Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1 | /*===---- avx2intrin.h - AVX2 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 <avx2intrin.h> directly; include <immintrin.h> instead." |
| 26 | #endif |
| 27 | |
Stephen Hines | 996e4dc | 2013-08-13 01:04:14 -0700 | [diff] [blame] | 28 | #ifndef __AVX2INTRIN_H |
| 29 | #define __AVX2INTRIN_H |
| 30 | |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 31 | /* SSE4 Multiple Packed Sums of Absolute Difference. */ |
| 32 | #define _mm256_mpsadbw_epu8(X, Y, M) __builtin_ia32_mpsadbw256((X), (Y), (M)) |
| 33 | |
| 34 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 35 | _mm256_abs_epi8(__m256i __a) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 36 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 37 | return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 38 | } |
| 39 | |
| 40 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 41 | _mm256_abs_epi16(__m256i __a) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 42 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 43 | return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 44 | } |
| 45 | |
| 46 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 47 | _mm256_abs_epi32(__m256i __a) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 48 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 49 | return (__m256i)__builtin_ia32_pabsd256((__v8si)__a); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 50 | } |
| 51 | |
| 52 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 53 | _mm256_packs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 54 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 55 | return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 56 | } |
| 57 | |
| 58 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 59 | _mm256_packs_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 60 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 61 | return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 62 | } |
| 63 | |
| 64 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 65 | _mm256_packus_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 66 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 67 | return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 68 | } |
| 69 | |
| 70 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 71 | _mm256_packus_epi32(__m256i __V1, __m256i __V2) |
| 72 | { |
| 73 | return (__m256i) __builtin_ia32_packusdw256((__v8si)__V1, (__v8si)__V2); |
| 74 | } |
| 75 | |
| 76 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 77 | _mm256_add_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 78 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 79 | return (__m256i)((__v32qi)__a + (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 80 | } |
| 81 | |
| 82 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 83 | _mm256_add_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 84 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 85 | return (__m256i)((__v16hi)__a + (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 86 | } |
| 87 | |
| 88 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 89 | _mm256_add_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 90 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 91 | return (__m256i)((__v8si)__a + (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 92 | } |
| 93 | |
| 94 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 95 | _mm256_add_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 96 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 97 | return __a + __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 98 | } |
| 99 | |
| 100 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 101 | _mm256_adds_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 102 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 103 | return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 104 | } |
| 105 | |
| 106 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 107 | _mm256_adds_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 108 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 109 | return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 110 | } |
| 111 | |
| 112 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 113 | _mm256_adds_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 114 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 115 | return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 116 | } |
| 117 | |
| 118 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 119 | _mm256_adds_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 120 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 121 | return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 122 | } |
| 123 | |
| 124 | #define _mm256_alignr_epi8(a, b, n) __extension__ ({ \ |
| 125 | __m256i __a = (a); \ |
| 126 | __m256i __b = (b); \ |
| 127 | (__m256i)__builtin_ia32_palignr256((__v32qi)__a, (__v32qi)__b, (n)); }) |
| 128 | |
| 129 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 130 | _mm256_and_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 131 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 132 | return __a & __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 133 | } |
| 134 | |
| 135 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 136 | _mm256_andnot_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 137 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 138 | return ~__a & __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 139 | } |
| 140 | |
| 141 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 142 | _mm256_avg_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 143 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 144 | return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 145 | } |
| 146 | |
| 147 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 148 | _mm256_avg_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 149 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 150 | return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 151 | } |
| 152 | |
| 153 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 154 | _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M) |
| 155 | { |
| 156 | return (__m256i)__builtin_ia32_pblendvb256((__v32qi)__V1, (__v32qi)__V2, |
| 157 | (__v32qi)__M); |
| 158 | } |
| 159 | |
| 160 | #define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \ |
| 161 | __m256i __V1 = (V1); \ |
| 162 | __m256i __V2 = (V2); \ |
Pirama Arumuga Nainar | db74586 | 2015-05-11 14:34:37 -0700 | [diff] [blame] | 163 | (__m256i)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \ |
Stephen Hines | e65db13 | 2014-05-30 13:26:31 -0700 | [diff] [blame] | 164 | (((M) & 0x01) ? 16 : 0), \ |
| 165 | (((M) & 0x02) ? 17 : 1), \ |
