sq_detailed_metrics 0.1.0

Sign up to get free protection for your applications and to get access to all the features.
Files changed (69) hide show
  1. checksums.yaml +7 -0
  2. data/extconf.rb +26 -0
  3. data/include/half.hpp +4575 -0
  4. data/include/msgpack.h +24 -0
  5. data/include/msgpack/fbuffer.h +42 -0
  6. data/include/msgpack/gcc_atomic.h +25 -0
  7. data/include/msgpack/object.h +118 -0
  8. data/include/msgpack/pack.h +174 -0
  9. data/include/msgpack/pack_define.h +18 -0
  10. data/include/msgpack/pack_template.h +952 -0
  11. data/include/msgpack/sbuffer.h +115 -0
  12. data/include/msgpack/sysdep.h +221 -0
  13. data/include/msgpack/timestamp.h +58 -0
  14. data/include/msgpack/unpack.h +281 -0
  15. data/include/msgpack/unpack_define.h +89 -0
  16. data/include/msgpack/unpack_template.h +471 -0
  17. data/include/msgpack/util.h +15 -0
  18. data/include/msgpack/version.h +38 -0
  19. data/include/msgpack/version_master.h +3 -0
  20. data/include/msgpack/vrefbuffer.h +144 -0
  21. data/include/msgpack/zbuffer.h +205 -0
  22. data/include/msgpack/zone.h +163 -0
  23. data/include/rapidjson/allocators.h +271 -0
  24. data/include/rapidjson/document.h +2575 -0
  25. data/include/rapidjson/encodedstream.h +299 -0
  26. data/include/rapidjson/encodings.h +716 -0
  27. data/include/rapidjson/error/en.h +74 -0
  28. data/include/rapidjson/error/error.h +155 -0
  29. data/include/rapidjson/filereadstream.h +99 -0
  30. data/include/rapidjson/filewritestream.h +104 -0
  31. data/include/rapidjson/fwd.h +151 -0
  32. data/include/rapidjson/internal/biginteger.h +290 -0
  33. data/include/rapidjson/internal/diyfp.h +258 -0
  34. data/include/rapidjson/internal/dtoa.h +245 -0
  35. data/include/rapidjson/internal/ieee754.h +78 -0
  36. data/include/rapidjson/internal/itoa.h +304 -0
  37. data/include/rapidjson/internal/meta.h +181 -0
  38. data/include/rapidjson/internal/pow10.h +55 -0
  39. data/include/rapidjson/internal/regex.h +701 -0
  40. data/include/rapidjson/internal/stack.h +230 -0
  41. data/include/rapidjson/internal/strfunc.h +55 -0
  42. data/include/rapidjson/internal/strtod.h +269 -0
  43. data/include/rapidjson/internal/swap.h +46 -0
  44. data/include/rapidjson/istreamwrapper.h +115 -0
  45. data/include/rapidjson/memorybuffer.h +70 -0
  46. data/include/rapidjson/memorystream.h +71 -0
  47. data/include/rapidjson/msinttypes/inttypes.h +316 -0
  48. data/include/rapidjson/msinttypes/stdint.h +300 -0
  49. data/include/rapidjson/ostreamwrapper.h +81 -0
  50. data/include/rapidjson/pointer.h +1358 -0
  51. data/include/rapidjson/prettywriter.h +255 -0
  52. data/include/rapidjson/rapidjson.h +615 -0
  53. data/include/rapidjson/reader.h +1879 -0
  54. data/include/rapidjson/schema.h +2006 -0
