Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 1 | /*===---- avx512vlintrin.h - AVX512VL intrinsics ---------------------------=== |
| 2 | * |
| 3 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 4 | * of this software and associated documentation files (the "Software"), to deal |
| 5 | * in the Software without restriction, including without limitation the rights |
| 6 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 7 | * copies of the Software, and to permit persons to whom the Software is |
| 8 | * furnished to do so, subject to the following conditions: |
| 9 | * |
| 10 | * The above copyright notice and this permission notice shall be included in |
| 11 | * all copies or substantial portions of the Software. |
| 12 | * |
| 13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 15 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 16 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 17 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 18 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 19 | * THE SOFTWARE. |
| 20 | * |
| 21 | *===-----------------------------------------------------------------------=== |
| 22 | */ |
| 23 | |
| 24 | #ifndef __IMMINTRIN_H |
| 25 | #error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead." |
| 26 | #endif |
| 27 | |
| 28 | #ifndef __AVX512VLINTRIN_H |
| 29 | #define __AVX512VLINTRIN_H |
| 30 | |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 31 | /* Define the default attributes for the functions in this file. */ |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 32 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"))) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 33 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 34 | /* Integer compare */ |
| 35 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 36 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 37 | _mm_cmpeq_epi32_mask(__m128i __a, __m128i __b) { |
| 38 | return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b, |
| 39 | (__mmask8)-1); |
| 40 | } |
| 41 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 42 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 43 | _mm_mask_cmpeq_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 44 | return (__mmask8)__builtin_ia32_pcmpeqd128_mask((__v4si)__a, (__v4si)__b, |
| 45 | __u); |
| 46 | } |
| 47 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 48 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 49 | _mm_cmpeq_epu32_mask(__m128i __a, __m128i __b) { |
| 50 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 0, |
| 51 | (__mmask8)-1); |
| 52 | } |
| 53 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 54 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 55 | _mm_mask_cmpeq_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 56 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 0, |
| 57 | __u); |
| 58 | } |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 59 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 60 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 61 | _mm256_cmpeq_epi32_mask(__m256i __a, __m256i __b) { |
| 62 | return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b, |
| 63 | (__mmask8)-1); |
| 64 | } |
| 65 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 66 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 67 | _mm256_mask_cmpeq_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 68 | return (__mmask8)__builtin_ia32_pcmpeqd256_mask((__v8si)__a, (__v8si)__b, |
| 69 | __u); |
| 70 | } |
| 71 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 72 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 73 | _mm256_cmpeq_epu32_mask(__m256i __a, __m256i __b) { |
| 74 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 0, |
| 75 | (__mmask8)-1); |
| 76 | } |
| 77 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 78 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 79 | _mm256_mask_cmpeq_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 80 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 0, |
| 81 | __u); |
| 82 | } |
| 83 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 84 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 85 | _mm_cmpeq_epi64_mask(__m128i __a, __m128i __b) { |
| 86 | return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b, |
| 87 | (__mmask8)-1); |
| 88 | } |
| 89 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 90 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 91 | _mm_mask_cmpeq_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 92 | return (__mmask8)__builtin_ia32_pcmpeqq128_mask((__v2di)__a, (__v2di)__b, |
| 93 | __u); |
| 94 | } |
| 95 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 96 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 97 | _mm_cmpeq_epu64_mask(__m128i __a, __m128i __b) { |
| 98 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 0, |
| 99 | (__mmask8)-1); |
| 100 | } |
| 101 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 102 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 103 | _mm_mask_cmpeq_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 104 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 0, |
| 105 | __u); |
| 106 | } |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 107 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 108 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 109 | _mm256_cmpeq_epi64_mask(__m256i __a, __m256i __b) { |
| 110 | return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b, |
| 111 | (__mmask8)-1); |
| 112 | } |
| 113 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 114 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 115 | _mm256_mask_cmpeq_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 116 | return (__mmask8)__builtin_ia32_pcmpeqq256_mask((__v4di)__a, (__v4di)__b, |
| 117 | __u); |
| 118 | } |
| 119 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 120 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 121 | _mm256_cmpeq_epu64_mask(__m256i __a, __m256i __b) { |
| 122 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 0, |
| 123 | (__mmask8)-1); |
| 124 | } |
| 125 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 126 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 127 | _mm256_mask_cmpeq_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 128 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 0, |
| 129 | __u); |
| 130 | } |
| 131 | |
| 132 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 133 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 134 | _mm_cmpge_epi32_mask(__m128i __a, __m128i __b) { |
| 135 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 5, |
| 136 | (__mmask8)-1); |
| 137 | } |
| 138 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 139 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 140 | _mm_mask_cmpge_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 141 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 5, |
| 142 | __u); |
| 143 | } |
| 144 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 145 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 146 | _mm_cmpge_epu32_mask(__m128i __a, __m128i __b) { |
| 147 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 5, |
| 148 | (__mmask8)-1); |
| 149 | } |
| 150 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 151 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 152 | _mm_mask_cmpge_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 153 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 5, |
| 154 | __u); |
| 155 | } |
| 156 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 157 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 158 | _mm256_cmpge_epi32_mask(__m256i __a, __m256i __b) { |
| 159 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 5, |
| 160 | (__mmask8)-1); |
| 161 | } |
| 162 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 163 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 164 | _mm256_mask_cmpge_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 165 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 5, |
| 166 | __u); |
| 167 | } |
| 168 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 169 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 170 | _mm256_cmpge_epu32_mask(__m256i __a, __m256i __b) { |
| 171 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 5, |
| 172 | (__mmask8)-1); |
| 173 | } |
| 174 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 175 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 176 | _mm256_mask_cmpge_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 177 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 5, |
| 178 | __u); |
| 179 | } |
| 180 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 181 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 182 | _mm_cmpge_epi64_mask(__m128i __a, __m128i __b) { |
| 183 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 5, |
| 184 | (__mmask8)-1); |
| 185 | } |
| 186 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 187 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 188 | _mm_mask_cmpge_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 189 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 5, |
| 190 | __u); |
| 191 | } |
| 192 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 193 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 194 | _mm_cmpge_epu64_mask(__m128i __a, __m128i __b) { |
| 195 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 5, |
| 196 | (__mmask8)-1); |
| 197 | } |
| 198 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 199 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 200 | _mm_mask_cmpge_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 201 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 5, |
| 202 | __u); |
| 203 | } |
| 204 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 205 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 206 | _mm256_cmpge_epi64_mask(__m256i __a, __m256i __b) { |
| 207 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 5, |
| 208 | (__mmask8)-1); |
| 209 | } |
| 210 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 211 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 212 | _mm256_mask_cmpge_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 213 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 5, |
| 214 | __u); |
| 215 | } |
| 216 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 217 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 218 | _mm256_cmpge_epu64_mask(__m256i __a, __m256i __b) { |
| 219 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 5, |
| 220 | (__mmask8)-1); |
| 221 | } |
| 222 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 223 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 224 | _mm256_mask_cmpge_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 225 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 5, |
| 226 | __u); |
| 227 | } |
| 228 | |
| 229 | |
| 230 | |
| 231 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 232 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 233 | _mm_cmpgt_epi32_mask(__m128i __a, __m128i __b) { |
| 234 | return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b, |
| 235 | (__mmask8)-1); |
| 236 | } |
| 237 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 238 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 239 | _mm_mask_cmpgt_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 240 | return (__mmask8)__builtin_ia32_pcmpgtd128_mask((__v4si)__a, (__v4si)__b, |
| 241 | __u); |
| 242 | } |
| 243 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 244 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 245 | _mm_cmpgt_epu32_mask(__m128i __a, __m128i __b) { |
| 246 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 6, |
| 247 | (__mmask8)-1); |
| 248 | } |
| 249 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 250 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 251 | _mm_mask_cmpgt_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 252 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 6, |
| 253 | __u); |
| 254 | } |
| 255 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 256 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 257 | _mm256_cmpgt_epi32_mask(__m256i __a, __m256i __b) { |
| 258 | return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b, |
| 259 | (__mmask8)-1); |
| 260 | } |
| 261 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 262 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 263 | _mm256_mask_cmpgt_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 264 | return (__mmask8)__builtin_ia32_pcmpgtd256_mask((__v8si)__a, (__v8si)__b, |
| 265 | __u); |
| 266 | } |
| 267 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 268 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 269 | _mm256_cmpgt_epu32_mask(__m256i __a, __m256i __b) { |
| 270 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 6, |
| 271 | (__mmask8)-1); |
| 272 | } |
| 273 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 274 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 275 | _mm256_mask_cmpgt_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 276 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 6, |
| 277 | __u); |
| 278 | } |
| 279 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 280 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 281 | _mm_cmpgt_epi64_mask(__m128i __a, __m128i __b) { |
| 282 | return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b, |
| 283 | (__mmask8)-1); |
| 284 | } |
| 285 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 286 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 287 | _mm_mask_cmpgt_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 288 | return (__mmask8)__builtin_ia32_pcmpgtq128_mask((__v2di)__a, (__v2di)__b, |
| 289 | __u); |
| 290 | } |
| 291 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 292 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 293 | _mm_cmpgt_epu64_mask(__m128i __a, __m128i __b) { |
| 294 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 6, |
| 295 | (__mmask8)-1); |
| 296 | } |
| 297 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 298 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 299 | _mm_mask_cmpgt_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 300 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 6, |
| 301 | __u); |
| 302 | } |
| 303 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 304 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 305 | _mm256_cmpgt_epi64_mask(__m256i __a, __m256i __b) { |
| 306 | return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b, |
| 307 | (__mmask8)-1); |
| 308 | } |
| 309 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 310 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 311 | _mm256_mask_cmpgt_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 312 | return (__mmask8)__builtin_ia32_pcmpgtq256_mask((__v4di)__a, (__v4di)__b, |
| 313 | __u); |
| 314 | } |
| 315 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 316 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 317 | _mm256_cmpgt_epu64_mask(__m256i __a, __m256i __b) { |
| 318 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 6, |
| 319 | (__mmask8)-1); |
| 320 | } |
| 321 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 322 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 323 | _mm256_mask_cmpgt_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 324 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 6, |
| 325 | __u); |
| 326 | } |
| 327 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 328 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 329 | _mm_cmple_epi32_mask(__m128i __a, __m128i __b) { |
| 330 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 2, |
| 331 | (__mmask8)-1); |
| 332 | } |
| 333 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 334 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 335 | _mm_mask_cmple_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 336 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 2, |
| 337 | __u); |
| 338 | } |
| 339 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 340 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 341 | _mm_cmple_epu32_mask(__m128i __a, __m128i __b) { |
| 342 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 2, |
| 343 | (__mmask8)-1); |
| 344 | } |
| 345 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 346 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 347 | _mm_mask_cmple_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 348 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 2, |
| 349 | __u); |
| 350 | } |
| 351 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 352 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 353 | _mm256_cmple_epi32_mask(__m256i __a, __m256i __b) { |
| 354 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 2, |
| 355 | (__mmask8)-1); |
| 356 | } |
| 357 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 358 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 359 | _mm256_mask_cmple_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 360 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 2, |
| 361 | __u); |
| 362 | } |
| 363 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 364 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 365 | _mm256_cmple_epu32_mask(__m256i __a, __m256i __b) { |
| 366 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 2, |
| 367 | (__mmask8)-1); |
| 368 | } |
| 369 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 370 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 371 | _mm256_mask_cmple_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 372 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 2, |
| 373 | __u); |
| 374 | } |
| 375 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 376 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 377 | _mm_cmple_epi64_mask(__m128i __a, __m128i __b) { |
| 378 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 2, |
| 379 | (__mmask8)-1); |
| 380 | } |
| 381 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 382 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 383 | _mm_mask_cmple_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 384 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 2, |
| 385 | __u); |
| 386 | } |
| 387 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 388 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 389 | _mm_cmple_epu64_mask(__m128i __a, __m128i __b) { |
| 390 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 2, |
| 391 | (__mmask8)-1); |
| 392 | } |
| 393 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 394 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 395 | _mm_mask_cmple_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 396 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 2, |
| 397 | __u); |
| 398 | } |
| 399 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 400 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 401 | _mm256_cmple_epi64_mask(__m256i __a, __m256i __b) { |
| 402 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 2, |
| 403 | (__mmask8)-1); |
| 404 | } |
| 405 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 406 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 407 | _mm256_mask_cmple_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 408 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 2, |
| 409 | __u); |
| 410 | } |
| 411 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 412 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 413 | _mm256_cmple_epu64_mask(__m256i __a, __m256i __b) { |
| 414 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 2, |
| 415 | (__mmask8)-1); |
| 416 | } |
| 417 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 418 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 419 | _mm256_mask_cmple_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 420 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 2, |
| 421 | __u); |
| 422 | } |
| 423 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 424 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 425 | _mm_cmplt_epi32_mask(__m128i __a, __m128i __b) { |
| 426 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 1, |
| 427 | (__mmask8)-1); |
| 428 | } |
| 429 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 430 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 431 | _mm_mask_cmplt_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 432 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 1, |
| 433 | __u); |
| 434 | } |
| 435 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 436 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 437 | _mm_cmplt_epu32_mask(__m128i __a, __m128i __b) { |
| 438 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 1, |
| 439 | (__mmask8)-1); |
| 440 | } |
| 441 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 442 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 443 | _mm_mask_cmplt_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 444 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 1, |
| 445 | __u); |
| 446 | } |
| 447 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 448 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 449 | _mm256_cmplt_epi32_mask(__m256i __a, __m256i __b) { |
| 450 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 1, |
| 451 | (__mmask8)-1); |
| 452 | } |
| 453 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 454 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 455 | _mm256_mask_cmplt_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 456 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 1, |
| 457 | __u); |
| 458 | } |
| 459 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 460 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 461 | _mm256_cmplt_epu32_mask(__m256i __a, __m256i __b) { |
| 462 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 1, |
| 463 | (__mmask8)-1); |
| 464 | } |
| 465 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 466 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 467 | _mm256_mask_cmplt_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 468 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 1, |
| 469 | __u); |
| 470 | } |
| 471 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 472 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 473 | _mm_cmplt_epi64_mask(__m128i __a, __m128i __b) { |
