| 1 | // (C) Copyright John Maddock 2008. |
| 2 | // Use, modification and distribution are subject to the |
| 3 | // Boost Software License, Version 1.0. (See accompanying file |
| 4 | // LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 5 | |
| 6 | #ifndef BOOST_MATH_SPECIAL_NEXT_HPP |
| 7 | #define BOOST_MATH_SPECIAL_NEXT_HPP |
| 8 | |
| 9 | #ifdef _MSC_VER |
| 10 | #pragma once |
| 11 | #endif |
| 12 | |
| 13 | #include <boost/math/special_functions/math_fwd.hpp> |
| 14 | #include <boost/math/policies/error_handling.hpp> |
| 15 | #include <boost/math/special_functions/fpclassify.hpp> |
| 16 | #include <boost/math/special_functions/sign.hpp> |
| 17 | #include <boost/math/special_functions/trunc.hpp> |
| 18 | |
| 19 | #include <float.h> |
| 20 | |
| 21 | #if !defined(_CRAYC) && !defined(__CUDACC__) && (!defined(__GNUC__) || (__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ > 3))) |
| 22 | #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || defined(__SSE2__) |
| 23 | #include "xmmintrin.h" |
| 24 | #define BOOST_MATH_CHECK_SSE2 |
| 25 | #endif |
| 26 | #endif |
| 27 | |
| 28 | namespace boost{ namespace math{ |
| 29 | |
| 30 | namespace concepts { |
| 31 | |
| 32 | class real_concept; |
| 33 | class std_real_concept; |
| 34 | |
| 35 | } |
| 36 | |
| 37 | namespace detail{ |
| 38 | |
| 39 | template <class T> |
| 40 | struct has_hidden_guard_digits; |
| 41 | template <> |
| 42 | struct has_hidden_guard_digits<float> : public boost::false_type {}; |
| 43 | template <> |
| 44 | struct has_hidden_guard_digits<double> : public boost::false_type {}; |
| 45 | template <> |
| 46 | struct has_hidden_guard_digits<long double> : public boost::false_type {}; |
| 47 | #ifdef BOOST_HAS_FLOAT128 |
| 48 | template <> |
| 49 | struct has_hidden_guard_digits<__float128> : public boost::false_type {}; |
| 50 | #endif |
| 51 | template <> |
| 52 | struct has_hidden_guard_digits<boost::math::concepts::real_concept> : public boost::false_type {}; |
| 53 | template <> |
| 54 | struct has_hidden_guard_digits<boost::math::concepts::std_real_concept> : public boost::false_type {}; |
| 55 | |
| 56 | template <class T, bool b> |
| 57 | struct has_hidden_guard_digits_10 : public boost::false_type {}; |
| 58 | template <class T> |
| 59 | struct has_hidden_guard_digits_10<T, true> : public boost::integral_constant<bool, (std::numeric_limits<T>::digits10 != std::numeric_limits<T>::max_digits10)> {}; |
| 60 | |
| 61 | template <class T> |
| 62 | struct has_hidden_guard_digits |
| 63 | : public has_hidden_guard_digits_10<T, |
| 64 | std::numeric_limits<T>::is_specialized |
| 65 | && (std::numeric_limits<T>::radix == 10) > |
| 66 | {}; |
| 67 | |
| 68 | template <class T> |
| 69 | inline const T& normalize_value(const T& val, const boost::false_type&) { return val; } |
| 70 | template <class T> |
| 71 | inline T normalize_value(const T& val, const boost::true_type&) |
| 72 | { |
| 73 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 74 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 75 | |
| 76 | boost::intmax_t shift = (boost::intmax_t)std::numeric_limits<T>::digits - (boost::intmax_t)ilogb(val) - 1; |
| 77 | T result = scalbn(val, shift); |
| 78 | result = round(result); |
| 79 | return scalbn(result, -shift); |
| 80 | } |
| 81 | |
| 82 | template <class T> |
| 83 | inline T get_smallest_value(boost::true_type const&) |
| 84 | { |
| 85 | // |
| 86 | // numeric_limits lies about denorms being present - particularly |
| 87 | // when this can be turned on or off at runtime, as is the case |
| 88 | // when using the SSE2 registers in DAZ or FTZ mode. |
| 89 | // |
| 90 | static const T m = std::numeric_limits<T>::denorm_min(); |
| 91 | #ifdef BOOST_MATH_CHECK_SSE2 |
| 92 | return (_mm_getcsr() & (_MM_FLUSH_ZERO_ON | 0x40)) ? tools::min_value<T>() : m;; |
| 93 | #else |
| 94 | return ((tools::min_value<T>() / 2) == 0) ? tools::min_value<T>() : m; |
| 95 | #endif |
| 96 | } |
| 97 | |
| 98 | template <class T> |
| 99 | inline T get_smallest_value(boost::false_type const&) |
| 100 | { |
| 101 | return tools::min_value<T>(); |
| 102 | } |
| 103 | |
