| 1 | /* Copyright (C) 2013 Povilas Kanapickas <povilas@radix.lt> |
| 2 | |
| 3 | Distributed under the Boost Software License, Version 1.0. |
| 4 | (See accompanying file LICENSE_1_0.txt or copy at |
| 5 | http://www.boost.org/LICENSE_1_0.txt) |
| 6 | */ |
| 7 | |
| 8 | #ifndef LIBSIMDPP_SIMD_SETUP_ARCH_H |
| 9 | #define LIBSIMDPP_SIMD_SETUP_ARCH_H |
| 10 | |
| 11 | #include <simdpp/detail/preprocessor.h> |
| 12 | #include <simdpp/detail/preprocessor/stringize.hpp> |
| 13 | |
| 14 | // Set up macros for current architecture. Note that this file may be included |
| 15 | // multiple times, more information on the caveats are within the file. |
| 16 | #include <simdpp/detail/preprocess_single_arch.h> |
| 17 | |
| 18 | // Set up main feature macros |
| 19 | #if SIMDPP_ARCH_PP_USE_NULL |
| 20 | #define SIMDPP_USE_NULL 1 |
| 21 | #else |
| 22 | #define SIMDPP_USE_NULL 0 |
| 23 | #endif |
| 24 | #if SIMDPP_ARCH_PP_USE_SSE2 |
| 25 | #define SIMDPP_USE_SSE2 1 |
| 26 | #else |
| 27 | #define SIMDPP_USE_SSE2 0 |
| 28 | #endif |
| 29 | #if SIMDPP_ARCH_PP_USE_SSE3 |
| 30 | #define SIMDPP_USE_SSE3 1 |
| 31 | #else |
| 32 | #define SIMDPP_USE_SSE3 0 |
| 33 | #endif |
| 34 | #if SIMDPP_ARCH_PP_USE_SSSE3 |
| 35 | #define SIMDPP_USE_SSSE3 1 |
| 36 | #else |
| 37 | #define SIMDPP_USE_SSSE3 0 |
| 38 | #endif |
| 39 | #if SIMDPP_ARCH_PP_USE_SSE4_1 |
| 40 | #define SIMDPP_USE_SSE4_1 1 |
| 41 | #else |
| 42 | #define SIMDPP_USE_SSE4_1 0 |
| 43 | #endif |
| 44 | #if SIMDPP_ARCH_PP_USE_X86_POPCNT_INSN |
| 45 | #define SIMDPP_USE_X86_POPCNT_INSN 1 |
| 46 | #else |
| 47 | #define SIMDPP_USE_X86_POPCNT_INSN 0 |
| 48 | #endif |
| 49 | #if SIMDPP_ARCH_PP_USE_AVX |
| 50 | #define SIMDPP_USE_AVX 1 |
| 51 | #else |
| 52 | #define SIMDPP_USE_AVX 0 |
| 53 | #endif |
| 54 | #if SIMDPP_ARCH_PP_USE_AVX2 |
| 55 | #define SIMDPP_USE_AVX2 1 |
| 56 | #else |
| 57 | #define SIMDPP_USE_AVX2 0 |
| 58 | #endif |
| 59 | #if SIMDPP_ARCH_PP_USE_FMA3 |
| 60 | #define SIMDPP_USE_FMA3 1 |
| 61 | #else |
| 62 | #define SIMDPP_USE_FMA3 0 |
| 63 | #endif |
| 64 | #if SIMDPP_ARCH_PP_USE_FMA4 |
| 65 | #define SIMDPP_USE_FMA4 1 |
| 66 | #else |
| 67 | #define SIMDPP_USE_FMA4 0 |
| 68 | #endif |
| 69 | #if SIMDPP_ARCH_PP_USE_XOP |
| 70 | #define SIMDPP_USE_XOP 1 |
| 71 | #else |
| 72 | #define SIMDPP_USE_XOP 0 |
| 73 | #endif |
| 74 | #if SIMDPP_ARCH_PP_USE_AVX512F |
| 75 | #define SIMDPP_USE_AVX512F 1 |
| 76 | #else |
| 77 | #define SIMDPP_USE_AVX512F 0 |
| 78 | #endif |
| 79 | #if SIMDPP_ARCH_PP_USE_AVX512BW |
| 80 | #define SIMDPP_USE_AVX512BW 1 |
| 81 | #else |
| 82 | #define SIMDPP_USE_AVX512BW 0 |
| 83 | #endif |
| 84 | #if SIMDPP_ARCH_PP_USE_AVX512DQ |
