| 1 | /* Copyright (C) 2003-2021 Free Software Foundation, Inc.
|
|---|
| 2 |
|
|---|
| 3 | This file is part of GCC.
|
|---|
| 4 |
|
|---|
| 5 | GCC is free software; you can redistribute it and/or modify
|
|---|
| 6 | it under the terms of the GNU General Public License as published by
|
|---|
| 7 | the Free Software Foundation; either version 3, or (at your option)
|
|---|
| 8 | any later version.
|
|---|
| 9 |
|
|---|
| 10 | GCC is distributed in the hope that it will be useful,
|
|---|
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|---|
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|---|
| 13 | GNU General Public License for more details.
|
|---|
| 14 |
|
|---|
| 15 | Under Section 7 of GPL version 3, you are granted additional
|
|---|
| 16 | permissions described in the GCC Runtime Library Exception, version
|
|---|
| 17 | 3.1, as published by the Free Software Foundation.
|
|---|
| 18 |
|
|---|
| 19 | You should have received a copy of the GNU General Public License and
|
|---|
| 20 | a copy of the GCC Runtime Library Exception along with this program;
|
|---|
| 21 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|---|
| 22 | <http://www.gnu.org/licenses/>. */
|
|---|
| 23 |
|
|---|
| 24 | /* Implemented from the specification included in the Intel C++ Compiler
|
|---|
| 25 | User Guide and Reference, version 9.0. */
|
|---|
| 26 |
|
|---|
| 27 | #ifndef _EMMINTRIN_H_INCLUDED
|
|---|
| 28 | #define _EMMINTRIN_H_INCLUDED
|
|---|
| 29 |
|
|---|
| 30 | /* We need definitions from the SSE header files*/
|
|---|
| 31 | #include <xmmintrin.h>
|
|---|
| 32 |
|
|---|
| 33 | #ifndef __SSE2__
|
|---|
| 34 | #pragma GCC push_options
|
|---|
| 35 | #pragma GCC target("sse2")
|
|---|
| 36 | #define __DISABLE_SSE2__
|
|---|
| 37 | #endif /* __SSE2__ */
|
|---|
| 38 |
|
|---|
| 39 | /* SSE2 */
|
|---|
| 40 | typedef double __v2df __attribute__ ((__vector_size__ (16)));
|
|---|
| 41 | typedef long long __v2di __attribute__ ((__vector_size__ (16)));
|
|---|
| 42 | typedef unsigned long long __v2du __attribute__ ((__vector_size__ (16)));
|
|---|
| 43 | typedef int __v4si __attribute__ ((__vector_size__ (16)));
|
|---|
| 44 | typedef unsigned int __v4su __attribute__ ((__vector_size__ (16)));
|
|---|
| 45 | typedef short __v8hi __attribute__ ((__vector_size__ (16)));
|
|---|
| 46 | typedef unsigned short __v8hu __attribute__ ((__vector_size__ (16)));
|
|---|
| 47 | typedef char __v16qi __attribute__ ((__vector_size__ (16)));
|
|---|
| 48 | typedef signed char __v16qs __attribute__ ((__vector_size__ (16)));
|
|---|
| 49 | typedef unsigned char __v16qu __attribute__ ((__vector_size__ (16)));
|
|---|
| 50 |
|
|---|
| 51 | /* The Intel API is flexible enough that we must allow aliasing with other
|
|---|
| 52 | vector types, and their scalar components. */
|
|---|
| 53 | typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
|
|---|
| 54 | typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
|
|---|
| 55 |
|
|---|
| 56 | /* Unaligned version of the same types. */
|
|---|
| 57 | typedef long long __m128i_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
|
|---|
| 58 | typedef double __m128d_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
|
|---|
| 59 |
|
|---|
| 60 | /* Create a selector for use with the SHUFPD instruction. */
|
|---|
| 61 | #define _MM_SHUFFLE2(fp1,fp0) \
|
|---|
| 62 | (((fp1) << 1) | (fp0))
|
|---|
| 63 |
|
|---|
| 64 | /* Create a vector with element 0 as F and the rest zero. */
|
|---|
| 65 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 66 | _mm_set_sd (double __F)
|
|---|
| 67 | {
|
|---|
| 68 | return __extension__ (__m128d){ __F, 0.0 };
|
|---|
| 69 | }
|
|---|
| 70 |
|
|---|
| 71 | /* Create a vector with both elements equal to F. */
|
|---|
| 72 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 73 | _mm_set1_pd (double __F)
|
|---|
| 74 | {
|
|---|
| 75 | return __extension__ (__m128d){ __F, __F };
|
|---|
| 76 | }
|
|---|
| 77 |
|
|---|
| 78 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 79 | _mm_set_pd1 (double __F)
|
|---|
| 80 | {
|
|---|
| 81 | return _mm_set1_pd (__F);
|
|---|
| 82 | }
|
|---|
| 83 |
|
|---|
| 84 | /* Create a vector with the lower value X and upper value W. */
|
|---|
| 85 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 86 | _mm_set_pd (double __W, double __X)
|
|---|
| 87 | {
|
|---|
| 88 | return __extension__ (__m128d){ __X, __W };
|
|---|
| 89 | }
|
|---|
| 90 |
|
|---|
| 91 | /* Create a vector with the lower value W and upper value X. */
|
|---|
| 92 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 93 | _mm_setr_pd (double __W, double __X)
|
|---|
| 94 | {
|
|---|
| 95 | return __extension__ (__m128d){ __W, __X };
|
|---|
| 96 | }
|
|---|
| 97 |
|
|---|
| 98 | /* Create an undefined vector. */
|
|---|
| 99 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 100 | _mm_undefined_pd (void)
|
|---|
| 101 | {
|
|---|
| 102 | __m128d __Y = __Y;
|
|---|
| 103 | return __Y;
|
|---|
| 104 | }
|
|---|
| 105 |
|
|---|
| 106 | /* Create a vector of zeros. */
|
|---|
| 107 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 108 | _mm_setzero_pd (void)
|
|---|
| 109 | {
|
|---|
| 110 | return __extension__ (__m128d){ 0.0, 0.0 };
|
|---|
| 111 | }
|
|---|
| 112 |
|
|---|
| 113 | /* Sets the low DPFP value of A from the low value of B. */
|
|---|
| 114 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 115 | _mm_move_sd (__m128d __A, __m128d __B)
|
|---|
| 116 | {
|
|---|
| 117 | return __extension__ (__m128d) __builtin_shuffle ((__v2df)__A, (__v2df)__B, (__v2di){2, 1});
|
|---|
| 118 | }
|
|---|
| 119 |
|
|---|
| 120 | /* Load two DPFP values from P. The address must be 16-byte aligned. */
|
|---|
| 121 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 122 | _mm_load_pd (double const *__P)
|
|---|
| 123 | {
|
|---|
| 124 | return *(__m128d *)__P;
|
|---|
| 125 | }
|
|---|
| 126 |
|
|---|
| 127 | /* Load two DPFP values from P. The address need not be 16-byte aligned. */
|
|---|
| 128 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 129 | _mm_loadu_pd (double const *__P)
|
|---|
| 130 | {
|
|---|
| 131 | return *(__m128d_u *)__P;
|
|---|
| 132 | }
|
|---|
| 133 |
|
|---|
| 134 | /* Create a vector with all two elements equal to *P. */
|
|---|
| 135 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 136 | _mm_load1_pd (double const *__P)
|
|---|
| 137 | {
|
|---|
| 138 | return _mm_set1_pd (*__P);
|
|---|
| 139 | }
|
|---|
| 140 |
|
|---|
| 141 | /* Create a vector with element 0 as *P and the rest zero. */
|
|---|
| 142 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 143 | _mm_load_sd (double const *__P)
|
|---|
| 144 | {
|
|---|
| 145 | return _mm_set_sd (*__P);
|
|---|
| 146 | }
|
|---|
| 147 |
|
|---|
| 148 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 149 | _mm_load_pd1 (double const *__P)
|
|---|
| 150 | {
|
|---|
| 151 | return _mm_load1_pd (__P);
|
|---|
| 152 | }
|
|---|
| 153 |
|
|---|
| 154 | /* Load two DPFP values in reverse order. The address must be aligned. */
|
|---|
| 155 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 156 | _mm_loadr_pd (double const *__P)
|
|---|
| 157 | {
|
|---|
| 158 | __m128d __tmp = _mm_load_pd (__P);
|
|---|
| 159 | return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
|
|---|
| 160 | }
|
|---|
| 161 |
|
|---|
| 162 | /* Store two DPFP values. The address must be 16-byte aligned. */
|
|---|
| 163 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 164 | _mm_store_pd (double *__P, __m128d __A)
|
|---|
| 165 | {
|
|---|
| 166 | *(__m128d *)__P = __A;
|
|---|
| 167 | }
|
|---|
| 168 |
|
|---|
| 169 | /* Store two DPFP values. The address need not be 16-byte aligned. */
|
|---|
| 170 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 171 | _mm_storeu_pd (double *__P, __m128d __A)
|
|---|
| 172 | {
|
|---|
| 173 | *(__m128d_u *)__P = __A;
|
|---|
| 174 | }
|
|---|
| 175 |
|
|---|
| 176 | /* Stores the lower DPFP value. */
|
|---|
| 177 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 178 | _mm_store_sd (double *__P, __m128d __A)
|
|---|
| 179 | {
|
|---|
| 180 | *__P = ((__v2df)__A)[0];
|
|---|
| 181 | }
|
|---|
| 182 |
|
|---|
| 183 | extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 184 | _mm_cvtsd_f64 (__m128d __A)
|
|---|
| 185 | {
|
|---|
| 186 | return ((__v2df)__A)[0];
|
|---|
| 187 | }
|
|---|
| 188 |
|
|---|
| 189 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 190 | _mm_storel_pd (double *__P, __m128d __A)
|
|---|
| 191 | {
|
|---|
| 192 | _mm_store_sd (__P, __A);
|
|---|
| 193 | }
|
|---|
| 194 |
|
|---|
| 195 | /* Stores the upper DPFP value. */
|
|---|
| 196 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 197 | _mm_storeh_pd (double *__P, __m128d __A)
|
|---|
| 198 | {
|
|---|
| 199 | *__P = ((__v2df)__A)[1];
|
|---|
| 200 | }
|
|---|
| 201 |
|
|---|
| 202 | /* Store the lower DPFP value across two words.
