Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | // you may not use this file except in compliance with the License. |
| 5 | // You may obtain a copy of the License at |
| 6 | // |
| 7 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | // |
| 9 | // Unless required by applicable law or agreed to in writing, software |
| 10 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | // See the License for the specific language governing permissions and |
| 13 | // limitations under the License. |
| 14 | |
JPEG XL | 226f84e | 2020-11-11 20:22:00 +0100 | [diff] [blame] | 15 | #ifndef HIGHWAY_HWY_BASE_H_ |
| 16 | #define HIGHWAY_HWY_BASE_H_ |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 17 | |
| 18 | // For SIMD module implementations and their callers, target-independent. |
| 19 | |
| 20 | #include <stddef.h> |
| 21 | #include <stdint.h> |
| 22 | |
| 23 | #include <atomic> |
| 24 | |
| 25 | // Add to #if conditions to prevent IDE from graying out code. |
| 26 | #if (defined __CDT_PARSER__) || (defined __INTELLISENSE__) || \ |
Jan Wassenberg | b97d18f | 2020-11-11 17:53:22 +0100 | [diff] [blame] | 27 | (defined Q_CREATOR_RUN) || (defined(__CLANGD__)) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 28 | #define HWY_IDE 1 |
| 29 | #else |
| 30 | #define HWY_IDE 0 |
| 31 | #endif |
| 32 | |
| 33 | //------------------------------------------------------------------------------ |
| 34 | // Detect compiler using predefined macros |
| 35 | |
| 36 | #ifdef _MSC_VER |
| 37 | #define HWY_COMPILER_MSVC _MSC_VER |
| 38 | #else |
| 39 | #define HWY_COMPILER_MSVC 0 |
| 40 | #endif |
| 41 | |
| 42 | #ifdef __INTEL_COMPILER |
| 43 | #define HWY_COMPILER_ICC __INTEL_COMPILER |
| 44 | #else |
| 45 | #define HWY_COMPILER_ICC 0 |
| 46 | #endif |
| 47 | |
| 48 | #ifdef __GNUC__ |
| 49 | #define HWY_COMPILER_GCC (__GNUC__ * 100 + __GNUC_MINOR__) |
| 50 | #else |
| 51 | #define HWY_COMPILER_GCC 0 |
| 52 | #endif |
| 53 | |
| 54 | // Clang can masquerade as MSVC/GCC, in which case both are set. |
| 55 | #ifdef __clang__ |
| 56 | #define HWY_COMPILER_CLANG (__clang_major__ * 100 + __clang_minor__) |
| 57 | #else |
| 58 | #define HWY_COMPILER_CLANG 0 |
| 59 | #endif |
| 60 | |
| 61 | // More than one may be nonzero, but we want at least one. |
| 62 | #if !HWY_COMPILER_MSVC && !HWY_COMPILER_ICC && !HWY_COMPILER_GCC && \ |
| 63 | !HWY_COMPILER_CLANG |
| 64 | #error "Unsupported compiler" |
| 65 | #endif |
| 66 | |
| 67 | //------------------------------------------------------------------------------ |
| 68 | // Compiler-specific definitions |
| 69 | |
| 70 | #define HWY_STR_IMPL(macro) #macro |
| 71 | #define HWY_STR(macro) HWY_STR_IMPL(macro) |
| 72 | |
| 73 | #if HWY_COMPILER_MSVC |
| 74 | |
| 75 | #include <intrin.h> |
| 76 | |
| 77 | #define HWY_RESTRICT __restrict |
| 78 | #define HWY_INLINE __forceinline |
| 79 | #define HWY_NOINLINE __declspec(noinline) |
| 80 | #define HWY_FLATTEN |
| 81 | #define HWY_NORETURN __declspec(noreturn) |
| 82 | #define HWY_LIKELY(expr) (expr) |
| 83 | #define HWY_UNLIKELY(expr) (expr) |