| 85 | #define SIMDPP_USE_AVX512DQ 1 |
| 86 | #else |
| 87 | #define SIMDPP_USE_AVX512DQ 0 |
| 88 | #endif |
| 89 | #if SIMDPP_ARCH_PP_USE_AVX512VL |
| 90 | #define SIMDPP_USE_AVX512VL 1 |
| 91 | #else |
| 92 | #define SIMDPP_USE_AVX512VL 0 |
| 93 | #endif |
| 94 | #if SIMDPP_ARCH_PP_USE_NEON |
| 95 | #define SIMDPP_USE_NEON 1 |
| 96 | #else |
| 97 | #define SIMDPP_USE_NEON 0 |
| 98 | #endif |
| 99 | #if SIMDPP_ARCH_PP_USE_NEON_FLT_SP |
| 100 | #define SIMDPP_USE_NEON_FLT_SP 1 |
| 101 | #else |
| 102 | #define SIMDPP_USE_NEON_FLT_SP 0 |
| 103 | #endif |
| 104 | #if SIMDPP_ARCH_PP_USE_ALTIVEC |
| 105 | #define SIMDPP_USE_ALTIVEC 1 |
| 106 | #else |
| 107 | #define SIMDPP_USE_ALTIVEC 0 |
| 108 | #endif |
| 109 | #if SIMDPP_ARCH_PP_USE_VSX_206 |
| 110 | #define SIMDPP_USE_VSX_206 1 |
| 111 | #else |
| 112 | #define SIMDPP_USE_VSX_206 0 |
| 113 | #endif |
| 114 | #if SIMDPP_ARCH_PP_USE_VSX_207 |
| 115 | #define SIMDPP_USE_VSX_207 1 |
| 116 | #else |
| 117 | #define SIMDPP_USE_VSX_207 0 |
| 118 | #endif |
| 119 | #if SIMDPP_ARCH_PP_USE_MSA |
| 120 | #define SIMDPP_USE_MSA 1 |
| 121 | #else |
| 122 | #define SIMDPP_USE_MSA 0 |
| 123 | #endif |
| 124 | |
| 125 | // Generate SIMDPP_ARCH_NAMESPACE. It's a human-readable identifier depending |
| 126 | // on the enabled instruction sets |
| 127 | #if SIMDPP_ARCH_PP_NS_USE_NULL |
| 128 | #define SIMDPP_NS_ID_NULL SIMDPP_INSN_ID_NULL |
| 129 | #else |
| 130 | #define SIMDPP_NS_ID_NULL |
| 131 | #endif |
| 132 | #if SIMDPP_ARCH_PP_NS_USE_SSE2 |
| 133 | #define SIMDPP_NS_ID_SSE2 SIMDPP_INSN_ID_SSE2 |
| 134 | #else |
| 135 | #define SIMDPP_NS_ID_SSE2 |
| 136 | #endif |
| 137 | #if SIMDPP_ARCH_PP_NS_USE_SSE3 |
| 138 | #define SIMDPP_NS_ID_SSE3 SIMDPP_INSN_ID_SSE3 |
| 139 | #else |
| 140 | #define SIMDPP_NS_ID_SSE3 |
| 141 | #endif |
| 142 | #if SIMDPP_ARCH_PP_NS_USE_SSSE3 |
| 143 | #define SIMDPP_NS_ID_SSSE3 SIMDPP_INSN_ID_SSSE3 |
| 144 | #else |
| 145 | #define SIMDPP_NS_ID_SSSE3 |
| 146 | #endif |
| 147 | #if SIMDPP_ARCH_PP_NS_USE_SSE4_1 |
| 148 | #define SIMDPP_NS_ID_SSE4_1 SIMDPP_INSN_ID_SSE4_1 |
| 149 | #else |
| 150 | #define SIMDPP_NS_ID_SSE4_1 |
| 151 | #endif |
| 152 | #if SIMDPP_ARCH_PP_NS_USE_POPCNT_INSN |
| 153 | #define SIMDPP_NS_ID_POPCNT_INSN SIMDPP_INSN_ID_POPCNT_INSN |
| 154 | #else |
| 155 | #define SIMDPP_NS_ID_POPCNT_INSN |
| 156 | #endif |
| 157 | #if SIMDPP_ARCH_PP_NS_USE_AVX |
| 158 | #define SIMDPP_NS_ID_AVX SIMDPP_INSN_ID_AVX |
| 159 | #else |
| 160 | #define SIMDPP_NS_ID_AVX |
| 161 | #endif |
| 162 | #if SIMDPP_ARCH_PP_NS_USE_AVX2 |
| 163 | #define SIMDPP_NS_ID_AVX2 SIMDPP_INSN_ID_AVX2 |
| 164 | #else |
