1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. 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 ==============================================================================*/ 15 16 #ifndef EIGEN_BFLOAT16_H 17 #define EIGEN_BFLOAT16_H 18 19 #define BF16_PACKET_FUNCTION(PACKET_F, PACKET_BF16, METHOD) \ 20 template <> \ 21 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED \ 22 PACKET_BF16 METHOD<PACKET_BF16>(const PACKET_BF16& _x) { \ 23 return F32ToBf16(METHOD<PACKET_F>(Bf16ToF32(_x))); \ 24 } 25 26 namespace Eigen { 27 28 struct bfloat16; 29 30 namespace bfloat16_impl { 31 32 // Make our own __bfloat16_raw definition. 33 struct __bfloat16_raw { __bfloat16_raw__bfloat16_raw34 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __bfloat16_raw() : value(0) {} __bfloat16_raw__bfloat16_raw35 explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __bfloat16_raw(unsigned short raw) : value(raw) {} 36 unsigned short value; 37 }; 38 39 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __bfloat16_raw raw_uint16_to_bfloat16(unsigned short value); 40 template <bool AssumeArgumentIsNormalOrInfinityOrZero> 41 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne(float ff); 42 // Forward declarations of template specializations, to avoid Visual C++ 2019 errors, saying: 43 // > error C2908: explicit specialization; 'float_to_bfloat16_rtne' has already been instantiated 44 template <> 45 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne<false>(float ff); 46 template <> 47 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne<true>(float ff); 48 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float bfloat16_to_float(__bfloat16_raw h); 49 50 struct bfloat16_base : public __bfloat16_raw { bfloat16_basebfloat16_base51 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16_base() {} bfloat16_basebfloat16_base52 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16_base(const __bfloat16_raw& h) : __bfloat16_raw(h) {} 53 }; 54 55 } // namespace bfloat16_impl 56 57 // Class definition. 58 struct bfloat16 : public bfloat16_impl::bfloat16_base { 59 60 typedef bfloat16_impl::__bfloat16_raw __bfloat16_raw; 61 bfloat16bfloat1662 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16() {} 63 bfloat16bfloat1664 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16(const __bfloat16_raw& h) : bfloat16_impl::bfloat16_base(h) {} 65 bfloat16bfloat1666 explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16(bool b) 67 : bfloat16_impl::bfloat16_base(bfloat16_impl::raw_uint16_to_bfloat16(b ? 0x3f80 : 0)) {} 68 69 template<class T> bfloat16bfloat1670 explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16(T val) 71 : bfloat16_impl::bfloat16_base(bfloat16_impl::float_to_bfloat16_rtne<internal::is_integral<T>::value>(static_cast<float>(val))) {} 72 bfloat16bfloat1673 explicit EIGEN_DEVICE_FUNC bfloat16(float f) 74 : bfloat16_impl::bfloat16_base(bfloat16_impl::float_to_bfloat16_rtne<false>(f)) {} 75 76 // Following the convention of numpy, converting between complex and 77 // float will lead to loss of imag value. 78 template<typename RealScalar> bfloat16bfloat1679 explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR bfloat16(const std::complex<RealScalar>& val) 80 : bfloat16_impl::bfloat16_base(bfloat16_impl::float_to_bfloat16_rtne<false>(static_cast<float>(val.real()))) {} 81 82 EIGEN_DEVICE_FUNC operator float() const { // NOLINT: Allow implicit conversion to float, because it is lossless. 