  55. data/include/rapidjson/stream.h +179 -0
  56. data/include/rapidjson/stringbuffer.h +117 -0
  57. data/include/rapidjson/writer.h +610 -0
  58. data/include/xxhash.h +328 -0
  59. data/json_conv.cpp +284 -0
  60. data/json_conv.hpp +17 -0
  61. data/metrics.cpp +239 -0
  62. data/metrics.hpp +84 -0
  63. data/msgpack/objectc.c +482 -0
  64. data/msgpack/unpack.c +703 -0
  65. data/msgpack/version.c +22 -0
  66. data/msgpack/vrefbuffer.c +250 -0
  67. data/msgpack/zone.c +222 -0
  68. data/sq_detailed_metrics.cpp +248 -0
  69. metadata +199 -0
@@ -0,0 +1,7 @@
1
+ ---
2
+ SHA256:
3
+ metadata.gz: a38eda9349f688e115d4e5aa5c33dde66dc0ada44e33e1a7d9a04446ef5a539c
4
+ data.tar.gz: edf3282941ac42ec326f6fe74bd7eda5a7b9520cfee3c33f1e3685aceae451bc
5
+ SHA512:
6
+ metadata.gz: 05c8c82fa5ca98839dabae7a2a04f4462fb784e85dba0349c05159662009885decbd73ac8c32243a3ffadeb8842cfb8d3de84ec51b78394c7f7dbc3f335c229a
7
+ data.tar.gz: 7b96e204233c774feb19f21cc21e012c073587a302760cf57fa9b4a1762e5c13ee2114197390bcca2acaa536634fb04238efa7267d413ffec05a67a36bfc144c
@@ -0,0 +1,26 @@
1
+ c = RbConfig::MAKEFILE_CONFIG
2
+
3
+ c['CC'] = ENV['LIBSQREEN_CC'] if ENV['LIBSQREEN_CC']
4
+ c['CFLAGS'] += ' ' + ENV['LIBSQREEN_CFLAGS'] if ENV['LIBSQREEN_CFLAGS']
5
+ c['LDFLAGS'] += ' ' + ENV['LIBSQREEN_LDFLAGS'] if ENV['LIBSQREEN_LDFLAGS']
6
+
7
+ require 'mkmf'
8
+ require 'shellwords'
9
+ require 'pathname'
10
+
11
+ dir = Pathname.new(File.dirname(__FILE__))
12
+ $srcs = Dir["#{dir}/msgpack/*.c", "#{dir}/*.c", "#{dir}/*.cpp"].map do |p|
13
+ Pathname.new(p).relative_path_from(dir).to_s
14
+ end
15
+ $VPATH << "$(srcdir)/msgpack"
16
+ $warnflags = '-Wall'
17
+ $INCFLAGS << " -I #{Shellwords.escape(dir.to_s)}/include"
18
+ if defined?($CXXFLAGS)
19
+ $CXXFLAGS += " -std=c++11 "
20
+ else
21
+ # old rubies
22
+ CONFIG['CXXFLAGS'] += " -std=c++11 "
23
+ end
24
+
25
+ create_header
26
+ create_makefile 'sq_detailed_metrics'
@@ -0,0 +1,4575 @@
1
+ // half - IEEE 754-based half-precision floating-point library.
2
+ //
3
+ // Copyright (c) 2012-2019 Christian Rau <rauy@users.sourceforge.net>
4
+ //
5
+ // Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation
6
+ // files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy,
7
+ // modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the
8
+ // Software is furnished to do so, subject to the following conditions:
9
+ //
10
+ // The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
11
+ //
12
+ // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE
13
+ // WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
14
+ // COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
15
+ // ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
16
+
17
+ // Version 2.1.0
18
+
19
+ /// \file
20
+ /// Main header file for half-precision functionality.