| 166 | (((M) & 0x04) ? 18 : 2), \ |
| 167 | (((M) & 0x08) ? 19 : 3), \ |
| 168 | (((M) & 0x10) ? 20 : 4), \ |
| 169 | (((M) & 0x20) ? 21 : 5), \ |
| 170 | (((M) & 0x40) ? 22 : 6), \ |
| 171 | (((M) & 0x80) ? 23 : 7), \ |
| 172 | (((M) & 0x01) ? 24 : 8), \ |
| 173 | (((M) & 0x02) ? 25 : 9), \ |
| 174 | (((M) & 0x04) ? 26 : 10), \ |
| 175 | (((M) & 0x08) ? 27 : 11), \ |
| 176 | (((M) & 0x10) ? 28 : 12), \ |
| 177 | (((M) & 0x20) ? 29 : 13), \ |
| 178 | (((M) & 0x40) ? 30 : 14), \ |
| 179 | (((M) & 0x80) ? 31 : 15)); }) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 180 | |
| 181 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 182 | _mm256_cmpeq_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 183 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 184 | return (__m256i)((__v32qi)__a == (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 185 | } |
| 186 | |
| 187 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 188 | _mm256_cmpeq_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 189 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 190 | return (__m256i)((__v16hi)__a == (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 191 | } |
| 192 | |
| 193 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 194 | _mm256_cmpeq_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 195 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 196 | return (__m256i)((__v8si)__a == (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 197 | } |
| 198 | |
| 199 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 200 | _mm256_cmpeq_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 201 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 202 | return (__m256i)(__a == __b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 203 | } |
| 204 | |
| 205 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 206 | _mm256_cmpgt_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 207 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 208 | return (__m256i)((__v32qi)__a > (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 209 | } |
| 210 | |
| 211 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 212 | _mm256_cmpgt_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 213 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 214 | return (__m256i)((__v16hi)__a > (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 215 | } |
| 216 | |
| 217 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 218 | _mm256_cmpgt_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 219 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 220 | return (__m256i)((__v8si)__a > (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 221 | } |
| 222 | |
| 223 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 224 | _mm256_cmpgt_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 225 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 226 | return (__m256i)(__a > __b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 227 | } |
| 228 | |
| 229 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 230 | _mm256_hadd_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 231 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 232 | return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 233 | } |
| 234 | |
| 235 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 236 | _mm256_hadd_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 237 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 238 | return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 239 | } |
| 240 | |
| 241 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 242 | _mm256_hadds_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 243 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 244 | return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 245 | } |
| 246 | |
| 247 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 248 | _mm256_hsub_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 249 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 250 | return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 251 | } |
| 252 | |
| 253 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 254 | _mm256_hsub_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 255 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 256 | return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 257 | } |
| 258 | |
| 259 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 260 | _mm256_hsubs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 261 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 262 | return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 263 | } |
| 264 | |
| 265 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 266 | _mm256_maddubs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 267 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 268 | return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 269 | } |
| 270 | |
| 271 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 272 | _mm256_madd_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 273 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 274 | return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 275 | } |
| 276 | |
| 277 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 278 | _mm256_max_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 279 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 280 | return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 281 | } |
| 282 | |
| 283 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 284 | _mm256_max_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 285 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 286 | return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 287 | } |
| 288 | |
| 289 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 290 | _mm256_max_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 291 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 292 | return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 293 | } |
| 294 | |
| 295 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 296 | _mm256_max_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 297 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 298 | return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 299 | } |
| 300 | |
| 301 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 302 | _mm256_max_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 303 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 304 | return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 305 | } |
| 306 | |
| 307 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 308 | _mm256_max_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 309 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 310 | return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 311 | } |