83 return bfloat16_impl::bfloat16_to_float(*this); 84 } 85 }; 86 } // namespace Eigen 87 88 namespace std { 89 template<> 90 struct numeric_limits<Eigen::bfloat16> { 91 static const bool is_specialized = true; 92 static const bool is_signed = true; 93 static const bool is_integer = false; 94 static const bool is_exact = false; 95 static const bool has_infinity = true; 96 static const bool has_quiet_NaN = true; 97 static const bool has_signaling_NaN = true; 98 static const float_denorm_style has_denorm = std::denorm_absent; 99 static const bool has_denorm_loss = false; 100 static const std::float_round_style round_style = numeric_limits<float>::round_style; 101 static const bool is_iec559 = false; 102 static const bool is_bounded = true; 103 static const bool is_modulo = false; 104 static const int digits = 8; 105 static const int digits10 = 2; 106 static const int max_digits10 = 4; 107 static const int radix = 2; 108 static const int min_exponent = numeric_limits<float>::min_exponent; 109 static const int min_exponent10 = numeric_limits<float>::min_exponent10; 110 static const int max_exponent = numeric_limits<float>::max_exponent; 111 static const int max_exponent10 = numeric_limits<float>::max_exponent10; 112 static const bool traps = numeric_limits<float>::traps; 113 static const bool tinyness_before = numeric_limits<float>::tinyness_before; 114 115 static Eigen::bfloat16 (min)() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x0080); } 116 static Eigen::bfloat16 lowest() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0xff7f); } 117 static Eigen::bfloat16 (max)() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7f7f); } 118 static Eigen::bfloat16 epsilon() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x3c00); } 119 static Eigen::bfloat16 round_error() { return Eigen::bfloat16(0x3f00); } 120 static Eigen::bfloat16 infinity() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7f80); } 121 static Eigen::bfloat16 quiet_NaN() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7fc0); } 122 static Eigen::bfloat16 signaling_NaN() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7f81); } 123 static Eigen::bfloat16 denorm_min() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x0001); } 124 }; 125 126 // If std::numeric_limits<T> is specialized, should also specialize 127 // std::numeric_limits<const T>, std::numeric_limits<volatile T>, and 128 // std::numeric_limits<const volatile T> 129 // https://stackoverflow.com/a/16519653/ 130 template<> 131 struct numeric_limits<const Eigen::bfloat16> : numeric_limits<Eigen::bfloat16> {}; 132 template<> 133 struct numeric_limits<volatile Eigen::bfloat16> : numeric_limits<Eigen::bfloat16> {}; 134 template<> 135 struct numeric_limits<const volatile Eigen::bfloat16> : numeric_limits<Eigen::bfloat16> {}; 136 } // namespace std 137 138 namespace Eigen { 139 140 namespace bfloat16_impl { 141 142 // We need to distinguish ‘clang as the CUDA compiler’ from ‘clang as the host compiler, 143 // invoked by NVCC’ (e.g. on MacOS). The former needs to see both host and device implementation 144 // of the functions, while the latter can only deal with one of them. 145 #if !defined(EIGEN_HAS_NATIVE_BF16) || (EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC) // Emulate support for bfloat16 floats 146 147 #if EIGEN_COMP_CLANG && defined(EIGEN_CUDACC) 148 // We need to provide emulated *host-side* BF16 operators for clang. 149 #pragma push_macro("EIGEN_DEVICE_FUNC") 150 #undef EIGEN_DEVICE_FUNC 151 #if defined(EIGEN_HAS_CUDA_BF16) && defined(EIGEN_HAS_NATIVE_BF16) 152 #define EIGEN_DEVICE_FUNC __host__ 153 #else // both host and device need emulated ops. 154 #define EIGEN_DEVICE_FUNC __host__ __device__ 155 #endif 156 #endif 157 158 // Definitions for CPUs, mostly working through conversion 159 // to/from fp32. 160 161 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator + (const bfloat16& a, const bfloat16& b) { 162 return bfloat16(float(a) + float(b)); 163 } 164 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator + (const bfloat16& a, const int& b) { 165 return bfloat16(float(a) + static_cast<float>(b)); 166 } 167 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator + (const int& a, const bfloat16& b) { 168 return bfloat16(static_cast<float>(a) + float(b)); 169 } 170 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator * (const bfloat16& a, const bfloat16& b) { 171 return bfloat16(float(a) * float(b)); 172 } 173 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator - (const bfloat16& a, const bfloat16& b) { 174 return bfloat16(float(a) - float(b)); 175 } 176 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator / (const bfloat16& a, const bfloat16& b) { 177 return bfloat16(float(a) / float(b)); 178 } 179 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator - (const bfloat16& a) { 180 bfloat16 result; 181 result.value = a.value ^ 0x8000; 182 return result; 183 } 184 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16& operator += (bfloat16& a, const bfloat16& b) { 185 a = bfloat16(float(a) + float(b)); 186 return a; 187 } 188 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16& operator *= (bfloat16& a, const bfloat16& b) { 189 a = bfloat16(float(a) * float(b)); 190 return a; 191 } 192 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16& operator -= (bfloat16& a, const bfloat16& b) { 193 a = bfloat16(float(a) - float(b)); 194 return a; 195 } 196 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16& operator /= (bfloat16& a, const bfloat16& b) { 197 a = bfloat16(float(a) / float(b)); 198 return a; 199 } 200 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator++(bfloat16& a) { 201 a += bfloat16(1); 202 return a; 203 } 204 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator--(bfloat16& a) { 205 a -= bfloat16(1); 206 return a; 207 } 208 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator++(bfloat16& a, int) { 209 bfloat16 original_value = a; 210 ++a; 211 return original_value; 212 } 213 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator--(bfloat16& a, int) { 214 bfloat16 original_value = a; 215 --a; 216 return original_value; 217 } 218 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const bfloat16& a, const bfloat16& b) { 219 return numext::equal_strict(float(a),float(b)); 220 } 221 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const bfloat16& a, const bfloat16& b) { 222 return numext::not_equal_strict(float(a), float(b)); 223 } 224 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const bfloat16& a, const bfloat16& b) { 225 return float(a) < float(b); 226 } 227 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const bfloat16& a, const bfloat16& b) { 228 return float(a) <= float(b); 229 } 230 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const bfloat16& a, const bfloat16& b) { 231 return float(a) > float(b); 232 } 233 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const bfloat16& a, const bfloat16& b) { 234 return float(a) >= float(b); 235 } 236 237 #if EIGEN_COMP_CLANG && defined(EIGEN_CUDACC) 238 #pragma pop_macro("EIGEN_DEVICE_FUNC") 239 #endif 240 #endif // Emulate support for bfloat16 floats 241 242 // Division by an index. Do it in full float precision to avoid accuracy 243 // issues in converting the denominator to bfloat16. 244 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator / (const bfloat16& a, Index b) { 245 return bfloat16(static_cast<float>(a) / static_cast<float>(b)); 246 } 247 248 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw truncate_to_bfloat16(const float v) { 249 __bfloat16_raw output; 250 if (Eigen::numext::isnan EIGEN_NOT_A_MACRO(v)) { 251 output.value = std::signbit(v) ? 0xFFC0: 0x7FC0; 252 return output; 253 } 254 const uint16_t* p = reinterpret_cast<const uint16_t*>(&v); 255 #if defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 256 output.value = p[0]; 257 #else 258 output.value = p[1]; 259 #endif 260 return output; 261 } 262 263 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __bfloat16_raw raw_uint16_to_bfloat16(numext::uint16_t value) { 264 return __bfloat16_raw(value); 265 } 266 267 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR numext::uint16_t raw_bfloat16_as_uint16(const __bfloat16_raw& bf) { 268 return bf.value; 269 } 270 271 // float_to_bfloat16_rtne template specialization that does not make any 272 // assumption about the value of its function argument (ff). 