Mercurial > hg > vamp-build-and-test
comparison DEPENDENCIES/generic/include/boost/math/special_functions/next.hpp @ 16:2665513ce2d3
Add boost headers
author | Chris Cannam |
---|---|
date | Tue, 05 Aug 2014 11:11:38 +0100 |
parents | |
children | c530137014c0 |
comparison
equal
deleted
inserted
replaced
15:663ca0da4350 | 16:2665513ce2d3 |
---|---|
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/policies/error_handling.hpp> | |
14 #include <boost/math/special_functions/fpclassify.hpp> | |
15 #include <boost/math/special_functions/sign.hpp> | |
16 #include <boost/math/special_functions/trunc.hpp> | |
17 | |
18 #ifdef BOOST_MSVC | |
19 #include <float.h> | |
20 #endif | |
21 | |
22 namespace boost{ namespace math{ | |
23 | |
24 namespace detail{ | |
25 | |
26 template <class T> | |
27 inline T get_smallest_value(mpl::true_ const&) | |
28 { | |
29 // | |
30 // numeric_limits lies about denorms being present - particularly | |
31 // when this can be turned on or off at runtime, as is the case | |
32 // when using the SSE2 registers in DAZ or FTZ mode. | |
33 // | |
34 static const T m = std::numeric_limits<T>::denorm_min(); | |
35 return ((tools::min_value<T>() - m) == tools::min_value<T>()) ? tools::min_value<T>() : m; | |
36 } | |
37 | |
38 template <class T> | |
39 inline T get_smallest_value(mpl::false_ const&) | |
40 { | |
41 return tools::min_value<T>(); | |
42 } | |
43 | |
44 template <class T> | |
45 inline T get_smallest_value() | |
46 { | |
47 #if defined(BOOST_MSVC) && (BOOST_MSVC <= 1310) | |
48 return get_smallest_value<T>(mpl::bool_<std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == 1)>()); | |
49 #else | |
50 return get_smallest_value<T>(mpl::bool_<std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == std::denorm_present)>()); | |
51 #endif | |
52 } | |
53 | |
54 // | |
55 // Returns the smallest value that won't generate denorms when | |
56 // we calculate the value of the least-significant-bit: | |
57 // | |
58 template <class T> | |
59 T get_min_shift_value(); | |
60 | |
61 template <class T> | |
62 struct min_shift_initializer | |
63 { | |
64 struct init | |
65 { | |
66 init() | |
67 { | |
68 do_init(); | |
69 } | |
70 static void do_init() | |
71 { | |
72 get_min_shift_value<T>(); | |
73 } | |
74 void force_instantiate()const{} | |
75 }; | |
76 static const init initializer; | |
77 static void force_instantiate() | |
78 { | |
79 initializer.force_instantiate(); | |
80 } | |
81 }; | |
82 | |
83 template <class T> | |
84 const typename min_shift_initializer<T>::init min_shift_initializer<T>::initializer; | |
85 | |
86 | |
87 template <class T> | |
88 inline T get_min_shift_value() | |
89 { | |
90 BOOST_MATH_STD_USING | |
91 static const T val = ldexp(tools::min_value<T>(), tools::digits<T>() + 1); | |
92 min_shift_initializer<T>::force_instantiate(); | |
93 | |
94 return val; | |
95 } | |
96 | |
97 template <class T, class Policy> | |
98 T float_next_imp(const T& val, const Policy& pol) | |
99 { | |
100 BOOST_MATH_STD_USING | |
101 int expon; | |
102 static const char* function = "float_next<%1%>(%1%)"; | |
103 | |
104 int fpclass = (boost::math::fpclassify)(val); | |
105 | |
106 if((fpclass == FP_NAN) || (fpclass == FP_INFINITE)) | |
107 { | |
108 if(val < 0) | |
109 return -tools::max_value<T>(); | |
110 return policies::raise_domain_error<T>( | |
111 function, | |
112 "Argument must be finite, but got %1%", val, pol); | |
113 } | |
114 | |
115 if(val >= tools::max_value<T>()) | |
116 return policies::raise_overflow_error<T>(function, 0, pol); | |
117 | |
118 if(val == 0) | |
119 return detail::get_smallest_value<T>(); | |
120 | |
121 if((fpclass != FP_SUBNORMAL) && (fpclass != FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>())) | |
122 { | |
123 // | |
124 // Special case: if the value of the least significant bit is a denorm, and the result | |
125 // would not be a denorm, then shift the input, increment, and shift back. | |
126 // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. | |
127 // | |
128 return ldexp(float_next(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>()); | |
129 } | |
