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); \ |
| 163 | (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); }) |
| 164 | |
| 165 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 166 | _mm256_cmpeq_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 167 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 168 | return (__m256i)((__v32qi)__a == (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 169 | } |
| 170 | |
| 171 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 172 | _mm256_cmpeq_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 173 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 174 | return (__m256i)((__v16hi)__a == (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 175 | } |
| 176 | |
| 177 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 178 | _mm256_cmpeq_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 179 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 180 | return (__m256i)((__v8si)__a == (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 181 | } |
| 182 | |
| 183 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 184 | _mm256_cmpeq_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 185 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 186 | return (__m256i)(__a == __b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 187 | } |
| 188 | |
| 189 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 190 | _mm256_cmpgt_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 191 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 192 | return (__m256i)((__v32qi)__a > (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 193 | } |
| 194 | |
| 195 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 196 | _mm256_cmpgt_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 197 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 198 | return (__m256i)((__v16hi)__a > (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 199 | } |
| 200 | |
| 201 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 202 | _mm256_cmpgt_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 203 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 204 | return (__m256i)((__v8si)__a > (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 205 | } |
| 206 | |
| 207 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 208 | _mm256_cmpgt_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 209 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 210 | return (__m256i)(__a > __b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 211 | } |
| 212 | |
| 213 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 214 | _mm256_hadd_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 215 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 216 | return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 217 | } |
| 218 | |
| 219 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 220 | _mm256_hadd_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 221 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 222 | return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 223 | } |
| 224 | |
| 225 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 226 | _mm256_hadds_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 227 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 228 | return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 229 | } |
| 230 | |
| 231 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 232 | _mm256_hsub_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 233 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 234 | return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 235 | } |
| 236 | |
| 237 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 238 | _mm256_hsub_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 239 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 240 | return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 241 | } |
| 242 | |
| 243 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 244 | _mm256_hsubs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 245 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 246 | return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 247 | } |
| 248 | |
| 249 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 250 | _mm256_maddubs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 251 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 252 | return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 253 | } |
| 254 | |
| 255 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 256 | _mm256_madd_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 257 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 258 | return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 259 | } |
| 260 | |
| 261 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 262 | _mm256_max_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 263 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 264 | return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 265 | } |
| 266 | |
| 267 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 268 | _mm256_max_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 269 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 270 | return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 271 | } |
| 272 | |
| 273 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 274 | _mm256_max_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 275 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 276 | return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 277 | } |
| 278 | |
| 279 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 280 | _mm256_max_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 281 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 282 | return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 283 | } |