| 474 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 1, |
| 475 | (__mmask8)-1); |
| 476 | } |
| 477 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 478 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 479 | _mm_mask_cmplt_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 480 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 1, |
| 481 | __u); |
| 482 | } |
| 483 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 484 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 485 | _mm_cmplt_epu64_mask(__m128i __a, __m128i __b) { |
| 486 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 1, |
| 487 | (__mmask8)-1); |
| 488 | } |
| 489 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 490 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 491 | _mm_mask_cmplt_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 492 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 1, |
| 493 | __u); |
| 494 | } |
| 495 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 496 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 497 | _mm256_cmplt_epi64_mask(__m256i __a, __m256i __b) { |
| 498 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 1, |
| 499 | (__mmask8)-1); |
| 500 | } |
| 501 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 502 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 503 | _mm256_mask_cmplt_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 504 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 1, |
| 505 | __u); |
| 506 | } |
| 507 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 508 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 509 | _mm256_cmplt_epu64_mask(__m256i __a, __m256i __b) { |
| 510 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 1, |
| 511 | (__mmask8)-1); |
| 512 | } |
| 513 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 514 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 515 | _mm256_mask_cmplt_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 516 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 1, |
| 517 | __u); |
| 518 | } |
| 519 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 520 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 521 | _mm_cmpneq_epi32_mask(__m128i __a, __m128i __b) { |
| 522 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 4, |
| 523 | (__mmask8)-1); |
| 524 | } |
| 525 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 526 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 527 | _mm_mask_cmpneq_epi32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 528 | return (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)__a, (__v4si)__b, 4, |
| 529 | __u); |
| 530 | } |
| 531 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 532 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 533 | _mm_cmpneq_epu32_mask(__m128i __a, __m128i __b) { |
| 534 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 4, |
| 535 | (__mmask8)-1); |
| 536 | } |
| 537 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 538 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 539 | _mm_mask_cmpneq_epu32_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 540 | return (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)__a, (__v4si)__b, 4, |
| 541 | __u); |
| 542 | } |
| 543 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 544 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 545 | _mm256_cmpneq_epi32_mask(__m256i __a, __m256i __b) { |
| 546 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 4, |
| 547 | (__mmask8)-1); |
| 548 | } |
| 549 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 550 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 551 | _mm256_mask_cmpneq_epi32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 552 | return (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)__a, (__v8si)__b, 4, |
| 553 | __u); |
| 554 | } |
| 555 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 556 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 557 | _mm256_cmpneq_epu32_mask(__m256i __a, __m256i __b) { |
| 558 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 4, |
| 559 | (__mmask8)-1); |
| 560 | } |
| 561 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 562 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 563 | _mm256_mask_cmpneq_epu32_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 564 | return (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)__a, (__v8si)__b, 4, |
| 565 | __u); |
| 566 | } |
| 567 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 568 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 569 | _mm_cmpneq_epi64_mask(__m128i __a, __m128i __b) { |
| 570 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 4, |
| 571 | (__mmask8)-1); |
| 572 | } |
| 573 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 574 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 575 | _mm_mask_cmpneq_epi64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 576 | return (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)__a, (__v2di)__b, 4, |
| 577 | __u); |
| 578 | } |
| 579 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 580 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 581 | _mm_cmpneq_epu64_mask(__m128i __a, __m128i __b) { |
| 582 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 4, |
| 583 | (__mmask8)-1); |
| 584 | } |
| 585 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 586 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 587 | _mm_mask_cmpneq_epu64_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 588 | return (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)__a, (__v2di)__b, 4, |
| 589 | __u); |
| 590 | } |
| 591 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 592 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 593 | _mm256_cmpneq_epi64_mask(__m256i __a, __m256i __b) { |
| 594 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 4, |
| 595 | (__mmask8)-1); |
| 596 | } |
| 597 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 598 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 599 | _mm256_mask_cmpneq_epi64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 600 | return (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)__a, (__v4di)__b, 4, |
| 601 | __u); |
| 602 | } |
| 603 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 604 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 605 | _mm256_cmpneq_epu64_mask(__m256i __a, __m256i __b) { |
| 606 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 4, |
| 607 | (__mmask8)-1); |
| 608 | } |
| 609 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 610 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 611 | _mm256_mask_cmpneq_epu64_mask(__mmask8 __u, __m256i __a, __m256i __b) { |
| 612 | return (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)__a, (__v4di)__b, 4, |
| 613 | __u); |
| 614 | } |
| 615 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 616 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 617 | _mm256_mask_add_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 618 | __m256i __B) |
| 619 | { |
| 620 | return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A, |
| 621 | (__v8si) __B, |
| 622 | (__v8si) __W, |
| 623 | (__mmask8) __U); |
| 624 | } |
| 625 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 626 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 627 | _mm256_maskz_add_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 628 | { |
| 629 | return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A, |
| 630 | (__v8si) __B, |
| 631 | (__v8si) |
| 632 | _mm256_setzero_si256 (), |
| 633 | (__mmask8) __U); |
| 634 | } |
| 635 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 636 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 637 | _mm256_mask_add_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 638 | __m256i __B) |
| 639 | { |
| 640 | return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A, |
| 641 | (__v4di) __B, |
| 642 | (__v4di) __W, |
| 643 | (__mmask8) __U); |
| 644 | } |
| 645 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 646 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 647 | _mm256_maskz_add_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 648 | { |
| 649 | return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A, |
| 650 | (__v4di) __B, |
| 651 | (__v4di) |
| 652 | _mm256_setzero_si256 (), |
| 653 | (__mmask8) __U); |
| 654 | } |
| 655 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 656 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 657 | _mm256_mask_sub_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 658 | __m256i __B) |
| 659 | { |
| 660 | return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A, |
| 661 | (__v8si) __B, |
| 662 | (__v8si) __W, |
| 663 | (__mmask8) __U); |
| 664 | } |
| 665 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 666 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 667 | _mm256_maskz_sub_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 668 | { |
| 669 | return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A, |
| 670 | (__v8si) __B, |
| 671 | (__v8si) |
| 672 | _mm256_setzero_si256 (), |
| 673 | (__mmask8) __U); |
| 674 | } |
| 675 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 676 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 677 | _mm256_mask_sub_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 678 | __m256i __B) |
| 679 | { |
| 680 | return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A, |
| 681 | (__v4di) __B, |
| 682 | (__v4di) __W, |
| 683 | (__mmask8) __U); |
| 684 | } |
| 685 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 686 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 687 | _mm256_maskz_sub_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 688 | { |
| 689 | return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A, |
| 690 | (__v4di) __B, |
| 691 | (__v4di) |
| 692 | _mm256_setzero_si256 (), |
| 693 | (__mmask8) __U); |
| 694 | } |
| 695 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 696 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 697 | _mm_mask_add_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 698 | __m128i __B) |
| 699 | { |
| 700 | return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A, |
| 701 | (__v4si) __B, |
| 702 | (__v4si) __W, |
| 703 | (__mmask8) __U); |
| 704 | } |
| 705 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 706 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 707 | _mm_maskz_add_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 708 | { |
| 709 | return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A, |
| 710 | (__v4si) __B, |
| 711 | (__v4si) |
| 712 | _mm_setzero_si128 (), |
| 713 | (__mmask8) __U); |
| 714 | } |
| 715 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 716 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 717 | _mm_mask_add_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 718 | __m128i __B) |
| 719 | { |
| 720 | return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A, |
| 721 | (__v2di) __B, |
| 722 | (__v2di) __W, |
| 723 | (__mmask8) __U); |
| 724 | } |
| 725 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 726 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 727 | _mm_maskz_add_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 728 | { |
| 729 | return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A, |
| 730 | (__v2di) __B, |
| 731 | (__v2di) |
| 732 | _mm_setzero_si128 (), |
| 733 | (__mmask8) __U); |
| 734 | } |
| 735 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 736 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 737 | _mm_mask_sub_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 738 | __m128i __B) |
| 739 | { |
| 740 | return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A, |
| 741 | (__v4si) __B, |
| 742 | (__v4si) __W, |
| 743 | (__mmask8) __U); |
| 744 | } |
| 745 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 746 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 747 | _mm_maskz_sub_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 748 | { |
| 749 | return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A, |
| 750 | (__v4si) __B, |
| 751 | (__v4si) |
| 752 | _mm_setzero_si128 (), |
| 753 | (__mmask8) __U); |
| 754 | } |
| 755 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 756 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 757 | _mm_mask_sub_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 758 | __m128i __B) |
| 759 | { |
| 760 | return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A, |
| 761 | (__v2di) __B, |
| 762 | (__v2di) __W, |
| 763 | (__mmask8) __U); |
| 764 | } |
| 765 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 766 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 767 | _mm_maskz_sub_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 768 | { |
| 769 | return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A, |
| 770 | (__v2di) __B, |
| 771 | (__v2di) |
| 772 | _mm_setzero_si128 (), |
| 773 | (__mmask8) __U); |
| 774 | } |
| 775 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 776 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 777 | _mm256_mask_mul_epi32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 778 | __m256i __Y) |
| 779 | { |
| 780 | return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, |
| 781 | (__v8si) __Y, |
| 782 | (__v4di) __W, __M); |
| 783 | } |
| 784 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 785 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 786 | _mm256_maskz_mul_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 787 | { |
| 788 | return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, |
| 789 | (__v8si) __Y, |
| 790 | (__v4di) |
| 791 | _mm256_setzero_si256 (), |
| 792 | __M); |
| 793 | } |
| 794 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 795 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 796 | _mm_mask_mul_epi32 (__m128i __W, __mmask8 __M, __m128i __X, |
| 797 | __m128i __Y) |
| 798 | { |
| 799 | return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, |
| 800 | (__v4si) __Y, |
| 801 | (__v2di) __W, __M); |
| 802 | } |
| 803 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 804 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 805 | _mm_maskz_mul_epi32 (__mmask8 __M, __m128i __X, __m128i __Y) |
| 806 | { |
| 807 | return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, |
| 808 | (__v4si) __Y, |
| 809 | (__v2di) |
| 810 | _mm_setzero_si128 (), |
| 811 | __M); |
| 812 | } |
| 813 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 814 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 815 | _mm256_mask_mul_epu32 (__m256i __W, __mmask8 __M, __m256i __X, |
| 816 | __m256i __Y) |
| 817 | { |
| 818 | return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, |
| 819 | (__v8si) __Y, |
| 820 | (__v4di) __W, __M); |
| 821 | } |
| 822 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 823 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 824 | _mm256_maskz_mul_epu32 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 825 | { |
| 826 | return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, |
| 827 | (__v8si) __Y, |
| 828 | (__v4di) |
| 829 | _mm256_setzero_si256 (), |
| 830 | __M); |
| 831 | } |
| 832 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 833 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 834 | _mm_mask_mul_epu32 (__m128i __W, __mmask8 __M, __m128i __X, |
| 835 | __m128i __Y) |
| 836 | { |
| 837 | return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, |
| 838 | (__v4si) __Y, |
| 839 | (__v2di) __W, __M); |
| 840 | } |
| 841 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 842 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 843 | _mm_maskz_mul_epu32 (__mmask8 __M, __m128i __X, __m128i __Y) |
| 844 | { |
| 845 | return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, |
| 846 | (__v4si) __Y, |
| 847 | (__v2di) |
| 848 | _mm_setzero_si128 (), |
| 849 | __M); |
| 850 | } |
| 851 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 852 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 853 | _mm256_maskz_mullo_epi32 (__mmask8 __M, __m256i __A, __m256i __B) |
| 854 | { |
| 855 | return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A, |
| 856 | (__v8si) __B, |
| 857 | (__v8si) |
| 858 | _mm256_setzero_si256 (), |
| 859 | __M); |
| 860 | } |
| 861 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 862 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 863 | _mm256_mask_mullo_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 864 | __m256i __B) |
| 865 | { |
| 866 | return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A, |
| 867 | (__v8si) __B, |
| 868 | (__v8si) __W, __M); |
| 869 | } |
| 870 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 871 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 872 | _mm_maskz_mullo_epi32 (__mmask8 __M, __m128i __A, __m128i __B) |
| 873 | { |
| 874 | return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A, |
| 875 | (__v4si) __B, |
| 876 | (__v4si) |
| 877 | _mm_setzero_si128 (), |
| 878 | __M); |
| 879 | } |
| 880 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 881 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 882 | _mm_mask_mullo_epi32 (__m128i __W, __mmask16 __M, __m128i __A, |
| 883 | __m128i __B) |
| 884 | { |
| 885 | return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A, |
| 886 | (__v4si) __B, |
| 887 | (__v4si) __W, __M); |
| 888 | } |
| 889 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 890 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 891 | _mm256_mask_and_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 892 | __m256i __B) |
| 893 | { |
| 894 | return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A, |
| 895 | (__v8si) __B, |
| 896 | (__v8si) __W, |
| 897 | (__mmask8) __U); |
| 898 | } |
| 899 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 900 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 901 | _mm256_maskz_and_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 902 | { |
| 903 | return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A, |
| 904 | (__v8si) __B, |
| 905 | (__v8si) |
| 906 | _mm256_setzero_si256 (), |
| 907 | (__mmask8) __U); |
| 908 | } |
| 909 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 910 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 911 | _mm_mask_and_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 912 | { |
| 913 | return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A, |
| 914 | (__v4si) __B, |
| 915 | (__v4si) __W, |
| 916 | (__mmask8) __U); |
| 917 | } |
| 918 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 919 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 920 | _mm_maskz_and_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 921 | { |
| 922 | return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A, |
| 923 | (__v4si) __B, |
| 924 | (__v4si) |
| 925 | _mm_setzero_si128 (), |
| 926 | (__mmask8) __U); |
| 927 | } |
| 928 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 929 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 930 | _mm256_mask_andnot_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 931 | __m256i __B) |
| 932 | { |
| 933 | return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A, |
| 934 | (__v8si) __B, |
| 935 | (__v8si) __W, |
| 936 | (__mmask8) __U); |
| 937 | } |
| 938 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 939 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 940 | _mm256_maskz_andnot_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 941 | { |
| 942 | return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A, |
| 943 | (__v8si) __B, |
| 944 | (__v8si) |
| 945 | _mm256_setzero_si256 (), |
| 946 | (__mmask8) __U); |
| 947 | } |
| 948 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 949 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 950 | _mm_mask_andnot_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 951 | __m128i __B) |
| 952 | { |
| 953 | return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A, |
| 954 | (__v4si) __B, |
| 955 | (__v4si) __W, |
| 956 | (__mmask8) __U); |
| 957 | } |
| 958 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 959 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 960 | _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 961 | { |
| 962 | return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A, |
| 963 | (__v4si) __B, |
| 964 | (__v4si) |
| 965 | _mm_setzero_si128 (), |
| 966 | (__mmask8) __U); |
| 967 | } |
| 968 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 969 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 970 | _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 971 | __m256i __B) |
| 972 | { |
| 973 | return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A, |
| 974 | (__v8si) __B, |
| 975 | (__v8si) __W, |
| 976 | (__mmask8) __U); |
| 977 | } |
| 978 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 979 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 980 | _mm256_maskz_or_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 981 | { |
| 982 | return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A, |
| 983 | (__v8si) __B, |
| 984 | (__v8si) |
| 985 | _mm256_setzero_si256 (), |
| 986 | (__mmask8) __U); |
| 987 | } |
| 988 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 989 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 990 | _mm_mask_or_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 991 | { |
| 992 | return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A, |
| 993 | (__v4si) __B, |
| 994 | (__v4si) __W, |
| 995 | (__mmask8) __U); |
| 996 | } |
| 997 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 998 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 999 | _mm_maskz_or_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1000 | { |
| 1001 | return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A, |
| 1002 | (__v4si) __B, |
| 1003 | (__v4si) |
| 1004 | _mm_setzero_si128 (), |
| 1005 | (__mmask8) __U); |
| 1006 | } |
| 1007 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1008 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1009 | _mm256_mask_xor_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 1010 | __m256i __B) |
| 1011 | { |
| 1012 | return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A, |
| 1013 | (__v8si) __B, |
| 1014 | (__v8si) __W, |
| 1015 | (__mmask8) __U); |