21
+
22
+ #ifndef HALF_HALF_HPP
23
+ #define HALF_HALF_HPP
24
+
25
+ #define HALF_GCC_VERSION (__GNUC__*100+__GNUC_MINOR__)
26
+
27
+ #if defined(__INTEL_COMPILER)
28
+ #define HALF_ICC_VERSION __INTEL_COMPILER
29
+ #elif defined(__ICC)
30
+ #define HALF_ICC_VERSION __ICC
31
+ #elif defined(__ICL)
32
+ #define HALF_ICC_VERSION __ICL
33
+ #else
34
+ #define HALF_ICC_VERSION 0
35
+ #endif
36
+
37
+ // check C++11 language features
38
+ #if defined(__clang__) // clang
39
+ #if __has_feature(cxx_static_assert) && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
40
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
41
+ #endif
42
+ #if __has_feature(cxx_constexpr) && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
43
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
44
+ #endif
45
+ #if __has_feature(cxx_noexcept) && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
46
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
47
+ #endif
48
+ #if __has_feature(cxx_user_literals) && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
49
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
50
+ #endif
51
+ #if __has_feature(cxx_thread_local) && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
52
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
53
+ #endif
54
+ #if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L) && !defined(HALF_ENABLE_CPP11_LONG_LONG)
55
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
56
+ #endif
57
+ #elif HALF_ICC_VERSION && defined(__INTEL_CXX11_MODE__) // Intel C++
58
+ #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
59
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
60
+ #endif
61
+ #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
62
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
63
+ #endif
64
+ #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
65
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
66
+ #endif
67
+ #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
68
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
69
+ #endif
70
+ #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
71
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
72
+ #endif
73
+ #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_LONG_LONG)
74
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
75
+ #endif
76
+ #elif defined(__GNUC__) // gcc
77
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L
78
+ #if HALF_GCC_VERSION >= 408 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
79
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
80
+ #endif
81
+ #if HALF_GCC_VERSION >= 407 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
82
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
83
+ #endif
84
+ #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
85
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
86
+ #endif
87
+ #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
88
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
89
+ #endif
90
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
91
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
92
+ #endif
93
+ #if !defined(HALF_ENABLE_CPP11_LONG_LONG)
94
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
95
+ #endif
96
+ #endif
97
+ #define HALF_TWOS_COMPLEMENT_INT 1
98
+ #elif defined(_MSC_VER) // Visual C++
99
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
100
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
101
+ #endif
102
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
103
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
104
+ #endif
105
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
106
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
107
+ #endif
108
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
109
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
110
+ #endif
111
+ #if _MSC_VER >= 1600 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
112
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
113
+ #endif
114
+ #if _MSC_VER >= 1310 && !defined(HALF_ENABLE_CPP11_LONG_LONG)
115
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
116
+ #endif
117
+ #define HALF_TWOS_COMPLEMENT_INT 1
118
+ #define HALF_POP_WARNINGS 1
119
+ #pragma warning(push)
120
+ #pragma warning(disable : 4099 4127 4146) //struct vs class, constant in if, negative unsigned
121
+ #endif
122
+
123
+ // check C++11 library features
124
+ #include <utility>
125
+ #if defined(_LIBCPP_VERSION) // libc++
126
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103
127
+ #ifndef HALF_ENABLE_CPP11_TYPE_TRAITS
128
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
129
+ #endif
130
+ #ifndef HALF_ENABLE_CPP11_CSTDINT
131
+ #define HALF_ENABLE_CPP11_CSTDINT 1
132
+ #endif
133
+ #ifndef HALF_ENABLE_CPP11_CMATH
134
+ #define HALF_ENABLE_CPP11_CMATH 1
135
+ #endif
136
+ #ifndef HALF_ENABLE_CPP11_HASH
137
+ #define HALF_ENABLE_CPP11_HASH 1
138
+ #endif
139
+ #ifndef HALF_ENABLE_CPP11_CFENV
140
+ #define HALF_ENABLE_CPP11_CFENV 1
141
+ #endif
142
+ #endif
143
+ #elif defined(__GLIBCXX__) // libstdc++
144
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103
145
+ #ifdef __clang__
146
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
147
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
148
+ #endif
149
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CSTDINT)
150
+ #define HALF_ENABLE_CPP11_CSTDINT 1
151
+ #endif
152
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CMATH)
153
+ #define HALF_ENABLE_CPP11_CMATH 1
154
+ #endif
155
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_HASH)
156
+ #define HALF_ENABLE_CPP11_HASH 1
157
+ #endif
158
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CFENV)
159
+ #define HALF_ENABLE_CPP11_CFENV 1
160
+ #endif
161
+ #else
162
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
163
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
164
+ #endif
165
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CSTDINT)
166
+ #define HALF_ENABLE_CPP11_CSTDINT 1
167
+ #endif
168
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CMATH)
169
+ #define HALF_ENABLE_CPP11_CMATH 1
170
+ #endif
171
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_HASH)
172
+ #define HALF_ENABLE_CPP11_HASH 1
173
+ #endif
174
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CFENV)
175
+ #define HALF_ENABLE_CPP11_CFENV 1
176
+ #endif
177
+ #endif
178
+ #endif
179
+ #elif defined(_CPPLIB_VER) // Dinkumware/Visual C++
180
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
181
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
182
+ #endif
183
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_CSTDINT)
184
+ #define HALF_ENABLE_CPP11_CSTDINT 1
185
+ #endif
186
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_HASH)
187
+ #define HALF_ENABLE_CPP11_HASH 1
188
+ #endif
189
+ #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CMATH)
190
+ #define HALF_ENABLE_CPP11_CMATH 1
191
+ #endif
192
+ #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CFENV)
193
+ #define HALF_ENABLE_CPP11_CFENV 1
194
+ #endif
195
+ #endif
196
+ #undef HALF_GCC_VERSION
197
+ #undef HALF_ICC_VERSION
198
+
199
+ // any error throwing C++ exceptions?