| 104 | template <class T> |
| 105 | inline T get_smallest_value() |
| 106 | { |
| 107 | #if defined(BOOST_MSVC) && (BOOST_MSVC <= 1310) |
| 108 | return get_smallest_value<T>(boost::integral_constant<bool, std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == 1)>()); |
| 109 | #else |
| 110 | return get_smallest_value<T>(boost::integral_constant<bool, std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == std::denorm_present)>()); |
| 111 | #endif |
| 112 | } |
| 113 | |
| 114 | // |
| 115 | // Returns the smallest value that won't generate denorms when |
| 116 | // we calculate the value of the least-significant-bit: |
| 117 | // |
| 118 | template <class T> |
| 119 | T get_min_shift_value(); |
| 120 | |
| 121 | template <class T> |
| 122 | struct min_shift_initializer |
| 123 | { |
| 124 | struct init |
| 125 | { |
| 126 | init() |
| 127 | { |
| 128 | do_init(); |
| 129 | } |
| 130 | static void do_init() |
| 131 | { |
| 132 | get_min_shift_value<T>(); |
| 133 | } |
| 134 | void force_instantiate()const{} |
| 135 | }; |
| 136 | static const init initializer; |
| 137 | static void force_instantiate() |
| 138 | { |
| 139 | initializer.force_instantiate(); |
| 140 | } |
| 141 | }; |
| 142 | |
| 143 | template <class T> |
| 144 | const typename min_shift_initializer<T>::init min_shift_initializer<T>::initializer; |
| 145 | |
| 146 | template <class T> |
| 147 | inline T calc_min_shifted(const boost::true_type&) |
| 148 | { |
| 149 | BOOST_MATH_STD_USING |
| 150 | return ldexp(tools::min_value<T>(), tools::digits<T>() + 1); |
| 151 | } |
| 152 | template <class T> |
| 153 | inline T calc_min_shifted(const boost::false_type&) |
| 154 | { |
| 155 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 156 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 157 | |
| 158 | return scalbn(tools::min_value<T>(), std::numeric_limits<T>::digits + 1); |
| 159 | } |
| 160 | |
| 161 | |
| 162 | template <class T> |
| 163 | inline T get_min_shift_value() |
| 164 | { |
| 165 | static const T val = calc_min_shifted<T>(boost::integral_constant<bool, !std::numeric_limits<T>::is_specialized || std::numeric_limits<T>::radix == 2>()); |
| 166 | min_shift_initializer<T>::force_instantiate(); |
| 167 | |
| 168 | return val; |
| 169 | } |
| 170 | |
| 171 | template <class T, class Policy> |
| 172 | T float_next_imp(const T& val, const boost::true_type&, const Policy& pol) |
| 173 | { |
| 174 | BOOST_MATH_STD_USING |
| 175 | int expon; |
| 176 | static const char* function = "float_next<%1%>(%1%)" ; |
| 177 | |
| 178 | int fpclass = (boost::math::fpclassify)(val); |
| 179 | |
| 180 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 181 | { |
| 182 | if(val < 0) |
| 183 | return -tools::max_value<T>(); |
| 184 | return policies::raise_domain_error<T>( |
| 185 | function, |
| 186 | "Argument must be finite, but got %1%" , val, pol); |
| 187 | } |
| 188 | |
| 189 | if(val >= tools::max_value<T>()) |
| 190 | return policies::raise_overflow_error<T>(function, 0, pol); |
| 191 | |
| 192 | if(val == 0) |
| 193 | return detail::get_smallest_value<T>(); |
| 194 | |
| 195 | if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>())) |
| 196 | { |
| 197 | // |
| 198 | // Special case: if the value of the least significant bit is a denorm, and the result |
| 199 | // would not be a denorm, then shift the input, increment, and shift back. |
| 200 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 201 | // |
| 202 | return ldexp(float_next(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>()); |
| 203 | } |
| 204 | |
| 205 | if(-0.5f == frexp(val, &expon)) |
| 206 | --expon; // reduce exponent when val is a power of two, and negative. |
| 207 | T diff = ldexp(T(1), expon - tools::digits<T>()); |
| 208 | if(diff == 0) |
| 209 | diff = detail::get_smallest_value<T>(); |
| 210 | return val + diff; |
| 211 | } // float_next_imp |
| 212 | // |
| 213 | // Special version for some base other than 2: |
| 214 | // |
| 215 | template <class T, class Policy> |
| 216 | T float_next_imp(const T& val, const boost::false_type&, const Policy& pol) |
| 217 | { |