| 1016 | } |
| 1017 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1018 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1019 | _mm256_maskz_xor_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 1020 | { |
| 1021 | return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A, |
| 1022 | (__v8si) __B, |
| 1023 | (__v8si) |
| 1024 | _mm256_setzero_si256 (), |
| 1025 | (__mmask8) __U); |
| 1026 | } |
| 1027 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1028 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1029 | _mm_mask_xor_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 1030 | __m128i __B) |
| 1031 | { |
| 1032 | return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A, |
| 1033 | (__v4si) __B, |
| 1034 | (__v4si) __W, |
| 1035 | (__mmask8) __U); |
| 1036 | } |
| 1037 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1038 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1039 | _mm_maskz_xor_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1040 | { |
| 1041 | return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A, |
| 1042 | (__v4si) __B, |
| 1043 | (__v4si) |
| 1044 | _mm_setzero_si128 (), |
| 1045 | (__mmask8) __U); |
| 1046 | } |
| 1047 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1048 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1049 | _mm256_mask_and_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 1050 | __m256i __B) |
| 1051 | { |
| 1052 | return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A, |
| 1053 | (__v4di) __B, |
| 1054 | (__v4di) __W, __U); |
| 1055 | } |
| 1056 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1057 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1058 | _mm256_maskz_and_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 1059 | { |
| 1060 | return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A, |
| 1061 | (__v4di) __B, |
| 1062 | (__v4di) |
| 1063 | _mm256_setzero_pd (), |
| 1064 | __U); |
| 1065 | } |
| 1066 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1067 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1068 | _mm_mask_and_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 1069 | __m128i __B) |
| 1070 | { |
| 1071 | return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A, |
| 1072 | (__v2di) __B, |
| 1073 | (__v2di) __W, __U); |
| 1074 | } |
| 1075 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1076 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1077 | _mm_maskz_and_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1078 | { |
| 1079 | return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A, |
| 1080 | (__v2di) __B, |
| 1081 | (__v2di) |
| 1082 | _mm_setzero_pd (), |
| 1083 | __U); |
| 1084 | } |
| 1085 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1086 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1087 | _mm256_mask_andnot_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 1088 | __m256i __B) |
| 1089 | { |
| 1090 | return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A, |
| 1091 | (__v4di) __B, |
| 1092 | (__v4di) __W, __U); |
| 1093 | } |
| 1094 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1095 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1096 | _mm256_maskz_andnot_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 1097 | { |
| 1098 | return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A, |
| 1099 | (__v4di) __B, |
| 1100 | (__v4di) |
| 1101 | _mm256_setzero_pd (), |
| 1102 | __U); |
| 1103 | } |
| 1104 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1105 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1106 | _mm_mask_andnot_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 1107 | __m128i __B) |
| 1108 | { |
| 1109 | return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A, |
| 1110 | (__v2di) __B, |
| 1111 | (__v2di) __W, __U); |
| 1112 | } |
| 1113 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1114 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1115 | _mm_maskz_andnot_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1116 | { |
| 1117 | return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A, |
| 1118 | (__v2di) __B, |
| 1119 | (__v2di) |
| 1120 | _mm_setzero_pd (), |
| 1121 | __U); |
| 1122 | } |
| 1123 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1124 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1125 | _mm256_mask_or_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 1126 | __m256i __B) |
| 1127 | { |
| 1128 | return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A, |
| 1129 | (__v4di) __B, |
| 1130 | (__v4di) __W, |
| 1131 | (__mmask8) __U); |
| 1132 | } |
| 1133 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1134 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1135 | _mm256_maskz_or_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 1136 | { |
| 1137 | return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A, |
| 1138 | (__v4di) __B, |
| 1139 | (__v4di) |
| 1140 | _mm256_setzero_si256 (), |
| 1141 | (__mmask8) __U); |
| 1142 | } |
| 1143 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1144 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1145 | _mm_mask_or_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 1146 | { |
| 1147 | return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A, |
| 1148 | (__v2di) __B, |
| 1149 | (__v2di) __W, |
| 1150 | (__mmask8) __U); |
| 1151 | } |
| 1152 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1153 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1154 | _mm_maskz_or_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1155 | { |
| 1156 | return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A, |
| 1157 | (__v2di) __B, |
| 1158 | (__v2di) |
| 1159 | _mm_setzero_si128 (), |
| 1160 | (__mmask8) __U); |
| 1161 | } |
| 1162 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1163 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1164 | _mm256_mask_xor_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 1165 | __m256i __B) |
| 1166 | { |
| 1167 | return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A, |
| 1168 | (__v4di) __B, |
| 1169 | (__v4di) __W, |
| 1170 | (__mmask8) __U); |
| 1171 | } |
| 1172 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1173 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1174 | _mm256_maskz_xor_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 1175 | { |
| 1176 | return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A, |
| 1177 | (__v4di) __B, |
| 1178 | (__v4di) |
| 1179 | _mm256_setzero_si256 (), |
| 1180 | (__mmask8) __U); |
| 1181 | } |
| 1182 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1183 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1184 | _mm_mask_xor_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 1185 | __m128i __B) |
| 1186 | { |
| 1187 | return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A, |
| 1188 | (__v2di) __B, |
| 1189 | (__v2di) __W, |
| 1190 | (__mmask8) __U); |
| 1191 | } |
| 1192 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1193 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 1194 | _mm_maskz_xor_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 1195 | { |
| 1196 | return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A, |
| 1197 | (__v2di) __B, |
| 1198 | (__v2di) |
| 1199 | _mm_setzero_si128 (), |
| 1200 | (__mmask8) __U); |
| 1201 | } |
| 1202 | |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 1203 | #define _mm_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1204 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
| 1205 | (__v4si)(__m128i)(b), \ |
| 1206 | (p), (__mmask8)-1); }) |
| 1207 | |
| 1208 | #define _mm_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1209 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
| 1210 | (__v4si)(__m128i)(b), \ |
| 1211 | (p), (__mmask8)(m)); }) |
| 1212 | |
| 1213 | #define _mm_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1214 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
| 1215 | (__v4si)(__m128i)(b), \ |
| 1216 | (p), (__mmask8)-1); }) |
| 1217 | |
| 1218 | #define _mm_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1219 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
| 1220 | (__v4si)(__m128i)(b), \ |
| 1221 | (p), (__mmask8)(m)); }) |
| 1222 | |
| 1223 | #define _mm256_cmp_epi32_mask(a, b, p) __extension__ ({ \ |
| 1224 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
| 1225 | (__v8si)(__m256i)(b), \ |
| 1226 | (p), (__mmask8)-1); }) |
| 1227 | |
| 1228 | #define _mm256_mask_cmp_epi32_mask(m, a, b, p) __extension__ ({ \ |
| 1229 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
| 1230 | (__v8si)(__m256i)(b), \ |
| 1231 | (p), (__mmask8)(m)); }) |
| 1232 | |
| 1233 | #define _mm256_cmp_epu32_mask(a, b, p) __extension__ ({ \ |
| 1234 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
| 1235 | (__v8si)(__m256i)(b), \ |
| 1236 | (p), (__mmask8)-1); }) |
| 1237 | |
| 1238 | #define _mm256_mask_cmp_epu32_mask(m, a, b, p) __extension__ ({ \ |
| 1239 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
| 1240 | (__v8si)(__m256i)(b), \ |
| 1241 | (p), (__mmask8)(m)); }) |
| 1242 | |
| 1243 | #define _mm_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1244 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
| 1245 | (__v2di)(__m128i)(b), \ |
| 1246 | (p), (__mmask8)-1); }) |
| 1247 | |
| 1248 | #define _mm_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1249 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
| 1250 | (__v2di)(__m128i)(b), \ |
| 1251 | (p), (__mmask8)(m)); }) |
| 1252 | |
| 1253 | #define _mm_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1254 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
| 1255 | (__v2di)(__m128i)(b), \ |
| 1256 | (p), (__mmask8)-1); }) |
| 1257 | |
| 1258 | #define _mm_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1259 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
| 1260 | (__v2di)(__m128i)(b), \ |
| 1261 | (p), (__mmask8)(m)); }) |
| 1262 | |
| 1263 | #define _mm256_cmp_epi64_mask(a, b, p) __extension__ ({ \ |
| 1264 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
| 1265 | (__v4di)(__m256i)(b), \ |
| 1266 | (p), (__mmask8)-1); }) |
| 1267 | |
| 1268 | #define _mm256_mask_cmp_epi64_mask(m, a, b, p) __extension__ ({ \ |
| 1269 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
| 1270 | (__v4di)(__m256i)(b), \ |
| 1271 | (p), (__mmask8)(m)); }) |
| 1272 | |
| 1273 | #define _mm256_cmp_epu64_mask(a, b, p) __extension__ ({ \ |
| 1274 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
| 1275 | (__v4di)(__m256i)(b), \ |
| 1276 | (p), (__mmask8)-1); }) |
| 1277 | |
| 1278 | #define _mm256_mask_cmp_epu64_mask(m, a, b, p) __extension__ ({ \ |
| 1279 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
| 1280 | (__v4di)(__m256i)(b), \ |
| 1281 | (p), (__mmask8)(m)); }) |
| 1282 | |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 1283 | #define _mm256_cmp_ps_mask(a, b, p) __extension__ ({ \ |
| 1284 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
| 1285 | (__v8sf)(__m256)(b), \ |
| 1286 | (p), (__mmask8)-1); }) |
| 1287 | |
| 1288 | #define _mm256_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
| 1289 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
| 1290 | (__v8sf)(__m256)(b), \ |
| 1291 | (p), (__mmask8)(m)); }) |
| 1292 | |
| 1293 | #define _mm256_cmp_pd_mask(a, b, p) __extension__ ({ \ |
| 1294 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \ |
| 1295 | (__v4df)(__m256)(b), \ |
| 1296 | (p), (__mmask8)-1); }) |
| 1297 | |
| 1298 | #define _mm256_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
| 1299 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256)(a), \ |
| 1300 | (__v4df)(__m256)(b), \ |
| 1301 | (p), (__mmask8)(m)); }) |
| 1302 | |
| 1303 | #define _mm128_cmp_ps_mask(a, b, p) __extension__ ({ \ |
| 1304 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
| 1305 | (__v4sf)(__m128)(b), \ |
| 1306 | (p), (__mmask8)-1); }) |
| 1307 | |
| 1308 | #define _mm128_mask_cmp_ps_mask(m, a, b, p) __extension__ ({ \ |
| 1309 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
| 1310 | (__v4sf)(__m128)(b), \ |
| 1311 | (p), (__mmask8)(m)); }) |
| 1312 | |
| 1313 | #define _mm128_cmp_pd_mask(a, b, p) __extension__ ({ \ |
| 1314 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \ |
| 1315 | (__v2df)(__m128)(b), \ |
| 1316 | (p), (__mmask8)-1); }) |
| 1317 | |
| 1318 | #define _mm128_mask_cmp_pd_mask(m, a, b, p) __extension__ ({ \ |
| 1319 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128)(a), \ |
| 1320 | (__v2df)(__m128)(b), \ |
| 1321 | (p), (__mmask8)(m)); }) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 1322 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1323 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1324 | _mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1325 | { |
| 1326 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1327 | (__v2df) __B, |
| 1328 | (__v2df) __C, |
| 1329 | (__mmask8) __U); |
| 1330 | } |
| 1331 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1332 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1333 | _mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1334 | { |
| 1335 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A, |
| 1336 | (__v2df) __B, |
| 1337 | (__v2df) __C, |
| 1338 | (__mmask8) __U); |
| 1339 | } |
| 1340 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1341 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1342 | _mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1343 | { |
| 1344 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1345 | (__v2df) __B, |
| 1346 | (__v2df) __C, |
| 1347 | (__mmask8) __U); |
| 1348 | } |
| 1349 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1350 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1351 | _mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1352 | { |
| 1353 | return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, |
| 1354 | (__v2df) __B, |
| 1355 | -(__v2df) __C, |
| 1356 | (__mmask8) __U); |
| 1357 | } |
| 1358 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1359 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1360 | _mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1361 | { |
| 1362 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, |
| 1363 | (__v2df) __B, |
| 1364 | -(__v2df) __C, |
| 1365 | (__mmask8) __U); |
| 1366 | } |
| 1367 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1368 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1369 | _mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1370 | { |
| 1371 | return (__m128d) __builtin_ia32_vfmaddpd128_mask3 (-(__v2df) __A, |
| 1372 | (__v2df) __B, |
| 1373 | (__v2df) __C, |
| 1374 | (__mmask8) __U); |
| 1375 | } |
| 1376 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1377 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1378 | _mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1379 | { |
| 1380 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1381 | (__v2df) __B, |
| 1382 | (__v2df) __C, |
| 1383 | (__mmask8) __U); |
| 1384 | } |
| 1385 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1386 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1387 | _mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1388 | { |
| 1389 | return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A, |
| 1390 | (__v2df) __B, |
| 1391 | -(__v2df) __C, |
| 1392 | (__mmask8) __U); |
| 1393 | } |
| 1394 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1395 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1396 | _mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1397 | { |
| 1398 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1399 | (__v4df) __B, |
| 1400 | (__v4df) __C, |
| 1401 | (__mmask8) __U); |
| 1402 | } |
| 1403 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1404 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1405 | _mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1406 | { |
| 1407 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A, |
| 1408 | (__v4df) __B, |
| 1409 | (__v4df) __C, |
| 1410 | (__mmask8) __U); |
| 1411 | } |
| 1412 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1413 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1414 | _mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1415 | { |
| 1416 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1417 | (__v4df) __B, |
| 1418 | (__v4df) __C, |
| 1419 | (__mmask8) __U); |
| 1420 | } |
| 1421 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1422 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1423 | _mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1424 | { |
| 1425 | return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, |
| 1426 | (__v4df) __B, |
| 1427 | -(__v4df) __C, |
| 1428 | (__mmask8) __U); |
| 1429 | } |
| 1430 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1431 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1432 | _mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1433 | { |
| 1434 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, |
| 1435 | (__v4df) __B, |
| 1436 | -(__v4df) __C, |
| 1437 | (__mmask8) __U); |
| 1438 | } |
| 1439 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1440 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1441 | _mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1442 | { |
| 1443 | return (__m256d) __builtin_ia32_vfmaddpd256_mask3 (-(__v4df) __A, |
| 1444 | (__v4df) __B, |
| 1445 | (__v4df) __C, |
| 1446 | (__mmask8) __U); |
| 1447 | } |
| 1448 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1449 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1450 | _mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1451 | { |
| 1452 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1453 | (__v4df) __B, |
| 1454 | (__v4df) __C, |
| 1455 | (__mmask8) __U); |
| 1456 | } |
| 1457 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1458 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1459 | _mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1460 | { |
| 1461 | return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A, |
| 1462 | (__v4df) __B, |
| 1463 | -(__v4df) __C, |
| 1464 | (__mmask8) __U); |
| 1465 | } |
| 1466 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1467 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1468 | _mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1469 | { |
| 1470 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1471 | (__v4sf) __B, |
| 1472 | (__v4sf) __C, |
| 1473 | (__mmask8) __U); |
| 1474 | } |
| 1475 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1476 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1477 | _mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1478 | { |
| 1479 | return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A, |
| 1480 | (__v4sf) __B, |
| 1481 | (__v4sf) __C, |
| 1482 | (__mmask8) __U); |
| 1483 | } |
| 1484 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1485 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1486 | _mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1487 | { |
| 1488 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1489 | (__v4sf) __B, |
| 1490 | (__v4sf) __C, |
| 1491 | (__mmask8) __U); |
| 1492 | } |
| 1493 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1494 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1495 | _mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1496 | { |
| 1497 | return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, |
| 1498 | (__v4sf) __B, |
| 1499 | -(__v4sf) __C, |
| 1500 | (__mmask8) __U); |
| 1501 | } |
| 1502 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1503 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1504 | _mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1505 | { |
| 1506 | return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, |
| 1507 | (__v4sf) __B, |
| 1508 | -(__v4sf) __C, |
| 1509 | (__mmask8) __U); |
| 1510 | } |
| 1511 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1512 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1513 | _mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1514 | { |
| 1515 | return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A, |
| 1516 | (__v4sf) __B, |
| 1517 | (__v4sf) __C, |
| 1518 | (__mmask8) __U); |
| 1519 | } |
| 1520 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1521 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1522 | _mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1523 | { |
| 1524 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1525 | (__v4sf) __B, |
| 1526 | (__v4sf) __C, |
| 1527 | (__mmask8) __U); |
| 1528 | } |
| 1529 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1530 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1531 | _mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1532 | { |
| 1533 | return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, |
| 1534 | (__v4sf) __B, |
| 1535 | -(__v4sf) __C, |
| 1536 | (__mmask8) __U); |
| 1537 | } |
| 1538 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1539 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1540 | _mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1541 | { |
| 1542 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1543 | (__v8sf) __B, |
| 1544 | (__v8sf) __C, |
| 1545 | (__mmask8) __U); |
| 1546 | } |
| 1547 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1548 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1549 | _mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1550 | { |
| 1551 | return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A, |
| 1552 | (__v8sf) __B, |
| 1553 | (__v8sf) __C, |
| 1554 | (__mmask8) __U); |
| 1555 | } |
| 1556 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1557 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1558 | _mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1559 | { |
| 1560 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1561 | (__v8sf) __B, |
| 1562 | (__v8sf) __C, |
| 1563 | (__mmask8) __U); |
| 1564 | } |
| 1565 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1566 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1567 | _mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1568 | { |
| 1569 | return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, |
| 1570 | (__v8sf) __B, |
| 1571 | -(__v8sf) __C, |
| 1572 | (__mmask8) __U); |
| 1573 | } |
