48 #define Vc_ICC __INTEL_COMPILER_BUILD_DATE 57 #define Vc_CLANG (__clang_major__ * 0x10000 + __clang_minor__ * 0x100 + __clang_patchlevel__) 66 #define Vc_APPLECLANG (__clang_major__ * 0x10000 + __clang_minor__ * 0x100 + __clang_patchlevel__) 75 #define Vc_GCC (__GNUC__ * 0x10000 + __GNUC_MINOR__ * 0x100 + __GNUC_PATCHLEVEL__) 83 #define Vc_MSVC _MSC_FULL_VER 90 #ifdef __INTEL_COMPILER 91 #define Vc_ICC __INTEL_COMPILER_BUILD_DATE 92 #elif defined(__clang__) && defined(__apple_build_version__) 93 #define Vc_APPLECLANG (__clang_major__ * 0x10000 + __clang_minor__ * 0x100 + __clang_patchlevel__) 94 #elif defined(__clang__) 95 #define Vc_CLANG (__clang_major__ * 0x10000 + __clang_minor__ * 0x100 + __clang_patchlevel__) 96 #elif defined(__GNUC__) 97 #define Vc_GCC (__GNUC__ * 0x10000 + __GNUC_MINOR__ * 0x100 + __GNUC_PATCHLEVEL__) 98 #elif defined(_MSC_VER) 99 #define Vc_MSVC _MSC_FULL_VER 101 #define Vc_UNSUPPORTED_COMPILER 1 104 #if defined Vc_GCC && Vc_GCC >= 0x60000 105 #define Vc_RESET_DIAGNOSTICS _Pragma("GCC diagnostic pop") 106 #pragma GCC diagnostic push 107 #pragma GCC diagnostic ignored "-Wignored-attributes" 109 #define Vc_RESET_DIAGNOSTICS 118 #pragma warning disable 2922 121 #if __cplusplus < 201103 && (!defined Vc_MSVC || _MSC_VER < 1900) 122 # error "Vc requires support for C++11." 123 #elif __cplusplus >= 201402L 125 # if __cplusplus > 201700L 130 #if defined(__GNUC__) && !defined(Vc_NO_INLINE_ASM) 135 # if Vc_GCC >= 0x70000 && defined __i386__ 139 # ifdef __GLIBC_PREREQ 140 # if __GLIBC_PREREQ(2,26) 141 # define Vc_HAVE_STD_MAX_ALIGN_T 1 144 # elif Vc_GCC >= 0x40900 145 # define Vc_HAVE_STD_MAX_ALIGN_T 1 147 # define Vc_HAVE_MAX_ALIGN_T 1 149 #elif !defined(Vc_CLANG) && !defined(Vc_ICC) 153 # define Vc_HAVE_STD_MAX_ALIGN_T 1 156 #if defined(Vc_GCC) || defined(Vc_CLANG) || defined Vc_APPLECLANG 157 #define Vc_USE_BUILTIN_VECTOR_TYPES 1 161 # define Vc_CDECL __cdecl 162 # define Vc_VDECL __vectorcall 172 #define Scalar 0x00100000 173 #define SSE 0x00200000 174 #define SSE2 0x00300000 175 #define SSE3 0x00400000 176 #define SSSE3 0x00500000 177 #define SSE4_1 0x00600000 178 #define SSE4_2 0x00700000 179 #define AVX 0x00800000 180 #define AVX2 0x00900000 182 #define XOP 0x00000001 183 #define FMA4 0x00000002 184 #define F16C 0x00000004 185 #define POPCNT 0x00000008 186 #define SSE4a 0x00000010 187 #define FMA 0x00000020 188 #define BMI2 0x00000040 190 #define IMPL_MASK 0xFFF00000 191 #define EXT_MASK 0x000FFFFF 205 # elif defined(_M_AMD64) 216 #if defined Vc_ICC && !defined __POPCNT__ 217 # if defined __SSE4_2__ || defined __SSE4A__ 218 # define __POPCNT__ 1 223 #error "You are using the old VC_IMPL macro. Since Vc 1.0 all Vc macros start with Vc_, i.e. a lower-case 'c'" 228 # if defined(__AVX2__) 229 # define Vc_IMPL_AVX2 1 230 # define Vc_IMPL_AVX 1 231 # elif defined(__AVX__) 232 # define Vc_IMPL_AVX 1 234 # if defined(__SSE4_2__) 235 # define Vc_IMPL_SSE 1 236 # define Vc_IMPL_SSE4_2 1 238 # if defined(__SSE4_1__) 239 # define Vc_IMPL_SSE 1 240 # define Vc_IMPL_SSE4_1 1 242 # if defined(__SSE3__) 243 # define Vc_IMPL_SSE 1 244 # define