|
|---|
| 203 | The address must be 16-byte aligned. */
|
|---|
| 204 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 205 | _mm_store1_pd (double *__P, __m128d __A)
|
|---|
| 206 | {
|
|---|
| 207 | _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
|
|---|
| 208 | }
|
|---|
| 209 |
|
|---|
| 210 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 211 | _mm_store_pd1 (double *__P, __m128d __A)
|
|---|
| 212 | {
|
|---|
| 213 | _mm_store1_pd (__P, __A);
|
|---|
| 214 | }
|
|---|
| 215 |
|
|---|
| 216 | /* Store two DPFP values in reverse order. The address must be aligned. */
|
|---|
| 217 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 218 | _mm_storer_pd (double *__P, __m128d __A)
|
|---|
| 219 | {
|
|---|
| 220 | _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
|
|---|
| 221 | }
|
|---|
| 222 |
|
|---|
| 223 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 224 | _mm_cvtsi128_si32 (__m128i __A)
|
|---|
| 225 | {
|
|---|
| 226 | return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
|
|---|
| 227 | }
|
|---|
| 228 |
|
|---|
| 229 | #ifdef __x86_64__
|
|---|
| 230 | /* Intel intrinsic. */
|
|---|
| 231 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 232 | _mm_cvtsi128_si64 (__m128i __A)
|
|---|
| 233 | {
|
|---|
| 234 | return ((__v2di)__A)[0];
|
|---|
| 235 | }
|
|---|
| 236 |
|
|---|
| 237 | /* Microsoft intrinsic. */
|
|---|
| 238 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 239 | _mm_cvtsi128_si64x (__m128i __A)
|
|---|
| 240 | {
|
|---|
| 241 | return ((__v2di)__A)[0];
|
|---|
| 242 | }
|
|---|
| 243 | #endif
|
|---|
| 244 |
|
|---|
| 245 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 246 | _mm_add_pd (__m128d __A, __m128d __B)
|
|---|
| 247 | {
|
|---|
| 248 | return (__m128d) ((__v2df)__A + (__v2df)__B);
|
|---|
| 249 | }
|
|---|
| 250 |
|
|---|
| 251 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 252 | _mm_add_sd (__m128d __A, __m128d __B)
|
|---|
| 253 | {
|
|---|
| 254 | return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 255 | }
|
|---|
| 256 |
|
|---|
| 257 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 258 | _mm_sub_pd (__m128d __A, __m128d __B)
|
|---|
| 259 | {
|
|---|
| 260 | return (__m128d) ((__v2df)__A - (__v2df)__B);
|
|---|
| 261 | }
|
|---|
| 262 |
|
|---|
| 263 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 264 | _mm_sub_sd (__m128d __A, __m128d __B)
|
|---|
| 265 | {
|
|---|
| 266 | return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 267 | }
|
|---|
| 268 |
|
|---|
| 269 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 270 | _mm_mul_pd (__m128d __A, __m128d __B)
|
|---|
| 271 | {
|
|---|
| 272 | return (__m128d) ((__v2df)__A * (__v2df)__B);
|
|---|
| 273 | }
|
|---|
| 274 |
|
|---|
| 275 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 276 | _mm_mul_sd (__m128d __A, __m128d __B)
|
|---|
| 277 | {
|
|---|
| 278 | return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 279 | }
|
|---|
| 280 |
|
|---|
| 281 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 282 | _mm_div_pd (__m128d __A, __m128d __B)
|
|---|
| 283 | {
|
|---|
| 284 | return (__m128d) ((__v2df)__A / (__v2df)__B);
|
|---|
| 285 | }
|
|---|
| 286 |
|
|---|
| 287 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 288 | _mm_div_sd (__m128d __A, __m128d __B)
|
|---|
| 289 | {
|
|---|
| 290 | return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 291 | }
|
|---|
| 292 |
|
|---|
| 293 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 294 | _mm_sqrt_pd (__m128d __A)
|
|---|
| 295 | {
|
|---|
| 296 | return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
|
|---|
| 297 | }
|
|---|
| 298 |
|
|---|
| 299 | /* Return pair {sqrt (B[0]), A[1]}. */
|
|---|
| 300 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 301 | _mm_sqrt_sd (__m128d __A, __m128d __B)
|
|---|
| 302 | {
|
|---|
| 303 | __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 304 | return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
|
|---|
| 305 | }
|
|---|
| 306 |
|
|---|
| 307 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 308 | _mm_min_pd (__m128d __A, __m128d __B)
|
|---|
| 309 | {
|
|---|
| 310 | return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 311 | }
|
|---|
| 312 |
|
|---|
| 313 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 314 | _mm_min_sd (__m128d __A, __m128d __B)
|
|---|
| 315 | {
|
|---|
| 316 | return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 317 | }
|
|---|
| 318 |
|
|---|
| 319 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 320 | _mm_max_pd (__m128d __A, __m128d __B)
|
|---|
| 321 | {
|
|---|
| 322 | return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 323 | }
|
|---|
| 324 |
|
|---|
| 325 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 326 | _mm_max_sd (__m128d __A, __m128d __B)
|
|---|
| 327 | {
|
|---|
| 328 | return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 329 | }
|
|---|
| 330 |
|
|---|
| 331 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 332 | _mm_and_pd (__m128d __A, __m128d __B)
|
|---|
| 333 | {
|
|---|
| 334 | return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 335 | }
|
|---|
| 336 |
|
|---|
| 337 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 338 | _mm_andnot_pd (__m128d __A, __m128d __B)
|
|---|
| 339 | {
|
|---|
| 340 | return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 341 | }
|
|---|
| 342 |
|
|---|
| 343 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 344 | _mm_or_pd (__m128d __A, __m128d __B)
|
|---|
| 345 | {
|
|---|
| 346 | return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 347 | }
|
|---|
| 348 |
|
|---|
| 349 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 350 | _mm_xor_pd (__m128d __A, __m128d __B)
|
|---|
| 351 | {
|
|---|
| 352 | return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 353 | }
|
|---|
| 354 |
|
|---|
| 355 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 356 | _mm_cmpeq_pd (__m128d __A, __m128d __B)
|
|---|
| 357 | {
|
|---|
| 358 | return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 359 | }
|
|---|
| 360 |
|
|---|
| 361 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 362 | _mm_cmplt_pd (__m128d __A, __m128d __B)
|
|---|
| 363 | {
|
|---|
| 364 | return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 365 | }
|
|---|
| 366 |
|
|---|
| 367 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 368 | _mm_cmple_pd (__m128d __A, __m128d __B)
|
|---|
| 369 | {
|
|---|
| 370 | return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
|
|---|
| 371 | }
|
|---|
| 372 |
|
|---|
| 373 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 374 | _mm_cmpgt_pd (__m128d __A, __m128d __B)
|
|---|
| 375 | {
|
|---|
| 376 | return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 377 | }
|
|---|
| 378 |
|
|---|
| 379 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 380 | _mm_cmpge_pd (__m128d __A, __m128d __B)
|
|---|
| 381 | {
|
|---|
| 382 | return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
|
|---|
| 383 | }
|
|---|
| 384 |
|
|---|
| 385 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 386 | _mm_cmpneq_pd (__m128d __A, __m128d __B)
|
|---|
| 387 | {
|
|---|
| 388 | return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 389 | }
|
|---|
| 390 |
|
|---|
| 391 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 392 | _mm_cmpnlt_pd (__m128d __A, __m128d __B)
|
|---|
| 393 | {
|
|---|
| 394 | return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 395 | }
|
|---|
| 396 |
|
|---|
| 397 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 398 | _mm_cmpnle_pd (__m128d __A, __m128d __B)
|
|---|
| 399 | {
|
|---|
| 400 | return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
|
|---|
| 401 | }
|
|---|
| 402 |
|
|---|
| 403 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 404 | _mm_cmpngt_pd (__m128d __A, __m128d __B)
|
|---|
| 405 | {
|
|---|
| 406 | return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 407 | }
|
|---|