| 1574 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1575 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1576 | _mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1577 | { |
| 1578 | return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, |
| 1579 | (__v8sf) __B, |
| 1580 | -(__v8sf) __C, |
| 1581 | (__mmask8) __U); |
| 1582 | } |
| 1583 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1584 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1585 | _mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1586 | { |
| 1587 | return (__m256) __builtin_ia32_vfmaddps256_mask3 (-(__v8sf) __A, |
| 1588 | (__v8sf) __B, |
| 1589 | (__v8sf) __C, |
| 1590 | (__mmask8) __U); |
| 1591 | } |
| 1592 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1593 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1594 | _mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1595 | { |
| 1596 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1597 | (__v8sf) __B, |
| 1598 | (__v8sf) __C, |
| 1599 | (__mmask8) __U); |
| 1600 | } |
| 1601 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1602 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1603 | _mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1604 | { |
| 1605 | return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A, |
| 1606 | (__v8sf) __B, |
| 1607 | -(__v8sf) __C, |
| 1608 | (__mmask8) __U); |
| 1609 | } |
| 1610 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1611 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1612 | _mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1613 | { |
| 1614 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1615 | (__v2df) __B, |
| 1616 | (__v2df) __C, |
| 1617 | (__mmask8) __U); |
| 1618 | } |
| 1619 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1620 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1621 | _mm_mask3_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1622 | { |
| 1623 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A, |
| 1624 | (__v2df) __B, |
| 1625 | (__v2df) __C, |
| 1626 | (__mmask8) |
| 1627 | __U); |
| 1628 | } |
| 1629 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1630 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1631 | _mm_maskz_fmaddsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1632 | { |
| 1633 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1634 | (__v2df) __B, |
| 1635 | (__v2df) __C, |
| 1636 | (__mmask8) |
| 1637 | __U); |
| 1638 | } |
| 1639 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1640 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1641 | _mm_mask_fmsubadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1642 | { |
| 1643 | return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, |
| 1644 | (__v2df) __B, |
| 1645 | -(__v2df) __C, |
| 1646 | (__mmask8) __U); |
| 1647 | } |
| 1648 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1649 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1650 | _mm_maskz_fmsubadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1651 | { |
| 1652 | return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, |
| 1653 | (__v2df) __B, |
| 1654 | -(__v2df) __C, |
| 1655 | (__mmask8) |
| 1656 | __U); |
| 1657 | } |
| 1658 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1659 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1660 | _mm256_mask_fmaddsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1661 | { |
| 1662 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1663 | (__v4df) __B, |
| 1664 | (__v4df) __C, |
| 1665 | (__mmask8) __U); |
| 1666 | } |
| 1667 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1668 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1669 | _mm256_mask3_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1670 | { |
| 1671 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A, |
| 1672 | (__v4df) __B, |
| 1673 | (__v4df) __C, |
| 1674 | (__mmask8) |
| 1675 | __U); |
| 1676 | } |
| 1677 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1678 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1679 | _mm256_maskz_fmaddsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1680 | { |
| 1681 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1682 | (__v4df) __B, |
| 1683 | (__v4df) __C, |
| 1684 | (__mmask8) |
| 1685 | __U); |
| 1686 | } |
| 1687 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1688 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1689 | _mm256_mask_fmsubadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1690 | { |
| 1691 | return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, |
| 1692 | (__v4df) __B, |
| 1693 | -(__v4df) __C, |
| 1694 | (__mmask8) __U); |
| 1695 | } |
| 1696 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1697 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1698 | _mm256_maskz_fmsubadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1699 | { |
| 1700 | return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, |
| 1701 | (__v4df) __B, |
| 1702 | -(__v4df) __C, |
| 1703 | (__mmask8) |
| 1704 | __U); |
| 1705 | } |
| 1706 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1707 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1708 | _mm_mask_fmaddsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1709 | { |
| 1710 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1711 | (__v4sf) __B, |
| 1712 | (__v4sf) __C, |
| 1713 | (__mmask8) __U); |
| 1714 | } |
| 1715 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1716 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1717 | _mm_mask3_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1718 | { |
| 1719 | return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A, |
| 1720 | (__v4sf) __B, |
| 1721 | (__v4sf) __C, |
| 1722 | (__mmask8) __U); |
| 1723 | } |
| 1724 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1725 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1726 | _mm_maskz_fmaddsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1727 | { |
| 1728 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1729 | (__v4sf) __B, |
| 1730 | (__v4sf) __C, |
| 1731 | (__mmask8) __U); |
| 1732 | } |
| 1733 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1734 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1735 | _mm_mask_fmsubadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1736 | { |
| 1737 | return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, |
| 1738 | (__v4sf) __B, |
| 1739 | -(__v4sf) __C, |
| 1740 | (__mmask8) __U); |
| 1741 | } |
| 1742 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1743 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1744 | _mm_maskz_fmsubadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1745 | { |
| 1746 | return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, |
| 1747 | (__v4sf) __B, |
| 1748 | -(__v4sf) __C, |
| 1749 | (__mmask8) __U); |
| 1750 | } |
| 1751 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1752 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1753 | _mm256_mask_fmaddsub_ps(__m256 __A, __mmask8 __U, __m256 __B, |
| 1754 | __m256 __C) |
| 1755 | { |
| 1756 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1757 | (__v8sf) __B, |
| 1758 | (__v8sf) __C, |
| 1759 | (__mmask8) __U); |
| 1760 | } |
| 1761 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1762 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1763 | _mm256_mask3_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1764 | { |
| 1765 | return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A, |
| 1766 | (__v8sf) __B, |
| 1767 | (__v8sf) __C, |
| 1768 | (__mmask8) __U); |
| 1769 | } |
| 1770 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1771 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1772 | _mm256_maskz_fmaddsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1773 | { |
| 1774 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1775 | (__v8sf) __B, |
| 1776 | (__v8sf) __C, |
| 1777 | (__mmask8) __U); |
| 1778 | } |
| 1779 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1780 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1781 | _mm256_mask_fmsubadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1782 | { |
| 1783 | return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, |
| 1784 | (__v8sf) __B, |
| 1785 | -(__v8sf) __C, |
| 1786 | (__mmask8) __U); |
| 1787 | } |
| 1788 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1789 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1790 | _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1791 | { |
| 1792 | return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, |
| 1793 | (__v8sf) __B, |
| 1794 | -(__v8sf) __C, |
| 1795 | (__mmask8) __U); |
| 1796 | } |
| 1797 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1798 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1799 | _mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1800 | { |
| 1801 | return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A, |
| 1802 | (__v2df) __B, |
| 1803 | (__v2df) __C, |
| 1804 | (__mmask8) __U); |
| 1805 | } |
| 1806 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1807 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1808 | _mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1809 | { |
| 1810 | return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A, |
| 1811 | (__v4df) __B, |
| 1812 | (__v4df) __C, |
| 1813 | (__mmask8) __U); |
| 1814 | } |
| 1815 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1816 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1817 | _mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1818 | { |
| 1819 | return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A, |
| 1820 | (__v4sf) __B, |
| 1821 | (__v4sf) __C, |
| 1822 | (__mmask8) __U); |
| 1823 | } |
| 1824 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1825 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1826 | _mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1827 | { |
| 1828 | return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A, |
| 1829 | (__v8sf) __B, |
| 1830 | (__v8sf) __C, |
| 1831 | (__mmask8) __U); |
| 1832 | } |
| 1833 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1834 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1835 | _mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1836 | { |
| 1837 | return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A, |
| 1838 | (__v2df) __B, |
| 1839 | (__v2df) __C, |
| 1840 | (__mmask8) |
| 1841 | __U); |
| 1842 | } |
| 1843 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1844 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1845 | _mm256_mask3_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1846 | { |
| 1847 | return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A, |
| 1848 | (__v4df) __B, |
| 1849 | (__v4df) __C, |
| 1850 | (__mmask8) |
| 1851 | __U); |
| 1852 | } |
| 1853 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1854 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1855 | _mm_mask3_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1856 | { |
| 1857 | return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A, |
| 1858 | (__v4sf) __B, |
| 1859 | (__v4sf) __C, |
| 1860 | (__mmask8) __U); |
| 1861 | } |
| 1862 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1863 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1864 | _mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1865 | { |
| 1866 | return (__m256) __builtin_ia32_vfmsubaddps256_mask3 ((__v8sf) __A, |
| 1867 | (__v8sf) __B, |
| 1868 | (__v8sf) __C, |
| 1869 | (__mmask8) __U); |
| 1870 | } |
| 1871 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1872 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1873 | _mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1874 | { |
| 1875 | return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A, |
| 1876 | (__v2df) __B, |
| 1877 | (__v2df) __C, |
| 1878 | (__mmask8) __U); |
| 1879 | } |
| 1880 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1881 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1882 | _mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1883 | { |
| 1884 | return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A, |
| 1885 | (__v4df) __B, |
| 1886 | (__v4df) __C, |
| 1887 | (__mmask8) __U); |
| 1888 | } |
| 1889 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1890 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1891 | _mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1892 | { |
| 1893 | return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A, |
| 1894 | (__v4sf) __B, |
| 1895 | (__v4sf) __C, |
| 1896 | (__mmask8) __U); |
| 1897 | } |
| 1898 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1899 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1900 | _mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1901 | { |
| 1902 | return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A, |
| 1903 | (__v8sf) __B, |
| 1904 | (__v8sf) __C, |
| 1905 | (__mmask8) __U); |
| 1906 | } |
| 1907 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1908 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1909 | _mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1910 | { |
| 1911 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A, |
| 1912 | (__v2df) __B, |
| 1913 | (__v2df) __C, |
| 1914 | (__mmask8) __U); |
| 1915 | } |
| 1916 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1917 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1918 | _mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1919 | { |
| 1920 | return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A, |
| 1921 | (__v2df) __B, |
| 1922 | (__v2df) __C, |
| 1923 | (__mmask8) __U); |
| 1924 | } |
| 1925 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1926 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1927 | _mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1928 | { |
| 1929 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A, |
| 1930 | (__v4df) __B, |
| 1931 | (__v4df) __C, |
| 1932 | (__mmask8) __U); |
| 1933 | } |
| 1934 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1935 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1936 | _mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1937 | { |
| 1938 | return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A, |
| 1939 | (__v4df) __B, |
| 1940 | (__v4df) __C, |
| 1941 | (__mmask8) __U); |
| 1942 | } |
| 1943 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1944 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1945 | _mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1946 | { |
| 1947 | return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A, |
| 1948 | (__v4sf) __B, |
| 1949 | (__v4sf) __C, |
| 1950 | (__mmask8) __U); |
| 1951 | } |
| 1952 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1953 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1954 | _mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1955 | { |
| 1956 | return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A, |
| 1957 | (__v4sf) __B, |
| 1958 | (__v4sf) __C, |
| 1959 | (__mmask8) __U); |
| 1960 | } |
| 1961 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1962 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1963 | _mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1964 | { |
| 1965 | return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A, |
| 1966 | (__v8sf) __B, |
| 1967 | (__v8sf) __C, |
| 1968 | (__mmask8) __U); |
| 1969 | } |
| 1970 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1971 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1972 | _mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1973 | { |
| 1974 | return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A, |
| 1975 | (__v8sf) __B, |
| 1976 | (__v8sf) __C, |
| 1977 | (__mmask8) __U); |
| 1978 | } |
| 1979 | |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1980 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1981 | _mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 1982 | return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, |
| 1983 | (__v2df) __B, |
| 1984 | (__v2df) __W, |
| 1985 | (__mmask8) __U); |
| 1986 | } |
| 1987 | |
| 1988 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1989 | _mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 1990 | return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, |
| 1991 | (__v2df) __B, |
| 1992 | (__v2df) |
| 1993 | _mm_setzero_pd (), |
| 1994 | (__mmask8) __U); |
| 1995 | } |
| 1996 | |
| 1997 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1998 | _mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 1999 | return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, |
| 2000 | (__v4df) __B, |
| 2001 | (__v4df) __W, |
| 2002 | (__mmask8) __U); |
| 2003 | } |
| 2004 | |
| 2005 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2006 | _mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 2007 | return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, |
| 2008 | (__v4df) __B, |
| 2009 | (__v4df) |
| 2010 | _mm256_setzero_pd (), |
| 2011 | (__mmask8) __U); |
| 2012 | } |
| 2013 | |
| 2014 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2015 | _mm_mask_add_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) { |
| 2016 | return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, |
| 2017 | (__v4sf) __B, |
| 2018 | (__v4sf) __W, |
| 2019 | (__mmask8) __U); |
| 2020 | } |
| 2021 | |
| 2022 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2023 | _mm_maskz_add_ps (__mmask16 __U, __m128 __A, __m128 __B) { |
| 2024 | return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, |
| 2025 | (__v4sf) __B, |
| 2026 | (__v4sf) |
| 2027 | _mm_setzero_ps (), |
| 2028 | (__mmask8) __U); |
| 2029 | } |
| 2030 | |
| 2031 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2032 | _mm256_mask_add_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) { |
| 2033 | return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, |
| 2034 | (__v8sf) __B, |
| 2035 | (__v8sf) __W, |
| 2036 | (__mmask8) __U); |
| 2037 | } |
| 2038 | |
| 2039 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2040 | _mm256_maskz_add_ps (__mmask16 __U, __m256 __A, __m256 __B) { |
| 2041 | return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, |
| 2042 | (__v8sf) __B, |
| 2043 | (__v8sf) |
| 2044 | _mm256_setzero_ps (), |
| 2045 | (__mmask8) __U); |
| 2046 | } |
| 2047 | |
| 2048 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2049 | _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) { |
| 2050 | return (__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) __A, |
| 2051 | (__v4si) __W, |
| 2052 | (__mmask8) __U); |
| 2053 | } |
| 2054 | |
| 2055 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2056 | _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) { |
| 2057 | return (__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) __A, |
| 2058 | (__v8si) __W, |
| 2059 | (__mmask8) __U); |
| 2060 | } |
| 2061 | |
| 2062 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2063 | _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) { |
| 2064 | return (__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) __A, |
| 2065 | (__v2df) __W, |
| 2066 | (__mmask8) __U); |
| 2067 | } |
| 2068 | |
| 2069 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2070 | _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) { |
| 2071 | return (__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) __A, |
| 2072 | (__v4df) __W, |
| 2073 | (__mmask8) __U); |
| 2074 | } |
| 2075 | |
| 2076 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2077 | _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) { |
| 2078 | return (__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) __A, |
| 2079 | (__v4sf) __W, |
| 2080 | (__mmask8) __U); |
| 2081 | } |
| 2082 | |
| 2083 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2084 | _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) { |
| 2085 | return (__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) __A, |
| 2086 | (__v8sf) __W, |
| 2087 | (__mmask8) __U); |
| 2088 | } |
| 2089 | |
| 2090 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2091 | _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) { |
| 2092 | return (__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) __A, |
| 2093 | (__v2di) __W, |
| 2094 | (__mmask8) __U); |
| 2095 | } |
| 2096 | |
| 2097 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2098 | _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) { |
| 2099 | return (__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) __A, |
| 2100 | (__v4di) __W, |
| 2101 | (__mmask8) __U); |
| 2102 | } |
| 2103 | |
| 2104 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2105 | _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2106 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 2107 | (__v2df) __W, |
| 2108 | (__mmask8) __U); |
| 2109 | } |
| 2110 | |
| 2111 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2112 | _mm_maskz_compress_pd (__mmask8 __U, __m128d __A) { |
| 2113 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 2114 | (__v2df) |
| 2115 | _mm_setzero_pd (), |
| 2116 | (__mmask8) __U); |
| 2117 | } |
| 2118 | |
| 2119 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2120 | _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2121 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 2122 | (__v4df) __W, |
| 2123 | (__mmask8) __U); |
| 2124 | } |
| 2125 | |
| 2126 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2127 | _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) { |
| 2128 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 2129 | (__v4df) |
| 2130 | _mm256_setzero_pd (), |
| 2131 | (__mmask8) __U); |
| 2132 | } |
| 2133 | |
| 2134 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2135 | _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2136 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2137 | (__v2di) __W, |
| 2138 | (__mmask8) __U); |
| 2139 | } |
| 2140 | |
| 2141 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2142 | _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) { |
| 2143 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 2144 | (__v2di) |
| 2145 | _mm_setzero_si128 (), |
| 2146 | (__mmask8) __U); |
| 2147 | } |
| 2148 | |
| 2149 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2150 | _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2151 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2152 | (__v4di) __W, |
| 2153 | (__mmask8) __U); |
| 2154 | } |
| 2155 | |