| 165 | #define SIMDPP_NS_ID_AVX2 |
| 166 | #endif |
| 167 | #if SIMDPP_ARCH_PP_NS_USE_FMA3 |
| 168 | #define SIMDPP_NS_ID_FMA3 SIMDPP_INSN_ID_FMA3 |
| 169 | #else |
| 170 | #define SIMDPP_NS_ID_FMA3 |
| 171 | #endif |
| 172 | #if SIMDPP_ARCH_PP_NS_USE_FMA4 |
| 173 | #define SIMDPP_NS_ID_FMA4 SIMDPP_INSN_ID_FMA4 |
| 174 | #else |
| 175 | #define SIMDPP_NS_ID_FMA4 |
| 176 | #endif |
| 177 | #if SIMDPP_ARCH_PP_NS_USE_XOP |
| 178 | #define SIMDPP_NS_ID_XOP SIMDPP_INSN_ID_XOP |
| 179 | #else |
| 180 | #define SIMDPP_NS_ID_XOP |
| 181 | #endif |
| 182 | #if SIMDPP_ARCH_PP_NS_USE_AVX512F |
| 183 | #define SIMDPP_NS_ID_AVX512F SIMDPP_INSN_ID_AVX512F |
| 184 | #else |
| 185 | #define SIMDPP_NS_ID_AVX512F |
| 186 | #endif |
| 187 | #if SIMDPP_ARCH_PP_NS_USE_AVX512BW |
| 188 | #define SIMDPP_NS_ID_AVX512BW SIMDPP_INSN_ID_AVX512BW |
| 189 | #else |
| 190 | #define SIMDPP_NS_ID_AVX512BW |
| 191 | #endif |
| 192 | #if SIMDPP_ARCH_PP_NS_USE_AVX512DQ |
| 193 | #define SIMDPP_NS_ID_AVX512DQ SIMDPP_INSN_ID_AVX512DQ |
| 194 | #else |
| 195 | #define SIMDPP_NS_ID_AVX512DQ |
| 196 | #endif |
| 197 | #if SIMDPP_ARCH_PP_NS_USE_AVX512VL |
| 198 | #define SIMDPP_NS_ID_AVX512VL SIMDPP_INSN_ID_AVX512VL |
| 199 | #else |
| 200 | #define SIMDPP_NS_ID_AVX512VL |
| 201 | #endif |
| 202 | #if SIMDPP_ARCH_PP_NS_USE_NEON |
| 203 | #define SIMDPP_NS_ID_NEON SIMDPP_INSN_ID_NEON |
| 204 | #else |
| 205 | #define SIMDPP_NS_ID_NEON |
| 206 | #endif |
| 207 | #if SIMDPP_ARCH_PP_NS_USE_NEON_FLT_SP |
| 208 | #define SIMDPP_NS_ID_NEON_FLT_SP SIMDPP_INSN_ID_NEON_FLT_SP |
| 209 | #else |
| 210 | #define SIMDPP_NS_ID_NEON_FLT_SP |
| 211 | #endif |
| 212 | #if SIMDPP_ARCH_PP_NS_USE_ALTIVEC |
| 213 | #define SIMDPP_NS_ID_ALTIVEC SIMDPP_INSN_ID_ALTIVEC |
| 214 | #else |
| 215 | #define SIMDPP_NS_ID_ALTIVEC |
| 216 | #endif |
| 217 | #if SIMDPP_ARCH_PP_NS_USE_VSX_206 |
| 218 | #define SIMDPP_NS_ID_VSX_206 SIMDPP_INSN_ID_VSX_206 |
| 219 | #else |
| 220 | #define SIMDPP_NS_ID_VSX_206 |
| 221 | #endif |
| 222 | #if SIMDPP_ARCH_PP_NS_USE_VSX_207 |
| 223 | #define SIMDPP_NS_ID_VSX_207 SIMDPP_INSN_ID_VSX_207 |
| 224 | #else |
| 225 | #define SIMDPP_NS_ID_VSX_207 |
| 226 | #endif |
| 227 | #if SIMDPP_ARCH_PP_NS_USE_MSA |
| 228 | #define SIMDPP_NS_ID_MSA SIMDPP_INSN_ID_MSA |
| 229 | #else |
| 230 | #define SIMDPP_NS_ID_MSA |
| 231 | #endif |
| 232 | |
| 233 | #define SIMDPP_ARCH_NAMESPACE SIMDPP_PP_PASTE22(arch, \ |
| 234 | SIMDPP_NS_ID_NULL, \ |
| 235 | SIMDPP_NS_ID_SSE2, \ |
| 236 | SIMDPP_NS_ID_SSE3, \ |
| 237 | SIMDPP_NS_ID_SSSE3, \ |
| 238 | SIMDPP_NS_ID_SSE4_1, \ |
| 239 | SIMDPP_NS_ID_POPCNT_INSN, \ |
| 240 | SIMDPP_NS_ID_AVX, \ |
| 241 | SIMDPP_NS_ID_AVX2, \ |