200
+ #if defined(HALF_ERRHANDLING_THROW_INVALID) || defined(HALF_ERRHANDLING_THROW_DIVBYZERO) || defined(HALF_ERRHANDLING_THROW_OVERFLOW) || defined(HALF_ERRHANDLING_THROW_UNDERFLOW) || defined(HALF_ERRHANDLING_THROW_INEXACT)
201
+ #define HALF_ERRHANDLING_THROWS 1
202
+ #endif
203
+
204
+ // any error handling enabled?
205
+ #define HALF_ERRHANDLING (HALF_ERRHANDLING_FLAGS||HALF_ERRHANDLING_ERRNO||HALF_ERRHANDLING_FENV||HALF_ERRHANDLING_THROWS)
206
+
207
+ #if HALF_ERRHANDLING
208
+ #define HALF_UNUSED_NOERR(name) name
209
+ #else
210
+ #define HALF_UNUSED_NOERR(name)
211
+ #endif
212
+
213
+ // support constexpr
214
+ #if HALF_ENABLE_CPP11_CONSTEXPR
215
+ #define HALF_CONSTEXPR constexpr
216
+ #define HALF_CONSTEXPR_CONST constexpr
217
+ #if HALF_ERRHANDLING
218
+ #define HALF_CONSTEXPR_NOERR
219
+ #else
220
+ #define HALF_CONSTEXPR_NOERR constexpr
221
+ #endif
222
+ #else
223
+ #define HALF_CONSTEXPR
224
+ #define HALF_CONSTEXPR_CONST const
225
+ #define HALF_CONSTEXPR_NOERR
226
+ #endif
227
+
228
+ // support noexcept
229
+ #if HALF_ENABLE_CPP11_NOEXCEPT
230
+ #define HALF_NOEXCEPT noexcept
231
+ #define HALF_NOTHROW noexcept
232
+ #else
233
+ #define HALF_NOEXCEPT
234
+ #define HALF_NOTHROW throw()
235
+ #endif
236
+
237
+ // support thread storage
238
+ #if HALF_ENABLE_CPP11_THREAD_LOCAL
239
+ #define HALF_THREAD_LOCAL thread_local
240
+ #else
241
+ #define HALF_THREAD_LOCAL static
242
+ #endif
243
+
244
+ #include <utility>
245
+ #include <algorithm>
246
+ #include <istream>
247
+ #include <ostream>
248
+ #include <limits>
249
+ #include <stdexcept>
250
+ #include <climits>
251
+ #include <cmath>
252
+ #include <cstring>
253
+ #include <cstdlib>
254
+ #if HALF_ENABLE_CPP11_TYPE_TRAITS
255
+ #include <type_traits>
256
+ #endif
257
+ #if HALF_ENABLE_CPP11_CSTDINT
258
+ #include <cstdint>
259
+ #endif
260
+ #if HALF_ERRHANDLING_ERRNO
261
+ #include <cerrno>
262
+ #endif
263
+ #if HALF_ENABLE_CPP11_CFENV
264
+ #include <cfenv>
265
+ #endif
266
+ #if HALF_ENABLE_CPP11_HASH
267
+ #include <functional>
268
+ #endif
269
+ #if HALF_ENABLE_F16C_INTRINSICS
270
+ #include <immintrin.h>
271
+ #endif
272
+
273
+
274
+ #ifndef HALF_ENABLE_F16C_INTRINSICS
275
+ /// Enable F16C intruction set intrinsics.