| 2156 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2157 | _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) { |
| 2158 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 2159 | (__v4di) |
| 2160 | _mm256_setzero_si256 (), |
| 2161 | (__mmask8) __U); |
| 2162 | } |
| 2163 | |
| 2164 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2165 | _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2166 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2167 | (__v4sf) __W, |
| 2168 | (__mmask8) __U); |
| 2169 | } |
| 2170 | |
| 2171 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2172 | _mm_maskz_compress_ps (__mmask8 __U, __m128 __A) { |
| 2173 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 2174 | (__v4sf) |
| 2175 | _mm_setzero_ps (), |
| 2176 | (__mmask8) __U); |
| 2177 | } |
| 2178 | |
| 2179 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2180 | _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2181 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2182 | (__v8sf) __W, |
| 2183 | (__mmask8) __U); |
| 2184 | } |
| 2185 | |
| 2186 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2187 | _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) { |
| 2188 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 2189 | (__v8sf) |
| 2190 | _mm256_setzero_ps (), |
| 2191 | (__mmask8) __U); |
| 2192 | } |
| 2193 | |
| 2194 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2195 | _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2196 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2197 | (__v4si) __W, |
| 2198 | (__mmask8) __U); |
| 2199 | } |
| 2200 | |
| 2201 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2202 | _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) { |
| 2203 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 2204 | (__v4si) |
| 2205 | _mm_setzero_si128 (), |
| 2206 | (__mmask8) __U); |
| 2207 | } |
| 2208 | |
| 2209 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2210 | _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2211 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2212 | (__v8si) __W, |
| 2213 | (__mmask8) __U); |
| 2214 | } |
| 2215 | |
| 2216 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2217 | _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) { |
| 2218 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 2219 | (__v8si) |
| 2220 | _mm256_setzero_si256 (), |
| 2221 | (__mmask8) __U); |
| 2222 | } |
| 2223 | |
| 2224 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2225 | _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) { |
| 2226 | __builtin_ia32_compressstoredf128_mask ((__v2df *) __P, |
| 2227 | (__v2df) __A, |
| 2228 | (__mmask8) __U); |
| 2229 | } |
| 2230 | |
| 2231 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2232 | _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) { |
| 2233 | __builtin_ia32_compressstoredf256_mask ((__v4df *) __P, |
| 2234 | (__v4df) __A, |
| 2235 | (__mmask8) __U); |
| 2236 | } |
| 2237 | |
| 2238 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2239 | _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) { |
| 2240 | __builtin_ia32_compressstoredi128_mask ((__v2di *) __P, |
| 2241 | (__v2di) __A, |
| 2242 | (__mmask8) __U); |
| 2243 | } |
| 2244 | |
| 2245 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2246 | _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) { |
| 2247 | __builtin_ia32_compressstoredi256_mask ((__v4di *) __P, |
| 2248 | (__v4di) __A, |
| 2249 | (__mmask8) __U); |
| 2250 | } |
| 2251 | |
| 2252 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2253 | _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) { |
| 2254 | __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, |
| 2255 | (__v4sf) __A, |
| 2256 | (__mmask8) __U); |
| 2257 | } |
| 2258 | |
| 2259 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2260 | _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) { |
| 2261 | __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P, |
| 2262 | (__v8sf) __A, |
| 2263 | (__mmask8) __U); |
| 2264 | } |
| 2265 | |
| 2266 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2267 | _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) { |
| 2268 | __builtin_ia32_compressstoresi128_mask ((__v4si *) __P, |
| 2269 | (__v4si) __A, |
| 2270 | (__mmask8) __U); |
| 2271 | } |
| 2272 | |
| 2273 | static __inline__ void __DEFAULT_FN_ATTRS |
| 2274 | _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) { |
| 2275 | __builtin_ia32_compressstoresi256_mask ((__v8si *) __P, |
| 2276 | (__v8si) __A, |
| 2277 | (__mmask8) __U); |
| 2278 | } |
| 2279 | |
| 2280 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2281 | _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2282 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2283 | (__v2df) __W, |
| 2284 | (__mmask8) __U); |
| 2285 | } |
| 2286 | |
| 2287 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2288 | _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2289 | return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, |
| 2290 | (__v2df) |
| 2291 | _mm_setzero_pd (), |
| 2292 | (__mmask8) __U); |
| 2293 | } |
| 2294 | |
| 2295 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2296 | _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2297 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2298 | (__v4df) __W, |
| 2299 | (__mmask8) __U); |
| 2300 | } |
| 2301 | |
| 2302 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2303 | _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
| 2304 | return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, |
| 2305 | (__v4df) |
| 2306 | _mm256_setzero_pd (), |
| 2307 | (__mmask8) __U); |
| 2308 | } |
| 2309 | |
| 2310 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2311 | _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2312 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2313 | (__v4sf) __W, |
| 2314 | (__mmask8) __U); |
| 2315 | } |
| 2316 | |
| 2317 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2318 | _mm_maskz_cvtepi32_ps (__mmask16 __U, __m128i __A) { |
| 2319 | return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, |
| 2320 | (__v4sf) |
| 2321 | _mm_setzero_ps (), |
| 2322 | (__mmask8) __U); |
| 2323 | } |
| 2324 | |
| 2325 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2326 | _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2327 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2328 | (__v8sf) __W, |
| 2329 | (__mmask8) __U); |
| 2330 | } |
| 2331 | |
| 2332 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2333 | _mm256_maskz_cvtepi32_ps (__mmask16 __U, __m256i __A) { |
| 2334 | return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, |
| 2335 | (__v8sf) |
| 2336 | _mm256_setzero_ps (), |
| 2337 | (__mmask8) __U); |
| 2338 | } |
| 2339 | |
| 2340 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2341 | _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2342 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2343 | (__v4si) __W, |
| 2344 | (__mmask8) __U); |
| 2345 | } |
| 2346 | |
| 2347 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2348 | _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2349 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 2350 | (__v4si) |
| 2351 | _mm_setzero_si128 (), |
| 2352 | (__mmask8) __U); |
| 2353 | } |
| 2354 | |
| 2355 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2356 | _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2357 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2358 | (__v4si) __W, |
| 2359 | (__mmask8) __U); |
| 2360 | } |
| 2361 | |
| 2362 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2363 | _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2364 | return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, |
| 2365 | (__v4si) |
| 2366 | _mm_setzero_si128 (), |
| 2367 | (__mmask8) __U); |
| 2368 | } |
| 2369 | |
| 2370 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2371 | _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) { |
| 2372 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2373 | (__v4sf) __W, |
| 2374 | (__mmask8) __U); |
| 2375 | } |
| 2376 | |
| 2377 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2378 | _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) { |
| 2379 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 2380 | (__v4sf) |
| 2381 | _mm_setzero_ps (), |
| 2382 | (__mmask8) __U); |
| 2383 | } |
| 2384 | |
| 2385 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2386 | _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) { |
| 2387 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2388 | (__v4sf) __W, |
| 2389 | (__mmask8) __U); |
| 2390 | } |
| 2391 | |
| 2392 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2393 | _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) { |
| 2394 | return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, |
| 2395 | (__v4sf) |
| 2396 | _mm_setzero_ps (), |
| 2397 | (__mmask8) __U); |
| 2398 | } |
| 2399 | |
| 2400 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2401 | _mm_cvtpd_epu32 (__m128d __A) { |
| 2402 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2403 | (__v4si) |
| 2404 | _mm_setzero_si128 (), |
| 2405 | (__mmask8) -1); |
| 2406 | } |
| 2407 | |
| 2408 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2409 | _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2410 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2411 | (__v4si) __W, |
| 2412 | (__mmask8) __U); |
| 2413 | } |
| 2414 | |
| 2415 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2416 | _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2417 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 2418 | (__v4si) |
| 2419 | _mm_setzero_si128 (), |
| 2420 | (__mmask8) __U); |
| 2421 | } |
| 2422 | |
| 2423 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2424 | _mm256_cvtpd_epu32 (__m256d __A) { |
| 2425 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2426 | (__v4si) |
| 2427 | _mm_setzero_si128 (), |
| 2428 | (__mmask8) -1); |
| 2429 | } |
| 2430 | |
| 2431 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2432 | _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2433 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2434 | (__v4si) __W, |
| 2435 | (__mmask8) __U); |
| 2436 | } |
| 2437 | |
| 2438 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2439 | _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2440 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 2441 | (__v4si) |
| 2442 | _mm_setzero_si128 (), |
| 2443 | (__mmask8) __U); |
| 2444 | } |
| 2445 | |
| 2446 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2447 | _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2448 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2449 | (__v4si) __W, |
| 2450 | (__mmask8) __U); |
| 2451 | } |
| 2452 | |
| 2453 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2454 | _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) { |
| 2455 | return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, |
| 2456 | (__v4si) |
| 2457 | _mm_setzero_si128 (), |
| 2458 | (__mmask8) __U); |
| 2459 | } |
| 2460 | |
| 2461 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2462 | _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2463 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2464 | (__v8si) __W, |
| 2465 | (__mmask8) __U); |
| 2466 | } |
| 2467 | |
| 2468 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2469 | _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) { |
| 2470 | return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, |
| 2471 | (__v8si) |
| 2472 | _mm256_setzero_si256 (), |
| 2473 | (__mmask8) __U); |
| 2474 | } |
| 2475 | |
| 2476 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2477 | _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) { |
| 2478 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2479 | (__v2df) __W, |
| 2480 | (__mmask8) __U); |
| 2481 | } |
| 2482 | |
| 2483 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2484 | _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2485 | return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, |
| 2486 | (__v2df) |
| 2487 | _mm_setzero_pd (), |
| 2488 | (__mmask8) __U); |
| 2489 | } |
| 2490 | |
| 2491 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2492 | _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) { |
| 2493 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2494 | (__v4df) __W, |
| 2495 | (__mmask8) __U); |
| 2496 | } |
| 2497 | |
| 2498 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2499 | _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
| 2500 | return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, |
| 2501 | (__v4df) |
| 2502 | _mm256_setzero_pd (), |
| 2503 | (__mmask8) __U); |
| 2504 | } |
| 2505 | |
| 2506 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2507 | _mm_cvtps_epu32 (__m128 __A) { |
| 2508 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2509 | (__v4si) |
| 2510 | _mm_setzero_si128 (), |
| 2511 | (__mmask8) -1); |
| 2512 | } |
| 2513 | |
| 2514 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2515 | _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2516 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2517 | (__v4si) __W, |
| 2518 | (__mmask8) __U); |
| 2519 | } |
| 2520 | |
| 2521 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2522 | _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) { |
| 2523 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2524 | (__v4si) |
| 2525 | _mm_setzero_si128 (), |
| 2526 | (__mmask8) __U); |
| 2527 | } |
| 2528 | |
| 2529 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2530 | _mm256_cvtps_epu32 (__m256 __A) { |
| 2531 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2532 | (__v8si) |
| 2533 | _mm256_setzero_si256 (), |
| 2534 | (__mmask8) -1); |
| 2535 | } |
| 2536 | |
| 2537 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2538 | _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2539 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2540 | (__v8si) __W, |
| 2541 | (__mmask8) __U); |
| 2542 | } |
| 2543 | |
| 2544 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2545 | _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) { |
| 2546 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2547 | (__v8si) |
| 2548 | _mm256_setzero_si256 (), |
| 2549 | (__mmask8) __U); |
| 2550 | } |
| 2551 | |
| 2552 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2553 | _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2554 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2555 | (__v4si) __W, |
| 2556 | (__mmask8) __U); |
| 2557 | } |
| 2558 | |
| 2559 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2560 | _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2561 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2562 | (__v4si) |
| 2563 | _mm_setzero_si128 (), |
| 2564 | (__mmask8) __U); |
| 2565 | } |
| 2566 | |
| 2567 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2568 | _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2569 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2570 | (__v4si) __W, |
| 2571 | (__mmask8) __U); |
| 2572 | } |
| 2573 | |
| 2574 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2575 | _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) { |
| 2576 | return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, |
| 2577 | (__v4si) |
| 2578 | _mm_setzero_si128 (), |
| 2579 | (__mmask8) __U); |
| 2580 | } |
| 2581 | |
| 2582 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2583 | _mm_cvttpd_epu32 (__m128d __A) { |
| 2584 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2585 | (__v4si) |
| 2586 | _mm_setzero_si128 (), |
| 2587 | (__mmask8) -1); |
| 2588 | } |
| 2589 | |
| 2590 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2591 | _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2592 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2593 | (__v4si) __W, |
| 2594 | (__mmask8) __U); |
| 2595 | } |
| 2596 | |
| 2597 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2598 | _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2599 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2600 | (__v4si) |
| 2601 | _mm_setzero_si128 (), |
| 2602 | (__mmask8) __U); |
| 2603 | } |
| 2604 | |
| 2605 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2606 | _mm256_cvttpd_epu32 (__m256d __A) { |
| 2607 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2608 | (__v4si) |
| 2609 | _mm_setzero_si128 (), |
| 2610 | (__mmask8) -1); |
| 2611 | } |
| 2612 | |
| 2613 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2614 | _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2615 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2616 | (__v4si) __W, |
| 2617 | (__mmask8) __U); |
| 2618 | } |
| 2619 | |
| 2620 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2621 | _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2622 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2623 | (__v4si) |
| 2624 | _mm_setzero_si128 (), |
| 2625 | (__mmask8) __U); |
| 2626 | } |
| 2627 | |
| 2628 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2629 | _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2630 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2631 | (__v4si) __W, |
| 2632 | (__mmask8) __U); |
| 2633 | } |
| 2634 | |
| 2635 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2636 | _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) { |
| 2637 | return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, |
| 2638 | (__v4si) |
| 2639 | _mm_setzero_si128 (), |
| 2640 | (__mmask8) __U); |
| 2641 | } |
| 2642 | |
| 2643 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2644 | _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2645 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2646 | (__v8si) __W, |
| 2647 | (__mmask8) __U); |
| 2648 | } |
| 2649 | |
| 2650 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2651 | _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) { |
| 2652 | return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, |
| 2653 | (__v8si) |
| 2654 | _mm256_setzero_si256 (), |
| 2655 | (__mmask8) __U); |
| 2656 | } |
| 2657 | |
| 2658 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2659 | _mm_cvttps_epu32 (__m128 __A) { |
| 2660 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2661 | (__v4si) |
| 2662 | _mm_setzero_si128 (), |
| 2663 | (__mmask8) -1); |
| 2664 | } |
| 2665 | |
| 2666 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2667 | _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2668 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2669 | (__v4si) __W, |
| 2670 | (__mmask8) __U); |
| 2671 | } |
| 2672 | |
| 2673 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2674 | _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) { |
| 2675 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2676 | (__v4si) |
| 2677 | _mm_setzero_si128 (), |
| 2678 | (__mmask8) __U); |
| 2679 | } |
| 2680 | |
| 2681 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2682 | _mm256_cvttps_epu32 (__m256 __A) { |
| 2683 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2684 | (__v8si) |
| 2685 | _mm256_setzero_si256 (), |
| 2686 | (__mmask8) -1); |
| 2687 | } |
| 2688 | |
| 2689 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2690 | _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2691 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2692 | (__v8si) __W, |
| 2693 | (__mmask8) __U); |
| 2694 | } |
| 2695 | |
| 2696 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2697 | _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) { |
| 2698 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2699 | (__v8si) |
| 2700 | _mm256_setzero_si256 (), |
| 2701 | (__mmask8) __U); |
| 2702 | } |
| 2703 | |
| 2704 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2705 | _mm_cvtepu32_pd (__m128i __A) { |
| 2706 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2707 | (__v2df) |
| 2708 | _mm_setzero_pd (), |
| 2709 | (__mmask8) -1); |
| 2710 | } |
| 2711 | |
| 2712 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2713 | _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
| 2714 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2715 | (__v2df) __W, |
| 2716 | (__mmask8) __U); |
| 2717 | } |
| 2718 | |
| 2719 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2720 | _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2721 | return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, |
| 2722 | (__v2df) |
| 2723 | _mm_setzero_pd (), |
| 2724 | (__mmask8) __U); |
| 2725 | } |
| 2726 | |
| 2727 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2728 | _mm256_cvtepu32_pd (__m128i __A) { |
| 2729 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2730 | (__v4df) |
| 2731 | _mm256_setzero_pd (), |
| 2732 | (__mmask8) -1); |
| 2733 | } |
| 2734 | |
| 2735 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2736 | _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
| 2737 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2738 | (__v4df) __W, |
| 2739 | (__mmask8) __U); |
| 2740 | } |
| 2741 | |
| 2742 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2743 | _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
| 2744 | return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, |
| 2745 | (__v4df) |
| 2746 | _mm256_setzero_pd (), |
| 2747 | (__mmask8) __U); |
| 2748 | } |
| 2749 | |
| 2750 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2751 | _mm_cvtepu32_ps (__m128i __A) { |
| 2752 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2753 | (__v4sf) |
| 2754 | _mm_setzero_ps (), |
| 2755 | (__mmask8) -1); |
| 2756 | } |
| 2757 | |
| 2758 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2759 | _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 2760 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2761 | (__v4sf) __W, |
| 2762 | (__mmask8) __U); |
| 2763 | } |
| 2764 | |
| 2765 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2766 | _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) { |
| 2767 | return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, |
| 2768 | (__v4sf) |
| 2769 | _mm_setzero_ps (), |
| 2770 | (__mmask8) __U); |
| 2771 | } |
| 2772 | |
| 2773 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2774 | _mm256_cvtepu32_ps (__m256i __A) { |