| 284 | |
| 285 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 286 | _mm256_max_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 287 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 288 | return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 289 | } |
| 290 | |
| 291 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 292 | _mm256_max_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 293 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 294 | return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 295 | } |
| 296 | |
| 297 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 298 | _mm256_min_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 299 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 300 | return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 301 | } |
| 302 | |
| 303 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 304 | _mm256_min_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 305 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 306 | return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 307 | } |
| 308 | |
| 309 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 310 | _mm256_min_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 311 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 312 | return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 313 | } |
| 314 | |
| 315 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 316 | _mm256_min_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 317 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 318 | return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 319 | } |
| 320 | |
| 321 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 322 | _mm256_min_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 323 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 324 | return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 325 | } |
| 326 | |
| 327 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 328 | _mm256_min_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 329 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 330 | return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 331 | } |
| 332 | |
| 333 | static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 334 | _mm256_movemask_epi8(__m256i __a) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 335 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 336 | return __builtin_ia32_pmovmskb256((__v32qi)__a); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 337 | } |
| 338 | |
| 339 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 340 | _mm256_cvtepi8_epi16(__m128i __V) |
| 341 | { |
| 342 | return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V); |
| 343 | } |
| 344 | |
| 345 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 346 | _mm256_cvtepi8_epi32(__m128i __V) |
| 347 | { |
| 348 | return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V); |
| 349 | } |
| 350 | |
| 351 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 352 | _mm256_cvtepi8_epi64(__m128i __V) |
| 353 | { |
| 354 | return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V); |
| 355 | } |
| 356 | |
| 357 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 358 | _mm256_cvtepi16_epi32(__m128i __V) |
| 359 | { |
| 360 | return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V); |
| 361 | } |
| 362 | |
| 363 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 364 | _mm256_cvtepi16_epi64(__m128i __V) |
| 365 | { |
| 366 | return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V); |
| 367 | } |
| 368 | |
| 369 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 370 | _mm256_cvtepi32_epi64(__m128i __V) |
| 371 | { |
| 372 | return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V); |
| 373 | } |
| 374 | |
| 375 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 376 | _mm256_cvtepu8_epi16(__m128i __V) |
| 377 | { |
| 378 | return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V); |
| 379 | } |
| 380 | |
| 381 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 382 | _mm256_cvtepu8_epi32(__m128i __V) |
| 383 | { |
| 384 | return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V); |
| 385 | } |
| 386 | |
| 387 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 388 | _mm256_cvtepu8_epi64(__m128i __V) |
| 389 | { |
| 390 | return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V); |
| 391 | } |
| 392 | |
| 393 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 394 | _mm256_cvtepu16_epi32(__m128i __V) |
| 395 | { |
| 396 | return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V); |
| 397 | } |
| 398 | |
| 399 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 400 | _mm256_cvtepu16_epi64(__m128i __V) |
| 401 | { |
| 402 | return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V); |
| 403 | } |
| 404 | |
| 405 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 406 | _mm256_cvtepu32_epi64(__m128i __V) |
| 407 | { |
| 408 | return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V); |
| 409 | } |
| 410 | |
| 411 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 412 | _mm256_mul_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 413 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 414 | return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 415 | } |
| 416 | |
| 417 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 418 | _mm256_mulhrs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 419 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 420 | return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 421 | } |
