1 /*===---- smmintrin.h - SSE4 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 _SMMINTRIN_H 25 #define _SMMINTRIN_H 26 27 #ifndef __SSE4_1__ 28 #error "SSE4.1 instruction set not enabled" 29 #else 30 31 #include <tmmintrin.h> 32 33 /* SSE4 Rounding macros. */ 34 #define _MM_FROUND_TO_NEAREST_INT 0x00 35 #define _MM_FROUND_TO_NEG_INF 0x01 36 #define _MM_FROUND_TO_POS_INF 0x02 37 #define _MM_FROUND_TO_ZERO 0x03 38 #define _MM_FROUND_CUR_DIRECTION 0x04 39 40 #define _MM_FROUND_RAISE_EXC 0x00 41 #define _MM_FROUND_NO_EXC 0x08 42 43 #define _MM_FROUND_NINT (_MM_FROUND_RAISE_EXC | _MM_FROUND_TO_NEAREST_INT) 44 #define _MM_FROUND_FLOOR (_MM_FROUND_RAISE_EXC | _MM_FROUND_TO_NEG_INF) 45 #define _MM_FROUND_CEIL (_MM_FROUND_RAISE_EXC | _MM_FROUND_TO_POS_INF) 46 #define _MM_FROUND_TRUNC (_MM_FROUND_RAISE_EXC | _MM_FROUND_TO_ZERO) 47 #define _MM_FROUND_RINT (_MM_FROUND_RAISE_EXC | _MM_FROUND_CUR_DIRECTION) 48 #define _MM_FROUND_NEARBYINT (_MM_FROUND_NO_EXC | _MM_FROUND_CUR_DIRECTION) 49 50 #define _mm_ceil_ps(X) _mm_round_ps((X), _MM_FROUND_CEIL) 51 #define _mm_ceil_pd(X) _mm_round_pd((X), _MM_FROUND_CEIL) 52 #define _mm_ceil_ss(X, Y) _mm_round_ss((X), (Y), _MM_FROUND_CEIL) 53 #define _mm_ceil_sd(X, Y) _mm_round_sd((X), (Y), _MM_FROUND_CEIL) 54 55 #define _mm_floor_ps(X) _mm_round_ps((X), _MM_FROUND_FLOOR) 56 #define _mm_floor_pd(X) _mm_round_pd((X), _MM_FROUND_FLOOR) 57 #define _mm_floor_ss(X, Y) _mm_round_ss((X), (Y), _MM_FROUND_FLOOR) 58 #define _mm_floor_sd(X, Y) _mm_round_sd((X), (Y), _MM_FROUND_FLOOR) 59 60 #define _mm_round_ps(X, M) __extension__ ({ \ 61 __m128 __X = (X); \ 62 (__m128) __builtin_ia32_roundps((__v4sf)__X, (M)); }) 63 64 #define _mm_round_ss(X, Y, M) __extension__ ({ \ 65 __m128 __X = (X); \ 66 __m128 __Y = (Y); \ 67 (__m128) __builtin_ia32_roundss((__v4sf)__X, (__v4sf)__Y, (M)); }) 68 69 #define _mm_round_pd(X, M) __extension__ ({ \ 70 __m128d __X = (X); \ 71 (__m128d) __builtin_ia32_roundpd((__v2df)__X, (M)); }) 72 73 #define _mm_round_sd(X, Y, M) __extension__ ({ \ 74 __m128d __X = (X); \ 75 __m128d __Y = (Y); \ 76 (__m128d) __builtin_ia32_roundsd((__v2df)__X, (__v2df)__Y, (M)); }) 77 78 /* SSE4 Packed Blending Intrinsics. */ 79 #define _mm_blend_pd(V1, V2, M) __extension__ ({ \ 80 __m128d __V1 = (V1); \ 81 __m128d __V2 = (V2); \ 82 (__m128d) __builtin_ia32_blendpd ((__v2df)__V1, (__v2df)__V2, (M)); }) 83 84 #define _mm_blend_ps(V1, V2, M) __extension__ ({ \ 85 __m128 __V1 = (V1); \ 86 __m128 __V2 = (V2); \ 87 (__m128) __builtin_ia32_blendps ((__v4sf)__V1, (__v4sf)__V2, (M)); }) 88 89 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) 90 _mm_blendv_pd (__m128d __V1, __m128d __V2, __m128d __M) 91 { 92 return (__m128d) __builtin_ia32_blendvpd ((__v2df)__V1, (__v2df)__V2, 93 (__v2df)__M); 94 } 95 96 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) 97 _mm_blendv_ps (__m128 __V1, __m128 __V2, __m128 __M) 98 { 99 return (__m128) __builtin_ia32_blendvps ((__v4sf)__V1, (__v4sf)__V2, 100 (__v4sf)__M); 101 } 102 103 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 104 _mm_blendv_epi8 (__m128i __V1, __m128i __V2, __m128i __M) 105 { 106 return (__m128i) __builtin_ia32_pblendvb128 ((__v16qi)__V1, (__v16qi)__V2, 107 (__v16qi)__M); 108 } 109 110 #define _mm_blend_epi16(V1, V2, M) __extension__ ({ \ 111 __m128i __V1 = (V1); \ 112 __m128i __V2 = (V2); \ 113 (__m128i) __builtin_ia32_pblendw128 ((__v8hi)__V1, (__v8hi)__V2, (M)); }) 114 115 /* SSE4 Dword Multiply Instructions. */ 116 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 117 _mm_mullo_epi32 (__m128i __V1, __m128i __V2) 118 { 119 return (__m128i) ((__v4si)__V1 * (__v4si)__V2); 120 } 121 122 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 123 _mm_mul_epi32 (__m128i __V1, __m128i __V2) 124 { 125 return (__m128i) __builtin_ia32_pmuldq128 ((__v4si)__V1, (__v4si)__V2); 126 } 127 128 /* SSE4 Floating Point Dot Product Instructions. */ 129 #define _mm_dp_ps(X, Y, M) __extension__ ({ \ 130 __m128 __X = (X); \ 131 __m128 __Y = (Y); \ 132 (__m128) __builtin_ia32_dpps((__v4sf)__X, (__v4sf)__Y, (M)); }) 133 134 #define _mm_dp_pd(X, Y, M) __extension__ ({\ 135 __m128d __X = (X); \ 136 __m128d __Y = (Y); \ 137 (__m128d) __builtin_ia32_dppd((__v2df)__X, (__v2df)__Y, (M)); }) 138 139 /* SSE4 Streaming Load Hint Instruction. */ 140 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 141 _mm_stream_load_si128 (__m128i *__V) 142 { 143 return (__m128i) __builtin_ia32_movntdqa ((__v2di *) __V); 144 } 145 146 /* SSE4 Packed Integer Min/Max Instructions. */ 147 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 148 _mm_min_epi8 (__m128i __V1, __m128i __V2) 149 { 150 return (__m128i) __builtin_ia32_pminsb128 ((__v16qi) __V1, (__v16qi) __V2); 151 } 152 153 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 154 _mm_max_epi8 (__m128i __V1, __m128i __V2) 155 { 156 return (__m128i) __builtin_ia32_pmaxsb128 ((__v16qi) __V1, (__v16qi) __V2); 157 } 158 159 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 160 _mm_min_epu16 (__m128i __V1, __m128i __V2) 161 { 162 return (__m128i) __builtin_ia32_pminuw128 ((__v8hi) __V1, (__v8hi) __V2); 163 } 164 165 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 166 _mm_max_epu16 (__m128i __V1, __m128i __V2) 167 { 168 return (__m128i) __builtin_ia32_pmaxuw128 ((__v8hi) __V1, (__v8hi) __V2); 169 } 170 171 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 172 _mm_min_epi32 (__m128i __V1, __m128i __V2) 173 { 174 return (__m128i) __builtin_ia32_pminsd128 ((__v4si) __V1, (__v4si) __V2); 175 } 176 177 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 178 _mm_max_epi32 (__m128i __V1, __m128i __V2) 179 { 180 return (__m128i) __builtin_ia32_pmaxsd128 ((__v4si) __V1, (__v4si) __V2); 181 } 182 183 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 184 _mm_min_epu32 (__m128i __V1, __m128i __V2) 185 { 186 return (__m128i) __builtin_ia32_pminud128((__v4si) __V1, (__v4si) __V2); 187 } 188 189 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 190 _mm_max_epu32 (__m128i __V1, __m128i __V2) 191 { 192 return (__m128i) __builtin_ia32_pmaxud128((__v4si) __V1, (__v4si) __V2); 193 } 194 195 /* SSE4 Insertion and Extraction from XMM Register Instructions. */ 196 #define _mm_insert_ps(X, Y, N) __builtin_ia32_insertps128((X), (Y), (N)) 197 #define _mm_extract_ps(X, N) (__extension__ \ 198 ({ union { int __i; float __f; } __t; \ 199 __v4sf __a = (__v4sf)(X); \ 200 __t.