Jan Wassenberg | acfaee6 | 2020-12-31 08:48:22 -0800 | [diff] [blame] | 84 | #define HWY_PRAGMA(tokens) __pragma(tokens) |
Jan Wassenberg | 994eb84 | 2020-12-31 06:12:45 -0800 | [diff] [blame] | 85 | #define HWY_DIAGNOSTICS(tokens) HWY_PRAGMA(warning(tokens)) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 86 | #define HWY_DIAGNOSTICS_OFF(msc, gcc) HWY_DIAGNOSTICS(msc) |
| 87 | #define HWY_MAYBE_UNUSED |
| 88 | #define HWY_HAS_ASSUME_ALIGNED 0 |
| 89 | #if (_MSC_VER >= 1700) |
| 90 | #define HWY_MUST_USE_RESULT _Check_return_ |
| 91 | #else |
| 92 | #define HWY_MUST_USE_RESULT |
| 93 | #endif |
| 94 | |
| 95 | #else |
| 96 | |
| 97 | #define HWY_RESTRICT __restrict__ |
| 98 | #define HWY_INLINE inline __attribute__((always_inline)) |
| 99 | #define HWY_NOINLINE __attribute__((noinline)) |
| 100 | #define HWY_FLATTEN __attribute__((flatten)) |
| 101 | #define HWY_NORETURN __attribute__((noreturn)) |
| 102 | #define HWY_LIKELY(expr) __builtin_expect(!!(expr), 1) |
| 103 | #define HWY_UNLIKELY(expr) __builtin_expect(!!(expr), 0) |
| 104 | #define HWY_PRAGMA(tokens) _Pragma(#tokens) |
| 105 | #define HWY_DIAGNOSTICS(tokens) HWY_PRAGMA(GCC diagnostic tokens) |
| 106 | #define HWY_DIAGNOSTICS_OFF(msc, gcc) HWY_DIAGNOSTICS(gcc) |
| 107 | // Encountered "attribute list cannot appear here" when using the C++17 |
| 108 | // [[maybe_unused]], so only use the old style attribute for now. |
| 109 | #define HWY_MAYBE_UNUSED __attribute__((unused)) |
| 110 | #define HWY_MUST_USE_RESULT __attribute__((warn_unused_result)) |
| 111 | |
| 112 | #endif // !HWY_COMPILER_MSVC |
| 113 | |
| 114 | //------------------------------------------------------------------------------ |
| 115 | // Builtin/attributes |
| 116 | |
| 117 | #ifdef __has_builtin |
| 118 | #define HWY_HAS_BUILTIN(name) __has_builtin(name) |
| 119 | #else |
| 120 | #define HWY_HAS_BUILTIN(name) 0 |
| 121 | #endif |
| 122 | |
| 123 | #ifdef __has_attribute |
| 124 | #define HWY_HAS_ATTRIBUTE(name) __has_attribute(name) |
| 125 | #else |
| 126 | #define HWY_HAS_ATTRIBUTE(name) 0 |
| 127 | #endif |
| 128 | |
| 129 | // Enables error-checking of format strings. |
| 130 | #if HWY_HAS_ATTRIBUTE(__format__) |
| 131 | #define HWY_FORMAT(idx_fmt, idx_arg) \ |
| 132 | __attribute__((__format__(__printf__, idx_fmt, idx_arg))) |
| 133 | #else |
| 134 | #define HWY_FORMAT(idx_fmt, idx_arg) |
| 135 | #endif |
| 136 | |
| 137 | // Returns a void* pointer which the compiler then assumes is N-byte aligned. |
| 138 | // Example: float* HWY_RESTRICT aligned = (float*)HWY_ASSUME_ALIGNED(in, 32); |
| 139 | // |
| 140 | // The assignment semantics are required by GCC/Clang. ICC provides an in-place |
| 141 | // __assume_aligned, whereas MSVC's __assume appears unsuitable. |
| 142 | #if HWY_HAS_BUILTIN(__builtin_assume_aligned) |
| 143 | #define HWY_ASSUME_ALIGNED(ptr, align) __builtin_assume_aligned((ptr), (align)) |
| 144 | #else |
| 145 | #define HWY_ASSUME_ALIGNED(ptr, align) (ptr) /* not supported */ |
| 146 | #endif |
| 147 | |
| 148 | // Clang and GCC require attributes on each function into which SIMD intrinsics |