| 422 | |
| 423 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 424 | _mm256_mulhi_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 425 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 426 | return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 427 | } |
| 428 | |
| 429 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 430 | _mm256_mulhi_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 431 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 432 | return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 433 | } |
| 434 | |
| 435 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 436 | _mm256_mullo_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 437 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 438 | return (__m256i)((__v16hi)__a * (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 439 | } |
| 440 | |
| 441 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 442 | _mm256_mullo_epi32 (__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 443 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 444 | return (__m256i)((__v8si)__a * (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 445 | } |
| 446 | |
| 447 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 448 | _mm256_mul_epu32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 449 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 450 | return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 451 | } |
| 452 | |
| 453 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 454 | _mm256_or_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 455 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 456 | return __a | __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 457 | } |
| 458 | |
| 459 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 460 | _mm256_sad_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 461 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 462 | return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 463 | } |
| 464 | |
| 465 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 466 | _mm256_shuffle_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 467 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 468 | return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 469 | } |
| 470 | |
| 471 | #define _mm256_shuffle_epi32(a, imm) __extension__ ({ \ |
| 472 | __m256i __a = (a); \ |
| 473 | (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \ |
| 474 | (imm) & 0x3, ((imm) & 0xc) >> 2, \ |
| 475 | ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \ |
| 476 | 4 + (((imm) & 0x03) >> 0), \ |
| 477 | 4 + (((imm) & 0x0c) >> 2), \ |
| 478 | 4 + (((imm) & 0x30) >> 4), \ |
| 479 | 4 + (((imm) & 0xc0) >> 6)); }) |
| 480 | |
| 481 | #define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \ |
| 482 | __m256i __a = (a); \ |
| 483 | (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \ |
| 484 | 0, 1, 2, 3, \ |
| 485 | 4 + (((imm) & 0x03) >> 0), \ |
| 486 | 4 + (((imm) & 0x0c) >> 2), \ |
| 487 | 4 + (((imm) & 0x30) >> 4), \ |
| 488 | 4 + (((imm) & 0xc0) >> 6), \ |
| 489 | 8, 9, 10, 11, \ |
| 490 | 12 + (((imm) & 0x03) >> 0), \ |
| 491 | 12 + (((imm) & 0x0c) >> 2), \ |
| 492 | 12 + (((imm) & 0x30) >> 4), \ |
| 493 | 12 + (((imm) & 0xc0) >> 6)); }) |
| 494 | |
| 495 | #define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \ |
| 496 | __m256i __a = (a); \ |
| 497 | (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \ |
| 498 | (imm) & 0x3,((imm) & 0xc) >> 2, \ |
| 499 | ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \ |
| 500 | 4, 5, 6, 7, \ |
| 501 | 8 + (((imm) & 0x03) >> 0), \ |
| 502 | 8 + (((imm) & 0x0c) >> 2), \ |
| 503 | 8 + (((imm) & 0x30) >> 4), \ |
| 504 | 8 + (((imm) & 0xc0) >> 6), \ |
| 505 | 12, 13, 14, 15); }) |
| 506 | |
| 507 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 508 | _mm256_sign_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 509 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 510 | return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 511 | } |
| 512 | |
| 513 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 514 | _mm256_sign_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 515 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 516 | return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 517 | } |
| 518 | |
| 519 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 520 | _mm256_sign_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 521 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 522 | return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 523 | } |
| 524 | |
| 525 | #define _mm256_slli_si256(a, count) __extension__ ({ \ |
| 526 | __m256i __a = (a); \ |
| 527 | (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); }) |
| 528 | |
| 529 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 530 | _mm256_slli_epi16(__m256i __a, int __count) |
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_psllwi256((__v16hi)__a, __count); |
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_sll_epi16(__m256i __a, __m128i __count) |
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_psllw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 539 | } |
| 540 | |
| 541 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 542 | _mm256_slli_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 543 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 544 | return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 545 | } |
| 546 | |
| 547 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 548 | _mm256_sll_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 549 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 550 | return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 551 | } |