| 312 | |
| 313 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 314 | _mm256_min_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 315 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 316 | return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 317 | } |
| 318 | |
| 319 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 320 | _mm256_min_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 321 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 322 | return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 323 | } |
| 324 | |
| 325 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 326 | _mm256_min_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 327 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 328 | return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 329 | } |
| 330 | |
| 331 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 332 | _mm256_min_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 333 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 334 | return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 335 | } |
| 336 | |
| 337 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 338 | _mm256_min_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 339 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 340 | return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 341 | } |
| 342 | |
| 343 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 344 | _mm256_min_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 345 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 346 | return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 347 | } |
| 348 | |
| 349 | static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 350 | _mm256_movemask_epi8(__m256i __a) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 351 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 352 | return __builtin_ia32_pmovmskb256((__v32qi)__a); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 353 | } |
| 354 | |
| 355 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 356 | _mm256_cvtepi8_epi16(__m128i __V) |
| 357 | { |
| 358 | return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V); |
| 359 | } |
| 360 | |
| 361 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 362 | _mm256_cvtepi8_epi32(__m128i __V) |
| 363 | { |
| 364 | return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V); |
| 365 | } |
| 366 | |
| 367 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 368 | _mm256_cvtepi8_epi64(__m128i __V) |
| 369 | { |
| 370 | return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V); |
| 371 | } |
| 372 | |
| 373 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 374 | _mm256_cvtepi16_epi32(__m128i __V) |
| 375 | { |
| 376 | return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V); |
| 377 | } |
| 378 | |
| 379 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 380 | _mm256_cvtepi16_epi64(__m128i __V) |
| 381 | { |
| 382 | return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V); |
| 383 | } |
| 384 | |
| 385 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 386 | _mm256_cvtepi32_epi64(__m128i __V) |
| 387 | { |
| 388 | return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V); |
| 389 | } |
| 390 | |
| 391 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 392 | _mm256_cvtepu8_epi16(__m128i __V) |
| 393 | { |
| 394 | return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V); |
| 395 | } |
| 396 | |
| 397 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 398 | _mm256_cvtepu8_epi32(__m128i __V) |
| 399 | { |
| 400 | return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V); |
| 401 | } |
| 402 | |
| 403 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 404 | _mm256_cvtepu8_epi64(__m128i __V) |
| 405 | { |
| 406 | return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V); |
| 407 | } |
| 408 | |
| 409 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 410 | _mm256_cvtepu16_epi32(__m128i __V) |
| 411 | { |
| 412 | return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V); |
| 413 | } |
| 414 | |
| 415 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 416 | _mm256_cvtepu16_epi64(__m128i __V) |
| 417 | { |
| 418 | return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V); |
| 419 | } |
| 420 | |
| 421 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 422 | _mm256_cvtepu32_epi64(__m128i __V) |
| 423 | { |
| 424 | return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V); |
| 425 | } |
| 426 | |
| 427 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 428 | _mm256_mul_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 429 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 430 | return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 431 | } |
| 432 | |
| 433 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 434 | _mm256_mulhrs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 435 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 436 | return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 437 | } |
| 438 | |
| 439 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 440 | _mm256_mulhi_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 441 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 442 | return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 443 | } |
| 444 | |
| 445 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 446 | _mm256_mulhi_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 447 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 448 | return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 449 | } |
| 450 | |
| 451 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 452 | _mm256_mullo_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 453 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 454 | return (__m256i)((__v16hi)__a * (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 455 | } |
| 456 | |
| 457 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 458 | _mm256_mullo_epi32 (__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 459 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 460 | return (__m256i)((__v8si)__a * (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 461 | } |
| 462 | |
| 463 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 464 | _mm256_mul_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 465 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 466 | return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 467 | } |
| 468 | |
| 469 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 470 | _mm256_or_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 471 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 472 | return __a | __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 473 | } |
| 474 | |