| 218 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 219 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 220 | |
| 221 | BOOST_MATH_STD_USING |
| 222 | boost::intmax_t expon; |
| 223 | static const char* function = "float_next<%1%>(%1%)" ; |
| 224 | |
| 225 | int fpclass = (boost::math::fpclassify)(val); |
| 226 | |
| 227 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 228 | { |
| 229 | if(val < 0) |
| 230 | return -tools::max_value<T>(); |
| 231 | return policies::raise_domain_error<T>( |
| 232 | function, |
| 233 | "Argument must be finite, but got %1%" , val, pol); |
| 234 | } |
| 235 | |
| 236 | if(val >= tools::max_value<T>()) |
| 237 | return policies::raise_overflow_error<T>(function, 0, pol); |
| 238 | |
| 239 | if(val == 0) |
| 240 | return detail::get_smallest_value<T>(); |
| 241 | |
| 242 | if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>())) |
| 243 | { |
| 244 | // |
| 245 | // Special case: if the value of the least significant bit is a denorm, and the result |
| 246 | // would not be a denorm, then shift the input, increment, and shift back. |
| 247 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 248 | // |
| 249 | return scalbn(float_next(T(scalbn(val, 2 * std::numeric_limits<T>::digits)), pol), -2 * std::numeric_limits<T>::digits); |
| 250 | } |
| 251 | |
| 252 | expon = 1 + ilogb(val); |
| 253 | if(-1 == scalbn(val, -expon) * std::numeric_limits<T>::radix) |
| 254 | --expon; // reduce exponent when val is a power of base, and negative. |
| 255 | T diff = scalbn(T(1), expon - std::numeric_limits<T>::digits); |
| 256 | if(diff == 0) |
| 257 | diff = detail::get_smallest_value<T>(); |
| 258 | return val + diff; |
| 259 | } // float_next_imp |
| 260 | |
| 261 | } // namespace detail |
| 262 | |
| 263 | template <class T, class Policy> |
| 264 | inline typename tools::promote_args<T>::type float_next(const T& val, const Policy& pol) |
| 265 | { |
| 266 | typedef typename tools::promote_args<T>::type result_type; |
| 267 | return detail::float_next_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), boost::integral_constant<bool, !std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol); |
| 268 | } |
| 269 | |
| 270 | #if 0 //def BOOST_MSVC |
| 271 | // |
| 272 | // We used to use ::_nextafter here, but doing so fails when using |
| 273 | // the SSE2 registers if the FTZ or DAZ flags are set, so use our own |
| 274 | // - albeit slower - code instead as at least that gives the correct answer. |
| 275 | // |
| 276 | template <class Policy> |
| 277 | inline double float_next(const double& val, const Policy& pol) |
| 278 | { |
| 279 | static const char* function = "float_next<%1%>(%1%)" ; |
| 280 | |
| 281 | if(!(boost::math::isfinite)(val) && (val > 0)) |
| 282 | return policies::raise_domain_error<double>( |
| 283 | function, |
| 284 | "Argument must be finite, but got %1%" , val, pol); |
| 285 | |
| 286 | if(val >= tools::max_value<double>()) |
| 287 | return policies::raise_overflow_error<double>(function, 0, pol); |
| 288 | |
| 289 | return ::_nextafter(val, tools::max_value<double>()); |
| 290 | } |
| 291 | #endif |
| 292 | |
| 293 | template <class T> |
| 294 | inline typename tools::promote_args<T>::type float_next(const T& val) |
| 295 | { |
| 296 | return float_next(val, policies::policy<>()); |
| 297 | } |
| 298 | |
| 299 | namespace detail{ |
| 300 | |
| 301 | template <class T, class Policy> |
| 302 | T float_prior_imp(const T& val, const boost::true_type&, const Policy& pol) |
| 303 | { |
| 304 | BOOST_MATH_STD_USING |
| 305 | int expon; |
| 306 | static const char* function = "float_prior<%1%>(%1%)" ; |
| 307 | |
| 308 | int fpclass = (boost::math::fpclassify)(val); |
| 309 | |
| 310 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 311 | { |
| 312 | if(val > 0) |
| 313 | return tools::max_value<T>(); |
| 314 | return policies::raise_domain_error<T>( |
| 315 | function, |
| 316 | "Argument must be finite, but got %1%" , val, pol); |
| 317 | } |
| 318 | |
| 319 | if(val <= -tools::max_value<T>()) |
| 320 | return -policies::raise_overflow_error<T>(function, 0, pol); |
| 321 | |
| 322 | if(val == 0) |