| 149 | // are inlined. Support both per-function annotation (HWY_ATTR) for lambdas and |
| 150 | // automatic annotation via pragmas. |
| 151 | #if HWY_COMPILER_CLANG |
| 152 | #define HWY_PUSH_ATTRIBUTES(targets_str) \ |
Jan Wassenberg | 2dc9ece | 2020-12-21 00:44:42 -0800 | [diff] [blame] | 153 | HWY_PRAGMA(clang attribute push(__attribute__((target(targets_str))), \ |
| 154 | apply_to = function)) |
| 155 | #define HWY_POP_ATTRIBUTES HWY_PRAGMA(clang attribute pop) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 156 | #elif HWY_COMPILER_GCC |
| 157 | #define HWY_PUSH_ATTRIBUTES(targets_str) \ |
Jan Wassenberg | 2dc9ece | 2020-12-21 00:44:42 -0800 | [diff] [blame] | 158 | HWY_PRAGMA(GCC push_options) HWY_PRAGMA(GCC target targets_str) |
| 159 | #define HWY_POP_ATTRIBUTES HWY_PRAGMA(GCC pop_options) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 160 | #else |
Highway | f999d0e | 2020-11-13 13:17:34 +0100 | [diff] [blame] | 161 | #define HWY_PUSH_ATTRIBUTES(targets_str) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 162 | #define HWY_POP_ATTRIBUTES |
| 163 | #endif |
| 164 | |
| 165 | //------------------------------------------------------------------------------ |
| 166 | // Detect architecture using predefined macros |
| 167 | |
| 168 | #if defined(__i386__) || defined(_M_IX86) |
| 169 | #define HWY_ARCH_X86_32 1 |
| 170 | #else |
| 171 | #define HWY_ARCH_X86_32 0 |
| 172 | #endif |
| 173 | |
| 174 | #if defined(__x86_64__) || defined(_M_X64) |
| 175 | #define HWY_ARCH_X86_64 1 |
| 176 | #else |
| 177 | #define HWY_ARCH_X86_64 0 |
| 178 | #endif |
| 179 | |
| 180 | #if HWY_ARCH_X86_32 || HWY_ARCH_X86_64 |
| 181 | #define HWY_ARCH_X86 1 |
| 182 | #else |
| 183 | #define HWY_ARCH_X86 0 |
| 184 | #endif |
| 185 | |
| 186 | #if defined(__powerpc64__) || defined(_M_PPC) |
| 187 | #define HWY_ARCH_PPC 1 |
| 188 | #else |
| 189 | #define HWY_ARCH_PPC 0 |
| 190 | #endif |
| 191 | |
| 192 | #if defined(__arm__) || defined(_M_ARM) || defined(__aarch64__) |
| 193 | #define HWY_ARCH_ARM 1 |
| 194 | #else |
| 195 | #define HWY_ARCH_ARM 0 |
| 196 | #endif |
| 197 | |
| 198 | // There isn't yet a standard __wasm or __wasm__. |
| 199 | #ifdef __EMSCRIPTEN__ |
| 200 | #define HWY_ARCH_WASM 1 |
| 201 | #else |
| 202 | #define HWY_ARCH_WASM 0 |
| 203 | #endif |
| 204 | |
Jan Wassenberg | 0034dac | 2021-01-07 01:18:02 -0800 | [diff] [blame] | 205 | #ifdef __riscv |
| 206 | #define HWY_ARCH_RVV 1 |
| 207 | #else |
| 208 | #define HWY_ARCH_RVV 0 |
| 209 | #endif |
| 210 | |
| 211 | #if (HWY_ARCH_X86 + HWY_ARCH_PPC + HWY_ARCH_ARM + HWY_ARCH_WASM + \ |
| 212 | HWY_ARCH_RVV) != 1 |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 213 | #error "Must detect exactly one platform" |
| 214 | #endif |
| 215 | |
| 216 | //------------------------------------------------------------------------------ |
| 217 | // Macros |
| 218 | |
Jan Wassenberg | 586d0c2 | 2021-01-06 09:09:54 -0800 | [diff] [blame] | 219 | #define HWY_API static HWY_INLINE HWY_FLATTEN HWY_MAYBE_UNUSED |