273 template <> 274 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne<false>(float ff) { 275 #if (defined(EIGEN_HAS_CUDA_BF16) && defined(EIGEN_HAS_HIP_BF16)) 276 // Nothing to do here 277 #else 278 __bfloat16_raw output; 279 280 if (Eigen::numext::isnan EIGEN_NOT_A_MACRO(ff)) { 281 // If the value is a NaN, squash it to a qNaN with msb of fraction set, 282 // this makes sure after truncation we don't end up with an inf. 283 // 284 // qNaN magic: All exponent bits set + most significant bit of fraction 285 // set. 286 output.value = std::signbit(ff) ? 0xFFC0: 0x7FC0; 287 } else { 288 // Fast rounding algorithm that rounds a half value to nearest even. This 289 // reduces expected error when we convert a large number of floats. Here 290 // is how it works: 291 // 292 // Definitions: 293 // To convert a float 32 to bfloat16, a float 32 can be viewed as 32 bits 294 // with the following tags: 295 // 296 // Sign | Exp (8 bits) | Frac (23 bits) 297 // S EEEEEEEE FFFFFFLRTTTTTTTTTTTTTTT 298 // 299 // S: Sign bit. 300 // E: Exponent bits. 301 // F: First 6 bits of fraction. 302 // L: Least significant bit of resulting bfloat16 if we truncate away the 303 // rest of the float32. This is also the 7th bit of fraction 304 // R: Rounding bit, 8th bit of fraction. 305 // T: Sticky bits, rest of fraction, 15 bits. 306 // 307 // To round half to nearest even, there are 3 cases where we want to round 308 // down (simply truncate the result of the bits away, which consists of 309 // rounding bit and sticky bits) and two cases where we want to round up 310 // (truncate then add one to the result). 311 // 312 // The fast converting algorithm simply adds lsb (L) to 0x7fff (15 bits of 313 // 1s) as the rounding bias, adds the rounding bias to the input, then 314 // truncates the last 16 bits away. 315 // 316 // To understand how it works, we can analyze this algorithm case by case: 317 // 318 // 1. L = 0, R = 0: 319 // Expect: round down, this is less than half value. 320 // 321 // Algorithm: 322 // - Rounding bias: 0x7fff + 0 = 0x7fff 323 // - Adding rounding bias to input may create any carry, depending on 324 // whether there is any value set to 1 in T bits. 325 // - R may be set to 1 if there is a carry. 326 // - L remains 0. 327 // - Note that this case also handles Inf and -Inf, where all fraction 328 // bits, including L, R and Ts are all 0. The output remains Inf after 329 // this algorithm. 330 // 331 // 2. L = 1, R = 0: 332 // Expect: round down, this is less than half value. 333 // 334 // Algorithm: 335 // - Rounding bias: 0x7fff + 1 = 0x8000 336 // - Adding rounding bias to input doesn't change sticky bits but 337 // adds 1 to rounding bit. 338 // - L remains 1. 339 // 340 // 3. L = 0, R = 1, all of T are 0: 341 // Expect: round down, this is exactly at half, the result is already 342 // even (L=0). 343 // 344 // Algorithm: 345 // - Rounding bias: 0x7fff + 0 = 0x7fff 346 // - Adding rounding bias to input sets all sticky bits to 1, but 347 // doesn't create a carry. 348 // - R remains 1. 349 // - L remains 0. 350 // 351 // 4. L = 1, R = 1: 352 // Expect: round up, this is exactly at half, the result needs to be 353 // round to the next even number. 354 // 355 // Algorithm: 356 // - Rounding bias: 0x7fff + 1 = 0x8000 357 // - Adding rounding bias to input doesn't change sticky bits, but 358 // creates a carry from rounding bit. 359 // - The carry sets L to 0, creates another carry bit and propagate 360 // forward to F bits. 361 // - If all the F bits are 1, a carry then propagates to the exponent 362 // bits, which then creates the minimum value with the next exponent 363 // value. Note that we won't have the case where exponents are all 1, 364 // since that's either a NaN (handled in the other if condition) or inf 365 // (handled in case 1). 366 // 367 // 5. L = 0, R = 1, any of T is 1: 368 // Expect: round up, this is greater than half. 369 // 370 // Algorithm: 371 // - Rounding bias: 0x7fff + 0 = 0x7fff 372 // - Adding rounding bias to input creates a carry from sticky bits, 373 // sets rounding bit to 0, then create another carry. 