130 | |
131 if(-0.5f == frexp(val, &expon)) | |
132 --expon; // reduce exponent when val is a power of two, and negative. | |
133 T diff = ldexp(T(1), expon - tools::digits<T>()); | |
134 if(diff == 0) | |
135 diff = detail::get_smallest_value<T>(); | |
136 return val + diff; | |
137 } | |
138 | |
139 } | |
140 | |
141 template <class T, class Policy> | |
142 inline typename tools::promote_args<T>::type float_next(const T& val, const Policy& pol) | |
143 { | |
144 typedef typename tools::promote_args<T>::type result_type; | |
145 return detail::float_next_imp(static_cast<result_type>(val), pol); | |
146 } | |
147 | |
148 #if 0 //def BOOST_MSVC | |
149 // | |
150 // We used to use ::_nextafter here, but doing so fails when using | |
151 // the SSE2 registers if the FTZ or DAZ flags are set, so use our own | |
152 // - albeit slower - code instead as at least that gives the correct answer. | |
153 // | |
154 template <class Policy> | |
155 inline double float_next(const double& val, const Policy& pol) | |
156 { | |
157 static const char* function = "float_next<%1%>(%1%)"; | |
158 | |
159 if(!(boost::math::isfinite)(val) && (val > 0)) | |
160 return policies::raise_domain_error<double>( | |
161 function, | |
162 "Argument must be finite, but got %1%", val, pol); | |
163 | |
164 if(val >= tools::max_value<double>()) | |
165 return policies::raise_overflow_error<double>(function, 0, pol); | |
166 | |
167 return ::_nextafter(val, tools::max_value<double>()); | |
168 } | |
169 #endif | |
170 | |
171 template <class T> | |
172 inline typename tools::promote_args<T>::type float_next(const T& val) | |
173 { | |
174 return float_next(val, policies::policy<>()); | |
175 } | |
176 | |
177 namespace detail{ | |
178 | |
179 template <class T, class Policy> | |
180 T float_prior_imp(const T& val, const Policy& pol) | |
181 { | |
182 BOOST_MATH_STD_USING | |
183 int expon; | |
184 static const char* function = "float_prior<%1%>(%1%)"; | |
185 | |
186 int fpclass = (boost::math::fpclassify)(val); | |
187 | |
188 if((fpclass == FP_NAN) || (fpclass == FP_INFINITE)) | |
189 { | |
190 if(val > 0) | |
191 return tools::max_value<T>(); | |
192 return policies::raise_domain_error<T>( | |
193 function, | |
194 "Argument must be finite, but got %1%", val, pol); | |
195 } | |
196 | |
197 if(val <= -tools::max_value<T>()) | |
198 return -policies::raise_overflow_error<T>(function, 0, pol); | |
199 | |
200 if(val == 0) | |
201 return -detail::get_smallest_value<T>(); | |
202 | |
203 if((fpclass != FP_SUBNORMAL) && (fpclass != FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>())) | |
204 { | |
205 // | |
206 // Special case: if the value of the least significant bit is a denorm, and the result | |
207 // would not be a denorm, then shift the input, increment, and shift back. | |
208 // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. | |
209 // | |
210 return ldexp(float_prior(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>()); | |
211 } | |
212 | |
213 T remain = frexp(val, &expon); | |
214 if(remain == 0.5) | |
215 --expon; // when val is a power of two we must reduce the exponent | |
216 T diff = ldexp(T(1), expon - tools::digits<T>()); | |
217 if(diff == 0) | |
218 diff = detail::get_smallest_value<T>(); | |
219 return val - diff; | |
220 } | |
221 | |
222 } | |
223 | |
224 template <class T, class Policy> | |
225 inline typename tools::promote_args<T>::type float_prior(const T& val, const Policy& pol) | |
226 { | |
227 typedef typename tools::promote_args<T>::type result_type; | |
228 return detail::float_prior_imp(static_cast<result_type>(val), pol); | |
229 } | |
230 | |
231 #if 0 //def BOOST_MSVC | |
232 // | |
233 // We used to use ::_nextafter here, but doing so fails when using | |
234 // the SSE2 registers if the FTZ or DAZ flags are set, so use our own | |
235 // - albeit slower - code instead as at least that gives the correct answer. | |
236 // | |
237 template <class Policy> | |
238 inline double float_prior(const double& val, const Policy& pol) | |
239 { | |
240 static const char* function = "float_prior<%1%>(%1%)"; | |
241 | |