| 323 | return -detail::get_smallest_value<T>(); |
| 324 | |
| 325 | if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>())) |
| 326 | { |
| 327 | // |
| 328 | // Special case: if the value of the least significant bit is a denorm, and the result |
| 329 | // would not be a denorm, then shift the input, increment, and shift back. |
| 330 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 331 | // |
| 332 | return ldexp(float_prior(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>()); |
| 333 | } |
| 334 | |
| 335 | T remain = frexp(val, &expon); |
| 336 | if(remain == 0.5f) |
| 337 | --expon; // when val is a power of two we must reduce the exponent |
| 338 | T diff = ldexp(T(1), expon - tools::digits<T>()); |
| 339 | if(diff == 0) |
| 340 | diff = detail::get_smallest_value<T>(); |
| 341 | return val - diff; |
| 342 | } // float_prior_imp |
| 343 | // |
| 344 | // Special version for bases other than 2: |
| 345 | // |
| 346 | template <class T, class Policy> |
| 347 | T float_prior_imp(const T& val, const boost::false_type&, const Policy& pol) |
| 348 | { |
| 349 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 350 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 351 | |
| 352 | BOOST_MATH_STD_USING |
| 353 | boost::intmax_t expon; |
| 354 | static const char* function = "float_prior<%1%>(%1%)" ; |
| 355 | |
| 356 | int fpclass = (boost::math::fpclassify)(val); |
| 357 | |
| 358 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 359 | { |
| 360 | if(val > 0) |
| 361 | return tools::max_value<T>(); |
| 362 | return policies::raise_domain_error<T>( |
| 363 | function, |
| 364 | "Argument must be finite, but got %1%" , val, pol); |
| 365 | } |
| 366 | |
| 367 | if(val <= -tools::max_value<T>()) |
| 368 | return -policies::raise_overflow_error<T>(function, 0, pol); |
| 369 | |
| 370 | if(val == 0) |
| 371 | return -detail::get_smallest_value<T>(); |
| 372 | |
| 373 | if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>())) |
| 374 | { |
| 375 | // |
| 376 | // Special case: if the value of the least significant bit is a denorm, and the result |
| 377 | // would not be a denorm, then shift the input, increment, and shift back. |
| 378 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 379 | // |
| 380 | return scalbn(float_prior(T(scalbn(val, 2 * std::numeric_limits<T>::digits)), pol), -2 * std::numeric_limits<T>::digits); |
| 381 | } |
| 382 | |
| 383 | expon = 1 + ilogb(val); |
| 384 | T remain = scalbn(val, -expon); |
| 385 | if(remain * std::numeric_limits<T>::radix == 1) |
| 386 | --expon; // when val is a power of two we must reduce the exponent |
| 387 | T diff = scalbn(T(1), expon - std::numeric_limits<T>::digits); |
| 388 | if(diff == 0) |
| 389 | diff = detail::get_smallest_value<T>(); |
| 390 | return val - diff; |
| 391 | } // float_prior_imp |
| 392 | |
| 393 | } // namespace detail |
| 394 | |
| 395 | template <class T, class Policy> |
| 396 | inline typename tools::promote_args<T>::type float_prior(const T& val, const Policy& pol) |
| 397 | { |
| 398 | typedef typename tools::promote_args<T>::type result_type; |
| 399 | return detail::float_prior_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), boost::integral_constant<bool, !std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol); |
| 400 | } |
| 401 | |
| 402 | #if 0 //def BOOST_MSVC |
| 403 | // |
| 404 | // We used to use ::_nextafter here, but doing so fails when using |
| 405 | // the SSE2 registers if the FTZ or DAZ flags are set, so use our own |
| 406 | // - albeit slower - code instead as at least that gives the correct answer. |
| 407 | // |
| 408 | template <class Policy> |
| 409 | inline double float_prior(const double& val, const Policy& pol) |
| 410 | { |
| 411 | static const char* function = "float_prior<%1%>(%1%)" ; |
| 412 | |
| 413 | if(!(boost::math::isfinite)(val) && (val < 0)) |
| 414 | return policies::raise_domain_error<double>( |
| 415 | function, |
| 416 | "Argument must be finite, but got %1%" , val, pol); |
| 417 | |
| 418 | if(val <= -tools::max_value<double>()) |
| 419 | return -policies::raise_overflow_error<double>(function, 0, pol); |