374 // - The second carry sets L to 1. 375 // 376 // Examples: 377 // 378 // Exact half value that is already even: 379 // Input: 380 // Sign | Exp (8 bit) | Frac (first 7 bit) | Frac (last 16 bit) 381 // S E E E E E E E E F F F F F F L RTTTTTTTTTTTTTTT 382 // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1000000000000000 383 // 384 // This falls into case 3. We truncate the rest of 16 bits and no 385 // carry is created into F and L: 386 // 387 // Output: 388 // Sign | Exp (8 bit) | Frac (first 7 bit) 389 // S E E E E E E E E F F F F F F L 390 // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 391 // 392 // Exact half value, round to next even number: 393 // Input: 394 // Sign | Exp (8 bit) | Frac (first 7 bit) | Frac (last 16 bit) 395 // S E E E E E E E E F F F F F F L RTTTTTTTTTTTTTTT 396 // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1000000000000000 397 // 398 // This falls into case 4. We create a carry from R and T, 399 // which then propagates into L and F: 400 // 401 // Output: 402 // Sign | Exp (8 bit) | Frac (first 7 bit) 403 // S E E E E E E E E F F F F F F L 404 // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 405 // 406 // 407 // Max denormal value round to min normal value: 408 // Input: 409 // Sign | Exp (8 bit) | Frac (first 7 bit) | Frac (last 16 bit) 410 // S E E E E E E E E F F F F F F L RTTTTTTTTTTTTTTT 411 // 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1111111111111111 412 // 413 // This falls into case 4. We create a carry from R and T, 414 // propagate into L and F, which then propagates into exponent 415 // bits: 416 // 417 // Output: 418 // Sign | Exp (8 bit) | Frac (first 7 bit) 419 // S E E E E E E E E F F F F F F L 420 // 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 421 // 422 // Max normal value round to Inf: 423 // Input: 424 // Sign | Exp (8 bit) | Frac (first 7 bit) | Frac (last 16 bit) 425 // S E E E E E E E E F F F F F F L RTTTTTTTTTTTTTTT 426 // 0 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1111111111111111 427 // 428 // This falls into case 4. We create a carry from R and T, 429 // propagate into L and F, which then propagates into exponent 430 // bits: 431 // 432 // Sign | Exp (8 bit) | Frac (first 7 bit) 433 // S E E E E E E E E F F F F F F L 434 // 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 435 436 // At this point, ff must be either a normal float, or +/-infinity. 437 output = float_to_bfloat16_rtne<true>(ff); 438 } 439 return output; 440 #endif 441 } 442 443 // float_to_bfloat16_rtne template specialization that assumes that its function 444 // argument (ff) is either a normal floating point number, or +/-infinity, or 445 // zero. Used to improve the runtime performance of conversion from an integer 446 // type to bfloat16. 447 template <> 448 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne<true>(float ff) { 449 #if (defined(EIGEN_HAS_CUDA_BF16) && defined(EIGEN_HAS_HIP_BF16)) 450 // Nothing to do here 451 #else 452 numext::uint32_t input = numext::bit_cast<numext::uint32_t>(ff); 453 __bfloat16_raw output; 454 455 // Least significant bit of resulting bfloat. 456 numext::uint32_t lsb = (input >> 16) & 1; 457 numext::uint32_t rounding_bias = 0x7fff + lsb; 458 input += rounding_bias; 459 output.value = static_cast<numext::uint16_t>(input >> 16); 460 return output; 461 #endif 462 } 463 464 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float bfloat16_to_float(__bfloat16_raw h) { 465 float result = 0; 466 unsigned short* q = reinterpret_cast<unsigned short*>(&result); 467 #if defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 468 q[0] = h.value; 469 #else 470 q[1] = h.value; 471 #endif 472 return result; 473 } 474 // --- standard functions --- 475 476 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const bfloat16& a) { 477 EIGEN_USING_STD(isinf); 478 return (isinf)(float(a)); 479 } 480 