| 220 | |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 221 | #define HWY_CONCAT_IMPL(a, b) a##b |
| 222 | #define HWY_CONCAT(a, b) HWY_CONCAT_IMPL(a, b) |
| 223 | |
| 224 | #define HWY_MIN(a, b) ((a) < (b) ? (a) : (b)) |
| 225 | #define HWY_MAX(a, b) ((a) < (b) ? (b) : (a)) |
| 226 | |
| 227 | // Alternative for asm volatile("" : : : "memory"), which has no effect. |
| 228 | #define HWY_FENCE std::atomic_thread_fence(std::memory_order_acq_rel) |
| 229 | |
| 230 | // 4 instances of a given literal value, useful as input to LoadDup128. |
| 231 | #define HWY_REP4(literal) literal, literal, literal, literal |
| 232 | |
Jan Wassenberg | b97d18f | 2020-11-11 17:53:22 +0100 | [diff] [blame] | 233 | #define HWY_ABORT(format, ...) \ |
| 234 | ::hwy::Abort(__FILE__, __LINE__, format, ##__VA_ARGS__) |
| 235 | |
| 236 | // Always enabled. |
| 237 | #define HWY_ASSERT(condition) \ |
| 238 | do { \ |
| 239 | if (!(condition)) { \ |
| 240 | HWY_ABORT("Assert %s", #condition); \ |
| 241 | } \ |
| 242 | } while (0) |
| 243 | |
| 244 | // Only for "debug" builds |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 245 | #if !defined(NDEBUG) || defined(ADDRESS_SANITIZER) || \ |
| 246 | defined(MEMORY_SANITIZER) || defined(THREAD_SANITIZER) |
Jan Wassenberg | b97d18f | 2020-11-11 17:53:22 +0100 | [diff] [blame] | 247 | #define HWY_DASSERT(condition) HWY_ASSERT(condition) |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 248 | #else |
| 249 | #define HWY_DASSERT(condition) \ |
| 250 | do { \ |
| 251 | } while (0) |
| 252 | #endif |
| 253 | |
| 254 | //------------------------------------------------------------------------------ |
| 255 | |
| 256 | namespace hwy { |
| 257 | |
Jan Wassenberg | b97d18f | 2020-11-11 17:53:22 +0100 | [diff] [blame] | 258 | // See also HWY_ALIGNMENT - aligned_allocator aligns to the larger of that and |
| 259 | // the vector size, whose upper bound is specified here. |
| 260 | |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 261 | #if HWY_ARCH_X86 |
| 262 | static constexpr size_t kMaxVectorSize = 64; // AVX-512 |
| 263 | #define HWY_ALIGN_MAX alignas(64) |
| 264 | #else |
| 265 | static constexpr size_t kMaxVectorSize = 16; |
| 266 | #define HWY_ALIGN_MAX alignas(16) |
| 267 | #endif |
| 268 | |
| 269 | HWY_NORETURN void HWY_FORMAT(3, 4) |
| 270 | Abort(const char* file, int line, const char* format, ...); |
| 271 | |
| 272 | template <typename T> |
| 273 | constexpr bool IsFloat() { |
| 274 | return T(1.25) != T(1); |
| 275 | } |
| 276 | |
| 277 | template <typename T> |
| 278 | constexpr bool IsSigned() { |
| 279 | return T(0) > T(-1); |
| 280 | } |
| 281 | |
| 282 | // Largest/smallest representable integer values. |
| 283 | template <typename T> |
| 284 | constexpr T LimitsMax() { |
| 285 | return IsSigned<T>() ? T((1ULL << (sizeof(T) * 8 - 1)) - 1) |
| 286 | : static_cast<T>(~0ull); |
| 287 | } |
| 288 | template <typename T> |
| 289 | constexpr T LimitsMin() { |
| 290 | return IsSigned<T>() ? T(-1) - LimitsMax<T>() : T(0); |
| 291 | } |
| 292 | |
| 293 | // Manual control of overload resolution (SFINAE). |