| 475 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 476 | _mm256_sad_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 477 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 478 | return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 479 | } |
| 480 | |
| 481 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 482 | _mm256_shuffle_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 483 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 484 | return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 485 | } |
| 486 | |
| 487 | #define _mm256_shuffle_epi32(a, imm) __extension__ ({ \ |
| 488 | __m256i __a = (a); \ |
| 489 | (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \ |
| 490 | (imm) & 0x3, ((imm) & 0xc) >> 2, \ |
| 491 | ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \ |
| 492 | 4 + (((imm) & 0x03) >> 0), \ |
| 493 | 4 + (((imm) & 0x0c) >> 2), \ |
| 494 | 4 + (((imm) & 0x30) >> 4), \ |
| 495 | 4 + (((imm) & 0xc0) >> 6)); }) |
| 496 | |
| 497 | #define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \ |
| 498 | __m256i __a = (a); \ |
| 499 | (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \ |
| 500 | 0, 1, 2, 3, \ |
| 501 | 4 + (((imm) & 0x03) >> 0), \ |
| 502 | 4 + (((imm) & 0x0c) >> 2), \ |
| 503 | 4 + (((imm) & 0x30) >> 4), \ |
| 504 | 4 + (((imm) & 0xc0) >> 6), \ |
| 505 | 8, 9, 10, 11, \ |
| 506 | 12 + (((imm) & 0x03) >> 0), \ |
| 507 | 12 + (((imm) & 0x0c) >> 2), \ |
| 508 | 12 + (((imm) & 0x30) >> 4), \ |
| 509 | 12 + (((imm) & 0xc0) >> 6)); }) |
| 510 | |
| 511 | #define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \ |
| 512 | __m256i __a = (a); \ |
| 513 | (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \ |
| 514 | (imm) & 0x3,((imm) & 0xc) >> 2, \ |
| 515 | ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \ |
| 516 | 4, 5, 6, 7, \ |
| 517 | 8 + (((imm) & 0x03) >> 0), \ |
| 518 | 8 + (((imm) & 0x0c) >> 2), \ |
| 519 | 8 + (((imm) & 0x30) >> 4), \ |
| 520 | 8 + (((imm) & 0xc0) >> 6), \ |
| 521 | 12, 13, 14, 15); }) |
| 522 | |
| 523 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 524 | _mm256_sign_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 525 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 526 | return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 527 | } |
| 528 | |
| 529 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 530 | _mm256_sign_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 531 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 532 | return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 533 | } |
| 534 | |
| 535 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 536 | _mm256_sign_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 537 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 538 | return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 539 | } |
| 540 | |
| 541 | #define _mm256_slli_si256(a, count) __extension__ ({ \ |
| 542 | __m256i __a = (a); \ |
| 543 | (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); }) |
| 544 | |
| 545 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 546 | _mm256_slli_epi16(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 547 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 548 | return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 549 | } |
| 550 | |
| 551 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 552 | _mm256_sll_epi16(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 553 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 554 | return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 555 | } |
| 556 | |
| 557 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 558 | _mm256_slli_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 559 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 560 | return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 561 | } |
| 562 | |
| 563 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 564 | _mm256_sll_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 565 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 566 | return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 567 | } |
| 568 | |
| 569 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 570 | _mm256_slli_epi64(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 571 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 572 | return __builtin_ia32_psllqi256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 573 | } |
| 574 | |
| 575 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 576 | _mm256_sll_epi64(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 577 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 578 | return __builtin_ia32_psllq256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 579 | } |
| 580 | |
| 581 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 582 | _mm256_srai_epi16(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 583 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 584 | return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 585 | } |
| 586 | |
| 587 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 588 | _mm256_sra_epi16(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 589 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 590 | return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 591 | } |
| 592 | |
| 593 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 594 | _mm256_srai_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 595 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 596 | return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 597 | } |
| 598 | |
| 599 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 600 | _mm256_sra_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 601 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 602 | return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 603 | } |
| 604 | |
| 605 | #define _mm256_srli_si256(a, count) __extension__ ({ \ |
| 606 | __m256i __a = (a); \ |
| 607 | (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); }) |
| 608 | |
| 609 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 610 | _mm256_srli_epi16(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 611 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 612 | return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 613 | } |
| 614 | |
| 615 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 616 | _mm256_srl_epi16(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 617 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 618 | return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 619 | } |
| 620 | |
| 621 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 622 | _mm256_srli_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 623 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 624 | return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 625 | } |