| 552 | |
| 553 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 554 | _mm256_slli_epi64(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 555 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 556 | return __builtin_ia32_psllqi256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 557 | } |
| 558 | |
| 559 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 560 | _mm256_sll_epi64(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 561 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 562 | return __builtin_ia32_psllq256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 563 | } |
| 564 | |
| 565 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 566 | _mm256_srai_epi16(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 567 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 568 | return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 569 | } |
| 570 | |
| 571 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 572 | _mm256_sra_epi16(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 573 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 574 | return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 575 | } |
| 576 | |
| 577 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 578 | _mm256_srai_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 579 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 580 | return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 581 | } |
| 582 | |
| 583 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 584 | _mm256_sra_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 585 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 586 | return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 587 | } |
| 588 | |
| 589 | #define _mm256_srli_si256(a, count) __extension__ ({ \ |
| 590 | __m256i __a = (a); \ |
| 591 | (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); }) |
| 592 | |
| 593 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 594 | _mm256_srli_epi16(__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_psrlwi256((__v16hi)__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_srl_epi16(__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_psrlw256((__v16hi)__a, (__v8hi)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 603 | } |
| 604 | |
| 605 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 606 | _mm256_srli_epi32(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 607 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 608 | return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 609 | } |
| 610 | |
| 611 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 612 | _mm256_srl_epi32(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 613 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 614 | return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 615 | } |
| 616 | |
| 617 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 618 | _mm256_srli_epi64(__m256i __a, int __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 619 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 620 | return __builtin_ia32_psrlqi256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 621 | } |
| 622 | |
| 623 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 624 | _mm256_srl_epi64(__m256i __a, __m128i __count) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 625 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 626 | return __builtin_ia32_psrlq256(__a, __count); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 627 | } |
| 628 | |
| 629 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 630 | _mm256_sub_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 631 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 632 | return (__m256i)((__v32qi)__a - (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 633 | } |
| 634 | |
| 635 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 636 | _mm256_sub_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 637 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 638 | return (__m256i)((__v16hi)__a - (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 639 | } |
| 640 | |
| 641 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 642 | _mm256_sub_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 643 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 644 | return (__m256i)((__v8si)__a - (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 645 | } |
| 646 | |
| 647 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 648 | _mm256_sub_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 649 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 650 | return __a - __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 651 | } |
| 652 | |
| 653 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 654 | _mm256_subs_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 655 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 656 | return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 657 | } |
| 658 | |
| 659 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 660 | _mm256_subs_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 661 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 662 | return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 663 | } |
| 664 | |
| 665 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 666 | _mm256_subs_epu8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 667 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 668 | return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 669 | } |
| 670 | |
| 671 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 672 | _mm256_subs_epu16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 673 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 674 | return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 675 | } |