| 242 | SIMDPP_NS_ID_AVX512F, \ |
| 243 | SIMDPP_NS_ID_AVX512BW, \ |
| 244 | SIMDPP_NS_ID_AVX512DQ, \ |
| 245 | SIMDPP_NS_ID_AVX512VL, \ |
| 246 | SIMDPP_NS_ID_FMA3, \ |
| 247 | SIMDPP_NS_ID_FMA4, \ |
| 248 | SIMDPP_NS_ID_XOP, \ |
| 249 | SIMDPP_NS_ID_NEON, \ |
| 250 | SIMDPP_NS_ID_NEON_FLT_SP, \ |
| 251 | SIMDPP_NS_ID_MSA, \ |
| 252 | SIMDPP_NS_ID_ALTIVEC, \ |
| 253 | SIMDPP_NS_ID_VSX_206, \ |
| 254 | SIMDPP_NS_ID_VSX_207) |
| 255 | |
| 256 | #define SIMDPP_ARCH_PP_THIS_COMPILE_ARCH_NAMESPACE SIMDPP_ARCH_NAMESPACE |
| 257 | #define SIMDPP_ARCH_PP_THIS_COMPILE_ARCH_FOR_DISPATCH 0 |
| 258 | #include <simdpp/dispatch/preprocess_single_compile_arch.h> |
| 259 | #undef SIMDPP_ARCH_PP_THIS_COMPILE_ARCH_FOR_DISPATCH |
| 260 | #undef SIMDPP_ARCH_PP_THIS_COMPILE_ARCH_NAMESPACE |
| 261 | |
| 262 | // Include headers relevant for the enabled instruction sets. |
| 263 | #if SIMDPP_USE_SSE2 |
| 264 | #include <xmmintrin.h> |
| 265 | #include <emmintrin.h> |
| 266 | #endif |
| 267 | |
| 268 | #if SIMDPP_USE_SSE3 |
| 269 | #include <pmmintrin.h> |
| 270 | #endif |
| 271 | |
| 272 | #if SIMDPP_USE_SSSE3 |
| 273 | #include <tmmintrin.h> |
| 274 | #endif |
| 275 | |
| 276 | #if SIMDPP_USE_SSE4_1 |
| 277 | #include <smmintrin.h> |
| 278 | #endif |
| 279 | |
| 280 | #if SIMDPP_USE_AVX |
| 281 | #include <immintrin.h> |
| 282 | #endif |
| 283 | |
| 284 | #if SIMDPP_USE_AVX2 |
| 285 | #include <immintrin.h> |
| 286 | #endif |
| 287 | |
| 288 | #if SIMDPP_USE_FMA3 |
| 289 | #include <immintrin.h> |
| 290 | #endif |
| 291 | |
| 292 | #if SIMDPP_USE_FMA4 |
| 293 | #include <x86intrin.h> |
| 294 | #if SIMDPP_USE_FMA3 |
| 295 | #error "X86_FMA3 and X86_FMA4 can't be used together" |
| 296 | #endif |
| 297 | #endif |
| 298 | |
| 299 | #if SIMDPP_USE_XOP |
| 300 | #include <x86intrin.h> |
| 301 | #endif |
| 302 | |
| 303 | #if SIMDPP_USE_AVX512F || SIMDPP_USE_AVX512BW |
| 304 | #include <immintrin.h> |
| 305 | #endif |
| 306 | |
| 307 | #if SIMDPP_USE_NEON || SIMDPP_USE_NEON_FLT_SP |
| 308 | #include <arm_neon.h> |
| 309 | #endif |
| 310 | |
| 311 | #if SIMDPP_USE_ALTIVEC |
| 312 | #include <altivec.h> |
| 313 | #undef vector |
| 314 | #undef pixel |
| 315 | #undef bool |
| 316 | #endif |
| 317 | |
| 318 | #if SIMDPP_USE_MSA |
| 319 | #include <msa.h> |
| 320 | #endif |
| 321 | |
| 322 | // helper macros |
| 323 | #if __amd64__ || __x86_64__ || _M_AMD64 || __aarch64__ || __powerpc64__ |
| 324 | #define SIMDPP_64_BITS 1 |
| 325 | #define SIMDPP_32_BITS 0 |
| 326 | #else |
| 327 | #define SIMDPP_32_BITS 1 |
| 328 | #define SIMDPP_64_BITS 0 |
| 329 | #endif |
| 330 | |
| 331 | #if SIMDPP_USE_NEON && SIMDPP_64_BITS |