| 626 | |
| 627 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 628 | _mm256_srl_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 629 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 630 | return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 631 | } |
| 632 | |
| 633 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 634 | _mm256_srli_epi64(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 635 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 636 | return __builtin_ia32_psrlqi256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 637 | } |
| 638 | |
| 639 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 640 | _mm256_srl_epi64(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 641 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 642 | return __builtin_ia32_psrlq256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 643 | } |
| 644 | |
| 645 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 646 | _mm256_sub_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 647 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 648 | return (__m256i)((__v32qi)__a - (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 649 | } |
| 650 | |
| 651 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 652 | _mm256_sub_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 653 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 654 | return (__m256i)((__v16hi)__a - (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 655 | } |
| 656 | |
| 657 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 658 | _mm256_sub_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 659 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 660 | return (__m256i)((__v8si)__a - (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 661 | } |
| 662 | |
| 663 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 664 | _mm256_sub_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 665 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 666 | return __a - __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 667 | } |
| 668 | |
| 669 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 670 | _mm256_subs_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 671 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 672 | return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 673 | } |
| 674 | |
| 675 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 676 | _mm256_subs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 677 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 678 | return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 679 | } |
| 680 | |
| 681 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 682 | _mm256_subs_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 683 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 684 | return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 685 | } |
| 686 | |
| 687 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 688 | _mm256_subs_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 689 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 690 | return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 691 | } |
| 692 | |
| 693 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 694 | _mm256_unpackhi_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 695 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 696 | return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 8, 32+8, 9, 32+9, 10, 32+10, 11, 32+11, 12, 32+12, 13, 32+13, 14, 32+14, 15, 32+15, 24, 32+24, 25, 32+25, 26, 32+26, 27, 32+27, 28, 32+28, 29, 32+29, 30, 32+30, 31, 32+31); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 697 | } |
| 698 | |
| 699 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 700 | _mm256_unpackhi_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 701 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 702 | return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 703 | } |
| 704 | |
| 705 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 706 | _mm256_unpackhi_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 707 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 708 | return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 709 | } |
| 710 | |
| 711 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 712 | _mm256_unpackhi_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 713 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 714 | return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 715 | } |
| 716 | |
| 717 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 718 | _mm256_unpacklo_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 719 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 720 | return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 0, 32+0, 1, 32+1, 2, 32+2, 3, 32+3, 4, 32+4, 5, 32+5, 6, 32+6, 7, 32+7, 16, 32+16, 17, 32+17, 18, 32+18, 19, 32+19, 20, 32+20, 21, 32+21, 22, 32+22, 23, 32+23); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 721 | } |
| 722 | |
| 723 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 724 | _mm256_unpacklo_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 725 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 726 | return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 727 | } |
| 728 | |
| 729 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 730 | _mm256_unpacklo_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 731 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 732 | return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 733 | } |
| 734 | |
| 735 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 736 | _mm256_unpacklo_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 737 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 738 | return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 739 | } |
| 740 | |
| 741 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 742 | _mm256_xor_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 743 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 744 | return __a ^ __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 745 | } |
| 746 | |
| 747 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 748 | _mm256_stream_load_si256(__m256i *__V) |
| 749 | { |
| 750 | return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V); |
| 751 | } |
| 752 | |
| 753 | static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 754 | _mm_broadcastss_ps(__m128 __X) |
| 755 | { |
| 756 | return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X); |
| 757 | } |
| 758 | |
| 759 | static __inline__ __m256 __attribute__((__always_inline__, __nodebug__)) |
| 760 | _mm256_broadcastss_ps(__m128 __X) |
| 761 | { |
| 762 | return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X); |