| 408 |
|
|---|
| 409 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 410 | _mm_cmpnge_pd (__m128d __A, __m128d __B)
|
|---|
| 411 | {
|
|---|
| 412 | return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
|
|---|
| 413 | }
|
|---|
| 414 |
|
|---|
| 415 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 416 | _mm_cmpord_pd (__m128d __A, __m128d __B)
|
|---|
| 417 | {
|
|---|
| 418 | return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 419 | }
|
|---|
| 420 |
|
|---|
| 421 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 422 | _mm_cmpunord_pd (__m128d __A, __m128d __B)
|
|---|
| 423 | {
|
|---|
| 424 | return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 425 | }
|
|---|
| 426 |
|
|---|
| 427 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 428 | _mm_cmpeq_sd (__m128d __A, __m128d __B)
|
|---|
| 429 | {
|
|---|
| 430 | return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 431 | }
|
|---|
| 432 |
|
|---|
| 433 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 434 | _mm_cmplt_sd (__m128d __A, __m128d __B)
|
|---|
| 435 | {
|
|---|
| 436 | return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 437 | }
|
|---|
| 438 |
|
|---|
| 439 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 440 | _mm_cmple_sd (__m128d __A, __m128d __B)
|
|---|
| 441 | {
|
|---|
| 442 | return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
|
|---|
| 443 | }
|
|---|
| 444 |
|
|---|
| 445 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 446 | _mm_cmpgt_sd (__m128d __A, __m128d __B)
|
|---|
| 447 | {
|
|---|
| 448 | return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
|
|---|
| 449 | (__v2df)
|
|---|
| 450 | __builtin_ia32_cmpltsd ((__v2df) __B,
|
|---|
| 451 | (__v2df)
|
|---|
| 452 | __A));
|
|---|
| 453 | }
|
|---|
| 454 |
|
|---|
| 455 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 456 | _mm_cmpge_sd (__m128d __A, __m128d __B)
|
|---|
| 457 | {
|
|---|
| 458 | return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
|
|---|
| 459 | (__v2df)
|
|---|
| 460 | __builtin_ia32_cmplesd ((__v2df) __B,
|
|---|
| 461 | (__v2df)
|
|---|
| 462 | __A));
|
|---|
| 463 | }
|
|---|
| 464 |
|
|---|
| 465 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 466 | _mm_cmpneq_sd (__m128d __A, __m128d __B)
|
|---|
| 467 | {
|
|---|
| 468 | return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 469 | }
|
|---|
| 470 |
|
|---|
| 471 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 472 | _mm_cmpnlt_sd (__m128d __A, __m128d __B)
|
|---|
| 473 | {
|
|---|
| 474 | return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 475 | }
|
|---|
| 476 |
|
|---|
| 477 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 478 | _mm_cmpnle_sd (__m128d __A, __m128d __B)
|
|---|
| 479 | {
|
|---|
| 480 | return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
|
|---|
| 481 | }
|
|---|
| 482 |
|
|---|
| 483 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 484 | _mm_cmpngt_sd (__m128d __A, __m128d __B)
|
|---|
| 485 | {
|
|---|
| 486 | return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
|
|---|
| 487 | (__v2df)
|
|---|
| 488 | __builtin_ia32_cmpnltsd ((__v2df) __B,
|
|---|
| 489 | (__v2df)
|
|---|
| 490 | __A));
|
|---|
| 491 | }
|
|---|
| 492 |
|
|---|
| 493 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 494 | _mm_cmpnge_sd (__m128d __A, __m128d __B)
|
|---|
| 495 | {
|
|---|
| 496 | return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
|
|---|
| 497 | (__v2df)
|
|---|
| 498 | __builtin_ia32_cmpnlesd ((__v2df) __B,
|
|---|
| 499 | (__v2df)
|
|---|
| 500 | __A));
|
|---|
| 501 | }
|
|---|
| 502 |
|
|---|
| 503 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 504 | _mm_cmpord_sd (__m128d __A, __m128d __B)
|
|---|
| 505 | {
|
|---|
| 506 | return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 507 | }
|
|---|
| 508 |
|
|---|
| 509 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 510 | _mm_cmpunord_sd (__m128d __A, __m128d __B)
|
|---|
| 511 | {
|
|---|
| 512 | return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
|
|---|
| 513 | }
|
|---|
| 514 |
|
|---|
| 515 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 516 | _mm_comieq_sd (__m128d __A, __m128d __B)
|
|---|
| 517 | {
|
|---|
| 518 | return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
|
|---|
| 519 | }
|
|---|
| 520 |
|
|---|
| 521 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 522 | _mm_comilt_sd (__m128d __A, __m128d __B)
|
|---|
| 523 | {
|
|---|
| 524 | return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
|
|---|
| 525 | }
|
|---|
| 526 |
|
|---|
| 527 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 528 | _mm_comile_sd (__m128d __A, __m128d __B)
|
|---|
| 529 | {
|
|---|
| 530 | return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
|
|---|
| 531 | }
|
|---|
| 532 |
|
|---|
| 533 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 534 | _mm_comigt_sd (__m128d __A, __m128d __B)
|
|---|
| 535 | {
|
|---|
| 536 | return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
|
|---|
| 537 | }
|
|---|
| 538 |
|
|---|
| 539 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 540 | _mm_comige_sd (__m128d __A, __m128d __B)
|
|---|
| 541 | {
|
|---|
| 542 | return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
|
|---|
| 543 | }
|
|---|
| 544 |
|
|---|
| 545 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 546 | _mm_comineq_sd (__m128d __A, __m128d __B)
|
|---|
| 547 | {
|
|---|
| 548 | return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
|
|---|
| 549 | }
|
|---|
| 550 |
|
|---|
| 551 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 552 | _mm_ucomieq_sd (__m128d __A, __m128d __B)
|
|---|
| 553 | {
|
|---|
| 554 | return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
|
|---|
| 555 | }
|
|---|
| 556 |
|
|---|
| 557 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 558 | _mm_ucomilt_sd (__m128d __A, __m128d __B)
|
|---|
| 559 | {
|
|---|
| 560 | return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
|
|---|
| 561 | }
|
|---|
| 562 |
|
|---|
| 563 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 564 | _mm_ucomile_sd (__m128d __A, __m128d __B)
|
|---|
| 565 | {
|
|---|
| 566 | return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
|
|---|
| 567 | }
|
|---|
| 568 |
|
|---|
| 569 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 570 | _mm_ucomigt_sd (__m128d __A, __m128d __B)
|
|---|
| 571 | {
|
|---|
| 572 | return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
|
|---|
| 573 | }
|
|---|
| 574 |
|
|---|
| 575 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 576 | _mm_ucomige_sd (__m128d __A, __m128d __B)
|
|---|
| 577 | {
|
|---|
| 578 | return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
|
|---|
| 579 | }
|
|---|
| 580 |
|
|---|
| 581 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 582 | _mm_ucomineq_sd (__m128d __A, __m128d __B)
|
|---|
| 583 | {
|
|---|
| 584 | return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
|
|---|
| 585 | }
|
|---|
| 586 |
|
|---|
| 587 | /* Create a vector of Qi, where i is the element number. */
|
|---|
| 588 |
|
|---|
| 589 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 590 | _mm_set_epi64x (long long __q1, long long __q0)
|
|---|
| 591 | {
|
|---|
| 592 | return __extension__ (__m128i)(__v2di){ __q0, __q1 };
|
|---|
| 593 | }
|
|---|
| 594 |
|
|---|
| 595 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 596 | _mm_set_epi64 (__m64 __q1, __m64 __q0)
|
|---|
| 597 | {
|
|---|
| 598 | return _mm_set_epi64x ((long long)__q1, (long long)__q0);
|
|---|
| 599 | }
|
|---|
| 600 |
|
|---|
| 601 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 602 | _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
|
|---|
| 603 | {
|
|---|
| 604 | return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
|
|---|
| 605 | }
|
|---|
| 606 |
|
|---|
| 607 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 608 | _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