| 420 | |
| 421 | return ::_nextafter(val, -tools::max_value<double>()); |
| 422 | } |
| 423 | #endif |
| 424 | |
| 425 | template <class T> |
| 426 | inline typename tools::promote_args<T>::type float_prior(const T& val) |
| 427 | { |
| 428 | return float_prior(val, policies::policy<>()); |
| 429 | } |
| 430 | |
| 431 | template <class T, class U, class Policy> |
| 432 | inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction, const Policy& pol) |
| 433 | { |
| 434 | typedef typename tools::promote_args<T, U>::type result_type; |
| 435 | return val < direction ? boost::math::float_next<result_type>(val, pol) : val == direction ? val : boost::math::float_prior<result_type>(val, pol); |
| 436 | } |
| 437 | |
| 438 | template <class T, class U> |
| 439 | inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction) |
| 440 | { |
| 441 | return nextafter(val, direction, policies::policy<>()); |
| 442 | } |
| 443 | |
| 444 | namespace detail{ |
| 445 | |
| 446 | template <class T, class Policy> |
| 447 | T float_distance_imp(const T& a, const T& b, const boost::true_type&, const Policy& pol) |
| 448 | { |
| 449 | BOOST_MATH_STD_USING |
| 450 | // |
| 451 | // Error handling: |
| 452 | // |
| 453 | static const char* function = "float_distance<%1%>(%1%, %1%)" ; |
| 454 | if(!(boost::math::isfinite)(a)) |
| 455 | return policies::raise_domain_error<T>( |
| 456 | function, |
| 457 | "Argument a must be finite, but got %1%" , a, pol); |
| 458 | if(!(boost::math::isfinite)(b)) |
| 459 | return policies::raise_domain_error<T>( |
| 460 | function, |
| 461 | "Argument b must be finite, but got %1%" , b, pol); |
| 462 | // |
| 463 | // Special cases: |
| 464 | // |
| 465 | if(a > b) |
| 466 | return -float_distance(b, a, pol); |
| 467 | if(a == b) |
| 468 | return T(0); |
| 469 | if(a == 0) |
| 470 | return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)); |
| 471 | if(b == 0) |
| 472 | return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); |
| 473 | if(boost::math::sign(a) != boost::math::sign(b)) |
| 474 | return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)) |
| 475 | + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); |
| 476 | // |
| 477 | // By the time we get here, both a and b must have the same sign, we want |
| 478 | // b > a and both positive for the following logic: |
| 479 | // |
| 480 | if(a < 0) |
| 481 | return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol); |
| 482 | |
| 483 | BOOST_ASSERT(a >= 0); |
| 484 | BOOST_ASSERT(b >= a); |
| 485 | |
| 486 | int expon; |
| 487 | // |
| 488 | // Note that if a is a denorm then the usual formula fails |
| 489 | // because we actually have fewer than tools::digits<T>() |
| 490 | // significant bits in the representation: |
| 491 | // |
| 492 | (void)frexp(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) ? tools::min_value<T>() : a, &expon); |
| 493 | T upper = ldexp(T(1), expon); |
| 494 | T result = T(0); |
| 495 | // |
| 496 | // If b is greater than upper, then we *must* split the calculation |
| 497 | // as the size of the ULP changes with each order of magnitude change: |
| 498 | // |
| 499 | if(b > upper) |
| 500 | { |
| 501 | int expon2; |
| 502 | (void)frexp(b, &expon2); |
| 503 | T upper2 = ldexp(T(0.5), expon2); |
| 504 | result = float_distance(upper2, b); |
| 505 | result += (expon2 - expon - 1) * ldexp(T(1), tools::digits<T>() - 1); |
| 506 | } |
| 507 | // |
| 508 | // Use compensated double-double addition to avoid rounding |
| 509 | // errors in the subtraction: |
| 510 | // |
| 511 | expon = tools::digits<T>() - expon; |
| 512 | T mb, x, y, z; |
| 513 | if(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) || (b - a < tools::min_value<T>())) |
| 514 | { |
| 515 | // |
| 516 | // Special case - either one end of the range is a denormal, or else the difference is. |
| 517 | // The regular code will fail if we're using the SSE2 registers on Intel and either |
| 518 | // the FTZ or DAZ flags are set. |
| 519 | // |
| 520 | T a2 = ldexp(a, tools::digits<T>()); |
| 521 | T b2 = ldexp(b, tools::digits<T>()); |
