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1303 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1304 | #define _mm512_mask_insertf64x2(W, U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1305 | ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1306 | (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1307 | (__v8df)(__m512d)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1308 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1309 | #define _mm512_maskz_insertf64x2(U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1310 | ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1311 | (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1312 | (__v8df)_mm512_setzero_pd())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1313 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1314 | #define _mm512_inserti32x8(A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1315 | ((__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \ |
| 1316 | (__v8si)(__m256i)(B), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1317 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1318 | #define _mm512_mask_inserti32x8(W, U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1319 | ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1320 | (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1321 | (__v16si)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1322 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1323 | #define _mm512_maskz_inserti32x8(U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1324 | ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1325 | (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1326 | (__v16si)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1327 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1328 | #define _mm512_inserti64x2(A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1329 | ((__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \ |
| 1330 | (__v2di)(__m128i)(B), (int)(imm))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1331 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1332 | #define _mm512_mask_inserti64x2(W, U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1333 | ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1334 | (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1335 | (__v8di)(__m512i)(W))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1336 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1337 | #define _mm512_maskz_inserti64x2(U, A, B, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1338 | ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1339 | (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1340 | (__v8di)_mm512_setzero_si512())) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1341 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1342 | #define _mm512_mask_fpclass_ps_mask(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1343 | ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ |
| 1344 | (int)(imm), (__mmask16)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1345 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1346 | #define _mm512_fpclass_ps_mask(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1347 | ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ |
| 1348 | (int)(imm), (__mmask16)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1349 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1350 | #define _mm512_mask_fpclass_pd_mask(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1351 | ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ |
| 1352 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1353 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1354 | #define _mm512_fpclass_pd_mask(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1355 | ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ |
| 1356 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1357 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1358 | #define _mm_fpclass_sd_mask(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1359 | ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1360 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1361 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1362 | #define _mm_mask_fpclass_sd_mask(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1363 | ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1364 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1365 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1366 | #define _mm_fpclass_ss_mask(A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1367 | ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1368 | (__mmask8)-1)) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1369 | |
Logan Chien | 55afb0a | 2018-10-15 10:42:14 +0800 | [diff] [blame] | 1370 | #define _mm_mask_fpclass_ss_mask(U, A, imm) \ |
Pirama Arumuga Nainar | 494f645 | 2021-12-02 10:42:14 -0800 | [diff] [blame] | 1371 | ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1372 | (__mmask8)(U))) |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1373 | |
Logan Chien | b0c8402 | 2018-11-09 16:19:54 +0800 | [diff] [blame] | 1374 | #undef __DEFAULT_FN_ATTRS512 |
Logan Chien | 2833ffb | 2018-10-09 10:03:24 +0800 | [diff] [blame] | 1375 | #undef __DEFAULT_FN_ATTRS |
| 1376 | |
| 1377 | #endif |