Vc_IMPL_SSE3 1 246 # if defined(__SSSE3__) 247 # define Vc_IMPL_SSE 1 248 # define Vc_IMPL_SSSE3 1 250 # if defined(__SSE2__) 251 # define Vc_IMPL_SSE 1 252 # define Vc_IMPL_SSE2 1 255 # if defined(Vc_IMPL_SSE) 258 # define Vc_IMPL_Scalar 1 261 # if !defined(Vc_IMPL_Scalar) 263 # define Vc_IMPL_FMA4 1 266 # define Vc_IMPL_XOP 1 269 # define Vc_IMPL_F16C 1 272 # define Vc_IMPL_POPCNT 1 275 # define Vc_IMPL_SSE4a 1 278 # define Vc_IMPL_FMA 1 281 # define Vc_IMPL_BMI2 1 287 # if (Vc_IMPL & IMPL_MASK) == AVX2 // AVX2 supersedes SSE 288 # define Vc_IMPL_AVX2 1 289 # define Vc_IMPL_AVX 1 290 # elif (Vc_IMPL & IMPL_MASK) == AVX // AVX supersedes SSE 291 # define Vc_IMPL_AVX 1 292 # elif (Vc_IMPL & IMPL_MASK) == Scalar 293 # define Vc_IMPL_Scalar 1 294 # elif (Vc_IMPL & IMPL_MASK) == SSE4_2 295 # define Vc_IMPL_SSE4_2 1 296 # define Vc_IMPL_SSE4_1 1 297 # define Vc_IMPL_SSSE3 1 298 # define Vc_IMPL_SSE3 1 299 # define Vc_IMPL_SSE2 1 300 # define Vc_IMPL_SSE 1 301 # elif (Vc_IMPL & IMPL_MASK) == SSE4_1 302 # define Vc_IMPL_SSE4_1 1 303 # define Vc_IMPL_SSSE3 1 304 # define Vc_IMPL_SSE3 1 305 # define Vc_IMPL_SSE2 1 306 # define Vc_IMPL_SSE 1 307 # elif (Vc_IMPL & IMPL_MASK) == SSSE3 308 # define Vc_IMPL_SSSE3 1 309 # define Vc_IMPL_SSE3 1 310 # define Vc_IMPL_SSE2 1 311 # define Vc_IMPL_SSE 1 312 # elif (Vc_IMPL & IMPL_MASK) == SSE3 313 # define Vc_IMPL_SSE3 1 314 # define Vc_IMPL_SSE2 1 315 # define Vc_IMPL_SSE 1 316 # elif (Vc_IMPL & IMPL_MASK) == SSE2 317 # define Vc_IMPL_SSE2 1 318 # define Vc_IMPL_SSE 1 319 # elif (Vc_IMPL & IMPL_MASK) == SSE 320 # define Vc_IMPL_SSE 1 321 # if defined(__SSE4_2__) 322 # define Vc_IMPL_SSE4_2 1 324 # if defined(__SSE4_1__) 325 # define Vc_IMPL_SSE4_1 1 327 # if defined(__SSE3__) 328 # define Vc_IMPL_SSE3 1 330 # if defined(__SSSE3__) 331 # define Vc_IMPL_SSSE3 1 333 # if defined(__SSE2__) 334 # define Vc_IMPL_SSE2 1 336 # elif (Vc_IMPL & IMPL_MASK) == 0 && (Vc_IMPL & SSE4a) 339 # define Vc_IMPL_SSE3 1 340 # define Vc_IMPL_SSE2 1 341 # define Vc_IMPL_SSE 1 344 # define Vc_IMPL_XOP 1 346 # if (Vc_IMPL & FMA4) 347 # define Vc_IMPL_FMA4 1 349 # if (Vc_IMPL & F16C) 350 # define Vc_IMPL_F16C 1 352 # if (!defined(Vc_IMPL_Scalar) && defined(__POPCNT__)) || (Vc_IMPL & POPCNT) 353 # define Vc_IMPL_POPCNT 1 355 # if (Vc_IMPL & SSE4a) 356 # define Vc_IMPL_SSE4a 1 359 # define Vc_IMPL_FMA 1 361 # if (Vc_IMPL & BMI2) 362 # define Vc_IMPL_BMI2 1 370 # define Vc_USE_VEX_CODING 1 375 # define Vc_IMPL_SSE4_2 1 376 # define Vc_IMPL_SSE4_1 1 377 # define Vc_IMPL_SSSE3 1 378 # define Vc_IMPL_SSE3 1 379 # define Vc_IMPL_SSE2 1 380 # define Vc_IMPL_SSE 1 383 #if defined(Vc_CLANG) && Vc_CLANG >= 0x30600 && Vc_CLANG < 0x30700 384 # if defined(Vc_IMPL_AVX) 385 # warning "clang 3.6.x miscompiles AVX code, frequently losing 50% of the data. Vc will fall back to SSE4 instead." 387 # if defined(Vc_IMPL_AVX2) 393 # if !defined(Vc_IMPL_Scalar) && !defined(Vc_IMPL_SSE) && !defined(Vc_IMPL_AVX) 394 # error "No suitable Vc implementation was selected! Probably Vc_IMPL was set to an invalid value." 395 # elif defined(Vc_IMPL_SSE) && !defined(Vc_IMPL_SSE2) 396 # error "SSE requested but no SSE2 support. Vc needs at least SSE2!" 420 #if defined Vc_IMPL_AVX2 421 #define Vc_DEFAULT_IMPL_AVX2 422 #elif defined Vc_IMPL_AVX 423 #define Vc_DEFAULT_IMPL_AVX 424 #elif defined Vc_IMPL_SSE 425 #define Vc_DEFAULT_IMPL_SSE 426 #elif defined Vc_IMPL_Scalar 427 #define Vc_DEFAULT_IMPL_Scalar 429 #error "Preprocessor logic broken. Please report a bug." 434 namespace Vc_VERSIONED_NAMESPACE