| 763 | } |
| 764 | |
| 765 | static __inline__ __m256d __attribute__((__always_inline__, __nodebug__)) |
| 766 | _mm256_broadcastsd_pd(__m128d __X) |
| 767 | { |
| 768 | return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X); |
| 769 | } |
| 770 | |
| 771 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 772 | _mm256_broadcastsi128_si256(__m128i __X) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 773 | { |
Stephen Hines | 3f86823 | 2015-04-10 09:22:19 -0700 | [diff] [blame] | 774 | return (__m256i)__builtin_shufflevector(__X, __X, 0, 1, 0, 1); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 775 | } |
| 776 | |
| 777 | #define _mm_blend_epi32(V1, V2, M) __extension__ ({ \ |
| 778 | __m128i __V1 = (V1); \ |
| 779 | __m128i __V2 = (V2); \ |
Stephen Hines | e65db13 | 2014-05-30 13:26:31 -0700 | [diff] [blame] | 780 | (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \ |
| 781 | (((M) & 0x01) ? 4 : 0), \ |
| 782 | (((M) & 0x02) ? 5 : 1), \ |
| 783 | (((M) & 0x04) ? 6 : 2), \ |
| 784 | (((M) & 0x08) ? 7 : 3)); }) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 785 | |
| 786 | #define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \ |
| 787 | __m256i __V1 = (V1); \ |
| 788 | __m256i __V2 = (V2); \ |
Stephen Hines | e65db13 | 2014-05-30 13:26:31 -0700 | [diff] [blame] | 789 | (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \ |
| 790 | (((M) & 0x01) ? 8 : 0), \ |
| 791 | (((M) & 0x02) ? 9 : 1), \ |
| 792 | (((M) & 0x04) ? 10 : 2), \ |
| 793 | (((M) & 0x08) ? 11 : 3), \ |
| 794 | (((M) & 0x10) ? 12 : 4), \ |
| 795 | (((M) & 0x20) ? 13 : 5), \ |
| 796 | (((M) & 0x40) ? 14 : 6), \ |
| 797 | (((M) & 0x80) ? 15 : 7)); }) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 798 | |
| 799 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 800 | _mm256_broadcastb_epi8(__m128i __X) |
| 801 | { |
| 802 | return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X); |
| 803 | } |
| 804 | |
| 805 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 806 | _mm256_broadcastw_epi16(__m128i __X) |
| 807 | { |
| 808 | return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X); |
| 809 | } |
| 810 | |
| 811 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 812 | _mm256_broadcastd_epi32(__m128i __X) |
| 813 | { |
| 814 | return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X); |
| 815 | } |
| 816 | |
| 817 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 818 | _mm256_broadcastq_epi64(__m128i __X) |
| 819 | { |
| 820 | return (__m256i)__builtin_ia32_pbroadcastq256(__X); |
| 821 | } |
| 822 | |
| 823 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 824 | _mm_broadcastb_epi8(__m128i __X) |
| 825 | { |
| 826 | return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X); |
| 827 | } |
| 828 | |
| 829 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 830 | _mm_broadcastw_epi16(__m128i __X) |
| 831 | { |
| 832 | return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X); |
| 833 | } |
| 834 | |
| 835 | |
| 836 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 837 | _mm_broadcastd_epi32(__m128i __X) |
| 838 | { |
| 839 | return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X); |
| 840 | } |
| 841 | |
| 842 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 843 | _mm_broadcastq_epi64(__m128i __X) |
| 844 | { |
| 845 | return (__m128i)__builtin_ia32_pbroadcastq128(__X); |
| 846 | } |
| 847 | |
| 848 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 849 | _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 850 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 851 | return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 852 | } |
| 853 | |
| 854 | #define _mm256_permute4x64_pd(V, M) __extension__ ({ \ |
| 855 | __m256d __V = (V); \ |
| 856 | (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \ |
| 857 | (M) & 0x3, ((M) & 0xc) >> 2, \ |
| 858 | ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); }) |
| 859 | |
| 860 | static __inline__ __m256 __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 861 | _mm256_permutevar8x32_ps(__m256 __a, __m256 __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 862 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 863 | return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 864 | } |
| 865 | |
| 866 | #define _mm256_permute4x64_epi64(V, M) __extension__ ({ \ |
| 867 | __m256i __V = (V); \ |
| 868 | (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \ |
| 869 | (M) & 0x3, ((M) & 0xc) >> 2, \ |
| 870 | ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); }) |
| 871 | |
| 872 | #define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \ |
| 873 | __m256i __V1 = (V1); \ |
| 874 | __m256i __V2 = (V2); \ |
| 875 | (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); }) |
| 876 | |
Stephen Hines | 3f86823 | 2015-04-10 09:22:19 -0700 | [diff] [blame] | 877 | #define _mm256_extracti128_si256(V, M) __extension__ ({ \ |
| 878 | (__m128i)__builtin_shufflevector( \ |
| 879 | (__v4di)(V), \ |
| 880 | (__v4di)(_mm256_setzero_si256()), \ |
| 881 | (((M) & 1) ? 2 : 0), \ |
| 882 | (((M) & 1) ? 3 : 1) );}) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 883 | |
Stephen Hines | 3f86823 | 2015-04-10 09:22:19 -0700 | [diff] [blame] | 884 | #define _mm256_inserti128_si256(V1, V2, M) __extension__ ({ \ |
| 885 | (__m256i)__builtin_shufflevector( \ |
| 886 | (__v4di)(V1), \ |
| 887 | (__v4di)_mm256_castsi128_si256((__m128i)(V2)), \ |
| 888 | (((M) & 1) ? 0 : 4), \ |
| 889 | (((M) & 1) ? 1 : 5), \ |
| 890 | (((M) & 1) ? 4 : 2), \ |
| 891 | (((M) & 1) ? 5 : 3) );}) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 892 | |
| 893 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 894 | _mm256_maskload_epi32(int const *__X, __m256i __M) |
| 895 | { |
| 896 | return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M); |
| 897 | } |
| 898 | |
| 899 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 900 | _mm256_maskload_epi64(long long const *__X, __m256i __M) |
| 901 | { |
| 902 | return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M); |
| 903 | } |
| 904 | |
| 905 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 906 | _mm_maskload_epi32(int const *__X, __m128i __M) |
| 907 | { |
| 908 | return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M); |
| 909 | } |
| 910 | |
| 911 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 912 | _mm_maskload_epi64(long long const *__X, __m128i __M) |
| 913 | { |
| 914 | return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M); |
| 915 | } |
| 916 | |
| 917 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 918 | _mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y) |
| 919 | { |
| 920 | __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y); |
| 921 | } |
| 922 | |