|
|---|
| 609 | short __q3, short __q2, short __q1, short __q0)
|
|---|
| 610 | {
|
|---|
| 611 | return __extension__ (__m128i)(__v8hi){
|
|---|
| 612 | __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
|
|---|
| 613 | }
|
|---|
| 614 |
|
|---|
| 615 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 616 | _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
|
|---|
| 617 | char __q11, char __q10, char __q09, char __q08,
|
|---|
| 618 | char __q07, char __q06, char __q05, char __q04,
|
|---|
| 619 | char __q03, char __q02, char __q01, char __q00)
|
|---|
| 620 | {
|
|---|
| 621 | return __extension__ (__m128i)(__v16qi){
|
|---|
| 622 | __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
|
|---|
| 623 | __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
|
|---|
| 624 | };
|
|---|
| 625 | }
|
|---|
| 626 |
|
|---|
| 627 | /* Set all of the elements of the vector to A. */
|
|---|
| 628 |
|
|---|
| 629 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 630 | _mm_set1_epi64x (long long __A)
|
|---|
| 631 | {
|
|---|
| 632 | return _mm_set_epi64x (__A, __A);
|
|---|
| 633 | }
|
|---|
| 634 |
|
|---|
| 635 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 636 | _mm_set1_epi64 (__m64 __A)
|
|---|
| 637 | {
|
|---|
| 638 | return _mm_set_epi64 (__A, __A);
|
|---|
| 639 | }
|
|---|
| 640 |
|
|---|
| 641 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 642 | _mm_set1_epi32 (int __A)
|
|---|
| 643 | {
|
|---|
| 644 | return _mm_set_epi32 (__A, __A, __A, __A);
|
|---|
| 645 | }
|
|---|
| 646 |
|
|---|
| 647 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 648 | _mm_set1_epi16 (short __A)
|
|---|
| 649 | {
|
|---|
| 650 | return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
|
|---|
| 651 | }
|
|---|
| 652 |
|
|---|
| 653 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 654 | _mm_set1_epi8 (char __A)
|
|---|
| 655 | {
|
|---|
| 656 | return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
|
|---|
| 657 | __A, __A, __A, __A, __A, __A, __A, __A);
|
|---|
| 658 | }
|
|---|
| 659 |
|
|---|
| 660 | /* Create a vector of Qi, where i is the element number.
|
|---|
| 661 | The parameter order is reversed from the _mm_set_epi* functions. */
|
|---|
| 662 |
|
|---|
| 663 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 664 | _mm_setr_epi64 (__m64 __q0, __m64 __q1)
|
|---|
| 665 | {
|
|---|
| 666 | return _mm_set_epi64 (__q1, __q0);
|
|---|
| 667 | }
|
|---|
| 668 |
|
|---|
| 669 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 670 | _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
|
|---|
| 671 | {
|
|---|
| 672 | return _mm_set_epi32 (__q3, __q2, __q1, __q0);
|
|---|
| 673 | }
|
|---|
| 674 |
|
|---|
| 675 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 676 | _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
|
|---|
| 677 | short __q4, short __q5, short __q6, short __q7)
|
|---|
| 678 | {
|
|---|
| 679 | return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
|
|---|
| 680 | }
|
|---|
| 681 |
|
|---|
| 682 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 683 | _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
|
|---|
| 684 | char __q04, char __q05, char __q06, char __q07,
|
|---|
| 685 | char __q08, char __q09, char __q10, char __q11,
|
|---|
| 686 | char __q12, char __q13, char __q14, char __q15)
|
|---|
| 687 | {
|
|---|
| 688 | return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
|
|---|
| 689 | __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
|
|---|
| 690 | }
|
|---|
| 691 |
|
|---|
| 692 | /* Create a vector with element 0 as *P and the rest zero. */
|
|---|
| 693 |
|
|---|
| 694 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 695 | _mm_load_si128 (__m128i const *__P)
|
|---|
| 696 | {
|
|---|
| 697 | return *__P;
|
|---|
| 698 | }
|
|---|
| 699 |
|
|---|
| 700 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 701 | _mm_loadu_si128 (__m128i_u const *__P)
|
|---|
| 702 | {
|
|---|
| 703 | return *__P;
|
|---|
| 704 | }
|
|---|
| 705 |
|
|---|
| 706 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 707 | _mm_loadl_epi64 (__m128i_u const *__P)
|
|---|
| 708 | {
|
|---|
| 709 | return _mm_set_epi64 ((__m64)0LL, *(__m64_u *)__P);
|
|---|
| 710 | }
|
|---|
| 711 |
|
|---|
| 712 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 713 | _mm_loadu_si64 (void const *__P)
|
|---|
| 714 | {
|
|---|
| 715 | return _mm_loadl_epi64 ((__m128i_u *)__P);
|
|---|
| 716 | }
|
|---|
| 717 |
|
|---|
| 718 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 719 | _mm_loadu_si32 (void const *__P)
|
|---|
| 720 | {
|
|---|
| 721 | return _mm_set_epi32 (*(int *)__P, (int)0, (int)0, (int)0);
|
|---|
| 722 | }
|
|---|
| 723 |
|
|---|
| 724 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 725 | _mm_loadu_si16 (void const *__P)
|
|---|
| 726 | {
|
|---|
| 727 | return _mm_set_epi16 (*(short *)__P, (short)0, (short)0, (short)0,
|
|---|
| 728 | (short)0, (short)0, (short)0, (short)0);
|
|---|
| 729 | }
|
|---|
| 730 |
|
|---|
| 731 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 732 | _mm_store_si128 (__m128i *__P, __m128i __B)
|
|---|
| 733 | {
|
|---|
| 734 | *__P = __B;
|
|---|
| 735 | }
|
|---|
| 736 |
|
|---|
| 737 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 738 | _mm_storeu_si128 (__m128i_u *__P, __m128i __B)
|
|---|
| 739 | {
|
|---|
| 740 | *__P = __B;
|
|---|
| 741 | }
|
|---|
| 742 |
|
|---|
| 743 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 744 | _mm_storel_epi64 (__m128i_u *__P, __m128i __B)
|
|---|
| 745 | {
|
|---|
| 746 | *(__m64_u *)__P = (__m64) ((__v2di)__B)[0];
|
|---|
| 747 | }
|
|---|
| 748 |
|
|---|
| 749 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 750 | _mm_storeu_si64 (void *__P, __m128i __B)
|
|---|
| 751 | {
|
|---|
| 752 | _mm_storel_epi64 ((__m128i_u *)__P, __B);
|
|---|
| 753 | }
|
|---|
| 754 |
|
|---|
| 755 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 756 | _mm_storeu_si32 (void *__P, __m128i __B)
|
|---|
| 757 | {
|
|---|
| 758 | *(__m32_u *)__P = (__m32) ((__v4si)__B)[0];
|
|---|
| 759 | }
|
|---|
| 760 |
|
|---|
| 761 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 762 | _mm_storeu_si16 (void *__P, __m128i __B)
|
|---|
| 763 | {
|
|---|
| 764 | *(__m16_u *)__P = (__m16) ((__v8hi)__B)[0];
|
|---|
| 765 | }
|
|---|
| 766 |
|
|---|
| 767 | extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 768 | _mm_movepi64_pi64 (__m128i __B)
|
|---|
| 769 | {
|
|---|
| 770 | return (__m64) ((__v2di)__B)[0];
|
|---|
| 771 | }
|
|---|
| 772 |
|
|---|
| 773 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 774 | _mm_movpi64_epi64 (__m64 __A)
|
|---|
| 775 | {
|
|---|
| 776 | return _mm_set_epi64 ((__m64)0LL, __A);
|
|---|
| 777 | }
|
|---|
| 778 |
|
|---|
| 779 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 780 | _mm_move_epi64 (__m128i __A)
|
|---|
| 781 | {
|
|---|
| 782 | return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
|
|---|
| 783 | }
|
|---|
| 784 |
|
|---|
| 785 | /* Create an undefined vector. */
|
|---|
| 786 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 787 | _mm_undefined_si128 (void)
|
|---|
| 788 | {
|
|---|
| 789 | __m128i __Y = __Y;
|
|---|
| 790 | return __Y;
|
|---|
| 791 | }
|
|---|
| 792 |
|
|---|
| 793 | /* Create a vector of zeros. */
|
|---|
| 794 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 795 | _mm_setzero_si128 (void)
|
|---|
| 796 | {
|
|---|
| 797 | return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
|
|---|
| 798 | }
|
|---|
| 799 |
|
|---|
| 800 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 801 | _mm_cvtepi32_pd (__m128i __A)
|
|---|
| 802 | {
|
|---|
| 803 | return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
|
|---|
| 804 | }
|
|---|
| 805 |
|
|---|
| 806 | extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 807 | _mm_cvtepi32_ps (__m128i __A)
|
|---|
| 808 | {
|
|---|
| 809 | return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
|
|---|
| 810 | }
|
|---|
