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_shufflevector((__v2df)__V1, (__v2df)__V2, \ 83 (((M) & 0x01) ? 2 : 0), \ 84 (((M) & 0x02) ? 3 : 1)); }) 85 86#define _mm_blend_ps(V1, V2, M) __extension__ ({ \ 87 __m128 __V1 = (V1); \ 88 __m128 __V2 = (V2); \ 89 (__m128)__builtin_shufflevector((__v4sf)__V1, (__v4sf)__V2, \ 90 (((M) & 0x01) ? 4 : 0), \ 91 (((M) & 0x02) ? 5 : 1), \ 92 (((M) & 0x04) ? 6 : 2), \ 93 (((M) & 0x08) ? 7 : 3)); }) 94 95static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) 96_mm_blendv_pd (__m128d __V1, __m128d __V2, __m128d __M) 97{ 98 return (__m128d) __builtin_ia32_blendvpd ((__v2df)__V1, (__v2df)__V2, 99 (__v2df)__M); 100} 101 102static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) 103_mm_blendv_ps (__m128 __V1, __m128 __V2, __m128 __M) 104{ 105 return (__m128) __builtin_ia32_blendvps ((__v4sf)__V1, (__v4sf)__V2, 106 (__v4sf)__M); 107} 108 109static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 110_mm_blendv_epi8 (__m128i __V1, __m128i __V2, __m128i __M) 111{ 112 return (__m128i) __builtin_ia32_pblendvb128 ((__v16qi)__V1, (__v16qi)__V2, 113 (__v16qi)__M); 114} 115 116#define _mm_blend_epi16(V1, V2, M) __extension__ ({ \ 117 __m128i __V1 = (V1); \ 118 __m128i __V2 = (V2); \ 119 (__m128i)__builtin_shufflevector((__v8hi)__V1, (__v8hi)__V2, \ 120 (((M) & 0x01) ? 8 : 0), \ 121 (((M) & 0x02) ? 9 : 1), \ 122 (((M) & 0x04) ? 10 : 2), \ 123 (((M) & 0x08) ? 11 : 3), \ 124 (((M) & 0x10) ? 12 : 4), \ 125 (((M) & 0x20) ? 13 : 5), \ 126 (((M) & 0x40) ? 14 : 6), \ 127 (((M) & 0x80) ? 15 : 7)); }) 128 129/* SSE4 Dword Multiply Instructions. */ 130static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 131_mm_mullo_epi32 (__m128i __V1, __m128i __V2) 132{ 133 return (__m128i) ((__v4si)__V1 * (__v4si)__V2); 134} 135 136static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 137_mm_mul_epi32 (__m128i __V1, __m128i __V2) 138{ 139 return (__m128i) __builtin_ia32_pmuldq128 ((__v4si)__V1, (__v4si)__V2); 140} 141 142/* SSE4 Floating Point Dot Product Instructions. */ 143#define _mm_dp_ps(X, Y, M) __extension__ ({ \ 144 __m128 __X = (X); \ 145 __m128 __Y = (Y); \ 146 (__m128) __builtin_ia32_dpps((__v4sf)__X, (__v4sf)__Y, (M)); }) 147 148#define _mm_dp_pd(X, Y, M) __extension__ ({\ 149 __m128d __X = (X); \ 150 __m128d __Y = (Y); \ 151 (__m128d) __builtin_ia32_dppd((__v2df)__X, (__v2df)__Y, (M)); }) 152 153/* SSE4 Streaming Load Hint Instruction. */ 154static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 155_mm_stream_load_si128 (__m128i *__V) 156{ 157 return (__m128i) __builtin_ia32_movntdqa ((__v2di *) __V); 158} 159 160/* SSE4 Packed Integer Min/Max Instructions. */ 161static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 162_mm_min_epi8 (__m128i __V1, __m128i __V2) 163{ 164 return (__m128i) __builtin_ia32_pminsb128 ((__v16qi) __V1, (__v16qi) __V2); 165} 166 167static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 168_mm_max_epi8 (__m128i __V1, __m128i __V2) 169{ 170 return (__m128i) __builtin_ia32_pmaxsb128 ((__v16qi) __V1, (__v16qi) __V2); 171} 172 173static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 174_mm_min_epu16 (__m128i __V1, __m128i __V2) 175{ 176 return (__m128i) __builtin_ia32_pminuw128 ((__v8hi) __V1, (__v8hi) __V2); 177} 178 179static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 180_mm_max_epu16 (__m128i __V1, __m128i __V2) 181{ 182 return (__m128i) __builtin_ia32_pmaxuw128 ((__v8hi) __V1, (__v8hi) __V2); 