| 923 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 924 | _mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y) |
| 925 | { |
| 926 | __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y); |
| 927 | } |
| 928 | |
| 929 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 930 | _mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y) |
| 931 | { |
| 932 | __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y); |
| 933 | } |
| 934 | |
| 935 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 936 | _mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y) |
| 937 | { |
| 938 | __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y); |
| 939 | } |
| 940 | |
| 941 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 942 | _mm256_sllv_epi32(__m256i __X, __m256i __Y) |
| 943 | { |
| 944 | return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y); |
| 945 | } |
| 946 | |
| 947 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 948 | _mm_sllv_epi32(__m128i __X, __m128i __Y) |
| 949 | { |
| 950 | return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y); |
| 951 | } |
| 952 | |
| 953 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 954 | _mm256_sllv_epi64(__m256i __X, __m256i __Y) |
| 955 | { |
| 956 | return (__m256i)__builtin_ia32_psllv4di(__X, __Y); |
| 957 | } |
| 958 | |
| 959 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 960 | _mm_sllv_epi64(__m128i __X, __m128i __Y) |
| 961 | { |
| 962 | return (__m128i)__builtin_ia32_psllv2di(__X, __Y); |
| 963 | } |
| 964 | |
| 965 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 966 | _mm256_srav_epi32(__m256i __X, __m256i __Y) |
| 967 | { |
| 968 | return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y); |
| 969 | } |
| 970 | |
| 971 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 972 | _mm_srav_epi32(__m128i __X, __m128i __Y) |
| 973 | { |
| 974 | return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y); |
| 975 | } |
| 976 | |
| 977 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 978 | _mm256_srlv_epi32(__m256i __X, __m256i __Y) |
| 979 | { |
| 980 | return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y); |
| 981 | } |
| 982 | |
| 983 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 984 | _mm_srlv_epi32(__m128i __X, __m128i __Y) |
| 985 | { |
| 986 | return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y); |
| 987 | } |
| 988 | |
| 989 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 990 | _mm256_srlv_epi64(__m256i __X, __m256i __Y) |
| 991 | { |
| 992 | return (__m256i)__builtin_ia32_psrlv4di(__X, __Y); |
| 993 | } |
| 994 | |
| 995 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 996 | _mm_srlv_epi64(__m128i __X, __m128i __Y) |
| 997 | { |
| 998 | return (__m128i)__builtin_ia32_psrlv2di(__X, __Y); |
| 999 | } |
| 1000 | |
| 1001 | #define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 1002 | __m128d __a = (a); \ |
| 1003 | double const *__m = (m); \ |
| 1004 | __m128i __i = (i); \ |
| 1005 | __m128d __mask = (mask); \ |
| 1006 | (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \ |
| 1007 | (__v4si)__i, (__v2df)__mask, (s)); }) |
| 1008 | |
| 1009 | #define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 1010 | __m256d __a = (a); \ |
| 1011 | double const *__m = (m); \ |
| 1012 | __m128i __i = (i); \ |
| 1013 | __m256d __mask = (mask); \ |
| 1014 | (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \ |
| 1015 | (__v4si)__i, (__v4df)__mask, (s)); }) |
| 1016 | |
| 1017 | #define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 1018 | __m128d __a = (a); \ |
| 1019 | double const *__m = (m); \ |
| 1020 | __m128i __i = (i); \ |
| 1021 | __m128d __mask = (mask); \ |
| 1022 | (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \ |
| 1023 | (__v2di)__i, (__v2df)__mask, (s)); }) |
| 1024 | |
| 1025 | #define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 1026 | __m256d __a = (a); \ |
| 1027 | double const *__m = (m); \ |
| 1028 | __m256i __i = (i); \ |
| 1029 | __m256d __mask = (mask); \ |
| 1030 | (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \ |
| 1031 | (__v4di)__i, (__v4df)__mask, (s)); }) |
| 1032 | |
| 1033 | #define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1034 | __m128 __a = (a); \ |
| 1035 | float const *__m = (m); \ |
| 1036 | __m128i __i = (i); \ |
| 1037 | __m128 __mask = (mask); \ |
| 1038 | (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \ |
| 1039 | (__v4si)__i, (__v4sf)__mask, (s)); }) |
| 1040 | |
| 1041 | #define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1042 | __m256 __a = (a); \ |
| 1043 | float const *__m = (m); \ |
| 1044 | __m256i __i = (i); \ |
| 1045 | __m256 __mask = (mask); \ |
| 1046 | (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \ |
| 1047 | (__v8si)__i, (__v8sf)__mask, (s)); }) |
| 1048 | |
| 1049 | #define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1050 | __m128 __a = (a); \ |
| 1051 | float const *__m = (m); \ |
| 1052 | __m128i __i = (i); \ |
| 1053 | __m128 __mask = (mask); \ |
| 1054 | (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \ |
| 1055 | (__v2di)__i, (__v4sf)__mask, (s)); }) |
| 1056 | |
| 1057 | #define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1058 | __m128 __a = (a); \ |
| 1059 | float const *__m = (m); \ |
| 1060 | __m256i __i = (i); \ |
| 1061 | __m128 __mask = (mask); \ |
| 1062 | (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \ |
| 1063 | (__v4di)__i, (__v4sf)__mask, (s)); }) |
| 1064 | |
| 1065 | #define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1066 | __m128i __a = (a); \ |
| 1067 | int const *__m = (m); \ |
| 1068 | __m128i __i = (i); \ |
| 1069 | __m128i __mask = (mask); \ |
| 1070 | (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \ |
| 1071 | (__v4si)__i, (__v4si)__mask, (s)); }) |
| 1072 | |
| 1073 | #define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1074 | __m256i __a = (a); \ |
| 1075 | int const *__m = (m); \ |
| 1076 | __m256i __i = (i); \ |
| 1077 | __m256i __mask = (mask); \ |
| 1078 | (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \ |
| 1079 | (__v8si)__i, (__v8si)__mask, (s)); }) |
| 1080 | |
| 1081 | #define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1082 | __m128i __a = (a); \ |
| 1083 | int const *__m = (m); \ |
| 1084 | __m128i __i = (i); \ |
| 1085 | __m128i __mask = (mask); \ |
| 1086 | (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \ |
| 1087 | (__v2di)__i, (__v4si)__mask, (s)); }) |
| 1088 | |
| 1089 | #define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1090 | __m128i __a = (a); \ |
| 1091 | int const *__m = (m); \ |
| 1092 | __m256i __i = (i); \ |
| 1093 | __m128i __mask = (mask); \ |
| 1094 | (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \ |
| 1095 | (__v4di)__i, (__v4si)__mask, (s)); }) |
| 1096 | |