| 811 |
|
|---|
| 812 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 813 | _mm_cvtpd_epi32 (__m128d __A)
|
|---|
| 814 | {
|
|---|
| 815 | return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
|
|---|
| 816 | }
|
|---|
| 817 |
|
|---|
| 818 | extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 819 | _mm_cvtpd_pi32 (__m128d __A)
|
|---|
| 820 | {
|
|---|
| 821 | return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
|
|---|
| 822 | }
|
|---|
| 823 |
|
|---|
| 824 | extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 825 | _mm_cvtpd_ps (__m128d __A)
|
|---|
| 826 | {
|
|---|
| 827 | return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
|
|---|
| 828 | }
|
|---|
| 829 |
|
|---|
| 830 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 831 | _mm_cvttpd_epi32 (__m128d __A)
|
|---|
| 832 | {
|
|---|
| 833 | return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
|
|---|
| 834 | }
|
|---|
| 835 |
|
|---|
| 836 | extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 837 | _mm_cvttpd_pi32 (__m128d __A)
|
|---|
| 838 | {
|
|---|
| 839 | return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
|
|---|
| 840 | }
|
|---|
| 841 |
|
|---|
| 842 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 843 | _mm_cvtpi32_pd (__m64 __A)
|
|---|
| 844 | {
|
|---|
| 845 | return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
|
|---|
| 846 | }
|
|---|
| 847 |
|
|---|
| 848 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 849 | _mm_cvtps_epi32 (__m128 __A)
|
|---|
| 850 | {
|
|---|
| 851 | return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
|
|---|
| 852 | }
|
|---|
| 853 |
|
|---|
| 854 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 855 | _mm_cvttps_epi32 (__m128 __A)
|
|---|
| 856 | {
|
|---|
| 857 | return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
|
|---|
| 858 | }
|
|---|
| 859 |
|
|---|
| 860 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 861 | _mm_cvtps_pd (__m128 __A)
|
|---|
| 862 | {
|
|---|
| 863 | return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
|
|---|
| 864 | }
|
|---|
| 865 |
|
|---|
| 866 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 867 | _mm_cvtsd_si32 (__m128d __A)
|
|---|
| 868 | {
|
|---|
| 869 | return __builtin_ia32_cvtsd2si ((__v2df) __A);
|
|---|
| 870 | }
|
|---|
| 871 |
|
|---|
| 872 | #ifdef __x86_64__
|
|---|
| 873 | /* Intel intrinsic. */
|
|---|
| 874 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 875 | _mm_cvtsd_si64 (__m128d __A)
|
|---|
| 876 | {
|
|---|
| 877 | return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
|
|---|
| 878 | }
|
|---|
| 879 |
|
|---|
| 880 | /* Microsoft intrinsic. */
|
|---|
| 881 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 882 | _mm_cvtsd_si64x (__m128d __A)
|
|---|
| 883 | {
|
|---|
| 884 | return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
|
|---|
| 885 | }
|
|---|
| 886 | #endif
|
|---|
| 887 |
|
|---|
| 888 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 889 | _mm_cvttsd_si32 (__m128d __A)
|
|---|
| 890 | {
|
|---|
| 891 | return __builtin_ia32_cvttsd2si ((__v2df) __A);
|
|---|
| 892 | }
|
|---|
| 893 |
|
|---|
| 894 | #ifdef __x86_64__
|
|---|
| 895 | /* Intel intrinsic. */
|
|---|
| 896 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 897 | _mm_cvttsd_si64 (__m128d __A)
|
|---|
| 898 | {
|
|---|
| 899 | return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
|
|---|
| 900 | }
|
|---|
| 901 |
|
|---|
| 902 | /* Microsoft intrinsic. */
|
|---|
| 903 | extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 904 | _mm_cvttsd_si64x (__m128d __A)
|
|---|
| 905 | {
|
|---|
| 906 | return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
|
|---|
| 907 | }
|
|---|
| 908 | #endif
|
|---|
| 909 |
|
|---|
| 910 | extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 911 | _mm_cvtsd_ss (__m128 __A, __m128d __B)
|
|---|
| 912 | {
|
|---|
| 913 | return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
|
|---|
| 914 | }
|
|---|
| 915 |
|
|---|
| 916 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 917 | _mm_cvtsi32_sd (__m128d __A, int __B)
|
|---|
| 918 | {
|
|---|
| 919 | return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
|
|---|
| 920 | }
|
|---|
| 921 |
|
|---|
| 922 | #ifdef __x86_64__
|
|---|
| 923 | /* Intel intrinsic. */
|
|---|
| 924 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 925 | _mm_cvtsi64_sd (__m128d __A, long long __B)
|
|---|
| 926 | {
|
|---|
| 927 | return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
|
|---|
| 928 | }
|
|---|
| 929 |
|
|---|
| 930 | /* Microsoft intrinsic. */
|
|---|
| 931 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 932 | _mm_cvtsi64x_sd (__m128d __A, long long __B)
|
|---|
| 933 | {
|
|---|
| 934 | return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
|
|---|
| 935 | }
|
|---|
| 936 | #endif
|
|---|
| 937 |
|
|---|
| 938 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 939 | _mm_cvtss_sd (__m128d __A, __m128 __B)
|
|---|
| 940 | {
|
|---|
| 941 | return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
|
|---|
| 942 | }
|
|---|
| 943 |
|
|---|
| 944 | #ifdef __OPTIMIZE__
|
|---|
| 945 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 946 | _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
|
|---|
| 947 | {
|
|---|
| 948 | return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
|
|---|
| 949 | }
|
|---|
| 950 | #else
|
|---|
| 951 | #define _mm_shuffle_pd(A, B, N) \
|
|---|
| 952 | ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A), \
|
|---|
| 953 | (__v2df)(__m128d)(B), (int)(N)))
|
|---|
| 954 | #endif
|
|---|
| 955 |
|
|---|
| 956 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 957 | _mm_unpackhi_pd (__m128d __A, __m128d __B)
|
|---|
| 958 | {
|
|---|
| 959 | return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 960 | }
|
|---|
| 961 |
|
|---|
| 962 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 963 | _mm_unpacklo_pd (__m128d __A, __m128d __B)
|
|---|
| 964 | {
|
|---|
| 965 | return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
|
|---|
| 966 | }
|
|---|
| 967 |
|
|---|
| 968 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 969 | _mm_loadh_pd (__m128d __A, double const *__B)
|
|---|
| 970 | {
|
|---|
| 971 | return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
|
|---|
| 972 | }
|
|---|
| 973 |
|
|---|
| 974 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 975 | _mm_loadl_pd (__m128d __A, double const *__B)
|
|---|
| 976 | {
|
|---|
| 977 | return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
|
|---|
| 978 | }
|
|---|
| 979 |
|
|---|
| 980 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 981 | _mm_movemask_pd (__m128d __A)
|
|---|
| 982 | {
|
|---|
| 983 | return __builtin_ia32_movmskpd ((__v2df)__A);
|
|---|
| 984 | }
|
|---|
| 985 |
|
|---|
| 986 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 987 | _mm_packs_epi16 (__m128i __A, __m128i __B)
|
|---|
| 988 | {
|
|---|
| 989 | return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 990 | }
|
|---|
| 991 |
|
|---|
| 992 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 993 | _mm_packs_epi32 (__m128i __A, __m128i __B)
|
|---|
| 994 | {
|
|---|
| 995 | return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 996 | }
|
|---|
| 997 |
|
|---|
| 998 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 999 | _mm_packus_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1000 | {
|
|---|
| 1001 | return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1002 | }
|
|---|
| 1003 |
|
|---|
| 1004 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1005 | _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1006 | {
|
|---|
| 1007 | return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1008 | }
|
|---|
| 1009 |
|
|---|
| 1010 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1011 | _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1012 | {
|
|---|
| 1013 | return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1014 | }