183} 184 185static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 186_mm_min_epi32 (__m128i __V1, __m128i __V2) 187{ 188 return (__m128i) __builtin_ia32_pminsd128 ((__v4si) __V1, (__v4si) __V2); 189} 190 191static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 192_mm_max_epi32 (__m128i __V1, __m128i __V2) 193{ 194 return (__m128i) __builtin_ia32_pmaxsd128 ((__v4si) __V1, (__v4si) __V2); 195} 196 197static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 198_mm_min_epu32 (__m128i __V1, __m128i __V2) 199{ 200 return (__m128i) __builtin_ia32_pminud128((__v4si) __V1, (__v4si) __V2); 201} 202 203static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 204_mm_max_epu32 (__m128i __V1, __m128i __V2) 205{ 206 return (__m128i) __builtin_ia32_pmaxud128((__v4si) __V1, (__v4si) __V2); 207} 208 209/* SSE4 Insertion and Extraction from XMM Register Instructions. */ 210#define _mm_insert_ps(X, Y, N) __builtin_ia32_insertps128((X), (Y), (N)) 211#define _mm_extract_ps(X, N) (__extension__ \ 212 ({ union { int __i; float __f; } __t; \ 213 __v4sf __a = (__v4sf)(X); \ 214 __t.__f = __a[(N) & 3]; \ 215 __t.__i;})) 216 217/* Miscellaneous insert and extract macros. */ 218/* Extract a single-precision float from X at index N into D. */ 219#define _MM_EXTRACT_FLOAT(D, X, N) (__extension__ ({ __v4sf __a = (__v4sf)(X); \ 220 (D) = __a[N]; })) 221 222/* Or together 2 sets of indexes (X and Y) with the zeroing bits (Z) to create 223 an index suitable for _mm_insert_ps. */ 224#define _MM_MK_INSERTPS_NDX(X, Y, Z) (((X) << 6) | ((Y) << 4) | (Z)) 225 226/* Extract a float from X at index N into the first index of the return. */ 227#define _MM_PICK_OUT_PS(X, N) _mm_insert_ps (_mm_setzero_ps(), (X), \ 228 _MM_MK_INSERTPS_NDX((N), 0, 0x0e)) 229 230/* Insert int into packed integer array at index. */ 231#define _mm_insert_epi8(X, I, N) (__extension__ ({ __v16qi __a = (__v16qi)(X); \ 232 __a[(N) & 15] = (I); \ 233 __a;})) 234#define _mm_insert_epi32(X, I, N) (__extension__ ({ __v4si __a = (__v4si)(X); \ 235 __a[(N) & 3] = (I); \ 236 __a;})) 237#ifdef __x86_64__ 238#define _mm_insert_epi64(X, I, N) (__extension__ ({ __v2di __a = (__v2di)(X); \ 239 __a[(N) & 1] = (I); \ 240 __a;})) 241#endif /* __x86_64__ */ 242 243/* Extract int from packed integer array at index. This returns the element 244 * as a zero extended value, so it is unsigned. 245 */ 246#define _mm_extract_epi8(X, N) (__extension__ ({ __v16qi __a = (__v16qi)(X); \ 247 (int)(unsigned char) \ 248 __a[(N) & 15];})) 249#define _mm_extract_epi32(X, N) (__extension__ ({ __v4si __a = (__v4si)(X); \ 250 __a[(N) & 3];})) 251#ifdef __x86_64__ 252#define _mm_extract_epi64(X, N) (__extension__ ({ __v2di __a = (__v2di)(X); \ 253 __a[(N) & 1];})) 254#endif /* __x86_64 */ 255 256/* SSE4 128-bit Packed Integer Comparisons. */ 257static __inline__ int __attribute__((__always_inline__, __nodebug__)) 258_mm_testz_si128(__m128i __M, __m128i __V) 259{ 260 return __builtin_ia32_ptestz128((__v2di)__M, (__v2di)__V); 261} 262 263static __inline__ int __attribute__((__always_inline__, __nodebug__)) 264_mm_testc_si128(__m128i __M, __m128i __V) 265{ 266 return __builtin_ia32_ptestc128((__v2di)__M, (__v2di)__V); 267} 268 269static __inline__ int __attribute__((__always_inline__, __nodebug__)) 270_mm_testnzc_si128(__m128i __M, __m128i __V) 271{ 272 return __builtin_ia32_ptestnzc128((__v2di)__M, (__v2di)__V); 273} 274 275#define _mm_test_all_ones(V) _mm_testc_si128((V), _mm_cmpeq_epi32((V), (V))) 