| 676 | |
| 677 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 678 | _mm256_unpackhi_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 679 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 680 | 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] | 681 | } |
| 682 | |
| 683 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 684 | _mm256_unpackhi_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 685 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 686 | 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] | 687 | } |
| 688 | |
| 689 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 690 | _mm256_unpackhi_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 691 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 692 | 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] | 693 | } |
| 694 | |
| 695 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 696 | _mm256_unpackhi_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 697 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 698 | return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 699 | } |
| 700 | |
| 701 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 702 | _mm256_unpacklo_epi8(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 703 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 704 | 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] | 705 | } |
| 706 | |
| 707 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 708 | _mm256_unpacklo_epi16(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 709 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 710 | 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] | 711 | } |
| 712 | |
| 713 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 714 | _mm256_unpacklo_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 715 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 716 | 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] | 717 | } |
| 718 | |
| 719 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 720 | _mm256_unpacklo_epi64(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 721 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 722 | return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 723 | } |
| 724 | |
| 725 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 726 | _mm256_xor_si256(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 727 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 728 | return __a ^ __b; |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 729 | } |
| 730 | |
| 731 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 732 | _mm256_stream_load_si256(__m256i *__V) |
| 733 | { |
| 734 | return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V); |
| 735 | } |
| 736 | |
| 737 | static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 738 | _mm_broadcastss_ps(__m128 __X) |
| 739 | { |
| 740 | return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X); |
| 741 | } |
| 742 | |
| 743 | static __inline__ __m256 __attribute__((__always_inline__, __nodebug__)) |
| 744 | _mm256_broadcastss_ps(__m128 __X) |
| 745 | { |
| 746 | return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X); |
| 747 | } |
| 748 | |
| 749 | static __inline__ __m256d __attribute__((__always_inline__, __nodebug__)) |
| 750 | _mm256_broadcastsd_pd(__m128d __X) |
| 751 | { |
| 752 | return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X); |
| 753 | } |
| 754 | |
| 755 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 756 | _mm256_broadcastsi128_si256(__m128i __X) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 757 | { |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 758 | return (__m256i)__builtin_ia32_vbroadcastsi256(__X); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 759 | } |
| 760 | |
| 761 | #define _mm_blend_epi32(V1, V2, M) __extension__ ({ \ |
| 762 | __m128i __V1 = (V1); \ |
| 763 | __m128i __V2 = (V2); \ |
| 764 | (__m128i)__builtin_ia32_pblendd128((__v4si)__V1, (__v4si)__V2, (M)); }) |
| 765 | |
| 766 | #define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \ |
| 767 | __m256i __V1 = (V1); \ |
| 768 | __m256i __V2 = (V2); \ |
| 769 | (__m256i)__builtin_ia32_pblendd256((__v8si)__V1, (__v8si)__V2, (M)); }) |
| 770 | |
| 771 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 772 | _mm256_broadcastb_epi8(__m128i __X) |
| 773 | { |
| 774 | return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X); |
| 775 | } |
| 776 | |
| 777 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 778 | _mm256_broadcastw_epi16(__m128i __X) |
| 779 | { |
| 780 | return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X); |
| 781 | } |
| 782 | |
| 783 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 784 | _mm256_broadcastd_epi32(__m128i __X) |
| 785 | { |
| 786 | return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X); |
| 787 | } |
| 788 | |
| 789 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 790 | _mm256_broadcastq_epi64(__m128i __X) |
| 791 | { |
| 792 | return (__m256i)__builtin_ia32_pbroadcastq256(__X); |
| 793 | } |
| 794 | |
| 795 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 796 | _mm_broadcastb_epi8(__m128i __X) |
| 797 | { |
| 798 | return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X); |
| 799 | } |
| 800 | |
| 801 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 802 | _mm_broadcastw_epi16(__m128i __X) |
| 803 | { |
| 804 | return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X); |
| 805 | } |
| 806 | |
| 807 | |
| 808 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 809 | _mm_broadcastd_epi32(__m128i __X) |