|
|---|
| 1015 |
|
|---|
| 1016 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1017 | _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1018 | {
|
|---|
| 1019 | return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 1020 | }
|
|---|
| 1021 |
|
|---|
| 1022 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1023 | _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1024 | {
|
|---|
| 1025 | return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
|
|---|
| 1026 | }
|
|---|
| 1027 |
|
|---|
| 1028 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1029 | _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1030 | {
|
|---|
| 1031 | return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1032 | }
|
|---|
| 1033 |
|
|---|
| 1034 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1035 | _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1036 | {
|
|---|
| 1037 | return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1038 | }
|
|---|
| 1039 |
|
|---|
| 1040 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1041 | _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1042 | {
|
|---|
| 1043 | return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 1044 | }
|
|---|
| 1045 |
|
|---|
| 1046 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1047 | _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1048 | {
|
|---|
| 1049 | return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
|
|---|
| 1050 | }
|
|---|
| 1051 |
|
|---|
| 1052 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1053 | _mm_add_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1054 | {
|
|---|
| 1055 | return (__m128i) ((__v16qu)__A + (__v16qu)__B);
|
|---|
| 1056 | }
|
|---|
| 1057 |
|
|---|
| 1058 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1059 | _mm_add_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1060 | {
|
|---|
| 1061 | return (__m128i) ((__v8hu)__A + (__v8hu)__B);
|
|---|
| 1062 | }
|
|---|
| 1063 |
|
|---|
| 1064 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1065 | _mm_add_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1066 | {
|
|---|
| 1067 | return (__m128i) ((__v4su)__A + (__v4su)__B);
|
|---|
| 1068 | }
|
|---|
| 1069 |
|
|---|
| 1070 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1071 | _mm_add_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1072 | {
|
|---|
| 1073 | return (__m128i) ((__v2du)__A + (__v2du)__B);
|
|---|
| 1074 | }
|
|---|
| 1075 |
|
|---|
| 1076 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1077 | _mm_adds_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1078 | {
|
|---|
| 1079 | return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1080 | }
|
|---|
| 1081 |
|
|---|
| 1082 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1083 | _mm_adds_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1084 | {
|
|---|
| 1085 | return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1086 | }
|
|---|
| 1087 |
|
|---|
| 1088 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1089 | _mm_adds_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1090 | {
|
|---|
| 1091 | return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1092 | }
|
|---|
| 1093 |
|
|---|
| 1094 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1095 | _mm_adds_epu16 (__m128i __A, __m128i __B)
|
|---|
| 1096 | {
|
|---|
| 1097 | return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1098 | }
|
|---|
| 1099 |
|
|---|
| 1100 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1101 | _mm_sub_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1102 | {
|
|---|
| 1103 | return (__m128i) ((__v16qu)__A - (__v16qu)__B);
|
|---|
| 1104 | }
|
|---|
| 1105 |
|
|---|
| 1106 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1107 | _mm_sub_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1108 | {
|
|---|
| 1109 | return (__m128i) ((__v8hu)__A - (__v8hu)__B);
|
|---|
| 1110 | }
|
|---|
| 1111 |
|
|---|
| 1112 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1113 | _mm_sub_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1114 | {
|
|---|
| 1115 | return (__m128i) ((__v4su)__A - (__v4su)__B);
|
|---|
| 1116 | }
|
|---|
| 1117 |
|
|---|
| 1118 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1119 | _mm_sub_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1120 | {
|
|---|
| 1121 | return (__m128i) ((__v2du)__A - (__v2du)__B);
|
|---|
| 1122 | }
|
|---|
| 1123 |
|
|---|
| 1124 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1125 | _mm_subs_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1126 | {
|
|---|
| 1127 | return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1128 | }
|
|---|
| 1129 |
|
|---|
| 1130 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1131 | _mm_subs_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1132 | {
|
|---|
| 1133 | return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1134 | }
|
|---|
| 1135 |
|
|---|
| 1136 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1137 | _mm_subs_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1138 | {
|
|---|
| 1139 | return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1140 | }
|
|---|
| 1141 |
|
|---|
| 1142 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1143 | _mm_subs_epu16 (__m128i __A, __m128i __B)
|
|---|
| 1144 | {
|
|---|
| 1145 | return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1146 | }
|
|---|
| 1147 |
|
|---|
| 1148 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1149 | _mm_madd_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1150 | {
|
|---|
| 1151 | return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1152 | }
|
|---|
| 1153 |
|
|---|
| 1154 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1155 | _mm_mulhi_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1156 | {
|
|---|
| 1157 | return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1158 | }
|
|---|
| 1159 |
|
|---|
| 1160 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1161 | _mm_mullo_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1162 | {
|
|---|
| 1163 | return (__m128i) ((__v8hu)__A * (__v8hu)__B);
|
|---|
| 1164 | }
|
|---|
| 1165 |
|
|---|
| 1166 | extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1167 | _mm_mul_su32 (__m64 __A, __m64 __B)
|
|---|
| 1168 | {
|
|---|
| 1169 | return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
|
|---|
| 1170 | }
|
|---|
| 1171 |
|
|---|
| 1172 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1173 | _mm_mul_epu32 (__m128i __A, __m128i __B)
|
|---|
| 1174 | {
|
|---|
| 1175 | return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 1176 | }
|
|---|
| 1177 |
|
|---|
| 1178 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1179 | _mm_slli_epi16 (__m128i __A, int __B)
|
|---|
| 1180 | {
|
|---|
| 1181 | return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
|
|---|
| 1182 | }
|
|---|
| 1183 |
|
|---|
| 1184 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1185 | _mm_slli_epi32 (__m128i __A, int __B)
|
|---|
| 1186 | {
|
|---|
| 1187 | return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
|
|---|
| 1188 | }
|
|---|
| 1189 |
|
|---|
| 1190 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1191 | _mm_slli_epi64 (__m128i __A, int __B)
|
|---|
| 1192 | {
|
|---|
| 1193 | return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
|
|---|
| 1194 | }
|
|---|
| 1195 |
|
|---|
| 1196 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1197 | _mm_srai_epi16 (__m128i __A, int __B)
|
|---|
| 1198 | {
|
|---|
| 1199 | return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
|
|---|
| 1200 | }
|
|---|
| 1201 |
|
|---|
| 1202 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1203 | _mm_srai_epi32 (__m128i __A, int __B)
|
|---|
| 1204 | {
|
|---|
| 1205 | return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
|
|---|
| 1206 | }
|
|---|
| 1207 |
|
|---|
| 1208 | #ifdef __OPTIMIZE__
|
|---|