| 294 | template <bool Condition, class T> |
| 295 | struct EnableIfT {}; |
| 296 | template <class T> |
| 297 | struct EnableIfT<true, T> { |
| 298 | using type = T; |
| 299 | }; |
| 300 | |
| 301 | template <bool Condition, class T = void> |
| 302 | using EnableIf = typename EnableIfT<Condition, T>::type; |
| 303 | |
Jan Wassenberg | b97d18f | 2020-11-11 17:53:22 +0100 | [diff] [blame] | 304 | template <typename T1, typename T2> |
| 305 | constexpr inline T1 DivCeil(T1 a, T2 b) { |
| 306 | return (a + b - 1) / b; |
| 307 | } |
| 308 | |
| 309 | // Works for any `align`; if a power of two, compiler emits ADD+AND. |
| 310 | constexpr inline size_t RoundUpTo(size_t what, size_t align) { |
| 311 | return DivCeil(what, align) * align; |
| 312 | } |
| 313 | |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 314 | // Undefined results for x == 0. |
Jan Wassenberg | 586d0c2 | 2021-01-06 09:09:54 -0800 | [diff] [blame] | 315 | HWY_API size_t Num0BitsBelowLS1Bit_Nonzero32(const uint32_t x) { |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 316 | #ifdef _MSC_VER |
Jan Wassenberg | 0034dac | 2021-01-07 01:18:02 -0800 | [diff] [blame] | 317 | unsigned long index; // NOLINT |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 318 | _BitScanForward(&index, x); |
| 319 | return index; |
| 320 | #else |
| 321 | return static_cast<size_t>(__builtin_ctz(x)); |
| 322 | #endif |
| 323 | } |
| 324 | |
Jan Wassenberg | 586d0c2 | 2021-01-06 09:09:54 -0800 | [diff] [blame] | 325 | HWY_API size_t PopCount(uint64_t x) { |
| 326 | #if HWY_COMPILER_CLANG || HWY_COMPILER_GCC |
| 327 | return static_cast<size_t>(__builtin_popcountll(x)); |
| 328 | #elif HWY_COMPILER_MSVC && HWY_ARCH_X86_64 |
| 329 | return _mm_popcnt_u64(x); |
| 330 | #elif HWY_COMPILER_MSVC |
| 331 | return _mm_popcnt_u32(uint32_t(x)) + _mm_popcnt_u32(uint32_t(x >> 32)); |
| 332 | #else |
| 333 | x -= ((x >> 1) & 0x55555555U); |
| 334 | x = (((x >> 2) & 0x33333333U) + (x & 0x33333333U)); |
| 335 | x = (((x >> 4) + x) & 0x0F0F0F0FU); |
| 336 | x += (x >> 8); |
| 337 | x += (x >> 16); |
| 338 | x += (x >> 32); |
| 339 | x = x & 0x0000007FU; |
| 340 | return (unsigned int)x; |
| 341 | #endif |
| 342 | } |
| 343 | |
| 344 | // The source/destination must not overlap/alias. |
| 345 | template <size_t kBytes, typename From, typename To> |
| 346 | HWY_API void CopyBytes(const From* from, To* to) { |
| 347 | #if HWY_COMPILER_MSVC |
| 348 | const uint8_t* HWY_RESTRICT from_bytes = |
| 349 | reinterpret_cast<const uint8_t*>(from); |
| 350 | uint8_t* HWY_RESTRICT to_bytes = reinterpret_cast<uint8_t*>(to); |
| 351 | for (size_t i = 0; i < kBytes; ++i) { |
| 352 | to_bytes[i] = from_bytes[i]; |
| 353 | } |
| 354 | #else |
| 355 | // Avoids horrible codegen on Clang (series of PINSRB) |
| 356 | __builtin_memcpy(to, from, kBytes); |
| 357 | #endif |
| 358 | } |
| 359 | |
Jan Wassenberg | 94a72d0 | 2020-10-29 18:04:03 +0100 | [diff] [blame] | 360 | } // namespace hwy |
| 361 | |
JPEG XL | 226f84e | 2020-11-11 20:22:00 +0100 | [diff] [blame] | 362 | #endif // HIGHWAY_HWY_BASE_H_ |