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const bfloat16& a) { 481 EIGEN_USING_STD(isnan); 482 return (isnan)(float(a)); 483 } 484 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const bfloat16& a) { 485 return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a)); 486 } 487 488 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 abs(const bfloat16& a) { 489 bfloat16 result; 490 result.value = a.value & 0x7FFF; 491 return result; 492 } 493 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 exp(const bfloat16& a) { 494 return bfloat16(::expf(float(a))); 495 } 496 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 expm1(const bfloat16& a) { 497 return bfloat16(numext::expm1(float(a))); 498 } 499 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log(const bfloat16& a) { 500 return bfloat16(::logf(float(a))); 501 } 502 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log1p(const bfloat16& a) { 503 return bfloat16(numext::log1p(float(a))); 504 } 505 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log10(const bfloat16& a) { 506 return bfloat16(::log10f(float(a))); 507 } 508 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log2(const bfloat16& a) { 509 return bfloat16(static_cast<float>(EIGEN_LOG2E) * ::logf(float(a))); 510 } 511 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 sqrt(const bfloat16& a) { 512 return bfloat16(::sqrtf(float(a))); 513 } 514 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 pow(const bfloat16& a, const bfloat16& b) { 515 return bfloat16(::powf(float(a), float(b))); 516 } 517 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 sin(const bfloat16& a) { 518 return bfloat16(::sinf(float(a))); 519 } 520 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 cos(const bfloat16& a) { 521 return bfloat16(::cosf(float(a))); 522 } 523 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 tan(const bfloat16& a) { 524 return bfloat16(::tanf(float(a))); 525 } 526 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 asin(const bfloat16& a) { 527 return bfloat16(::asinf(float(a))); 528 } 529 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 acos(const bfloat16& a) { 530 return bfloat16(::acosf(float(a))); 531 } 532 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 atan(const bfloat16& a) { 533 return bfloat16(::atanf(float(a))); 534 } 535 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 sinh(const bfloat16& a) { 536 return bfloat16(::sinhf(float(a))); 537 } 538 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 cosh(const bfloat16& a) { 539 return bfloat16(::coshf(float(a))); 540 } 541 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 tanh(const bfloat16& a) { 542 return bfloat16(::tanhf(float(a))); 543 } 544 #if EIGEN_HAS_CXX11_MATH 545 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 asinh(const bfloat16& a) { 546 return bfloat16(::asinhf(float(a))); 547 } 548 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 acosh(const bfloat16& a) { 549 return bfloat16(::acoshf(float(a))); 550 } 551 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 atanh(const bfloat16& a) { 552 return bfloat16(::atanhf(float(a))); 553 } 554 #endif 555 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 floor(const bfloat16& a) { 556 return bfloat16(::floorf(float(a))); 557 } 558 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 ceil(const bfloat16& a) { 559 return bfloat16(::ceilf(float(a))); 560 } 561 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 rint(const bfloat16& a) { 562 return bfloat16(::rintf(float(a))); 563 } 564 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 round(const bfloat16& a) { 565 return bfloat16(::roundf(float(a))); 566 } 567 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 fmod(const bfloat16& a, const bfloat16& b) { 568 return bfloat16(::fmodf(float(a), float(b))); 569 } 570 571 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 (min)(const bfloat16& a, const bfloat16& b) { 572 const float f1 = static_cast<float>(a); 573 const