| 1209 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1210 | _mm_bsrli_si128 (__m128i __A, const int __N)
|
|---|
| 1211 | {
|
|---|
| 1212 | return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
|
|---|
| 1213 | }
|
|---|
| 1214 |
|
|---|
| 1215 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1216 | _mm_bslli_si128 (__m128i __A, const int __N)
|
|---|
| 1217 | {
|
|---|
| 1218 | return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
|
|---|
| 1219 | }
|
|---|
| 1220 |
|
|---|
| 1221 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1222 | _mm_srli_si128 (__m128i __A, const int __N)
|
|---|
| 1223 | {
|
|---|
| 1224 | return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
|
|---|
| 1225 | }
|
|---|
| 1226 |
|
|---|
| 1227 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1228 | _mm_slli_si128 (__m128i __A, const int __N)
|
|---|
| 1229 | {
|
|---|
| 1230 | return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
|
|---|
| 1231 | }
|
|---|
| 1232 | #else
|
|---|
| 1233 | #define _mm_bsrli_si128(A, N) \
|
|---|
| 1234 | ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
|
|---|
| 1235 | #define _mm_bslli_si128(A, N) \
|
|---|
| 1236 | ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
|
|---|
| 1237 | #define _mm_srli_si128(A, N) \
|
|---|
| 1238 | ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
|
|---|
| 1239 | #define _mm_slli_si128(A, N) \
|
|---|
| 1240 | ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
|
|---|
| 1241 | #endif
|
|---|
| 1242 |
|
|---|
| 1243 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1244 | _mm_srli_epi16 (__m128i __A, int __B)
|
|---|
| 1245 | {
|
|---|
| 1246 | return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
|
|---|
| 1247 | }
|
|---|
| 1248 |
|
|---|
| 1249 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1250 | _mm_srli_epi32 (__m128i __A, int __B)
|
|---|
| 1251 | {
|
|---|
| 1252 | return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
|
|---|
| 1253 | }
|
|---|
| 1254 |
|
|---|
| 1255 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1256 | _mm_srli_epi64 (__m128i __A, int __B)
|
|---|
| 1257 | {
|
|---|
| 1258 | return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
|
|---|
| 1259 | }
|
|---|
| 1260 |
|
|---|
| 1261 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1262 | _mm_sll_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1263 | {
|
|---|
| 1264 | return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1265 | }
|
|---|
| 1266 |
|
|---|
| 1267 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1268 | _mm_sll_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1269 | {
|
|---|
| 1270 | return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
|
|---|
| 1271 | }
|
|---|
| 1272 |
|
|---|
| 1273 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1274 | _mm_sll_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1275 | {
|
|---|
| 1276 | return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
|
|---|
| 1277 | }
|
|---|
| 1278 |
|
|---|
| 1279 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1280 | _mm_sra_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1281 | {
|
|---|
| 1282 | return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1283 | }
|
|---|
| 1284 |
|
|---|
| 1285 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1286 | _mm_sra_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1287 | {
|
|---|
| 1288 | return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 1289 | }
|
|---|
| 1290 |
|
|---|
| 1291 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1292 | _mm_srl_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1293 | {
|
|---|
| 1294 | return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1295 | }
|
|---|
| 1296 |
|
|---|
| 1297 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1298 | _mm_srl_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1299 | {
|
|---|
| 1300 | return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
|
|---|
| 1301 | }
|
|---|
| 1302 |
|
|---|
| 1303 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1304 | _mm_srl_epi64 (__m128i __A, __m128i __B)
|
|---|
| 1305 | {
|
|---|
| 1306 | return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
|
|---|
| 1307 | }
|
|---|
| 1308 |
|
|---|
| 1309 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1310 | _mm_and_si128 (__m128i __A, __m128i __B)
|
|---|
| 1311 | {
|
|---|
| 1312 | return (__m128i) ((__v2du)__A & (__v2du)__B);
|
|---|
| 1313 | }
|
|---|
| 1314 |
|
|---|
| 1315 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1316 | _mm_andnot_si128 (__m128i __A, __m128i __B)
|
|---|
| 1317 | {
|
|---|
| 1318 | return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
|
|---|
| 1319 | }
|
|---|
| 1320 |
|
|---|
| 1321 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1322 | _mm_or_si128 (__m128i __A, __m128i __B)
|
|---|
| 1323 | {
|
|---|
| 1324 | return (__m128i) ((__v2du)__A | (__v2du)__B);
|
|---|
| 1325 | }
|
|---|
| 1326 |
|
|---|
| 1327 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1328 | _mm_xor_si128 (__m128i __A, __m128i __B)
|
|---|
| 1329 | {
|
|---|
| 1330 | return (__m128i) ((__v2du)__A ^ (__v2du)__B);
|
|---|
| 1331 | }
|
|---|
| 1332 |
|
|---|
| 1333 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1334 | _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1335 | {
|
|---|
| 1336 | return (__m128i) ((__v16qi)__A == (__v16qi)__B);
|
|---|
| 1337 | }
|
|---|
| 1338 |
|
|---|
| 1339 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1340 | _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1341 | {
|
|---|
| 1342 | return (__m128i) ((__v8hi)__A == (__v8hi)__B);
|
|---|
| 1343 | }
|
|---|
| 1344 |
|
|---|
| 1345 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1346 | _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1347 | {
|
|---|
| 1348 | return (__m128i) ((__v4si)__A == (__v4si)__B);
|
|---|
| 1349 | }
|
|---|
| 1350 |
|
|---|
| 1351 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1352 | _mm_cmplt_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1353 | {
|
|---|
| 1354 | return (__m128i) ((__v16qs)__A < (__v16qs)__B);
|
|---|
| 1355 | }
|
|---|
| 1356 |
|
|---|
| 1357 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1358 | _mm_cmplt_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1359 | {
|
|---|
| 1360 | return (__m128i) ((__v8hi)__A < (__v8hi)__B);
|
|---|
| 1361 | }
|
|---|
| 1362 |
|
|---|
| 1363 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1364 | _mm_cmplt_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1365 | {
|
|---|
| 1366 | return (__m128i) ((__v4si)__A < (__v4si)__B);
|
|---|
| 1367 | }
|
|---|
| 1368 |
|
|---|
| 1369 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1370 | _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
|
|---|
| 1371 | {
|
|---|
| 1372 | return (__m128i) ((__v16qs)__A > (__v16qs)__B);
|
|---|
| 1373 | }
|
|---|
| 1374 |
|
|---|
| 1375 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1376 | _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1377 | {
|
|---|
| 1378 | return (__m128i) ((__v8hi)__A > (__v8hi)__B);
|
|---|
| 1379 | }
|
|---|
| 1380 |
|
|---|
| 1381 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1382 | _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
|
|---|
| 1383 | {
|
|---|
| 1384 | return (__m128i) ((__v4si)__A > (__v4si)__B);
|
|---|
| 1385 | }
|
|---|
| 1386 |
|
|---|
| 1387 | #ifdef __OPTIMIZE__
|
|---|
| 1388 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1389 | _mm_extract_epi16 (__m128i const __A, int const __N)
|
|---|
| 1390 | {
|
|---|
| 1391 | return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
|
|---|
| 1392 | }
|
|---|
| 1393 |
|
|---|
| 1394 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1395 | _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
|
|---|
| 1396 | {
|
|---|
| 1397 | return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
|
|---|
| 1398 | }
|
|---|
| 1399 | #else
|
|---|
| 1400 | #define _mm_extract_epi16(A, N) \