437 typedef signed char int8_t;
438 typedef unsigned char uint8_t;
439 typedef signed short int16_t;
440 typedef unsigned short uint16_t;
441 typedef signed int int32_t;
442 typedef unsigned int uint32_t;
443 typedef signed long long int64_t;
444 typedef unsigned long long uint64_t;
501 ImplementationMask = 0xfff
534 ExtraInstructionsMask = 0xfffff000u
550 return static_cast<Implementation>(Features & ImplementationMask);
555 return static_cast<unsigned int>(impl) == current();
563 return static_cast<unsigned int>(low) <= current() &&
564 static_cast<unsigned int>(high) >= current();
569 static constexpr
bool runs_on(
unsigned int extraInstructions)
571 return (extraInstructions & Features & ExtraInstructionsMask) ==
572 (Features & ExtraInstructionsMask);
582 #ifdef Vc_IMPL_Scalar 584 #elif defined(Vc_IMPL_AVX2) 586 #elif defined(Vc_IMPL_AVX) 588 #elif defined(Vc_IMPL_SSE4_2) 590 #elif defined(Vc_IMPL_SSE4_1) 592 #elif defined(Vc_IMPL_SSSE3) 594 #elif defined(Vc_IMPL_SSE3) 596 #elif defined(Vc_IMPL_SSE2) 608 #ifdef Vc_IMPL_POPCNT 617 #ifdef Vc_USE_VEX_CODING 626 #endif // VC_GLOBAL_H_
ExtraInstructions
The list of available instructions is not easily described by a linear list of instruction sets...
Align on boundary of page sizes (e.g.
static constexpr bool is_between(Implementation low, Implementation high)
Returns whether the current Vc::Implementation implements at least low and at most high...
Support for FMA instructions (3 operand variant)
Implementation
Enum to identify a certain SIMD instruction set.
static constexpr bool runs_on(unsigned int extraInstructions)
Returns whether the current code would run on a CPU providing extraInstructions.
This class identifies the specific implementation Vc uses in the current translation unit in terms of...
Support for BMI2 instructions.
Support for XOP instructions.
MallocAlignment
Enum that specifies the alignment and padding restrictions to use for memory allocation with Vc::mall...
Support for the population count instruction.
Support for SSE4a instructions.
static constexpr bool is(Implementation impl)
Returns whether impl is the current Vc::Implementation.
ImplementationT< > CurrentImplementation
Identifies the Vc implementation used in the current translation unit.
x86 SSE + SSE2 + SSE3 + SSSE3 + SSE4.1 + SSE4.2
Align on boundary of cache line sizes (e.g.
Support for ternary instruction coding (VEX)
x86 SSE + SSE2 + SSE3 + SSSE3
Support for FMA4 instructions.
static constexpr Implementation current()
Returns the currently used Vc::Implementation.
uses only fundamental types
Align on boundary of vector sizes (e.g.
Support for float16 conversions in hardware.
x86 SSE + SSE2 + SSE3 + SSSE3 + SSE4.1