| 810 | { |
| 811 | return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X); |
| 812 | } |
| 813 | |
| 814 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 815 | _mm_broadcastq_epi64(__m128i __X) |
| 816 | { |
| 817 | return (__m128i)__builtin_ia32_pbroadcastq128(__X); |
| 818 | } |
| 819 | |
| 820 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 821 | _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 822 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 823 | return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 824 | } |
| 825 | |
| 826 | #define _mm256_permute4x64_pd(V, M) __extension__ ({ \ |
| 827 | __m256d __V = (V); \ |
| 828 | (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \ |
| 829 | (M) & 0x3, ((M) & 0xc) >> 2, \ |
| 830 | ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); }) |
| 831 | |
| 832 | static __inline__ __m256 __attribute__((__always_inline__, __nodebug__)) |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 833 | _mm256_permutevar8x32_ps(__m256 __a, __m256 __b) |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 834 | { |
Stephen Hines | c6ee7df | 2013-04-02 18:41:57 -0700 | [diff] [blame] | 835 | return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b); |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 836 | } |
| 837 | |
| 838 | #define _mm256_permute4x64_epi64(V, M) __extension__ ({ \ |
| 839 | __m256i __V = (V); \ |
| 840 | (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \ |
| 841 | (M) & 0x3, ((M) & 0xc) >> 2, \ |
| 842 | ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); }) |
| 843 | |
| 844 | #define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \ |
| 845 | __m256i __V1 = (V1); \ |
| 846 | __m256i __V2 = (V2); \ |
| 847 | (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); }) |
| 848 | |
| 849 | #define _mm256_extracti128_si256(A, O) __extension__ ({ \ |
| 850 | __m256i __A = (A); \ |
| 851 | (__m128i)__builtin_ia32_extract128i256(__A, (O)); }) |
| 852 | |
| 853 | #define _mm256_inserti128_si256(V1, V2, O) __extension__ ({ \ |
| 854 | __m256i __V1 = (V1); \ |
| 855 | __m128i __V2 = (V2); \ |
| 856 | (__m256i)__builtin_ia32_insert128i256(__V1, __V2, (O)); }) |
| 857 | |
| 858 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 859 | _mm256_maskload_epi32(int const *__X, __m256i __M) |
| 860 | { |
| 861 | return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M); |
| 862 | } |
| 863 | |
| 864 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 865 | _mm256_maskload_epi64(long long const *__X, __m256i __M) |
| 866 | { |
| 867 | return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M); |
| 868 | } |
| 869 | |
| 870 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 871 | _mm_maskload_epi32(int const *__X, __m128i __M) |
| 872 | { |
| 873 | return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M); |
| 874 | } |
| 875 | |
| 876 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 877 | _mm_maskload_epi64(long long const *__X, __m128i __M) |
| 878 | { |
| 879 | return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M); |
| 880 | } |
| 881 | |
| 882 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 883 | _mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y) |
| 884 | { |
| 885 | __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y); |
| 886 | } |
| 887 | |
| 888 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 889 | _mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y) |
| 890 | { |
| 891 | __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y); |
| 892 | } |
| 893 | |
| 894 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 895 | _mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y) |
| 896 | { |
| 897 | __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y); |
| 898 | } |
| 899 | |
| 900 | static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 901 | _mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y) |
| 902 | { |
| 903 | __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y); |
| 904 | } |
| 905 | |
| 906 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 907 | _mm256_sllv_epi32(__m256i __X, __m256i __Y) |
| 908 | { |
| 909 | return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y); |
| 910 | } |
| 911 | |
| 912 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 913 | _mm_sllv_epi32(__m128i __X, __m128i __Y) |
| 914 | { |
| 915 | return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y); |
| 916 | } |
| 917 | |
| 918 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 919 | _mm256_sllv_epi64(__m256i __X, __m256i __Y) |
| 920 | { |
| 921 | return (__m256i)__builtin_ia32_psllv4di(__X, __Y); |
| 922 | } |
| 923 | |
| 924 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 925 | _mm_sllv_epi64(__m128i __X, __m128i __Y) |
| 926 | { |
| 927 | return (__m128i)__builtin_ia32_psllv2di(__X, __Y); |
| 928 | } |
| 929 | |
| 930 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 931 | _mm256_srav_epi32(__m256i __X, __m256i __Y) |
| 932 | { |
| 933 | return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y); |
| 934 | } |
| 935 | |
| 936 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 937 | _mm_srav_epi32(__m128i __X, __m128i __Y) |
| 938 | { |
| 939 | return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y); |
| 940 | } |
| 941 | |
| 942 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 943 | _mm256_srlv_epi32(__m256i __X, __m256i __Y) |
| 944 | { |
| 945 | return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y); |
| 946 | } |
| 947 | |
| 948 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 949 | _mm_srlv_epi32(__m128i __X, __m128i __Y) |
| 950 | { |
| 951 | return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y); |