__f = __a[N]; \ 201 __t.__i;})) 202 203 /* Miscellaneous insert and extract macros. */ 204 /* Extract a single-precision float from X at index N into D. */ 205 #define _MM_EXTRACT_FLOAT(D, X, N) (__extension__ ({ __v4sf __a = (__v4sf)(X); \ 206 (D) = __a[N]; })) 207 208 /* Or together 2 sets of indexes (X and Y) with the zeroing bits (Z) to create 209 an index suitable for _mm_insert_ps. */ 210 #define _MM_MK_INSERTPS_NDX(X, Y, Z) (((X) << 6) | ((Y) << 4) | (Z)) 211 212 /* Extract a float from X at index N into the first index of the return. */ 213 #define _MM_PICK_OUT_PS(X, N) _mm_insert_ps (_mm_setzero_ps(), (X), \ 214 _MM_MK_INSERTPS_NDX((N), 0, 0x0e)) 215 216 /* Insert int into packed integer array at index. */ 217 #define _mm_insert_epi8(X, I, N) (__extension__ ({ __v16qi __a = (__v16qi)(X); \ 218 __a[(N)] = (I); \ 219 __a;})) 220 #define _mm_insert_epi32(X, I, N) (__extension__ ({ __v4si __a = (__v4si)(X); \ 221 __a[(N)] = (I); \ 222 __a;})) 223 #ifdef __x86_64__ 224 #define _mm_insert_epi64(X, I, N) (__extension__ ({ __v2di __a = (__v2di)(X); \ 225 __a[(N)] = (I); \ 226 __a;})) 227 #endif /* __x86_64__ */ 228 229 /* Extract int from packed integer array at index. This returns the element 230 * as a zero extended value, so it is unsigned. 231 */ 232 #define _mm_extract_epi8(X, N) (__extension__ ({ __v16qi __a = (__v16qi)(X); \ 233 (unsigned char)__a[(N)];})) 234 #define _mm_extract_epi32(X, N) (__extension__ ({ __v4si __a = (__v4si)(X); \ 235 (unsigned)__a[(N)];})) 236 #ifdef __x86_64__ 237 #define _mm_extract_epi64(X, N) (__extension__ ({ __v2di __a = (__v2di)(X); \ 238 __a[(N)];})) 239 #endif /* __x86_64 */ 240 241 /* SSE4 128-bit Packed Integer Comparisons. */ 242 static __inline__ int __attribute__((__always_inline__, __nodebug__)) 243 _mm_testz_si128(__m128i __M, __m128i __V) 244 { 245 return __builtin_ia32_ptestz128((__v2di)__M, (__v2di)__V); 246 } 247 248 static __inline__ int __attribute__((__always_inline__, __nodebug__)) 249 _mm_testc_si128(__m128i __M, __m128i __V) 250 { 251 return __builtin_ia32_ptestc128((__v2di)__M, (__v2di)__V); 252 } 253 254 static __inline__ int __attribute__((__always_inline__, __nodebug__)) 255 _mm_testnzc_si128(__m128i __M, __m128i __V) 256 { 257 return __builtin_ia32_ptestnzc128((__v2di)__M, (__v2di)__V); 258 } 259 260 #define _mm_test_all_ones(V) _mm_testc_si128((V), _mm_cmpeq_epi32((V), (V))) 261 #define _mm_test_mix_ones_zeros(M, V) _mm_testnzc_si128((M), (V)) 262 #define _mm_test_all_zeros(M, V) _mm_testz_si128 ((M), (V)) 263 264 /* SSE4 64-bit Packed Integer Comparisons. */ 265 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 266 _mm_cmpeq_epi64(__m128i __V1, __m128i __V2) 267 { 268 return (__m128i)((__v2di)__V1 == (__v2di)__V2); 269 } 270 271 /* SSE4 Packed Integer Sign-Extension. */ 272 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 273 _mm_cvtepi8_epi16(__m128i __V) 274 { 275 return (__m128i) __builtin_ia32_pmovsxbw128((__v16qi) __V); 276 } 277 278 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 