276
+ /// Defining this to 1 enables the use of [F16C compiler intrinsics](https://en.wikipedia.org/wiki/F16C) for converting between
277
+ /// half-precision and single-precision values which may result in improved performance. This will not perform additional checks
278
+ /// for support of the F16C instruction set, so an appropriate target platform is required when enabling this feature.
279
+ ///
280
+ /// Unless predefined it will be enabled automatically when the `__F16C__` symbol is defined, which some compilers do on supporting platforms.
281
+ #define HALF_ENABLE_F16C_INTRINSICS __F16C__
282
+ #endif
283
+
284
+ #ifdef HALF_DOXYGEN_ONLY
285
+ /// Type for internal floating-point computations.
286
+ /// This can be predefined to a built-in floating-point type (`float`, `double` or `long double`) to override the internal
287
+ /// half-precision implementation to use this type for computing arithmetic operations and mathematical function (if available).
288
+ /// This can result in improved performance for arithmetic operators and mathematical functions but might cause results to
289
+ /// deviate from the specified half-precision rounding mode and inhibits proper detection of half-precision exceptions.
290
+ #define HALF_ARITHMETIC_TYPE (undefined)
291
+
292
+ /// Enable internal exception flags.
293
+ /// Defining this to 1 causes operations on half-precision values to raise internal floating-point exception flags according to
294
+ /// the IEEE 754 standard. These can then be cleared and checked with clearexcept(), testexcept().
295
+ #define HALF_ERRHANDLING_FLAGS 0
296
+
297
+ /// Enable exception propagation to `errno`.
298
+ /// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to
299
+ /// [errno](https://en.cppreference.com/w/cpp/error/errno) from `<cerrno>`. Specifically this will propagate domain errors as
300
+ /// [EDOM](https://en.cppreference.com/w/cpp/error/errno_macros) and pole, overflow and underflow errors as
301
+ /// [ERANGE](https://en.cppreference.com/w/cpp/error/errno_macros). Inexact errors won't be propagated.
302
+ #define HALF_ERRHANDLING_ERRNO 0
303
+
304
+ /// Enable exception propagation to built-in floating-point platform.
305
+ /// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to the built-in
306
+ /// single- and double-precision implementation's exception flags using the
307
+ /// [C++11 floating-point environment control](https://en.cppreference.com/w/cpp/numeric/fenv) from `<cfenv>`. However, this
308
+ /// does not work in reverse and single- or double-precision exceptions will not raise the corresponding half-precision
309
+ /// exception flags, nor will explicitly clearing flags clear the corresponding built-in flags.
310
+ #define HALF_ERRHANDLING_FENV 0
311
+
312
+ /// Throw C++ exception on domain errors.
313
+ /// Defining this to a string literal causes operations on half-precision values to throw a
314
+ /// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on domain errors.
315
+ #define HALF_ERRHANDLING_THROW_INVALID (undefined)
316
+
317
+ /// Throw C++ exception on pole errors.
318
+ /// Defining this to a string literal causes operations on half-precision values to throw a
319
+ /// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on pole errors.
320
+ #define HALF_ERRHANDLING_THROW_DIVBYZERO (undefined)
321
+
322
+ /// Throw C++ exception on overflow errors.
323
+ /// Defining this to a string literal causes operations on half-precision values to throw a
324
+ /// [std::overflow_error](https://en.cppreference.com/w/cpp/error/overflow_error) with the specified message on overflows.
325
+ #define HALF_ERRHANDLING_THROW_OVERFLOW (undefined)
326
+
327
+ /// Throw C++ exception on underflow errors.