| 332 | #undef SIMDPP_USE_NEON_FLT_SP |
| 333 | #define SIMDPP_USE_NEON_FLT_SP 1 |
| 334 | #endif |
| 335 | |
| 336 | #if SIMDPP_USE_ALTIVEC |
| 337 | #ifndef __BYTE_ORDER__ |
| 338 | #error "Could not determine byte order" |
| 339 | #elif __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
| 340 | #define SIMDPP_LITTLE_ENDIAN 1 |
| 341 | #define SIMDPP_BIG_ENDIAN 0 |
| 342 | #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
| 343 | #define SIMDPP_LITTLE_ENDIAN 0 |
| 344 | #define SIMDPP_BIG_ENDIAN 1 |
| 345 | #else |
| 346 | #error "Could not determine byte order" |
| 347 | #endif |
| 348 | #endif |
| 349 | |
| 350 | #define SIMDPP_USE_NEON32 (SIMDPP_USE_NEON && SIMDPP_32_BITS) |
| 351 | #define SIMDPP_USE_NEON64 (SIMDPP_USE_NEON && SIMDPP_64_BITS) |
| 352 | #define SIMDPP_USE_NEON32_FLT_SP (SIMDPP_USE_NEON_FLT_SP && SIMDPP_32_BITS) |
| 353 | #define SIMDPP_USE_NEON_NO_FLT_SP (SIMDPP_USE_NEON && !SIMDPP_USE_NEON_FLT_SP) |
| 354 | |
| 355 | #if __i386__ || __i386 || _M_IX86 || __amd64__ || __x64_64__ || _M_AMD64 || _M_X64 |
| 356 | #define SIMDPP_X86 1 |
| 357 | #elif _M_ARM || __arm__ || __aarch64__ |
| 358 | #define SIMDPP_ARM 1 |
| 359 | #elif __powerpc__ || __powerpc64__ |
| 360 | #define SIMDPP_PPC 1 |
| 361 | #elif __mips__ |
| 362 | #define SIMDPP_MIPS 1 |
| 363 | #endif |
| 364 | |
| 365 | /** @def SIMDPP_ARCH_NAME |
| 366 | Usable in contexts where a string is required |
| 367 | */ |
| 368 | #define SIMDPP_ARCH_NAME SIMDPP_PP_STRINGIZE(SIMDPP_ARCH_NAMESPACE) |
| 369 | |
| 370 | // misc macros |
| 371 | #if __GNUC__ |
| 372 | #define SIMDPP_INL __attribute__((__always_inline__)) inline |
| 373 | #elif _MSC_VER |
| 374 | #define SIMDPP_INL __forceinline |
| 375 | #else |
| 376 | #define SIMDPP_INL inline |
| 377 | #endif |
| 378 | |
| 379 | #if defined(__GNUC__) || defined(__clang__) |
| 380 | #define SIMDPP_DEPRECATED(msg) __attribute__ ((deprecated(msg))) |
| 381 | #else |
| 382 | #define SIMDPP_DEPRECATED(msg) |
| 383 | #endif |
| 384 | |
| 385 | #if __GNUC__ |
| 386 | #define SIMDPP_ALIGN(X) __attribute__((__aligned__(X))) |
| 387 | #elif _MSC_VER |
| 388 | #define SIMDPP_ALIGN(X) __declspec(align(X)) |
| 389 | #else |
| 390 | #error "Unsupported compiler" |
| 391 | #endif |
| 392 | |
| 393 | #define SIMDPP_LIBRARY_VERSION_CXX11 1 |
| 394 | #define SIMDPP_LIBRARY_VERSION_CXX98 0 |
| 395 | |
| 396 | #include <simdpp/detail/workarounds.h> |
| 397 | #include <simdpp/deprecations.h> |
| 398 | |
| 399 | // #define SIMDPP_EXPR_DEBUG 1 |
| 400 | |
| 401 | // FIXME: unused (workarounds for AMD CPUs) |
| 402 | // #define SIMDPP_USE_AMD |
| 403 | |
| 404 | #endif |
| 405 | |