| 522 | mb = -(std::min)(T(ldexp(upper, tools::digits<T>())), b2); |
| 523 | x = a2 + mb; |
| 524 | z = x - a2; |
| 525 | y = (a2 - (x - z)) + (mb - z); |
| 526 | |
| 527 | expon -= tools::digits<T>(); |
| 528 | } |
| 529 | else |
| 530 | { |
| 531 | mb = -(std::min)(upper, b); |
| 532 | x = a + mb; |
| 533 | z = x - a; |
| 534 | y = (a - (x - z)) + (mb - z); |
| 535 | } |
| 536 | if(x < 0) |
| 537 | { |
| 538 | x = -x; |
| 539 | y = -y; |
| 540 | } |
| 541 | result += ldexp(x, expon) + ldexp(y, expon); |
| 542 | // |
| 543 | // Result must be an integer: |
| 544 | // |
| 545 | BOOST_ASSERT(result == floor(result)); |
| 546 | return result; |
| 547 | } // float_distance_imp |
| 548 | // |
| 549 | // Special versions for bases other than 2: |
| 550 | // |
| 551 | template <class T, class Policy> |
| 552 | T float_distance_imp(const T& a, const T& b, const boost::false_type&, const Policy& pol) |
| 553 | { |
| 554 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 555 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 556 | |
| 557 | BOOST_MATH_STD_USING |
| 558 | // |
| 559 | // Error handling: |
| 560 | // |
| 561 | static const char* function = "float_distance<%1%>(%1%, %1%)" ; |
| 562 | if(!(boost::math::isfinite)(a)) |
| 563 | return policies::raise_domain_error<T>( |
| 564 | function, |
| 565 | "Argument a must be finite, but got %1%" , a, pol); |
| 566 | if(!(boost::math::isfinite)(b)) |
| 567 | return policies::raise_domain_error<T>( |
| 568 | function, |
| 569 | "Argument b must be finite, but got %1%" , b, pol); |
| 570 | // |
| 571 | // Special cases: |
| 572 | // |
| 573 | if(a > b) |
| 574 | return -float_distance(b, a, pol); |
| 575 | if(a == b) |
| 576 | return T(0); |
| 577 | if(a == 0) |
| 578 | return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)); |
| 579 | if(b == 0) |
| 580 | return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); |
| 581 | if(boost::math::sign(a) != boost::math::sign(b)) |
| 582 | return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)) |
| 583 | + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); |
| 584 | // |
| 585 | // By the time we get here, both a and b must have the same sign, we want |
| 586 | // b > a and both positive for the following logic: |
| 587 | // |
| 588 | if(a < 0) |
| 589 | return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol); |
| 590 | |
| 591 | BOOST_ASSERT(a >= 0); |
| 592 | BOOST_ASSERT(b >= a); |
| 593 | |
| 594 | boost::intmax_t expon; |
| 595 | // |
| 596 | // Note that if a is a denorm then the usual formula fails |
| 597 | // because we actually have fewer than tools::digits<T>() |
| 598 | // significant bits in the representation: |
| 599 | // |
| 600 | expon = 1 + ilogb(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) ? tools::min_value<T>() : a); |
| 601 | T upper = scalbn(T(1), expon); |
| 602 | T result = T(0); |
| 603 | // |
| 604 | // If b is greater than upper, then we *must* split the calculation |
| 605 | // as the size of the ULP changes with each order of magnitude change: |
| 606 | // |
| 607 | if(b > upper) |
| 608 | { |
| 609 | boost::intmax_t expon2 = 1 + ilogb(b); |
| 610 | T upper2 = scalbn(T(1), expon2 - 1); |
| 611 | result = float_distance(upper2, b); |
| 612 | result += (expon2 - expon - 1) * scalbn(T(1), std::numeric_limits<T>::digits - 1); |
| 613 | } |
| 614 | // |
| 615 | // Use compensated double-double addition to avoid rounding |
| 616 | // errors in the subtraction: |
| 617 | // |
| 618 | expon = std::numeric_limits<T>::digits - expon; |
| 619 | T mb, x, y, z; |
| 620 | if(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) || (b - a < tools::min_value<T>())) |
| 621 | { |
| 622 | // |
| 623 | // Special case - either one end of the range is a denormal, or else the difference is. |
| 624 | // The regular code will fail if we're using the SSE2 registers on Intel and either |
| 625 | // the FTZ or DAZ flags are set. |
| 626 | // |
| 627 | T a2 = scalbn(a, std::numeric_limits<T>::digits); |
| 628 | T b2 = scalbn(b, std::numeric_limits<T>::digits); |
| 629 | mb = -(std::min)(T(scalbn(upper, std::numeric_limits<T>::digits)), b2); |