276#define _mm_test_mix_ones_zeros(M, V) _mm_testnzc_si128((M), (V)) 277#define _mm_test_all_zeros(M, V) _mm_testz_si128 ((M), (V)) 278 279/* SSE4 64-bit Packed Integer Comparisons. */ 280static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 281_mm_cmpeq_epi64(__m128i __V1, __m128i __V2) 282{ 283 return (__m128i)((__v2di)__V1 == (__v2di)__V2); 284} 285 286/* SSE4 Packed Integer Sign-Extension. */ 287static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 288_mm_cvtepi8_epi16(__m128i __V) 289{ 290 return (__m128i) __builtin_ia32_pmovsxbw128((__v16qi) __V); 291} 292 293static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 294_mm_cvtepi8_epi32(__m128i __V) 295{ 296 return (__m128i) __builtin_ia32_pmovsxbd128((__v16qi) __V); 297} 298 299static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 300_mm_cvtepi8_epi64(__m128i __V) 301{ 302 return (__m128i) __builtin_ia32_pmovsxbq128((__v16qi) __V); 303} 304 305static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 306_mm_cvtepi16_epi32(__m128i __V) 307{ 308 return (__m128i) __builtin_ia32_pmovsxwd128((__v8hi) __V); 309} 310 311static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 312_mm_cvtepi16_epi64(__m128i __V) 313{ 314 return (__m128i) __builtin_ia32_pmovsxwq128((__v8hi)__V); 315} 316 317static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 318_mm_cvtepi32_epi64(__m128i __V) 319{ 320 return (__m128i) __builtin_ia32_pmovsxdq128((__v4si)__V); 321} 322 323/* SSE4 Packed Integer Zero-Extension. */ 324static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 325_mm_cvtepu8_epi16(__m128i __V) 326{ 327 return (__m128i) __builtin_ia32_pmovzxbw128((__v16qi) __V); 328} 329 330static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 331_mm_cvtepu8_epi32(__m128i __V) 332{ 333 return (__m128i) __builtin_ia32_pmovzxbd128((__v16qi)__V); 334} 335 336static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 337_mm_cvtepu8_epi64(__m128i __V) 338{ 339 return (__m128i) __builtin_ia32_pmovzxbq128((__v16qi)__V); 340} 341 342static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 343_mm_cvtepu16_epi32(__m128i __V) 344{ 345 return (__m128i) __builtin_ia32_pmovzxwd128((__v8hi)__V); 346} 347 348static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 349_mm_cvtepu16_epi64(__m128i __V) 350{ 351 return (__m128i) __builtin_ia32_pmovzxwq128((__v8hi)__V); 352} 353 354static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 355_mm_cvtepu32_epi64(__m128i __V) 356{ 357 return (__m128i) __builtin_ia32_pmovzxdq128((__v4si)__V); 358} 359 360/* SSE4 Pack with Unsigned Saturation. */ 361static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 362_mm_packus_epi32(__m128i __V1, __m128i __V2) 363{ 364 return (__m128i) __builtin_ia32_packusdw128((__v4si)__V1, (__v4si)__V2); 365} 366 367/* SSE4 Multiple Packed Sums of Absolute Difference. */ 368#define _mm_mpsadbw_epu8(X, Y, M) __extension__ ({ \ 369 __m128i __X = (X); \ 370 __m128i __Y = (Y); \ 371 (__m128i) __builtin_ia32_mpsadbw128((__v16qi)__X, (__v16qi)__Y, (M)); }) 372 373static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 374_mm_minpos_epu16(__m128i __V) 375{ 376 return (__m128i) __builtin_ia32_phminposuw128((__v8hi)__V); 377} 378 379/* These definitions are normally in nmmintrin.h, but gcc puts them in here 380 so we'll do the same. */ 381#ifdef __SSE4_2__ 382 383/* These specify the type of data that we're comparing. */ 384#define _SIDD_UBYTE_OPS 