| 1097 | #define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1098 | __m128i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1099 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1100 | __m128i __i = (i); \ |
| 1101 | __m128i __mask = (mask); \ |
| 1102 | (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \ |
| 1103 | (__v4si)__i, (__v2di)__mask, (s)); }) |
| 1104 | |
| 1105 | #define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1106 | __m256i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1107 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1108 | __m128i __i = (i); \ |
| 1109 | __m256i __mask = (mask); \ |
| 1110 | (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \ |
| 1111 | (__v4si)__i, (__v4di)__mask, (s)); }) |
| 1112 | |
| 1113 | #define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1114 | __m128i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1115 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1116 | __m128i __i = (i); \ |
| 1117 | __m128i __mask = (mask); \ |
| 1118 | (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \ |
| 1119 | (__v2di)__i, (__v2di)__mask, (s)); }) |
| 1120 | |
| 1121 | #define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1122 | __m256i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1123 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1124 | __m256i __i = (i); \ |
| 1125 | __m256i __mask = (mask); \ |
| 1126 | (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \ |
| 1127 | (__v4di)__i, (__v4di)__mask, (s)); }) |
| 1128 | |
| 1129 | #define _mm_i32gather_pd(m, i, s) __extension__ ({ \ |
| 1130 | double const *__m = (m); \ |
| 1131 | __m128i __i = (i); \ |
| 1132 | (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \ |
| 1133 | (const __v2df *)__m, (__v4si)__i, \ |
| 1134 | (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); }) |
| 1135 | |
| 1136 | #define _mm256_i32gather_pd(m, i, s) __extension__ ({ \ |
| 1137 | double const *__m = (m); \ |
| 1138 | __m128i __i = (i); \ |
| 1139 | (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \ |
| 1140 | (const __v4df *)__m, (__v4si)__i, \ |
| 1141 | (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); }) |
| 1142 | |
| 1143 | #define _mm_i64gather_pd(m, i, s) __extension__ ({ \ |
| 1144 | double const *__m = (m); \ |
| 1145 | __m128i __i = (i); \ |
| 1146 | (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \ |
| 1147 | (const __v2df *)__m, (__v2di)__i, \ |
| 1148 | (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); }) |
| 1149 | |
| 1150 | #define _mm256_i64gather_pd(m, i, s) __extension__ ({ \ |
| 1151 | double const *__m = (m); \ |
| 1152 | __m256i __i = (i); \ |
| 1153 | (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \ |
| 1154 | (const __v4df *)__m, (__v4di)__i, \ |
| 1155 | (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); }) |
| 1156 | |
| 1157 | #define _mm_i32gather_ps(m, i, s) __extension__ ({ \ |
| 1158 | float const *__m = (m); \ |
| 1159 | __m128i __i = (i); \ |
| 1160 | (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \ |
| 1161 | (const __v4sf *)__m, (__v4si)__i, \ |
| 1162 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1163 | |
| 1164 | #define _mm256_i32gather_ps(m, i, s) __extension__ ({ \ |
| 1165 | float const *__m = (m); \ |
| 1166 | __m256i __i = (i); \ |
| 1167 | (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \ |
| 1168 | (const __v8sf *)__m, (__v8si)__i, \ |
| 1169 | (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); }) |
| 1170 | |
| 1171 | #define _mm_i64gather_ps(m, i, s) __extension__ ({ \ |
| 1172 | float const *__m = (m); \ |
| 1173 | __m128i __i = (i); \ |
| 1174 | (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \ |
| 1175 | (const __v4sf *)__m, (__v2di)__i, \ |
| 1176 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1177 | |
| 1178 | #define _mm256_i64gather_ps(m, i, s) __extension__ ({ \ |
| 1179 | float const *__m = (m); \ |
| 1180 | __m256i __i = (i); \ |
| 1181 | (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \ |
| 1182 | (const __v4sf *)__m, (__v4di)__i, \ |
| 1183 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1184 | |
| 1185 | #define _mm_i32gather_epi32(m, i, s) __extension__ ({ \ |
| 1186 | int const *__m = (m); \ |
| 1187 | __m128i __i = (i); \ |
| 1188 | (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \ |
| 1189 | (const __v4si *)__m, (__v4si)__i, \ |
| 1190 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1191 | |
| 1192 | #define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \ |
| 1193 | int const *__m = (m); \ |
| 1194 | __m256i __i = (i); \ |
| 1195 | (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \ |
| 1196 | (const __v8si *)__m, (__v8si)__i, \ |
| 1197 | (__v8si)_mm256_set1_epi32(-1), (s)); }) |
| 1198 | |
| 1199 | #define _mm_i64gather_epi32(m, i, s) __extension__ ({ \ |
| 1200 | int const *__m = (m); \ |
| 1201 | __m128i __i = (i); \ |
| 1202 | (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \ |
| 1203 | (const __v4si *)__m, (__v2di)__i, \ |
| 1204 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1205 | |
| 1206 | #define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \ |
| 1207 | int const *__m = (m); \ |
| 1208 | __m256i __i = (i); \ |
| 1209 | (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \ |
| 1210 | (const __v4si *)__m, (__v4di)__i, \ |
| 1211 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1212 | |
| 1213 | #define _mm_i32gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1214 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1215 | __m128i __i = (i); \ |
| 1216 | (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \ |
| 1217 | (const __v2di *)__m, (__v4si)__i, \ |
| 1218 | (__v2di)_mm_set1_epi64x(-1), (s)); }) |
| 1219 | |
| 1220 | #define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1221 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1222 | __m128i __i = (i); \ |
| 1223 | (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \ |
| 1224 | (const __v4di *)__m, (__v4si)__i, \ |
| 1225 | (__v4di)_mm256_set1_epi64x(-1), (s)); }) |
| 1226 | |
| 1227 | #define _mm_i64gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1228 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1229 | __m128i __i = (i); \ |
| 1230 | (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \ |
| 1231 | (const __v2di *)__m, (__v2di)__i, \ |
| 1232 | (__v2di)_mm_set1_epi64x(-1), (s)); }) |
| 1233 | |
| 1234 | #define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame] | 1235 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1236 | __m256i __i = (i); \ |
| 1237 | (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \ |
| 1238 | (const __v4di *)__m, (__v4di)__i, \ |
| 1239 | (__v4di)_mm256_set1_epi64x(-1), (s)); }) |
Stephen Hines | 996e4dc | 2013-08-13 01:04:14 -0700 | [diff] [blame] | 1240 | |
| 1241 | #endif /* __AVX2INTRIN_H */ |