| 630 | x = a2 + mb; |
| 631 | z = x - a2; |
| 632 | y = (a2 - (x - z)) + (mb - z); |
| 633 | |
| 634 | expon -= std::numeric_limits<T>::digits; |
| 635 | } |
| 636 | else |
| 637 | { |
| 638 | mb = -(std::min)(upper, b); |
| 639 | x = a + mb; |
| 640 | z = x - a; |
| 641 | y = (a - (x - z)) + (mb - z); |
| 642 | } |
| 643 | if(x < 0) |
| 644 | { |
| 645 | x = -x; |
| 646 | y = -y; |
| 647 | } |
| 648 | result += scalbn(x, expon) + scalbn(y, expon); |
| 649 | // |
| 650 | // Result must be an integer: |
| 651 | // |
| 652 | BOOST_ASSERT(result == floor(result)); |
| 653 | return result; |
| 654 | } // float_distance_imp |
| 655 | |
| 656 | } // namespace detail |
| 657 | |
| 658 | template <class T, class U, class Policy> |
| 659 | inline typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b, const Policy& pol) |
| 660 | { |
| 661 | typedef typename tools::promote_args<T, U>::type result_type; |
| 662 | return detail::float_distance_imp(detail::normalize_value(static_cast<result_type>(a), typename detail::has_hidden_guard_digits<result_type>::type()), detail::normalize_value(static_cast<result_type>(b), typename detail::has_hidden_guard_digits<result_type>::type()), boost::integral_constant<bool, !std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol); |
| 663 | } |
| 664 | |
| 665 | template <class T, class U> |
| 666 | typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b) |
| 667 | { |
| 668 | return boost::math::float_distance(a, b, policies::policy<>()); |
| 669 | } |
| 670 | |
| 671 | namespace detail{ |
| 672 | |
| 673 | template <class T, class Policy> |
| 674 | T float_advance_imp(T val, int distance, const boost::true_type&, const Policy& pol) |
| 675 | { |
| 676 | BOOST_MATH_STD_USING |
| 677 | // |
| 678 | // Error handling: |
| 679 | // |
| 680 | static const char* function = "float_advance<%1%>(%1%, int)" ; |
| 681 | |
| 682 | int fpclass = (boost::math::fpclassify)(val); |
| 683 | |
| 684 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 685 | return policies::raise_domain_error<T>( |
| 686 | function, |
| 687 | "Argument val must be finite, but got %1%" , val, pol); |
| 688 | |
| 689 | if(val < 0) |
| 690 | return -float_advance(-val, -distance, pol); |
| 691 | if(distance == 0) |
| 692 | return val; |
| 693 | if(distance == 1) |
| 694 | return float_next(val, pol); |
| 695 | if(distance == -1) |
| 696 | return float_prior(val, pol); |
| 697 | |
| 698 | if(fabs(val) < detail::get_min_shift_value<T>()) |
| 699 | { |
| 700 | // |
| 701 | // Special case: if the value of the least significant bit is a denorm, |
| 702 | // implement in terms of float_next/float_prior. |
| 703 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 704 | // |
| 705 | if(distance > 0) |
| 706 | { |
| 707 | do{ val = float_next(val, pol); } while(--distance); |
| 708 | } |
| 709 | else |
| 710 | { |
| 711 | do{ val = float_prior(val, pol); } while(++distance); |
| 712 | } |
| 713 | return val; |
| 714 | } |
| 715 | |
| 716 | int expon; |
| 717 | (void)frexp(val, &expon); |
| 718 | T limit = ldexp((distance < 0 ? T(0.5f) : T(1)), expon); |
| 719 | if(val <= tools::min_value<T>()) |
| 720 | { |
| 721 | limit = sign(T(distance)) * tools::min_value<T>(); |
| 722 | } |
| 723 | T limit_distance = float_distance(val, limit); |
| 724 | while(fabs(limit_distance) < abs(distance)) |
| 725 | { |
| 726 | distance -= itrunc(limit_distance); |
| 727 | val = limit; |
| 728 | if(distance < 0) |
| 729 | { |
| 730 | limit /= 2; |
| 731 | expon--; |
| 732 | } |
| 733 | else |
| 734 | { |
| 735 | limit *= 2; |
| 736 | expon++; |
| 737 | } |
| 738 | limit_distance = float_distance(val, limit); |
| 739 | if(distance && (limit_distance == 0)) |
| 740 | { |
| 741 | return policies::raise_evaluation_error<T>(function, "Internal logic failed while trying to increment floating point value %1%: most likely your FPU is in non-IEEE conforming mode." , val, pol); |
| 742 | } |
| 743 | } |
| 744 | if((0.5f == frexp(val, &expon)) && (distance < 0)) |
| 745 | --expon; |
| 746 | T diff = 0; |
| 747 | if(val != 0) |