| 2775 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2776 | (__v8sf) |
| 2777 | _mm256_setzero_ps (), |
| 2778 | (__mmask8) -1); |
| 2779 | } |
| 2780 | |
| 2781 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2782 | _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
| 2783 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2784 | (__v8sf) __W, |
| 2785 | (__mmask8) __U); |
| 2786 | } |
| 2787 | |
| 2788 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2789 | _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) { |
| 2790 | return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, |
| 2791 | (__v8sf) |
| 2792 | _mm256_setzero_ps (), |
| 2793 | (__mmask8) __U); |
| 2794 | } |
| 2795 | |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2796 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2797 | _mm_mask_div_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2798 | return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, |
| 2799 | (__v2df) __B, |
| 2800 | (__v2df) __W, |
| 2801 | (__mmask8) __U); |
| 2802 | } |
| 2803 | |
| 2804 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2805 | _mm_maskz_div_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 2806 | return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, |
| 2807 | (__v2df) __B, |
| 2808 | (__v2df) |
| 2809 | _mm_setzero_pd (), |
| 2810 | (__mmask8) __U); |
| 2811 | } |
| 2812 | |
| 2813 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2814 | _mm256_mask_div_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 2815 | __m256d __B) { |
| 2816 | return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, |
| 2817 | (__v4df) __B, |
| 2818 | (__v4df) __W, |
| 2819 | (__mmask8) __U); |
| 2820 | } |
| 2821 | |
| 2822 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2823 | _mm256_maskz_div_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 2824 | return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, |
| 2825 | (__v4df) __B, |
| 2826 | (__v4df) |
| 2827 | _mm256_setzero_pd (), |
| 2828 | (__mmask8) __U); |
| 2829 | } |
| 2830 | |
| 2831 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2832 | _mm_mask_div_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2833 | return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, |
| 2834 | (__v4sf) __B, |
| 2835 | (__v4sf) __W, |
| 2836 | (__mmask8) __U); |
| 2837 | } |
| 2838 | |
| 2839 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2840 | _mm_maskz_div_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 2841 | return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, |
| 2842 | (__v4sf) __B, |
| 2843 | (__v4sf) |
| 2844 | _mm_setzero_ps (), |
| 2845 | (__mmask8) __U); |
| 2846 | } |
| 2847 | |
| 2848 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2849 | _mm256_mask_div_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2850 | return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, |
| 2851 | (__v8sf) __B, |
| 2852 | (__v8sf) __W, |
| 2853 | (__mmask8) __U); |
| 2854 | } |
| 2855 | |
| 2856 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2857 | _mm256_maskz_div_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 2858 | return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, |
| 2859 | (__v8sf) __B, |
| 2860 | (__v8sf) |
| 2861 | _mm256_setzero_ps (), |
| 2862 | (__mmask8) __U); |
| 2863 | } |
| 2864 | |
| 2865 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2866 | _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2867 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2868 | (__v2df) __W, |
| 2869 | (__mmask8) __U); |
| 2870 | } |
| 2871 | |
| 2872 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2873 | _mm_maskz_expand_pd (__mmask8 __U, __m128d __A) { |
| 2874 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2875 | (__v2df) |
| 2876 | _mm_setzero_pd (), |
| 2877 | (__mmask8) __U); |
| 2878 | } |
| 2879 | |
| 2880 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2881 | _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2882 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2883 | (__v4df) __W, |
| 2884 | (__mmask8) __U); |
| 2885 | } |
| 2886 | |
| 2887 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2888 | _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) { |
| 2889 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2890 | (__v4df) |
| 2891 | _mm256_setzero_pd (), |
| 2892 | (__mmask8) __U); |
| 2893 | } |
| 2894 | |
| 2895 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2896 | _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2897 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2898 | (__v2di) __W, |
| 2899 | (__mmask8) __U); |
| 2900 | } |
| 2901 | |
| 2902 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2903 | _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) { |
| 2904 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2905 | (__v2di) |
| 2906 | _mm_setzero_si128 (), |
| 2907 | (__mmask8) __U); |
| 2908 | } |
| 2909 | |
| 2910 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2911 | _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2912 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2913 | (__v4di) __W, |
| 2914 | (__mmask8) __U); |
| 2915 | } |
| 2916 | |
| 2917 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2918 | _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) { |
| 2919 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2920 | (__v4di) |
| 2921 | _mm256_setzero_si256 (), |
| 2922 | (__mmask8) __U); |
| 2923 | } |
| 2924 | |
| 2925 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2926 | _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) { |
| 2927 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2928 | (__v2df) __W, |
| 2929 | (__mmask8) |
| 2930 | __U); |
| 2931 | } |
| 2932 | |
| 2933 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2934 | _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2935 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2936 | (__v2df) |
| 2937 | _mm_setzero_pd (), |
| 2938 | (__mmask8) |
| 2939 | __U); |
| 2940 | } |
| 2941 | |
| 2942 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2943 | _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) { |
| 2944 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2945 | (__v4df) __W, |
| 2946 | (__mmask8) |
| 2947 | __U); |
| 2948 | } |
| 2949 | |
| 2950 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2951 | _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2952 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2953 | (__v4df) |
| 2954 | _mm256_setzero_pd (), |
| 2955 | (__mmask8) |
| 2956 | __U); |
| 2957 | } |
| 2958 | |
| 2959 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2960 | _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2961 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2962 | (__v2di) __W, |
| 2963 | (__mmask8) |
| 2964 | __U); |
| 2965 | } |
| 2966 | |
| 2967 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2968 | _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2969 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2970 | (__v2di) |
| 2971 | _mm_setzero_si128 (), |
| 2972 | (__mmask8) |
| 2973 | __U); |
| 2974 | } |
| 2975 | |
| 2976 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2977 | _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U, |
| 2978 | void const *__P) { |
| 2979 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2980 | (__v4di) __W, |
| 2981 | (__mmask8) |
| 2982 | __U); |
| 2983 | } |
| 2984 | |
| 2985 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2986 | _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2987 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2988 | (__v4di) |
| 2989 | _mm256_setzero_si256 (), |
| 2990 | (__mmask8) |
| 2991 | __U); |
| 2992 | } |
| 2993 | |
| 2994 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2995 | _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) { |
| 2996 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2997 | (__v4sf) __W, |
| 2998 | (__mmask8) __U); |
| 2999 | } |
| 3000 | |
| 3001 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3002 | _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 3003 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 3004 | (__v4sf) |
| 3005 | _mm_setzero_ps (), |
| 3006 | (__mmask8) |
| 3007 | __U); |
| 3008 | } |
| 3009 | |
| 3010 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3011 | _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) { |
| 3012 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 3013 | (__v8sf) __W, |
| 3014 | (__mmask8) __U); |
| 3015 | } |
| 3016 | |
| 3017 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3018 | _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 3019 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 3020 | (__v8sf) |
| 3021 | _mm256_setzero_ps (), |
| 3022 | (__mmask8) |
| 3023 | __U); |
| 3024 | } |
| 3025 | |
| 3026 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3027 | _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) { |
| 3028 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 3029 | (__v4si) __W, |
| 3030 | (__mmask8) |
| 3031 | __U); |
| 3032 | } |
| 3033 | |
| 3034 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3035 | _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 3036 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 3037 | (__v4si) |
| 3038 | _mm_setzero_si128 (), |
| 3039 | (__mmask8) __U); |
| 3040 | } |
| 3041 | |
| 3042 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3043 | _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U, |
| 3044 | void const *__P) { |
| 3045 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 3046 | (__v8si) __W, |
| 3047 | (__mmask8) |
| 3048 | __U); |
| 3049 | } |
| 3050 | |
| 3051 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3052 | _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 3053 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 3054 | (__v8si) |
| 3055 | _mm256_setzero_si256 (), |
| 3056 | (__mmask8) |
| 3057 | __U); |
| 3058 | } |
| 3059 | |
| 3060 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3061 | _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 3062 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 3063 | (__v4sf) __W, |
| 3064 | (__mmask8) __U); |
| 3065 | } |
| 3066 | |
| 3067 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3068 | _mm_maskz_expand_ps (__mmask8 __U, __m128 __A) { |
| 3069 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 3070 | (__v4sf) |
| 3071 | _mm_setzero_ps (), |
| 3072 | (__mmask8) __U); |
| 3073 | } |
| 3074 | |
| 3075 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3076 | _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 3077 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 3078 | (__v8sf) __W, |
| 3079 | (__mmask8) __U); |
| 3080 | } |
| 3081 | |
| 3082 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3083 | _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) { |
| 3084 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 3085 | (__v8sf) |
| 3086 | _mm256_setzero_ps (), |
| 3087 | (__mmask8) __U); |
| 3088 | } |
| 3089 | |
| 3090 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3091 | _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3092 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 3093 | (__v4si) __W, |
| 3094 | (__mmask8) __U); |
| 3095 | } |
| 3096 | |
| 3097 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3098 | _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) { |
| 3099 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 3100 | (__v4si) |
| 3101 | _mm_setzero_si128 (), |
| 3102 | (__mmask8) __U); |
| 3103 | } |
| 3104 | |
| 3105 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3106 | _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3107 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 3108 | (__v8si) __W, |
| 3109 | (__mmask8) __U); |
| 3110 | } |
| 3111 | |
| 3112 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3113 | _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) { |
| 3114 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 3115 | (__v8si) |
| 3116 | _mm256_setzero_si256 (), |
| 3117 | (__mmask8) __U); |
| 3118 | } |
| 3119 | |
| 3120 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3121 | _mm_getexp_pd (__m128d __A) { |
| 3122 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3123 | (__v2df) |
| 3124 | _mm_setzero_pd (), |
| 3125 | (__mmask8) -1); |
| 3126 | } |
| 3127 | |
| 3128 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3129 | _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 3130 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3131 | (__v2df) __W, |
| 3132 | (__mmask8) __U); |
| 3133 | } |
| 3134 | |
| 3135 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3136 | _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) { |
| 3137 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 3138 | (__v2df) |
| 3139 | _mm_setzero_pd (), |
| 3140 | (__mmask8) __U); |
| 3141 | } |
| 3142 | |
| 3143 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3144 | _mm256_getexp_pd (__m256d __A) { |
| 3145 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3146 | (__v4df) |
| 3147 | _mm256_setzero_pd (), |
| 3148 | (__mmask8) -1); |
| 3149 | } |
| 3150 | |
| 3151 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3152 | _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 3153 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3154 | (__v4df) __W, |
| 3155 | (__mmask8) __U); |
| 3156 | } |
| 3157 | |
| 3158 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3159 | _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) { |
| 3160 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 3161 | (__v4df) |
| 3162 | _mm256_setzero_pd (), |
| 3163 | (__mmask8) __U); |
| 3164 | } |
| 3165 | |
| 3166 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3167 | _mm_getexp_ps (__m128 __A) { |
| 3168 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3169 | (__v4sf) |
| 3170 | _mm_setzero_ps (), |
| 3171 | (__mmask8) -1); |
| 3172 | } |
| 3173 | |
| 3174 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3175 | _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 3176 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3177 | (__v4sf) __W, |
| 3178 | (__mmask8) __U); |
| 3179 | } |
| 3180 | |
| 3181 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3182 | _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) { |
| 3183 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 3184 | (__v4sf) |
| 3185 | _mm_setzero_ps (), |
| 3186 | (__mmask8) __U); |
| 3187 | } |
| 3188 | |
| 3189 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3190 | _mm256_getexp_ps (__m256 __A) { |
| 3191 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3192 | (__v8sf) |
| 3193 | _mm256_setzero_ps (), |
| 3194 | (__mmask8) -1); |
| 3195 | } |
| 3196 | |
| 3197 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3198 | _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 3199 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3200 | (__v8sf) __W, |
| 3201 | (__mmask8) __U); |
| 3202 | } |
| 3203 | |
| 3204 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3205 | _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) { |
| 3206 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 3207 | (__v8sf) |
| 3208 | _mm256_setzero_ps (), |
| 3209 | (__mmask8) __U); |
| 3210 | } |
| 3211 | |
| 3212 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3213 | _mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3214 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3215 | (__v2df) __B, |
| 3216 | (__v2df) __W, |
| 3217 | (__mmask8) __U); |
| 3218 | } |
| 3219 | |
| 3220 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3221 | _mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3222 | return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, |
| 3223 | (__v2df) __B, |
| 3224 | (__v2df) |
| 3225 | _mm_setzero_pd (), |
| 3226 | (__mmask8) __U); |
| 3227 | } |
| 3228 | |
| 3229 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3230 | _mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3231 | __m256d __B) { |
| 3232 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3233 | (__v4df) __B, |
| 3234 | (__v4df) __W, |
| 3235 | (__mmask8) __U); |
| 3236 | } |
| 3237 | |
| 3238 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3239 | _mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3240 | return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, |
| 3241 | (__v4df) __B, |
| 3242 | (__v4df) |
| 3243 | _mm256_setzero_pd (), |
| 3244 | (__mmask8) __U); |
| 3245 | } |
| 3246 | |
| 3247 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3248 | _mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3249 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3250 | (__v4sf) __B, |
| 3251 | (__v4sf) __W, |
| 3252 | (__mmask8) __U); |
| 3253 | } |
| 3254 | |
| 3255 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3256 | _mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3257 | return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, |
| 3258 | (__v4sf) __B, |
| 3259 | (__v4sf) |
| 3260 | _mm_setzero_ps (), |
| 3261 | (__mmask8) __U); |
| 3262 | } |
| 3263 | |
| 3264 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3265 | _mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3266 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3267 | (__v8sf) __B, |
| 3268 | (__v8sf) __W, |
| 3269 | (__mmask8) __U); |
| 3270 | } |
| 3271 | |
| 3272 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3273 | _mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3274 | return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, |
| 3275 | (__v8sf) __B, |
| 3276 | (__v8sf) |
| 3277 | _mm256_setzero_ps (), |
| 3278 | (__mmask8) __U); |
| 3279 | } |
| 3280 | |
| 3281 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3282 | _mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3283 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3284 | (__v2df) __B, |
| 3285 | (__v2df) __W, |
| 3286 | (__mmask8) __U); |
| 3287 | } |
| 3288 | |
| 3289 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3290 | _mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3291 | return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, |
| 3292 | (__v2df) __B, |
| 3293 | (__v2df) |
| 3294 | _mm_setzero_pd (), |
| 3295 | (__mmask8) __U); |
| 3296 | } |
| 3297 | |
| 3298 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3299 | _mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3300 | __m256d __B) { |
| 3301 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3302 | (__v4df) __B, |
| 3303 | (__v4df) __W, |
| 3304 | (__mmask8) __U); |
| 3305 | } |
| 3306 | |
| 3307 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3308 | _mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3309 | return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, |
| 3310 | (__v4df) __B, |
| 3311 | (__v4df) |
| 3312 | _mm256_setzero_pd (), |
| 3313 | (__mmask8) __U); |
| 3314 | } |
| 3315 | |
| 3316 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3317 | _mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3318 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3319 | (__v4sf) __B, |
| 3320 | (__v4sf) __W, |
| 3321 | (__mmask8) __U); |
| 3322 | } |
| 3323 | |
| 3324 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3325 | _mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3326 | return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, |
| 3327 | (__v4sf) __B, |
| 3328 | (__v4sf) |
| 3329 | _mm_setzero_ps (), |
| 3330 | (__mmask8) __U); |
| 3331 | } |
| 3332 | |
| 3333 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3334 | _mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3335 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3336 | (__v8sf) __B, |
| 3337 | (__v8sf) __W, |
| 3338 | (__mmask8) __U); |
| 3339 | } |
| 3340 | |
| 3341 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3342 | _mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3343 | return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, |
| 3344 | (__v8sf) __B, |
| 3345 | (__v8sf) |
| 3346 | _mm256_setzero_ps (), |
| 3347 | (__mmask8) __U); |
| 3348 | } |
| 3349 | |
| 3350 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3351 | _mm_mask_mul_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3352 | return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, |
| 3353 | (__v2df) __B, |
| 3354 | (__v2df) __W, |
| 3355 | (__mmask8) __U); |
| 3356 | } |
| 3357 | |
| 3358 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3359 | _mm_maskz_mul_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3360 | return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, |
| 3361 | (__v2df) __B, |
| 3362 | (__v2df) |
| 3363 | _mm_setzero_pd (), |
| 3364 | (__mmask8) __U); |
| 3365 | } |
| 3366 | |
| 3367 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3368 | _mm256_mask_mul_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3369 | __m256d __B) { |
| 3370 | return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, |
| 3371 | (__v4df) __B, |
| 3372 | (__v4df) __W, |
| 3373 | (__mmask8) __U); |
| 3374 | } |
| 3375 | |
| 3376 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3377 | _mm256_maskz_mul_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3378 | return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, |
| 3379 | (__v4df) __B, |
| 3380 | (__v4df) |
| 3381 | _mm256_setzero_pd (), |
| 3382 | (__mmask8) __U); |
| 3383 | } |
| 3384 | |
| 3385 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3386 | _mm_mask_mul_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3387 | return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, |
| 3388 | (__v4sf) __B, |
| 3389 | (__v4sf) __W, |
| 3390 | (__mmask8) __U); |
| 3391 | } |
| 3392 | |
| 3393 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3394 | _mm_maskz_mul_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3395 | return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, |
| 3396 | (__v4sf) __B, |
| 3397 | (__v4sf) |
| 3398 | _mm_setzero_ps (), |
| 3399 | (__mmask8) __U); |
| 3400 | } |
| 3401 | |
| 3402 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3403 | _mm256_mask_mul_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3404 | return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, |
| 3405 | (__v8sf) __B, |
| 3406 | (__v8sf) __W, |