|
|---|
| 1401 | ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
|
|---|
| 1402 | #define _mm_insert_epi16(A, D, N) \
|
|---|
| 1403 | ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
|
|---|
| 1404 | (int)(D), (int)(N)))
|
|---|
| 1405 | #endif
|
|---|
| 1406 |
|
|---|
| 1407 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1408 | _mm_max_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1409 | {
|
|---|
| 1410 | return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1411 | }
|
|---|
| 1412 |
|
|---|
| 1413 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1414 | _mm_max_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1415 | {
|
|---|
| 1416 | return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1417 | }
|
|---|
| 1418 |
|
|---|
| 1419 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1420 | _mm_min_epi16 (__m128i __A, __m128i __B)
|
|---|
| 1421 | {
|
|---|
| 1422 | return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1423 | }
|
|---|
| 1424 |
|
|---|
| 1425 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1426 | _mm_min_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1427 | {
|
|---|
| 1428 | return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1429 | }
|
|---|
| 1430 |
|
|---|
| 1431 | extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1432 | _mm_movemask_epi8 (__m128i __A)
|
|---|
| 1433 | {
|
|---|
| 1434 | return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
|
|---|
| 1435 | }
|
|---|
| 1436 |
|
|---|
| 1437 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1438 | _mm_mulhi_epu16 (__m128i __A, __m128i __B)
|
|---|
| 1439 | {
|
|---|
| 1440 | return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1441 | }
|
|---|
| 1442 |
|
|---|
| 1443 | #ifdef __OPTIMIZE__
|
|---|
| 1444 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1445 | _mm_shufflehi_epi16 (__m128i __A, const int __mask)
|
|---|
| 1446 | {
|
|---|
| 1447 | return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
|
|---|
| 1448 | }
|
|---|
| 1449 |
|
|---|
| 1450 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1451 | _mm_shufflelo_epi16 (__m128i __A, const int __mask)
|
|---|
| 1452 | {
|
|---|
| 1453 | return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
|
|---|
| 1454 | }
|
|---|
| 1455 |
|
|---|
| 1456 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1457 | _mm_shuffle_epi32 (__m128i __A, const int __mask)
|
|---|
| 1458 | {
|
|---|
| 1459 | return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
|
|---|
| 1460 | }
|
|---|
| 1461 | #else
|
|---|
| 1462 | #define _mm_shufflehi_epi16(A, N) \
|
|---|
| 1463 | ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
|
|---|
| 1464 | #define _mm_shufflelo_epi16(A, N) \
|
|---|
| 1465 | ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
|
|---|
| 1466 | #define _mm_shuffle_epi32(A, N) \
|
|---|
| 1467 | ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
|
|---|
| 1468 | #endif
|
|---|
| 1469 |
|
|---|
| 1470 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1471 | _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
|
|---|
| 1472 | {
|
|---|
| 1473 | __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
|
|---|
| 1474 | }
|
|---|
| 1475 |
|
|---|
| 1476 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1477 | _mm_avg_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1478 | {
|
|---|
| 1479 | return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1480 | }
|
|---|
| 1481 |
|
|---|
| 1482 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1483 | _mm_avg_epu16 (__m128i __A, __m128i __B)
|
|---|
| 1484 | {
|
|---|
| 1485 | return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
|
|---|
| 1486 | }
|
|---|
| 1487 |
|
|---|
| 1488 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1489 | _mm_sad_epu8 (__m128i __A, __m128i __B)
|
|---|
| 1490 | {
|
|---|
| 1491 | return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
|
|---|
| 1492 | }
|
|---|
| 1493 |
|
|---|
| 1494 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1495 | _mm_stream_si32 (int *__A, int __B)
|
|---|
| 1496 | {
|
|---|
| 1497 | __builtin_ia32_movnti (__A, __B);
|
|---|
| 1498 | }
|
|---|
| 1499 |
|
|---|
| 1500 | #ifdef __x86_64__
|
|---|
| 1501 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1502 | _mm_stream_si64 (long long int *__A, long long int __B)
|
|---|
| 1503 | {
|
|---|
| 1504 | __builtin_ia32_movnti64 (__A, __B);
|
|---|
| 1505 | }
|
|---|
| 1506 | #endif
|
|---|
| 1507 |
|
|---|
| 1508 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1509 | _mm_stream_si128 (__m128i *__A, __m128i __B)
|
|---|
| 1510 | {
|
|---|
| 1511 | __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
|
|---|
| 1512 | }
|
|---|
| 1513 |
|
|---|
| 1514 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1515 | _mm_stream_pd (double *__A, __m128d __B)
|
|---|
| 1516 | {
|
|---|
| 1517 | __builtin_ia32_movntpd (__A, (__v2df)__B);
|
|---|
| 1518 | }
|
|---|
| 1519 |
|
|---|
| 1520 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1521 | _mm_clflush (void const *__A)
|
|---|
| 1522 | {
|
|---|
| 1523 | __builtin_ia32_clflush (__A);
|
|---|
| 1524 | }
|
|---|
| 1525 |
|
|---|
| 1526 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1527 | _mm_lfence (void)
|
|---|
| 1528 | {
|
|---|
| 1529 | __builtin_ia32_lfence ();
|
|---|
| 1530 | }
|
|---|
| 1531 |
|
|---|
| 1532 | extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1533 | _mm_mfence (void)
|
|---|
| 1534 | {
|
|---|
| 1535 | __builtin_ia32_mfence ();
|
|---|
| 1536 | }
|
|---|
| 1537 |
|
|---|
| 1538 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1539 | _mm_cvtsi32_si128 (int __A)
|
|---|
| 1540 | {
|
|---|
| 1541 | return _mm_set_epi32 (0, 0, 0, __A);
|
|---|
| 1542 | }
|
|---|
| 1543 |
|
|---|
| 1544 | #ifdef __x86_64__
|
|---|
| 1545 | /* Intel intrinsic. */
|
|---|
| 1546 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1547 | _mm_cvtsi64_si128 (long long __A)
|
|---|
| 1548 | {
|
|---|
| 1549 | return _mm_set_epi64x (0, __A);
|
|---|
| 1550 | }
|
|---|
| 1551 |
|
|---|
| 1552 | /* Microsoft intrinsic. */
|
|---|
| 1553 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1554 | _mm_cvtsi64x_si128 (long long __A)
|
|---|
| 1555 | {
|
|---|
| 1556 | return _mm_set_epi64x (0, __A);
|
|---|
| 1557 | }
|
|---|
| 1558 | #endif
|
|---|
| 1559 |
|
|---|
| 1560 | /* Casts between various SP, DP, INT vector types. Note that these do no
|
|---|
| 1561 | conversion of values, they just change the type. */
|
|---|
| 1562 | extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1563 | _mm_castpd_ps(__m128d __A)
|
|---|
| 1564 | {
|
|---|
| 1565 | return (__m128) __A;
|
|---|
| 1566 | }
|
|---|
| 1567 |
|
|---|
| 1568 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1569 | _mm_castpd_si128(__m128d __A)
|
|---|
| 1570 | {
|
|---|
| 1571 | return (__m128i) __A;
|
|---|
| 1572 | }
|
|---|
| 1573 |
|
|---|
| 1574 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1575 | _mm_castps_pd(__m128 __A)
|
|---|
| 1576 | {
|
|---|
| 1577 | return (__m128d) __A;
|
|---|
| 1578 | }
|
|---|
| 1579 |
|
|---|
| 1580 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1581 | _mm_castps_si128(__m128 __A)
|
|---|
| 1582 | {
|
|---|
| 1583 | return (__m128i) __A;
|
|---|
| 1584 | }
|
|---|
| 1585 |
|
|---|
| 1586 | extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1587 | _mm_castsi128_ps(__m128i __A)
|
|---|
| 1588 | {
|
|---|
| 1589 | return (__m128) __A;
|
|---|
| 1590 | }
|
|---|
| 1591 |
|
|---|
| 1592 | extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
|
|---|
| 1593 | _mm_castsi128_pd(__m128i __A)
|
|---|
| 1594 | {
|
|---|
| 1595 | return (__m128d) __A;
|
|---|
| 1596 | }
|
|---|
| 1597 |
|
|---|
| 1598 | #ifdef __DISABLE_SSE2__
|
|---|
| 1599 | #undef __DISABLE_SSE2__
|
|---|
| 1600 | #pragma GCC pop_options
|
|---|
| 1601 | #endif /* __DISABLE_SSE2__ */
|
|---|
| 1602 |
|
|---|
| 1603 | #endif /* _EMMINTRIN_H_INCLUDED */
|
|---|