| 748 | diff = distance * ldexp(T(1), expon - tools::digits<T>()); |
| 749 | if(diff == 0) |
| 750 | diff = distance * detail::get_smallest_value<T>(); |
| 751 | return val += diff; |
| 752 | } // float_advance_imp |
| 753 | // |
| 754 | // Special version for bases other than 2: |
| 755 | // |
| 756 | template <class T, class Policy> |
| 757 | T float_advance_imp(T val, int distance, const boost::false_type&, const Policy& pol) |
| 758 | { |
| 759 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized); |
| 760 | BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2); |
| 761 | |
| 762 | BOOST_MATH_STD_USING |
| 763 | // |
| 764 | // Error handling: |
| 765 | // |
| 766 | static const char* function = "float_advance<%1%>(%1%, int)" ; |
| 767 | |
| 768 | int fpclass = (boost::math::fpclassify)(val); |
| 769 | |
| 770 | if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE)) |
| 771 | return policies::raise_domain_error<T>( |
| 772 | function, |
| 773 | "Argument val must be finite, but got %1%" , val, pol); |
| 774 | |
| 775 | if(val < 0) |
| 776 | return -float_advance(-val, -distance, pol); |
| 777 | if(distance == 0) |
| 778 | return val; |
| 779 | if(distance == 1) |
| 780 | return float_next(val, pol); |
| 781 | if(distance == -1) |
| 782 | return float_prior(val, pol); |
| 783 | |
| 784 | if(fabs(val) < detail::get_min_shift_value<T>()) |
| 785 | { |
| 786 | // |
| 787 | // Special case: if the value of the least significant bit is a denorm, |
| 788 | // implement in terms of float_next/float_prior. |
| 789 | // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. |
| 790 | // |
| 791 | if(distance > 0) |
| 792 | { |
| 793 | do{ val = float_next(val, pol); } while(--distance); |
| 794 | } |
| 795 | else |
| 796 | { |
| 797 | do{ val = float_prior(val, pol); } while(++distance); |
| 798 | } |
| 799 | return val; |
| 800 | } |
| 801 | |
| 802 | boost::intmax_t expon = 1 + ilogb(val); |
| 803 | T limit = scalbn(T(1), distance < 0 ? expon - 1 : expon); |
| 804 | if(val <= tools::min_value<T>()) |
| 805 | { |
| 806 | limit = sign(T(distance)) * tools::min_value<T>(); |
| 807 | } |
| 808 | T limit_distance = float_distance(val, limit); |
| 809 | while(fabs(limit_distance) < abs(distance)) |
| 810 | { |
| 811 | distance -= itrunc(limit_distance); |
| 812 | val = limit; |
| 813 | if(distance < 0) |
| 814 | { |
| 815 | limit /= std::numeric_limits<T>::radix; |
| 816 | expon--; |
| 817 | } |
| 818 | else |
| 819 | { |
| 820 | limit *= std::numeric_limits<T>::radix; |
| 821 | expon++; |
| 822 | } |
| 823 | limit_distance = float_distance(val, limit); |
| 824 | if(distance && (limit_distance == 0)) |
| 825 | { |
| 826 | return policies::raise_evaluation_error<T>(function, "Internal logic failed while trying to increment floating point value %1%: most likely your FPU is in non-IEEE conforming mode." , val, pol); |
| 827 | } |
| 828 | } |
| 829 | /*expon = 1 + ilogb(val); |
| 830 | if((1 == scalbn(val, 1 + expon)) && (distance < 0)) |
| 831 | --expon;*/ |
| 832 | T diff = 0; |
| 833 | if(val != 0) |
| 834 | diff = distance * scalbn(T(1), expon - std::numeric_limits<T>::digits); |
| 835 | if(diff == 0) |
| 836 | diff = distance * detail::get_smallest_value<T>(); |
| 837 | return val += diff; |
| 838 | } // float_advance_imp |
| 839 | |
| 840 | } // namespace detail |
| 841 | |
| 842 | template <class T, class Policy> |
| 843 | inline typename tools::promote_args<T>::type float_advance(T val, int distance, const Policy& pol) |
| 844 | { |
| 845 | typedef typename tools::promote_args<T>::type result_type; |
| 846 | return detail::float_advance_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), distance, boost::integral_constant<bool, !std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol); |
| 847 | } |
| 848 | |
| 849 | template <class T> |
| 850 | inline typename tools::promote_args<T>::type float_advance(const T& val, int distance) |
| 851 | { |
| 852 | return boost::math::float_advance(val, distance, policies::policy<>()); |
| 853 | } |
| 854 | |
| 855 | }} // boost math namespaces |
| 856 | |
| 857 | #endif // BOOST_MATH_SPECIAL_NEXT_HPP |
| 858 | |
| 859 | |