| 3407 | (__mmask8) __U); |
| 3408 | } |
| 3409 | |
| 3410 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3411 | _mm256_maskz_mul_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3412 | return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, |
| 3413 | (__v8sf) __B, |
| 3414 | (__v8sf) |
| 3415 | _mm256_setzero_ps (), |
| 3416 | (__mmask8) __U); |
| 3417 | } |
| 3418 | |
| 3419 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3420 | _mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3421 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3422 | (__v4si) __W, |
| 3423 | (__mmask8) __U); |
| 3424 | } |
| 3425 | |
| 3426 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3427 | _mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A) { |
| 3428 | return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, |
| 3429 | (__v4si) |
| 3430 | _mm_setzero_si128 (), |
| 3431 | (__mmask8) __U); |
| 3432 | } |
| 3433 | |
| 3434 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3435 | _mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3436 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3437 | (__v8si) __W, |
| 3438 | (__mmask8) __U); |
| 3439 | } |
| 3440 | |
| 3441 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3442 | _mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A) { |
| 3443 | return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, |
| 3444 | (__v8si) |
| 3445 | _mm256_setzero_si256 (), |
| 3446 | (__mmask8) __U); |
| 3447 | } |
| 3448 | |
| 3449 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3450 | _mm_abs_epi64 (__m128i __A) { |
| 3451 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3452 | (__v2di) |
| 3453 | _mm_setzero_si128 (), |
| 3454 | (__mmask8) -1); |
| 3455 | } |
| 3456 | |
| 3457 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3458 | _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 3459 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3460 | (__v2di) __W, |
| 3461 | (__mmask8) __U); |
| 3462 | } |
| 3463 | |
| 3464 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3465 | _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) { |
| 3466 | return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, |
| 3467 | (__v2di) |
| 3468 | _mm_setzero_si128 (), |
| 3469 | (__mmask8) __U); |
| 3470 | } |
| 3471 | |
| 3472 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3473 | _mm256_abs_epi64 (__m256i __A) { |
| 3474 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3475 | (__v4di) |
| 3476 | _mm256_setzero_si256 (), |
| 3477 | (__mmask8) -1); |
| 3478 | } |
| 3479 | |
| 3480 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3481 | _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 3482 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3483 | (__v4di) __W, |
| 3484 | (__mmask8) __U); |
| 3485 | } |
| 3486 | |
| 3487 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3488 | _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) { |
| 3489 | return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, |
| 3490 | (__v4di) |
| 3491 | _mm256_setzero_si256 (), |
| 3492 | (__mmask8) __U); |
| 3493 | } |
| 3494 | |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3495 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3496 | _mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3497 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3498 | (__v4si) __B, |
| 3499 | (__v4si) |
| 3500 | _mm_setzero_si128 (), |
| 3501 | __M); |
| 3502 | } |
| 3503 | |
| 3504 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3505 | _mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3506 | __m128i __B) { |
| 3507 | return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, |
| 3508 | (__v4si) __B, |
| 3509 | (__v4si) __W, __M); |
| 3510 | } |
| 3511 | |
| 3512 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3513 | _mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3514 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3515 | (__v8si) __B, |
| 3516 | (__v8si) |
| 3517 | _mm256_setzero_si256 (), |
| 3518 | __M); |
| 3519 | } |
| 3520 | |
| 3521 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3522 | _mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3523 | __m256i __B) { |
| 3524 | return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, |
| 3525 | (__v8si) __B, |
| 3526 | (__v8si) __W, __M); |
| 3527 | } |
| 3528 | |
| 3529 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3530 | _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3531 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3532 | (__v2di) __B, |
| 3533 | (__v2di) |
| 3534 | _mm_setzero_si128 (), |
| 3535 | __M); |
| 3536 | } |
| 3537 | |
| 3538 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3539 | _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3540 | __m128i __B) { |
| 3541 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3542 | (__v2di) __B, |
| 3543 | (__v2di) __W, __M); |
| 3544 | } |
| 3545 | |
| 3546 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3547 | _mm_max_epi64 (__m128i __A, __m128i __B) { |
| 3548 | return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, |
| 3549 | (__v2di) __B, |
| 3550 | (__v2di) |
| 3551 | _mm_setzero_si128 (), |
| 3552 | (__mmask8) -1); |
| 3553 | } |
| 3554 | |
| 3555 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3556 | _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3557 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3558 | (__v4di) __B, |
| 3559 | (__v4di) |
| 3560 | _mm256_setzero_si256 (), |
| 3561 | __M); |
| 3562 | } |
| 3563 | |
| 3564 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3565 | _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3566 | __m256i __B) { |
| 3567 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3568 | (__v4di) __B, |
| 3569 | (__v4di) __W, __M); |
| 3570 | } |
| 3571 | |
| 3572 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3573 | _mm256_max_epi64 (__m256i __A, __m256i __B) { |
| 3574 | return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, |
| 3575 | (__v4di) __B, |
| 3576 | (__v4di) |
| 3577 | _mm256_setzero_si256 (), |
| 3578 | (__mmask8) -1); |
| 3579 | } |
| 3580 | |
| 3581 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3582 | _mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3583 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3584 | (__v4si) __B, |
| 3585 | (__v4si) |
| 3586 | _mm_setzero_si128 (), |
| 3587 | __M); |
| 3588 | } |
| 3589 | |
| 3590 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3591 | _mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3592 | __m128i __B) { |
| 3593 | return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, |
| 3594 | (__v4si) __B, |
| 3595 | (__v4si) __W, __M); |
| 3596 | } |
| 3597 | |
| 3598 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3599 | _mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3600 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3601 | (__v8si) __B, |
| 3602 | (__v8si) |
| 3603 | _mm256_setzero_si256 (), |
| 3604 | __M); |
| 3605 | } |
| 3606 | |
| 3607 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3608 | _mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3609 | __m256i __B) { |
| 3610 | return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, |
| 3611 | (__v8si) __B, |
| 3612 | (__v8si) __W, __M); |
| 3613 | } |
| 3614 | |
| 3615 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3616 | _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3617 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3618 | (__v2di) __B, |
| 3619 | (__v2di) |
| 3620 | _mm_setzero_si128 (), |
| 3621 | __M); |
| 3622 | } |
| 3623 | |
| 3624 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3625 | _mm_max_epu64 (__m128i __A, __m128i __B) { |
| 3626 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3627 | (__v2di) __B, |
| 3628 | (__v2di) |
| 3629 | _mm_setzero_si128 (), |
| 3630 | (__mmask8) -1); |
| 3631 | } |
| 3632 | |
| 3633 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3634 | _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3635 | __m128i __B) { |
| 3636 | return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, |
| 3637 | (__v2di) __B, |
| 3638 | (__v2di) __W, __M); |
| 3639 | } |
| 3640 | |
| 3641 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3642 | _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3643 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3644 | (__v4di) __B, |
| 3645 | (__v4di) |
| 3646 | _mm256_setzero_si256 (), |
| 3647 | __M); |
| 3648 | } |
| 3649 | |
| 3650 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3651 | _mm256_max_epu64 (__m256i __A, __m256i __B) { |
| 3652 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3653 | (__v4di) __B, |
| 3654 | (__v4di) |
| 3655 | _mm256_setzero_si256 (), |
| 3656 | (__mmask8) -1); |
| 3657 | } |
| 3658 | |
| 3659 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3660 | _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3661 | __m256i __B) { |
| 3662 | return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, |
| 3663 | (__v4di) __B, |
| 3664 | (__v4di) __W, __M); |
| 3665 | } |
| 3666 | |
| 3667 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3668 | _mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3669 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3670 | (__v4si) __B, |
| 3671 | (__v4si) |
| 3672 | _mm_setzero_si128 (), |
| 3673 | __M); |
| 3674 | } |
| 3675 | |
| 3676 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3677 | _mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3678 | __m128i __B) { |
| 3679 | return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, |
| 3680 | (__v4si) __B, |
| 3681 | (__v4si) __W, __M); |
| 3682 | } |
| 3683 | |
| 3684 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3685 | _mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3686 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3687 | (__v8si) __B, |
| 3688 | (__v8si) |
| 3689 | _mm256_setzero_si256 (), |
| 3690 | __M); |
| 3691 | } |
| 3692 | |
| 3693 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3694 | _mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3695 | __m256i __B) { |
| 3696 | return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, |
| 3697 | (__v8si) __B, |
| 3698 | (__v8si) __W, __M); |
| 3699 | } |
| 3700 | |
| 3701 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3702 | _mm_min_epi64 (__m128i __A, __m128i __B) { |
| 3703 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3704 | (__v2di) __B, |
| 3705 | (__v2di) |
| 3706 | _mm_setzero_si128 (), |
| 3707 | (__mmask8) -1); |
| 3708 | } |
| 3709 | |
| 3710 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3711 | _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3712 | __m128i __B) { |
| 3713 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3714 | (__v2di) __B, |
| 3715 | (__v2di) __W, __M); |
| 3716 | } |
| 3717 | |
| 3718 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3719 | _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3720 | return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, |
| 3721 | (__v2di) __B, |
| 3722 | (__v2di) |
| 3723 | _mm_setzero_si128 (), |
| 3724 | __M); |
| 3725 | } |
| 3726 | |
| 3727 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3728 | _mm256_min_epi64 (__m256i __A, __m256i __B) { |
| 3729 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3730 | (__v4di) __B, |
| 3731 | (__v4di) |
| 3732 | _mm256_setzero_si256 (), |
| 3733 | (__mmask8) -1); |
| 3734 | } |
| 3735 | |
| 3736 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3737 | _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3738 | __m256i __B) { |
| 3739 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3740 | (__v4di) __B, |
| 3741 | (__v4di) __W, __M); |
| 3742 | } |
| 3743 | |
| 3744 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3745 | _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3746 | return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, |
| 3747 | (__v4di) __B, |
| 3748 | (__v4di) |
| 3749 | _mm256_setzero_si256 (), |
| 3750 | __M); |
| 3751 | } |
| 3752 | |
| 3753 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3754 | _mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3755 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3756 | (__v4si) __B, |
| 3757 | (__v4si) |
| 3758 | _mm_setzero_si128 (), |
| 3759 | __M); |
| 3760 | } |
| 3761 | |
| 3762 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3763 | _mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3764 | __m128i __B) { |
| 3765 | return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, |
| 3766 | (__v4si) __B, |
| 3767 | (__v4si) __W, __M); |
| 3768 | } |
| 3769 | |
| 3770 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3771 | _mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3772 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3773 | (__v8si) __B, |
| 3774 | (__v8si) |
| 3775 | _mm256_setzero_si256 (), |
| 3776 | __M); |
| 3777 | } |
| 3778 | |
| 3779 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3780 | _mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3781 | __m256i __B) { |
| 3782 | return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, |
| 3783 | (__v8si) __B, |
| 3784 | (__v8si) __W, __M); |
| 3785 | } |
| 3786 | |
| 3787 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3788 | _mm_min_epu64 (__m128i __A, __m128i __B) { |
| 3789 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3790 | (__v2di) __B, |
| 3791 | (__v2di) |
| 3792 | _mm_setzero_si128 (), |
| 3793 | (__mmask8) -1); |
| 3794 | } |
| 3795 | |
| 3796 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3797 | _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, |
| 3798 | __m128i __B) { |
| 3799 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3800 | (__v2di) __B, |
| 3801 | (__v2di) __W, __M); |
| 3802 | } |
| 3803 | |
| 3804 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3805 | _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3806 | return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, |
| 3807 | (__v2di) __B, |
| 3808 | (__v2di) |
| 3809 | _mm_setzero_si128 (), |
| 3810 | __M); |
| 3811 | } |
| 3812 | |
| 3813 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3814 | _mm256_min_epu64 (__m256i __A, __m256i __B) { |
| 3815 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3816 | (__v4di) __B, |
| 3817 | (__v4di) |
| 3818 | _mm256_setzero_si256 (), |
| 3819 | (__mmask8) -1); |
| 3820 | } |
| 3821 | |
| 3822 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3823 | _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, |
| 3824 | __m256i __B) { |
| 3825 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3826 | (__v4di) __B, |
| 3827 | (__v4di) __W, __M); |
| 3828 | } |
| 3829 | |
| 3830 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3831 | _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3832 | return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, |
| 3833 | (__v4di) __B, |
| 3834 | (__v4di) |
| 3835 | _mm256_setzero_si256 (), |
| 3836 | __M); |
| 3837 | } |
| 3838 | |
| 3839 | #define _mm_roundscale_pd(__A, __imm) __extension__ ({ \ |
| 3840 | (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, \ |
| 3841 | __imm, (__v2df) _mm_setzero_pd (), (__mmask8) -1); }) |
| 3842 | |
| 3843 | |
| 3844 | #define _mm_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \ |
| 3845 | (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \ |
| 3846 | (__v2df) __W, (__mmask8) __U); }) |
| 3847 | |
| 3848 | |
| 3849 | #define _mm_maskz_roundscale_pd(__U, __A, __imm) __extension__ ({ \ |
| 3850 | (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, __imm, \ |
| 3851 | (__v2df) _mm_setzero_pd (), (__mmask8) __U); }) |
| 3852 | |
| 3853 | |
| 3854 | #define _mm256_roundscale_pd(__A, __imm) __extension__ ({ \ |
| 3855 | (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \ |
| 3856 | (__v4df) _mm256_setzero_pd (), (__mmask8) -1); }) |
| 3857 | |
| 3858 | |
| 3859 | #define _mm256_mask_roundscale_pd(__W, __U, __A, __imm) __extension__ ({ \ |
| 3860 | (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \ |
| 3861 | (__v4df) __W, (__mmask8) __U); }) |
| 3862 | |
| 3863 | |
| 3864 | #define _mm256_maskz_roundscale_pd(__U, __A, __imm) __extension__ ({ \ |
| 3865 | (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, __imm, \ |
| 3866 | (__v4df) _mm256_setzero_pd(), (__mmask8) __U); }) |
| 3867 | |
| 3868 | #define _mm_roundscale_ps(__A, __imm) __extension__ ({ \ |
| 3869 | (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \ |
| 3870 | (__v4sf) _mm_setzero_ps(), (__mmask8) -1); }) |
| 3871 | |
| 3872 | |
| 3873 | #define _mm_mask_roundscale_ps(__W, __U, __A, __imm) __extension__ ({ \ |
| 3874 | (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \ |
| 3875 | (__v4sf) __W, (__mmask8) __U); }) |
| 3876 | |
| 3877 | |
| 3878 | #define _mm_maskz_roundscale_ps(__U, __A, __imm) __extension__ ({ \ |
| 3879 | (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, __imm, \ |
| 3880 | (__v4sf) _mm_setzero_ps(), (__mmask8) __U); }) |
| 3881 | |
| 3882 | #define _mm256_roundscale_ps(__A, __imm) __extension__ ({ \ |
| 3883 | (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,__imm, \ |
| 3884 | (__v8sf) _mm256_setzero_ps(), (__mmask8) -1); }) |
| 3885 | |
| 3886 | #define _mm256_mask_roundscale_ps(__W, __U, __A,__imm) __extension__ ({ \ |
| 3887 | (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \ |
| 3888 | (__v8sf) __W, (__mmask8) __U); }) |
| 3889 | |
| 3890 | |
| 3891 | #define _mm256_maskz_roundscale_ps(__U, __A, __imm) __extension__ ({ \ |
| 3892 | (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, __imm, \ |
| 3893 | (__v8sf) _mm256_setzero_ps(), (__mmask8) __U); }) |
| 3894 | |
| 3895 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3896 | _mm_scalef_pd (__m128d __A, __m128d __B) { |
| 3897 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3898 | (__v2df) __B, |
| 3899 | (__v2df) |
| 3900 | _mm_setzero_pd (), |
| 3901 | (__mmask8) -1); |
| 3902 | } |
| 3903 | |
| 3904 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3905 | _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 3906 | __m128d __B) { |
| 3907 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3908 | (__v2df) __B, |
| 3909 | (__v2df) __W, |
| 3910 | (__mmask8) __U); |
| 3911 | } |
| 3912 | |
| 3913 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3914 | _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3915 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3916 | (__v2df) __B, |
| 3917 | (__v2df) |
| 3918 | _mm_setzero_pd (), |
| 3919 | (__mmask8) __U); |
| 3920 | } |
| 3921 | |
| 3922 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3923 | _mm256_scalef_pd (__m256d __A, __m256d __B) { |
| 3924 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3925 | (__v4df) __B, |
| 3926 | (__v4df) |
| 3927 | _mm256_setzero_pd (), |
| 3928 | (__mmask8) -1); |
| 3929 | } |
| 3930 | |
| 3931 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3932 | _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3933 | __m256d __B) { |
| 3934 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3935 | (__v4df) __B, |
| 3936 | (__v4df) __W, |
| 3937 | (__mmask8) __U); |
| 3938 | } |
| 3939 | |
| 3940 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3941 | _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3942 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3943 | (__v4df) __B, |
| 3944 | (__v4df) |
| 3945 | _mm256_setzero_pd (), |
| 3946 | (__mmask8) __U); |
| 3947 | } |
| 3948 | |
| 3949 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3950 | _mm_scalef_ps (__m128 __A, __m128 __B) { |
| 3951 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3952 | (__v4sf) __B, |
| 3953 | (__v4sf) |
| 3954 | _mm_setzero_ps (), |
| 3955 | (__mmask8) -1); |
| 3956 | } |
| 3957 | |
| 3958 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3959 | _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3960 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3961 | (__v4sf) __B, |
| 3962 | (__v4sf) __W, |
| 3963 | (__mmask8) __U); |
| 3964 | } |
| 3965 | |
| 3966 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3967 | _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3968 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3969 | (__v4sf) __B, |
| 3970 | (__v4sf) |
| 3971 | _mm_setzero_ps (), |
| 3972 | (__mmask8) __U); |
| 3973 | } |
| 3974 | |
| 3975 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3976 | _mm256_scalef_ps (__m256 __A, __m256 __B) { |
| 3977 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3978 | (__v8sf) __B, |
| 3979 | (__v8sf) |
| 3980 | _mm256_setzero_ps (), |
| 3981 | (__mmask8) -1); |
| 3982 | } |
| 3983 | |
| 3984 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3985 | _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 3986 | __m256 __B) { |
| 3987 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3988 | (__v8sf) __B, |
| 3989 | (__v8sf) __W, |
| 3990 | (__mmask8) __U); |
| 3991 | } |
| 3992 | |
| 3993 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3994 | _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3995 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3996 | (__v8sf) __B, |
| 3997 | (__v8sf) |
| 3998 | _mm256_setzero_ps (), |
| 3999 | (__mmask8) __U); |
| 4000 | } |
| 4001 | |
| 4002 | #define _mm_i64scatter_pd(__addr,__index, __v1, __scale) __extension__ ({ \ |
| 4003 | __builtin_ia32_scatterdiv2df(__addr, (__mmask8) 0xFF, (__v2di) __index, \ |
| 4004 | (__v2df) __v1, __scale); }) |
| 4005 | |
| 4006 | #define _mm_mask_i64scatter_pd(__addr, __mask, __index, __v1, \ |
| 4007 | __scale) __extension__ ({ \ |
| 4008 | __builtin_ia32_scatterdiv2df (__addr, __mask, (__v2di) __index, \ |
| 4009 | (__v2df) __v1, __scale); }) |
| 4010 | |
| 4011 | |
| 4012 | #define _mm_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4013 | __builtin_ia32_scatterdiv2di (__addr, (__mmask8) 0xFF, \ |
| 4014 | (__v2di) __index, (__v2di) __v1, __scale); }) |
| 4015 | |
| 4016 | #define _mm_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\ |
| 4017 | __scale) __extension__ ({ \ |
| 4018 | __builtin_ia32_scatterdiv2di (__addr, __mask, (__v2di) __index,\ |
| 4019 | (__v2di) __v1, __scale); }) |
| 4020 | |
| 4021 | #define _mm256_i64scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4022 | __builtin_ia32_scatterdiv4df (__addr, (__mmask8) 0xFF,\ |