| 952 | } |
| 953 | |
| 954 | static __inline__ __m256i __attribute__((__always_inline__, __nodebug__)) |
| 955 | _mm256_srlv_epi64(__m256i __X, __m256i __Y) |
| 956 | { |
| 957 | return (__m256i)__builtin_ia32_psrlv4di(__X, __Y); |
| 958 | } |
| 959 | |
| 960 | static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 961 | _mm_srlv_epi64(__m128i __X, __m128i __Y) |
| 962 | { |
| 963 | return (__m128i)__builtin_ia32_psrlv2di(__X, __Y); |
| 964 | } |
| 965 | |
| 966 | #define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 967 | __m128d __a = (a); \ |
| 968 | double const *__m = (m); \ |
| 969 | __m128i __i = (i); \ |
| 970 | __m128d __mask = (mask); \ |
| 971 | (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \ |
| 972 | (__v4si)__i, (__v2df)__mask, (s)); }) |
| 973 | |
| 974 | #define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 975 | __m256d __a = (a); \ |
| 976 | double const *__m = (m); \ |
| 977 | __m128i __i = (i); \ |
| 978 | __m256d __mask = (mask); \ |
| 979 | (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \ |
| 980 | (__v4si)__i, (__v4df)__mask, (s)); }) |
| 981 | |
| 982 | #define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 983 | __m128d __a = (a); \ |
| 984 | double const *__m = (m); \ |
| 985 | __m128i __i = (i); \ |
| 986 | __m128d __mask = (mask); \ |
| 987 | (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \ |
| 988 | (__v2di)__i, (__v2df)__mask, (s)); }) |
| 989 | |
| 990 | #define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \ |
| 991 | __m256d __a = (a); \ |
| 992 | double const *__m = (m); \ |
| 993 | __m256i __i = (i); \ |
| 994 | __m256d __mask = (mask); \ |
| 995 | (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \ |
| 996 | (__v4di)__i, (__v4df)__mask, (s)); }) |
| 997 | |
| 998 | #define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 999 | __m128 __a = (a); \ |
| 1000 | float const *__m = (m); \ |
| 1001 | __m128i __i = (i); \ |
| 1002 | __m128 __mask = (mask); \ |
| 1003 | (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \ |
| 1004 | (__v4si)__i, (__v4sf)__mask, (s)); }) |
| 1005 | |
| 1006 | #define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1007 | __m256 __a = (a); \ |
| 1008 | float const *__m = (m); \ |
| 1009 | __m256i __i = (i); \ |
| 1010 | __m256 __mask = (mask); \ |
| 1011 | (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \ |
| 1012 | (__v8si)__i, (__v8sf)__mask, (s)); }) |
| 1013 | |
| 1014 | #define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1015 | __m128 __a = (a); \ |
| 1016 | float const *__m = (m); \ |
| 1017 | __m128i __i = (i); \ |
| 1018 | __m128 __mask = (mask); \ |
| 1019 | (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \ |
| 1020 | (__v2di)__i, (__v4sf)__mask, (s)); }) |
| 1021 | |
| 1022 | #define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \ |
| 1023 | __m128 __a = (a); \ |
| 1024 | float const *__m = (m); \ |
| 1025 | __m256i __i = (i); \ |
| 1026 | __m128 __mask = (mask); \ |
| 1027 | (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \ |
| 1028 | (__v4di)__i, (__v4sf)__mask, (s)); }) |
| 1029 | |
| 1030 | #define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1031 | __m128i __a = (a); \ |
| 1032 | int const *__m = (m); \ |
| 1033 | __m128i __i = (i); \ |
| 1034 | __m128i __mask = (mask); \ |
| 1035 | (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \ |
| 1036 | (__v4si)__i, (__v4si)__mask, (s)); }) |
| 1037 | |
| 1038 | #define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1039 | __m256i __a = (a); \ |
| 1040 | int const *__m = (m); \ |
| 1041 | __m256i __i = (i); \ |
| 1042 | __m256i __mask = (mask); \ |
| 1043 | (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \ |
| 1044 | (__v8si)__i, (__v8si)__mask, (s)); }) |
| 1045 | |
| 1046 | #define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1047 | __m128i __a = (a); \ |
| 1048 | int const *__m = (m); \ |
| 1049 | __m128i __i = (i); \ |
| 1050 | __m128i __mask = (mask); \ |
| 1051 | (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \ |
| 1052 | (__v2di)__i, (__v4si)__mask, (s)); }) |
| 1053 | |
| 1054 | #define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \ |
| 1055 | __m128i __a = (a); \ |
| 1056 | int const *__m = (m); \ |
| 1057 | __m256i __i = (i); \ |
| 1058 | __m128i __mask = (mask); \ |
| 1059 | (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \ |
| 1060 | (__v4di)__i, (__v4si)__mask, (s)); }) |
| 1061 | |
| 1062 | #define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1063 | __m128i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1064 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1065 | __m128i __i = (i); \ |
| 1066 | __m128i __mask = (mask); \ |
| 1067 | (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \ |
| 1068 | (__v4si)__i, (__v2di)__mask, (s)); }) |
| 1069 | |
| 1070 | #define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1071 | __m256i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1072 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1073 | __m128i __i = (i); \ |
| 1074 | __m256i __mask = (mask); \ |
| 1075 | (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \ |
| 1076 | (__v4si)__i, (__v4di)__mask, (s)); }) |
| 1077 | |
| 1078 | #define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1079 | __m128i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1080 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1081 | __m128i __i = (i); \ |
| 1082 | __m128i __mask = (mask); \ |
| 1083 | (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \ |
| 1084 | (__v2di)__i, (__v2di)__mask, (s)); }) |
| 1085 | |
| 1086 | #define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \ |
| 1087 | __m256i __a = (a); \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1088 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1089 | __m256i __i = (i); \ |
| 1090 | __m256i __mask = (mask); \ |
| 1091 | (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \ |
| 1092 | (__v4di)__i, (__v4di)__mask, (s)); }) |
| 1093 | |
| 1094 | #define _mm_i32gather_pd(m, i, s) __extension__ ({ \ |
| 1095 | double const *__m = (m); \ |
| 1096 | __m128i __i = (i); \ |
| 1097 | (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \ |
| 1098 | (const __v2df *)__m, (__v4si)__i, \ |
| 1099 | (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); }) |
| 1100 | |
| 1101 | #define _mm256_i32gather_pd(m, i, s) __extension__ ({ \ |
| 1102 | double const *__m = (m); \ |
| 1103 | __m128i __i = (i); \ |
| 1104 | (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \ |
| 1105 | (const __v4df *)__m, (__v4si)__i, \ |
| 1106 | (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); }) |
| 1107 | |
| 1108 | #define _mm_i64gather_pd(m, i, s) __extension__ ({ \ |
| 1109 | double const *__m = (m); \ |
| 1110 | __m128i __i = (i); \ |
| 1111 | (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \ |
| 1112 | (const __v2df *)__m, (__v2di)__i, \ |
| 1113 | (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); }) |
| 1114 | |
| 1115 | #define _mm256_i64gather_pd(m, i, s) __extension__ ({ \ |
| 1116 | double const *__m = (m); \ |
| 1117 | __m256i __i = (i); \ |
| 1118 | (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \ |
| 1119 | (const __v4df *)__m, (__v4di)__i, \ |
| 1120 | (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); }) |
| 1121 | |
| 1122 | #define _mm_i32gather_ps(m, i, s) __extension__ ({ \ |
| 1123 | float const *__m = (m); \ |
| 1124 | __m128i __i = (i); \ |
| 1125 | (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \ |
| 1126 | (const __v4sf *)__m, (__v4si)__i, \ |
| 1127 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1128 | |
| 1129 | #define _mm256_i32gather_ps(m, i, s) __extension__ ({ \ |
| 1130 | float const *__m = (m); \ |
| 1131 | __m256i __i = (i); \ |
| 1132 | (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \ |
| 1133 | (const __v8sf *)__m, (__v8si)__i, \ |
| 1134 | (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); }) |
| 1135 | |
| 1136 | #define _mm_i64gather_ps(m, i, s) __extension__ ({ \ |
| 1137 | float const *__m = (m); \ |
| 1138 | __m128i __i = (i); \ |
| 1139 | (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \ |
| 1140 | (const __v4sf *)__m, (__v2di)__i, \ |
| 1141 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1142 | |
| 1143 | #define _mm256_i64gather_ps(m, i, s) __extension__ ({ \ |
| 1144 | float const *__m = (m); \ |
| 1145 | __m256i __i = (i); \ |
| 1146 | (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \ |
| 1147 | (const __v4sf *)__m, (__v4di)__i, \ |
| 1148 | (__v4sf)_mm_set1_ps((float)(int)-1), (s)); }) |
| 1149 | |
| 1150 | #define _mm_i32gather_epi32(m, i, s) __extension__ ({ \ |
| 1151 | int const *__m = (m); \ |
| 1152 | __m128i __i = (i); \ |
| 1153 | (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \ |
| 1154 | (const __v4si *)__m, (__v4si)__i, \ |
| 1155 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1156 | |
| 1157 | #define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \ |
| 1158 | int const *__m = (m); \ |
| 1159 | __m256i __i = (i); \ |
| 1160 | (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \ |
| 1161 | (const __v8si *)__m, (__v8si)__i, \ |
| 1162 | (__v8si)_mm256_set1_epi32(-1), (s)); }) |
| 1163 | |
| 1164 | #define _mm_i64gather_epi32(m, i, s) __extension__ ({ \ |
| 1165 | int const *__m = (m); \ |
| 1166 | __m128i __i = (i); \ |
| 1167 | (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \ |
| 1168 | (const __v4si *)__m, (__v2di)__i, \ |
| 1169 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1170 | |
| 1171 | #define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \ |
| 1172 | int const *__m = (m); \ |
| 1173 | __m256i __i = (i); \ |
| 1174 | (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \ |
| 1175 | (const __v4si *)__m, (__v4di)__i, \ |
| 1176 | (__v4si)_mm_set1_epi32(-1), (s)); }) |
| 1177 | |
| 1178 | #define _mm_i32gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1179 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1180 | __m128i __i = (i); \ |
| 1181 | (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \ |
| 1182 | (const __v2di *)__m, (__v4si)__i, \ |
| 1183 | (__v2di)_mm_set1_epi64x(-1), (s)); }) |
| 1184 | |
| 1185 | #define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1186 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1187 | __m128i __i = (i); \ |
| 1188 | (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \ |
| 1189 | (const __v4di *)__m, (__v4si)__i, \ |
| 1190 | (__v4di)_mm256_set1_epi64x(-1), (s)); }) |
| 1191 | |
| 1192 | #define _mm_i64gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1193 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1194 | __m128i __i = (i); \ |
| 1195 | (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \ |
| 1196 | (const __v2di *)__m, (__v2di)__i, \ |
| 1197 | (__v2di)_mm_set1_epi64x(-1), (s)); }) |
| 1198 | |
| 1199 | #define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \ |
Stephen Hines | 803c468 | 2014-02-19 18:30:17 -0800 | [diff] [blame^] | 1200 | long long const *__m = (m); \ |
Ying Wang | 6099914 | 2013-01-07 13:59:36 -0800 | [diff] [blame] | 1201 | __m256i __i = (i); \ |
| 1202 | (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \ |
| 1203 | (const __v4di *)__m, (__v4di)__i, \ |
| 1204 | (__v4di)_mm256_set1_epi64x(-1), (s)); }) |
Stephen Hines | 996e4dc | 2013-08-13 01:04:14 -0700 | [diff] [blame] | 1205 | |
| 1206 | #endif /* __AVX2INTRIN_H */ |