242 if(!(boost::math::isfinite)(val) && (val < 0)) | |
243 return policies::raise_domain_error<double>( | |
244 function, | |
245 "Argument must be finite, but got %1%", val, pol); | |
246 | |
247 if(val <= -tools::max_value<double>()) | |
248 return -policies::raise_overflow_error<double>(function, 0, pol); | |
249 | |
250 return ::_nextafter(val, -tools::max_value<double>()); | |
251 } | |
252 #endif | |
253 | |
254 template <class T> | |
255 inline typename tools::promote_args<T>::type float_prior(const T& val) | |
256 { | |
257 return float_prior(val, policies::policy<>()); | |
258 } | |
259 | |
260 template <class T, class U, class Policy> | |
261 inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction, const Policy& pol) | |
262 { | |
263 typedef typename tools::promote_args<T, U>::type result_type; | |
264 return val < direction ? boost::math::float_next<result_type>(val, pol) : val == direction ? val : boost::math::float_prior<result_type>(val, pol); | |
265 } | |
266 | |
267 template <class T, class U> | |
268 inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction) | |
269 { | |
270 return nextafter(val, direction, policies::policy<>()); | |
271 } | |
272 | |
273 namespace detail{ | |
274 | |
275 template <class T, class Policy> | |
276 T float_distance_imp(const T& a, const T& b, const Policy& pol) | |
277 { | |
278 BOOST_MATH_STD_USING | |
279 // | |
280 // Error handling: | |
281 // | |
282 static const char* function = "float_distance<%1%>(%1%, %1%)"; | |
283 if(!(boost::math::isfinite)(a)) | |
284 return policies::raise_domain_error<T>( | |
285 function, | |
286 "Argument a must be finite, but got %1%", a, pol); | |
287 if(!(boost::math::isfinite)(b)) | |
288 return policies::raise_domain_error<T>( | |
289 function, | |
290 "Argument b must be finite, but got %1%", b, pol); | |
291 // | |
292 // Special cases: | |
293 // | |
294 if(a > b) | |
295 return -float_distance(b, a, pol); | |
296 if(a == b) | |
297 return 0; | |
298 if(a == 0) | |
299 return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)); | |
300 if(b == 0) | |
301 return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); | |
302 if(boost::math::sign(a) != boost::math::sign(b)) | |
303 return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol)) | |
304 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol)); | |
305 // | |
306 // By the time we get here, both a and b must have the same sign, we want | |
307 // b > a and both postive for the following logic: | |
308 // | |
309 if(a < 0) | |
310 return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol); | |
311 | |
312 BOOST_ASSERT(a >= 0); | |
313 BOOST_ASSERT(b >= a); | |
314 | |
315 int expon; | |
316 // | |
317 // Note that if a is a denorm then the usual formula fails | |
318 // because we actually have fewer than tools::digits<T>() | |
319 // significant bits in the representation: | |
320 // | |
321 frexp(((boost::math::fpclassify)(a) == FP_SUBNORMAL) ? tools::min_value<T>() : a, &expon); | |
322 T upper = ldexp(T(1), expon); | |
323 T result = 0; | |
324 expon = tools::digits<T>() - expon; | |
325 // | |
326 // If b is greater than upper, then we *must* split the calculation | |
327 // as the size of the ULP changes with each order of magnitude change: | |
328 // | |
329 if(b > upper) | |
330 { | |
331 result = float_distance(upper, b); | |
332 } | |
333 // | |
334 // Use compensated double-double addition to avoid rounding | |
335 // errors in the subtraction: | |
336 // | |
337 T mb, x, y, z; | |
338 if(((boost::math::fpclassify)(a) == FP_SUBNORMAL) || (b - a < tools::min_value<T>())) | |
339 { | |
340 // | |
341 // Special case - either one end of the range is a denormal, or else the difference is. | |
342 // The regular code will fail if we're using the SSE2 registers on Intel and either | |
343 // the FTZ or DAZ flags are set. | |
344 // | |
345 T a2 = ldexp(a, tools::digits<T>()); | |
346 T b2 = ldexp(b, tools::digits<T>()); | |
347 mb = -(std::min)(T(ldexp(upper, tools::digits<T>())), b2); | |
348 x = a2 + mb; | |
349 z = x - a2; | |
350 y = (a2 - (x - z)) + (mb - z); | |
351 | |