0x00 385#define _SIDD_UWORD_OPS 0x01 386#define _SIDD_SBYTE_OPS 0x02 387#define _SIDD_SWORD_OPS 0x03 388 389/* These specify the type of comparison operation. */ 390#define _SIDD_CMP_EQUAL_ANY 0x00 391#define _SIDD_CMP_RANGES 0x04 392#define _SIDD_CMP_EQUAL_EACH 0x08 393#define _SIDD_CMP_EQUAL_ORDERED 0x0c 394 395/* These macros specify the polarity of the operation. */ 396#define _SIDD_POSITIVE_POLARITY 0x00 397#define _SIDD_NEGATIVE_POLARITY 0x10 398#define _SIDD_MASKED_POSITIVE_POLARITY 0x20 399#define _SIDD_MASKED_NEGATIVE_POLARITY 0x30 400 401/* These macros are used in _mm_cmpXstri() to specify the return. */ 402#define _SIDD_LEAST_SIGNIFICANT 0x00 403#define _SIDD_MOST_SIGNIFICANT 0x40 404 405/* These macros are used in _mm_cmpXstri() to specify the return. */ 406#define _SIDD_BIT_MASK 0x00 407#define _SIDD_UNIT_MASK 0x40 408 409/* SSE4.2 Packed Comparison Intrinsics. */ 410#define _mm_cmpistrm(A, B, M) __builtin_ia32_pcmpistrm128((A), (B), (M)) 411#define _mm_cmpistri(A, B, M) __builtin_ia32_pcmpistri128((A), (B), (M)) 412 413#define _mm_cmpestrm(A, LA, B, LB, M) \ 414 __builtin_ia32_pcmpestrm128((A), (LA), (B), (LB), (M)) 415#define _mm_cmpestri(A, LA, B, LB, M) \ 416 __builtin_ia32_pcmpestri128((A), (LA), (B), (LB), (M)) 417 418/* SSE4.2 Packed Comparison Intrinsics and EFlag Reading. */ 419#define _mm_cmpistra(A, B, M) \ 420 __builtin_ia32_pcmpistria128((A), (B), (M)) 421#define _mm_cmpistrc(A, B, M) \ 422 __builtin_ia32_pcmpistric128((A), (B), (M)) 423#define _mm_cmpistro(A, B, M) \ 424 __builtin_ia32_pcmpistrio128((A), (B), (M)) 425#define _mm_cmpistrs(A, B, M) \ 426 __builtin_ia32_pcmpistris128((A), (B), (M)) 427#define _mm_cmpistrz(A, B, M) \ 428 __builtin_ia32_pcmpistriz128((A), (B), (M)) 429 430#define _mm_cmpestra(A, LA, B, LB, M) \ 431 __builtin_ia32_pcmpestria128((A), (LA), (B), (LB), (M)) 432#define _mm_cmpestrc(A, LA, B, LB, M) \ 433 __builtin_ia32_pcmpestric128((A), (LA), (B), (LB), (M)) 434#define _mm_cmpestro(A, LA, B, LB, M) \ 435 __builtin_ia32_pcmpestrio128((A), (LA), (B), (LB), (M)) 436#define _mm_cmpestrs(A, LA, B, LB, M) \ 437 __builtin_ia32_pcmpestris128((A), (LA), (B), (LB), (M)) 438#define _mm_cmpestrz(A, LA, B, LB, M) \ 439 __builtin_ia32_pcmpestriz128((A), (LA), (B), (LB), (M)) 440 441/* SSE4.2 Compare Packed Data -- Greater Than. */ 442static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) 443_mm_cmpgt_epi64(__m128i __V1, __m128i __V2) 444{ 445 return (__m128i)((__v2di)__V1 > (__v2di)__V2); 446} 447 448/* SSE4.2 Accumulate CRC32. */ 449static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 450_mm_crc32_u8(unsigned int __C, unsigned char __D) 451{ 452 return __builtin_ia32_crc32qi(__C, __D); 453} 454 455static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 456_mm_crc32_u16(unsigned int __C, unsigned short __D) 457{ 458 return __builtin_ia32_crc32hi(__C, __D); 459} 460 461static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__)) 462_mm_crc32_u32(unsigned int __C, unsigned int __D) 463{ 464 return __builtin_ia32_crc32si(__C, __D); 465} 466 467#ifdef __x86_64__ 468static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__)) 469_mm_crc32_u64(unsigned long long __C, unsigned long long __D) 470{ 471 return __builtin_ia32_crc32di(__C, __D); 472} 473#endif /* __x86_64__ */ 474 475#ifdef __POPCNT__ 476#include <popcntintrin.h> 477#endif 478 479#endif /* __SSE4_2__ */ 480#endif /* __SSE4_1__ */ 481 482#endif /* _SMMINTRIN_H */ 483