| 4023 | (__v4di) __index, (__v4df) __v1, __scale); }) |
| 4024 | |
| 4025 | #define _mm256_mask_i64scatter_pd(__addr, __mask, __index, __v1,\ |
| 4026 | __scale) __extension__ ({ \ |
| 4027 | __builtin_ia32_scatterdiv4df (__addr, __mask, (__v4di) __index,\ |
| 4028 | (__v4df) __v1, __scale); }) |
| 4029 | |
| 4030 | #define _mm256_i64scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4031 | __builtin_ia32_scatterdiv4di (__addr, (__mmask8) 0xFF, (__v4di) __index,\ |
| 4032 | (__v4di) __v1, __scale); }) |
| 4033 | |
| 4034 | #define _mm256_mask_i64scatter_epi64(__addr, __mask, __index, __v1,\ |
| 4035 | __scale) __extension__ ({ \ |
| 4036 | __builtin_ia32_scatterdiv4di (__addr, __mask, (__v4di) __index,\ |
| 4037 | (__v4di) __v1, __scale); }) |
| 4038 | |
| 4039 | #define _mm_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4040 | __builtin_ia32_scatterdiv4sf (__addr, (__mmask8) 0xFF,\ |
| 4041 | (__v2di) __index, (__v4sf) __v1, __scale); }) |
| 4042 | |
| 4043 | #define _mm_mask_i64scatter_ps(__addr, __mask, __index, __v1, \ |
| 4044 | __scale) __extension__ ({ \ |
| 4045 | __builtin_ia32_scatterdiv4sf (__addr, __mask, (__v2di) __index,\ |
| 4046 | (__v4sf) __v1, __scale); }) |
| 4047 | |
| 4048 | #define _mm_i64scatter_epi32(__addr, __index, __v1, \ |
| 4049 | __scale) __extension__ ({ \ |
| 4050 | __builtin_ia32_scatterdiv4si (__addr, (__mmask8) 0xFF,\ |
| 4051 | (__v2di) __index, (__v4si) __v1, __scale); }) |
| 4052 | |
| 4053 | #define _mm_mask_i64scatter_epi32(__addr, __mask, __index, __v1,\ |
| 4054 | __scale) __extension__ ({ \ |
| 4055 | __builtin_ia32_scatterdiv4si (__addr, __mask, (__v2di) __index,\ |
| 4056 | (__v4si) __v1, __scale); }) |
| 4057 | |
| 4058 | #define _mm256_i64scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4059 | __builtin_ia32_scatterdiv8sf (__addr, (__mmask8) 0xFF, (__v4di) __index, \ |
| 4060 | (__v4sf) __v1, __scale); }) |
| 4061 | |
| 4062 | #define _mm256_mask_i64scatter_ps(__addr, __mask, __index, __v1, \ |
| 4063 | __scale) __extension__ ({ \ |
| 4064 | __builtin_ia32_scatterdiv8sf (__addr, __mask, (__v4di) __index, \ |
| 4065 | (__v4sf) __v1, __scale); }) |
| 4066 | |
| 4067 | #define _mm256_i64scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4068 | __builtin_ia32_scatterdiv8si (__addr, (__mmask8) 0xFF, \ |
| 4069 | (__v4di) __index, (__v4si) __v1, __scale); }) |
| 4070 | |
| 4071 | #define _mm256_mask_i64scatter_epi32(__addr, __mask, __index, __v1, \ |
| 4072 | __scale) __extension__ ({ \ |
| 4073 | __builtin_ia32_scatterdiv8si(__addr, __mask, (__v4di) __index, \ |
| 4074 | (__v4si) __v1, __scale); }) |
| 4075 | |
| 4076 | #define _mm_i32scatter_pd(__addr, __index, __v1, \ |
| 4077 | __scale) __extension__ ({ \ |
| 4078 | __builtin_ia32_scattersiv2df (__addr, (__mmask8) 0xFF, \ |
| 4079 | (__v4si) __index, (__v2df) __v1, __scale); }) |
| 4080 | |
| 4081 | #define _mm_mask_i32scatter_pd(__addr, __mask, __index, __v1, \ |
| 4082 | __scale) __extension__ ({ \ |
| 4083 | __builtin_ia32_scattersiv2df (__addr, __mask, (__v4si) __index,\ |
| 4084 | (__v2df) __v1, __scale); }) |
| 4085 | |
| 4086 | #define _mm_i32scatter_epi64(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4087 | __builtin_ia32_scattersiv2di (__addr, (__mmask8) 0xFF, \ |
| 4088 | (__v4si) __index, (__v2di) __v1, __scale); }) |
| 4089 | |
| 4090 | #define _mm_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \ |
| 4091 | __scale) __extension__ ({ \ |
| 4092 | __builtin_ia32_scattersiv2di (__addr, __mask, (__v4si) __index, \ |
| 4093 | (__v2di) __v1, __scale); }) |
| 4094 | |
| 4095 | #define _mm256_i32scatter_pd(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4096 | __builtin_ia32_scattersiv4df (__addr, (__mmask8) 0xFF, \ |
| 4097 | (__v4si) __index, (__v4df) __v1, __scale); }) |
| 4098 | |
| 4099 | #define _mm256_mask_i32scatter_pd(__addr, __mask, __index, __v1, \ |
| 4100 | __scale) __extension__ ({ \ |
| 4101 | __builtin_ia32_scattersiv4df (__addr, __mask, (__v4si) __index, \ |
| 4102 | (__v4df) __v1, __scale); }) |
| 4103 | |
| 4104 | #define _mm256_i32scatter_epi64(__addr, __index, __v1, \ |
| 4105 | __scale) __extension__ ({ \ |
| 4106 | __builtin_ia32_scattersiv4di (__addr, (__mmask8) 0xFF, \ |
| 4107 | (__v4si) __index, (__v4di) __v1, __scale); }) |
| 4108 | |
| 4109 | #define _mm256_mask_i32scatter_epi64(__addr, __mask, __index, __v1, \ |
| 4110 | __scale) __extension__ ({ \ |
| 4111 | __builtin_ia32_scattersiv4di (__addr, __mask, (__v4si) __index, \ |
| 4112 | (__v4di) __v1, __scale); }) |
| 4113 | |
| 4114 | #define _mm_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4115 | __builtin_ia32_scattersiv4sf (__addr, (__mmask8) 0xFF, \ |
| 4116 | (__v4si) __index, (__v4sf) __v1, __scale); }) |
| 4117 | |
| 4118 | #define _mm_mask_i32scatter_ps(__addr, __mask, __index, __v1, \ |
| 4119 | __scale) __extension__ ({ \ |
| 4120 | __builtin_ia32_scattersiv4sf (__addr, __mask, (__v4si) __index, \ |
| 4121 | (__v4sf) __v1, __scale); }) |
| 4122 | |
| 4123 | #define _mm_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4124 | __builtin_ia32_scattersiv4si (__addr, (__mmask8) 0xFF, \ |
| 4125 | (__v4si) __index, (__v4si) __v1, __scale); }) |
| 4126 | |
| 4127 | #define _mm_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \ |
| 4128 | __scale) __extension__ ({ \ |
| 4129 | __builtin_ia32_scattersiv4si (__addr, __mask, (__v4si) __index,\ |
| 4130 | (__v4si) __v1, __scale); }) |
| 4131 | |
| 4132 | #define _mm256_i32scatter_ps(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4133 | __builtin_ia32_scattersiv8sf (__addr, (__mmask8) 0xFF, \ |
| 4134 | (__v8si) __index, (__v8sf) __v1, __scale); }) |
| 4135 | |
| 4136 | #define _mm256_mask_i32scatter_ps(__addr, __mask, __index, __v1, \ |
| 4137 | __scale) __extension__ ({ \ |
| 4138 | __builtin_ia32_scattersiv8sf (__addr, __mask, (__v8si) __index,\ |
| 4139 | (__v8sf) __v1, __scale); }) |
| 4140 | |
| 4141 | #define _mm256_i32scatter_epi32(__addr, __index, __v1, __scale) __extension__ ({ \ |
| 4142 | __builtin_ia32_scattersiv8si (__addr, (__mmask8) 0xFF, \ |
| 4143 | (__v8si) __index, (__v8si) __v1, __scale); }) |
| 4144 | |
| 4145 | #define _mm256_mask_i32scatter_epi32(__addr, __mask, __index, __v1, \ |
| 4146 | __scale) __extension__ ({ \ |
| 4147 | __builtin_ia32_scattersiv8si (__addr, __mask, (__v8si) __index, \ |
| 4148 | (__v8si) __v1, __scale); }) |
| 4149 | |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame^] | 4150 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4151 | _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 4152 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4153 | (__v2df) __W, |
| 4154 | (__mmask8) __U); |
| 4155 | } |
| 4156 | |
| 4157 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4158 | _mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A) { |
| 4159 | return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, |
| 4160 | (__v2df) |
| 4161 | _mm_setzero_pd (), |
| 4162 | (__mmask8) __U); |
| 4163 | } |
| 4164 | |
| 4165 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4166 | _mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 4167 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4168 | (__v4df) __W, |
| 4169 | (__mmask8) __U); |
| 4170 | } |
| 4171 | |
| 4172 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4173 | _mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A) { |
| 4174 | return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, |
| 4175 | (__v4df) |
| 4176 | _mm256_setzero_pd (), |
| 4177 | (__mmask8) __U); |
| 4178 | } |
| 4179 | |
| 4180 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4181 | _mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 4182 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4183 | (__v4sf) __W, |
| 4184 | (__mmask8) __U); |
| 4185 | } |
| 4186 | |
| 4187 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4188 | _mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A) { |
| 4189 | return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, |
| 4190 | (__v4sf) |
| 4191 | _mm_setzero_ps (), |
| 4192 | (__mmask8) __U); |
| 4193 | } |
| 4194 | |
| 4195 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4196 | _mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 4197 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4198 | (__v8sf) __W, |
| 4199 | (__mmask8) __U); |
| 4200 | } |
| 4201 | |
| 4202 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4203 | _mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A) { |
| 4204 | return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, |
| 4205 | (__v8sf) |
| 4206 | _mm256_setzero_ps (), |
| 4207 | (__mmask8) __U); |
| 4208 | } |
| 4209 | |
| 4210 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4211 | _mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 4212 | return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, |
| 4213 | (__v2df) __B, |
| 4214 | (__v2df) __W, |
| 4215 | (__mmask8) __U); |
| 4216 | } |
| 4217 | |
| 4218 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4219 | _mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 4220 | return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, |
| 4221 | (__v2df) __B, |
| 4222 | (__v2df) |
| 4223 | _mm_setzero_pd (), |
| 4224 | (__mmask8) __U); |
| 4225 | } |
| 4226 | |
| 4227 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4228 | _mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 4229 | __m256d __B) { |
| 4230 | return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, |
| 4231 | (__v4df) __B, |
| 4232 | (__v4df) __W, |
| 4233 | (__mmask8) __U); |
| 4234 | } |
| 4235 | |
| 4236 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4237 | _mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 4238 | return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, |
| 4239 | (__v4df) __B, |
| 4240 | (__v4df) |
| 4241 | _mm256_setzero_pd (), |
| 4242 | (__mmask8) __U); |
| 4243 | } |
| 4244 | |
| 4245 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4246 | _mm_mask_sub_ps (__m128 __W, __mmask16 __U, __m128 __A, __m128 __B) { |
| 4247 | return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, |
| 4248 | (__v4sf) __B, |
| 4249 | (__v4sf) __W, |
| 4250 | (__mmask8) __U); |
| 4251 | } |
| 4252 | |
| 4253 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4254 | _mm_maskz_sub_ps (__mmask16 __U, __m128 __A, __m128 __B) { |
| 4255 | return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, |
| 4256 | (__v4sf) __B, |
| 4257 | (__v4sf) |
| 4258 | _mm_setzero_ps (), |
| 4259 | (__mmask8) __U); |
| 4260 | } |
| 4261 | |
| 4262 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4263 | _mm256_mask_sub_ps (__m256 __W, __mmask16 __U, __m256 __A, __m256 __B) { |
| 4264 | return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, |
| 4265 | (__v8sf) __B, |
| 4266 | (__v8sf) __W, |
| 4267 | (__mmask8) __U); |
| 4268 | } |
| 4269 | |
| 4270 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4271 | _mm256_maskz_sub_ps (__mmask16 __U, __m256 __A, __m256 __B) { |
| 4272 | return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, |
| 4273 | (__v8sf) __B, |
| 4274 | (__v8sf) |
| 4275 | _mm256_setzero_ps (), |
| 4276 | (__mmask8) __U); |
| 4277 | } |
| 4278 | |
| 4279 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4280 | _mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4281 | __m128i __B) { |
| 4282 | return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A, |
| 4283 | (__v4si) __I |
| 4284 | /* idx */ , |
| 4285 | (__v4si) __B, |
| 4286 | (__mmask8) __U); |
| 4287 | } |
| 4288 | |
| 4289 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4290 | _mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I, |
| 4291 | __mmask8 __U, __m256i __B) { |
| 4292 | return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A, |
| 4293 | (__v8si) __I |
| 4294 | /* idx */ , |
| 4295 | (__v8si) __B, |
| 4296 | (__mmask8) __U); |
| 4297 | } |
| 4298 | |
| 4299 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4300 | _mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U, |
| 4301 | __m128d __B) { |
| 4302 | return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A, |
| 4303 | (__v2di) __I |
| 4304 | /* idx */ , |
| 4305 | (__v2df) __B, |
| 4306 | (__mmask8) |
| 4307 | __U); |
| 4308 | } |
| 4309 | |
| 4310 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4311 | _mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U, |
| 4312 | __m256d __B) { |
| 4313 | return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A, |
| 4314 | (__v4di) __I |
| 4315 | /* idx */ , |
| 4316 | (__v4df) __B, |
| 4317 | (__mmask8) |
| 4318 | __U); |
| 4319 | } |
| 4320 | |
| 4321 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4322 | _mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U, |
| 4323 | __m128 __B) { |
| 4324 | return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A, |
| 4325 | (__v4si) __I |
| 4326 | /* idx */ , |
| 4327 | (__v4sf) __B, |
| 4328 | (__mmask8) __U); |
| 4329 | } |
| 4330 | |
| 4331 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4332 | _mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U, |
| 4333 | __m256 __B) { |
| 4334 | return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A, |
| 4335 | (__v8si) __I |
| 4336 | /* idx */ , |
| 4337 | (__v8sf) __B, |
| 4338 | (__mmask8) __U); |
| 4339 | } |
| 4340 | |
| 4341 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4342 | _mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U, |
| 4343 | __m128i __B) { |
| 4344 | return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A, |
| 4345 | (__v2di) __I |
| 4346 | /* idx */ , |
| 4347 | (__v2di) __B, |
| 4348 | (__mmask8) __U); |
| 4349 | } |
| 4350 | |
| 4351 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4352 | _mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I, |
| 4353 | __mmask8 __U, __m256i __B) { |
| 4354 | return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A, |
| 4355 | (__v4di) __I |
| 4356 | /* idx */ , |
| 4357 | (__v4di) __B, |
| 4358 | (__mmask8) __U); |
| 4359 | } |
| 4360 | |
| 4361 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4362 | _mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) { |
| 4363 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4364 | /* idx */ , |
| 4365 | (__v4si) __A, |
| 4366 | (__v4si) __B, |
| 4367 | (__mmask8) -1); |
| 4368 | } |
| 4369 | |
| 4370 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4371 | _mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4372 | __m128i __B) { |
| 4373 | return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I |
| 4374 | /* idx */ , |
| 4375 | (__v4si) __A, |
| 4376 | (__v4si) __B, |
| 4377 | (__mmask8) __U); |
| 4378 | } |
| 4379 | |
| 4380 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4381 | _mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4382 | __m128i __B) { |
| 4383 | return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I |
| 4384 | /* idx */ , |
| 4385 | (__v4si) __A, |
| 4386 | (__v4si) __B, |
| 4387 | (__mmask8) |
| 4388 | __U); |
| 4389 | } |
| 4390 | |
| 4391 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4392 | _mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) { |
| 4393 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4394 | /* idx */ , |
| 4395 | (__v8si) __A, |
| 4396 | (__v8si) __B, |
| 4397 | (__mmask8) -1); |
| 4398 | } |
| 4399 | |
| 4400 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4401 | _mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4402 | __m256i __B) { |
| 4403 | return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I |
| 4404 | /* idx */ , |
| 4405 | (__v8si) __A, |
| 4406 | (__v8si) __B, |
| 4407 | (__mmask8) __U); |
| 4408 | } |
| 4409 | |
| 4410 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4411 | _mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A, |
| 4412 | __m256i __I, __m256i __B) { |
| 4413 | return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I |
| 4414 | /* idx */ , |
| 4415 | (__v8si) __A, |
| 4416 | (__v8si) __B, |
| 4417 | (__mmask8) |
| 4418 | __U); |
| 4419 | } |
| 4420 | |
| 4421 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4422 | _mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) { |
| 4423 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4424 | /* idx */ , |
| 4425 | (__v2df) __A, |
| 4426 | (__v2df) __B, |
| 4427 | (__mmask8) - |
| 4428 | 1); |
| 4429 | } |
| 4430 | |
| 4431 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4432 | _mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I, |
| 4433 | __m128d __B) { |
| 4434 | return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I |
| 4435 | /* idx */ , |
| 4436 | (__v2df) __A, |
| 4437 | (__v2df) __B, |
| 4438 | (__mmask8) |
| 4439 | __U); |
| 4440 | } |
| 4441 | |
| 4442 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 4443 | _mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I, |
| 4444 | __m128d __B) { |
| 4445 | return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I |
| 4446 | /* idx */ , |
| 4447 | (__v2df) __A, |
| 4448 | (__v2df) __B, |
| 4449 | (__mmask8) |
| 4450 | __U); |
| 4451 | } |
| 4452 | |
| 4453 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4454 | _mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) { |
| 4455 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4456 | /* idx */ , |
| 4457 | (__v4df) __A, |
| 4458 | (__v4df) __B, |
| 4459 | (__mmask8) - |
| 4460 | 1); |
| 4461 | } |
| 4462 | |
| 4463 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4464 | _mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I, |
| 4465 | __m256d __B) { |
| 4466 | return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I |
| 4467 | /* idx */ , |
| 4468 | (__v4df) __A, |
| 4469 | (__v4df) __B, |
| 4470 | (__mmask8) |
| 4471 | __U); |
| 4472 | } |
| 4473 | |
| 4474 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 4475 | _mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I, |
| 4476 | __m256d __B) { |
| 4477 | return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I |
| 4478 | /* idx */ , |
| 4479 | (__v4df) __A, |
| 4480 | (__v4df) __B, |
| 4481 | (__mmask8) |
| 4482 | __U); |
| 4483 | } |
| 4484 | |
| 4485 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4486 | _mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) { |
| 4487 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4488 | /* idx */ , |
| 4489 | (__v4sf) __A, |
| 4490 | (__v4sf) __B, |
| 4491 | (__mmask8) -1); |
| 4492 | } |
| 4493 | |
| 4494 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4495 | _mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I, |
| 4496 | __m128 __B) { |
| 4497 | return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I |
| 4498 | /* idx */ , |
| 4499 | (__v4sf) __A, |
| 4500 | (__v4sf) __B, |
| 4501 | (__mmask8) __U); |
| 4502 | } |
| 4503 | |
| 4504 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 4505 | _mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I, |
| 4506 | __m128 __B) { |
| 4507 | return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I |
| 4508 | /* idx */ , |
| 4509 | (__v4sf) __A, |
| 4510 | (__v4sf) __B, |
| 4511 | (__mmask8) |
| 4512 | __U); |
| 4513 | } |
| 4514 | |
| 4515 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4516 | _mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) { |
| 4517 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4518 | /* idx */ , |
| 4519 | (__v8sf) __A, |
| 4520 | (__v8sf) __B, |
| 4521 | (__mmask8) -1); |
| 4522 | } |
| 4523 | |
| 4524 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4525 | _mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I, |
| 4526 | __m256 __B) { |
| 4527 | return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I |
| 4528 | /* idx */ , |
| 4529 | (__v8sf) __A, |
| 4530 | (__v8sf) __B, |
| 4531 | (__mmask8) __U); |
| 4532 | } |
| 4533 | |
| 4534 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 4535 | _mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I, |
| 4536 | __m256 __B) { |
| 4537 | return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I |
| 4538 | /* idx */ , |
| 4539 | (__v8sf) __A, |
| 4540 | (__v8sf) __B, |
| 4541 | (__mmask8) |
| 4542 | __U); |
| 4543 | } |
| 4544 | |
| 4545 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4546 | _mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) { |
| 4547 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4548 | /* idx */ , |
| 4549 | (__v2di) __A, |
| 4550 | (__v2di) __B, |
| 4551 | (__mmask8) -1); |
| 4552 | } |
| 4553 | |
| 4554 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4555 | _mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I, |
| 4556 | __m128i __B) { |
| 4557 | return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I |
| 4558 | /* idx */ , |
| 4559 | (__v2di) __A, |
| 4560 | (__v2di) __B, |
| 4561 | (__mmask8) __U); |
| 4562 | } |
| 4563 | |
| 4564 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4565 | _mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I, |
| 4566 | __m128i __B) { |
| 4567 | return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I |
| 4568 | /* idx */ , |
| 4569 | (__v2di) __A, |
| 4570 | (__v2di) __B, |
| 4571 | (__mmask8) |
| 4572 | __U); |
| 4573 | } |
| 4574 | |
| 4575 | |
| 4576 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4577 | _mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) { |
| 4578 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4579 | /* idx */ , |
| 4580 | (__v4di) __A, |
| 4581 | (__v4di) __B, |
| 4582 | (__mmask8) -1); |
| 4583 | } |
| 4584 | |
| 4585 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4586 | _mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I, |
| 4587 | __m256i __B) { |
| 4588 | return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I |
| 4589 | /* idx */ , |
| 4590 | (__v4di) __A, |
| 4591 | (__v4di) __B, |
| 4592 | (__mmask8) __U); |
| 4593 | } |
| 4594 | |
| 4595 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4596 | _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A, |
| 4597 | __m256i __I, __m256i __B) { |
| 4598 | return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I |
| 4599 | /* idx */ , |
| 4600 | (__v4di) __A, |
| 4601 | (__v4di) __B, |
| 4602 | (__mmask8) |
| 4603 | __U); |
| 4604 | } |
| 4605 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 4606 | #undef __DEFAULT_FN_ATTRS |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 4607 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 4608 | #endif /* __AVX512VLINTRIN_H */ |