352 expon -= tools::digits<T>(); | |
353 } | |
354 else | |
355 { | |
356 mb = -(std::min)(upper, b); | |
357 x = a + mb; | |
358 z = x - a; | |
359 y = (a - (x - z)) + (mb - z); | |
360 } | |
361 if(x < 0) | |
362 { | |
363 x = -x; | |
364 y = -y; | |
365 } | |
366 result += ldexp(x, expon) + ldexp(y, expon); | |
367 // | |
368 // Result must be an integer: | |
369 // | |
370 BOOST_ASSERT(result == floor(result)); | |
371 return result; | |
372 } | |
373 | |
374 } | |
375 | |
376 template <class T, class U, class Policy> | |
377 inline typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b, const Policy& pol) | |
378 { | |
379 typedef typename tools::promote_args<T, U>::type result_type; | |
380 return detail::float_distance_imp(static_cast<result_type>(a), static_cast<result_type>(b), pol); | |
381 } | |
382 | |
383 template <class T, class U> | |
384 typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b) | |
385 { | |
386 return boost::math::float_distance(a, b, policies::policy<>()); | |
387 } | |
388 | |
389 namespace detail{ | |
390 | |
391 template <class T, class Policy> | |
392 T float_advance_imp(T val, int distance, const Policy& pol) | |
393 { | |
394 BOOST_MATH_STD_USING | |
395 // | |
396 // Error handling: | |
397 // | |
398 static const char* function = "float_advance<%1%>(%1%, int)"; | |
399 | |
400 int fpclass = (boost::math::fpclassify)(val); | |
401 | |
402 if((fpclass == FP_NAN) || (fpclass == FP_INFINITE)) | |
403 return policies::raise_domain_error<T>( | |
404 function, | |
405 "Argument val must be finite, but got %1%", val, pol); | |
406 | |
407 if(val < 0) | |
408 return -float_advance(-val, -distance, pol); | |
409 if(distance == 0) | |
410 return val; | |
411 if(distance == 1) | |
412 return float_next(val, pol); | |
413 if(distance == -1) | |
414 return float_prior(val, pol); | |
415 | |
416 if(fabs(val) < detail::get_min_shift_value<T>()) | |
417 { | |
418 // | |
419 // Special case: if the value of the least significant bit is a denorm, | |
420 // implement in terms of float_next/float_prior. | |
421 // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set. | |
422 // | |
423 if(distance > 0) | |
424 { | |
425 do{ val = float_next(val, pol); } while(--distance); | |
426 } | |
427 else | |
428 { | |
429 do{ val = float_prior(val, pol); } while(++distance); | |
430 } | |
431 return val; | |
432 } | |
433 | |
434 int expon; | |
435 frexp(val, &expon); | |
436 T limit = ldexp((distance < 0 ? T(0.5f) : T(1)), expon); | |
437 if(val <= tools::min_value<T>()) | |
438 { | |
439 limit = sign(T(distance)) * tools::min_value<T>(); | |
440 } | |
441 T limit_distance = float_distance(val, limit); | |
442 while(fabs(limit_distance) < abs(distance)) | |
443 { | |
444 distance -= itrunc(limit_distance); | |
445 val = limit; | |
446 if(distance < 0) | |
447 { | |
448 limit /= 2; | |
449 expon--; | |
450 } | |
451 else | |
452 { | |
453 limit *= 2; | |
454 expon++; | |
455 } | |
456 limit_distance = float_distance(val, limit); | |
457 if(distance && (limit_distance == 0)) | |
458 { | |
459 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); | |
460 } | |
461 } | |
462 if((0.5f == frexp(val, &expon)) && (distance < 0)) | |
463 --expon; | |
464 T diff = 0; | |
465 if(val != 0) | |
466 diff = distance * ldexp(T(1), expon - tools::digits<T>()); | |
467 if(diff == 0) | |
468 diff = distance * detail::get_smallest_value<T>(); | |
469 return val += diff; | |
470 } | |
471 | |
472 } | |
473 | |
474 template <class T, class Policy> | |
475 inline typename tools::promote_args<T>::type float_advance(T val, int distance, const Policy& pol) | |
476 { | |
477 typedef typename tools::promote_args<T>::type result_type; | |
478 return detail::float_advance_imp(static_cast<result_type>(val), distance, pol); | |
479 } | |
480 | |
481 template <class T> | |
482 inline typename tools::promote_args<T>::type float_advance(const T& val, int distance) | |
483 { | |
484 return boost::math::float_advance(val, distance, policies::policy<>()); | |
485 } | |
486 | |
487 }} // namespaces | |
488 | |
489 #endif // BOOST_MATH_SPECIAL_NEXT_HPP | |
490 |