279 _mm_cvtepi8_epi32(__m128i __V) 280 { 281 return (__m128i) __builtin_ia32_pmovsxbd128((__v16qi) __V); 282 } 283 284 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 285 _mm_cvtepi8_epi64(__m128i __V) 286 { 287 return (__m128i) __builtin_ia32_pmovsxbq128((__v16qi) __V); 288 } 289 290 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 291 _mm_cvtepi16_epi32(__m128i __V) 292 { 293 return (__m128i) __builtin_ia32_pmovsxwd128((__v8hi) __V); 294 } 295 296 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 297 _mm_cvtepi16_epi64(__m128i __V) 298 { 299 return (__m128i) __builtin_ia32_pmovsxwq128((__v8hi)__V); 300 } 301 302 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 303 _mm_cvtepi32_epi64(__m128i __V) 304 { 305 return (__m128i) __builtin_ia32_pmovsxdq128((__v4si)__V); 306 } 307 308 /* SSE4 Packed Integer Zero-Extension. */ 309 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 310 _mm_cvtepu8_epi16(__m128i __V) 311 { 312 return (__m128i) __builtin_ia32_pmovzxbw128((__v16qi) __V); 313 } 314 315 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 316 _mm_cvtepu8_epi32(__m128i __V) 317 { 318 return (__m128i) __builtin_ia32_pmovzxbd128((__v16qi)__V); 319 } 320 321 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 322 _mm_cvtepu8_epi64(__m128i __V) 323 { 324 return (__m128i) __builtin_ia32_pmovzxbq128((__v16qi)__V); 325 } 326 327 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 328 _mm_cvtepu16_epi32(__m128i __V) 329 { 330 return (__m128i) __builtin_ia32_pmovzxwd128((__v8hi)__V); 331 } 332 333 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 334 _mm_cvtepu16_epi64(__m128i __V) 335 { 336 return (__m128i) __builtin_ia32_pmovzxwq128((__v8hi)__V); 337 } 338 339 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 340 _mm_cvtepu32_epi64(__m128i __V) 341 { 342 return (__m128i) __builtin_ia32_pmovzxdq128((__v4si)__V); 343 } 344 345 /* SSE4 Pack with Unsigned Saturation. */ 346 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 347 _mm_packus_epi32(__m128i __V1, __m128i __V2) 348 { 349 return (__m128i) __builtin_ia32_packusdw128((__v4si)__V1, (__v4si)__V2); 350 } 351 352 /* SSE4 Multiple Packed Sums of Absolute Difference. */ 353 #define _mm_mpsadbw_epu8(X, Y, M) __extension__ ({ \ 354 __m128i __X = (X); \ 355 __m128i __Y = (Y); \ 356 (__m128i) __builtin_ia32_mpsadbw128((__v16qi)__X, (__v16qi)__Y, (M)); }) 357 358 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 359 _mm_minpos_epu16(__m128i __V) 360 { 361 return (__m128i) __builtin_ia32_phminposuw128((__v8hi)__V); 362 } 363 364 /* These definitions are normally in nmmintrin.h, but gcc puts them in here 365 so we'll do the same. */ 366 #ifdef __SSE4_2__ 367 368 /* These specify the type of data that we're comparing. */ 369 #define _SIDD_UBYTE_OPS 0x00 370 #define _SIDD_UWORD_OPS 0x01 371 #define _SIDD_SBYTE_OPS 0x02 372 #define _SIDD_SWORD_OPS 0x03 373 374 /* These specify the type of comparison operation. */ 375 #define _SIDD_CMP_EQUAL_ANY 0x00 376 #define _SIDD_CMP_RANGES 0x04 377 #define _SIDD_CMP_EQUAL_EACH 0x08 378 #define _SIDD_CMP_EQUAL_ORDERED 0x0c 379 380 /* These macros specify