float f2 = static_cast<float>(b); 574 return f2 < f1 ? b : a; 575 } 576 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 (max)(const bfloat16& a, const bfloat16& b) { 577 const float f1 = static_cast<float>(a); 578 const float f2 = static_cast<float>(b); 579 return f1 < f2 ? b : a; 580 } 581 582 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 fmin(const bfloat16& a, const bfloat16& b) { 583 const float f1 = static_cast<float>(a); 584 const float f2 = static_cast<float>(b); 585 return bfloat16(::fminf(f1, f2)); 586 } 587 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 fmax(const bfloat16& a, const bfloat16& b) { 588 const float f1 = static_cast<float>(a); 589 const float f2 = static_cast<float>(b); 590 return bfloat16(::fmaxf(f1, f2)); 591 } 592 593 #ifndef EIGEN_NO_IO 594 EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const bfloat16& v) { 595 os << static_cast<float>(v); 596 return os; 597 } 598 #endif 599 600 } // namespace bfloat16_impl 601 602 namespace internal { 603 604 template<> 605 struct random_default_impl<bfloat16, false, false> 606 { 607 static inline bfloat16 run(const bfloat16& x, const bfloat16& y) 608 { 609 return x + (y-x) * bfloat16(float(std::rand()) / float(RAND_MAX)); 610 } 611 static inline bfloat16 run() 612 { 613 return run(bfloat16(-1.f), bfloat16(1.f)); 614 } 615 }; 616 617 template<> struct is_arithmetic<bfloat16> { enum { value = true }; }; 618 619 } // namespace internal 620 621 template<> struct NumTraits<Eigen::bfloat16> 622 : GenericNumTraits<Eigen::bfloat16> 623 { 624 enum { 625 IsSigned = true, 626 IsInteger = false, 627 IsComplex = false, 628 RequireInitialization = false 629 }; 630 631 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 epsilon() { 632 return bfloat16_impl::raw_uint16_to_bfloat16(0x3c00); 633 } 634 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 dummy_precision() { 635 return bfloat16_impl::raw_uint16_to_bfloat16(0x3D4D); // bfloat16(5e-2f); 636 637 } 638 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 highest() { 639 return bfloat16_impl::raw_uint16_to_bfloat16(0x7F7F); 640 } 641 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 lowest() { 642 return bfloat16_impl::raw_uint16_to_bfloat16(0xFF7F); 643 } 644 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 infinity() { 645 return bfloat16_impl::raw_uint16_to_bfloat16(0x7f80); 646 } 647 EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::bfloat16 quiet_NaN() { 648 return bfloat16_impl::raw_uint16_to_bfloat16(0x7fc0); 649 } 650 }; 651 652 } // namespace Eigen 653 654 namespace Eigen { 655 namespace numext { 656 657 template<> 658 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE 659 bool (isnan)(const Eigen::bfloat16& h) { 660 return (bfloat16_impl::isnan)(h); 661 } 662 663 template<> 664 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE 665 bool (isinf)(const Eigen::bfloat16& h) { 666 return (bfloat16_impl::isinf)(h); 667 } 668 669 template<> 670 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE 671 bool (isfinite)(const Eigen::bfloat16& h) { 672 return (bfloat16_impl::isfinite)(h); 673 } 674 675 template <> 676 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::bfloat16 bit_cast<Eigen::bfloat16, uint16_t>(const uint16_t& src) { 677 return Eigen::bfloat16(Eigen::bfloat16_impl::raw_uint16_to_bfloat16(src)); 678 } 679 680 template <> 681 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC uint16_t bit_cast<uint16_t, Eigen::bfloat16>(const Eigen::bfloat16& src) { 682 return Eigen::bfloat16_impl::raw_bfloat16_as_uint16(src); 683 } 684 685 } // namespace numext 686 } // namespace Eigen 687 688 #if EIGEN_HAS_STD_HASH 689 namespace std { 690 template <> 691 struct hash<Eigen::bfloat16> { 692 EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::bfloat16& a) const { 693 return static_cast<std::size_t>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(a)); 694 } 695 }; 696 } // namespace std 697 #endif 698 699 700 #endif // EIGEN_BFLOAT16_H 701