the polarity of the operation. */ 381 #define _SIDD_POSITIVE_POLARITY 0x00 382 #define _SIDD_NEGATIVE_POLARITY 0x10 383 #define _SIDD_MASKED_POSITIVE_POLARITY 0x20 384 #define _SIDD_MASKED_NEGATIVE_POLARITY 0x30 385 386 /* These macros are used in _mm_cmpXstri() to specify the return. */ 387 #define _SIDD_LEAST_SIGNIFICANT 0x00 388 #define _SIDD_MOST_SIGNIFICANT 0x40 389 390 /* These macros are used in _mm_cmpXstri() to specify the return. */ 391 #define _SIDD_BIT_MASK 0x00 392 #define _SIDD_UNIT_MASK 0x40 393 394 /* SSE4.2 Packed Comparison Intrinsics. */ 395 #define _mm_cmpistrm(A, B, M) __builtin_ia32_pcmpistrm128((A), (B), (M)) 396 #define _mm_cmpistri(A, B, M) __builtin_ia32_pcmpistri128((A), (B), (M)) 397 398 #define _mm_cmpestrm(A, LA, B, LB, M) \ 399 __builtin_ia32_pcmpestrm128((A), (LA), (B), (LB), (M)) 400 #define _mm_cmpestri(A, LA, B, LB, M) \ 401 __builtin_ia32_pcmpestri128((A), (LA), (B), (LB), (M)) 402 403 /* SSE4.2 Packed Comparison Intrinsics and EFlag Reading. */ 404 #define _mm_cmpistra(A, B, M) \ 405 __builtin_ia32_pcmpistria128((A), (B), (M)) 406 #define _mm_cmpistrc(A, B, M) \ 407 __builtin_ia32_pcmpistric128((A), (B), (M)) 408 #define _mm_cmpistro(A, B, M) \ 409 __builtin_ia32_pcmpistrio128((A), (B), (M)) 410 #define _mm_cmpistrs(A, B, M) \ 411 __builtin_ia32_pcmpistris128((A), (B), (M)) 412 #define _mm_cmpistrz(A, B, M) \ 413 __builtin_ia32_pcmpistriz128((A), (B), (M)) 414 415 #define _mm_cmpestra(A, LA, B, LB, M) \ 416 __builtin_ia32_pcmpestria128((A), (LA), (B), (LB), (M)) 417 #define _mm_cmpestrc(A, LA, B, LB, M) \ 418 __builtin_ia32_pcmpestric128((A), (LA), (B), (LB), (M)) 419 #define _mm_cmpestro(A, LA, B, LB, M) \ 420 __builtin_ia32_pcmpestrio128((A), (LA), (B), (LB), (M)) 421 #define _mm_cmpestrs(A, LA, B, LB, M) \ 422 __builtin_ia32_pcmpestris128((A), (LA), (B), (LB), (M)) 423 #define _mm_cmpestrz(A, LA, B, LB, M) \ 424 __builtin_ia32_pcmpestriz128((A), (LA), (B), (LB), (M)) 425 426 /* SSE4.2 Compare Packed Data -- Greater Than. */ 427 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 428 _mm_cmpgt_epi64(__m128i __V1, __m128i __V2) 429 { 430 return (__m128i)((__v2di)__V1 > (__v2di)__V2); 431 } 432 433 /* SSE4.2 Accumulate CRC32. */ 434 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 435 _mm_crc32_u8(unsigned int __C, unsigned char __D) 436 { 437 return __builtin_ia32_crc32qi(__C, __D); 438 } 439 440 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 441 _mm_crc32_u16(unsigned int __C, unsigned short __D) 442 { 443 return __builtin_ia32_crc32hi(__C, __D); 444 } 445 446 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 447 _mm_crc32_u32(unsigned int __C, unsigned int __D) 448 { 449 return __builtin_ia32_crc32si(__C, __D); 450 } 451 452 #ifdef __x86_64__ 453 static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__)) 454 _mm_crc32_u64(unsigned long long __C, unsigned long long __D) 455 { 456 return __builtin_ia32_crc32di(__C, __D); 457 } 458 #endif /* __x86_64__ */ 459 460 #ifdef __POPCNT__ 461 #include <popcntintrin.h> 462 #endif 463 464 #endif /* __SSE4_2__ */ 465 #endif /* __SSE4_1__ */ 466 467 #endif /* _SMMINTRIN_H */ 468