328
+ /// Defining this to a string literal causes operations on half-precision values to throw a
329
+ /// [std::underflow_error](https://en.cppreference.com/w/cpp/error/underflow_error) with the specified message on underflows.
330
+ #define HALF_ERRHANDLING_THROW_UNDERFLOW (undefined)
331
+
332
+ /// Throw C++ exception on rounding errors.
333
+ /// Defining this to 1 causes operations on half-precision values to throw a
334
+ /// [std::range_error](https://en.cppreference.com/w/cpp/error/range_error) with the specified message on general rounding errors.
335
+ #define HALF_ERRHANDLING_THROW_INEXACT (undefined)
336
+ #endif
337
+
338
+ #ifndef HALF_ERRHANDLING_OVERFLOW_TO_INEXACT
339
+ /// Raise INEXACT exception on overflow.
340
+ /// Defining this to 1 (default) causes overflow errors to automatically raise inexact exceptions in addition.
341
+ /// These will be raised after any possible handling of the underflow exception.
342
+ #define HALF_ERRHANDLING_OVERFLOW_TO_INEXACT 1
343
+ #endif
344
+
345
+ #ifndef HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
346
+ /// Raise INEXACT exception on underflow.
347
+ /// Defining this to 1 (default) causes underflow errors to automatically raise inexact exceptions in addition.
348
+ /// These will be raised after any possible handling of the underflow exception.
349
+ ///
350
+ /// **Note:** This will actually cause underflow (and the accompanying inexact) exceptions to be raised *only* when the result
351
+ /// is inexact, while if disabled bare underflow errors will be raised for *any* (possibly exact) subnormal result.
352
+ #define HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT 1
353
+ #endif
354
+
355
+ /// Default rounding mode.
356
+ /// This specifies the rounding mode used for all conversions between [half](\ref half_float::half)s and more precise types
357
+ /// (unless using half_cast() and specifying the rounding mode directly) as well as in arithmetic operations and mathematical
358
+ /// functions. It can be redefined (before including half.hpp) to one of the standard rounding modes using their respective
359
+ /// constants or the equivalent values of
360
+ /// [std::float_round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/float_round_style):
361
+ ///
362
+ /// `std::float_round_style` | value | rounding
363
+ /// ---------------------------------|-------|-------------------------
364
+ /// `std::round_indeterminate` | -1 | fastest
365
+ /// `std::round_toward_zero` | 0 | toward zero
366
+ /// `std::round_to_nearest` | 1 | to nearest (default)
367
+ /// `std::round_toward_infinity` | 2 | toward positive infinity
368
+ /// `std::round_toward_neg_infinity` | 3 | toward negative infinity
369
+ ///
370
+ /// By default this is set to `1` (`std::round_to_nearest`), which rounds results to the nearest representable value. It can even
371
+ /// be set to [std::numeric_limits<float>::round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/round_style) to synchronize
372
+ /// the rounding mode with that of the built-in single-precision implementation (which is likely `std::round_to_nearest`, though).
373
+ #ifndef HALF_ROUND_STYLE
374
+ #define HALF_ROUND_STYLE 1 // = std::round_to_nearest
375
+ #endif
376
+
377
+ /// Value signaling overflow.
378
+ /// In correspondence with `HUGE_VAL[F|L]` from `<cmath>` this symbol expands to a positive value signaling the overflow of an
379
+ /// operation, in particular it just evaluates to positive infinity.
380
+ ///
381
+ /// **See also:** Documentation for [HUGE_VAL](https://en.cppreference.com/w/cpp/numeric/math/HUGE_VAL)
382
+ #define HUGE_VALH std::numeric_limits<half_float::half>::infinity()
383
+
384
+ /// Fast half-precision fma function.
385
+ /// This symbol is defined if the fma() function generally executes as fast as, or faster than, a separate
386
+ /// half-precision multiplication followed by an addition, which is always the case.
387
+ ///
388
+ /// **See also:** Documentation for [FP_FAST_FMA](https://en.cppreference.com/w/cpp/numeric/math/fma)
389
+ #define FP_FAST_FMAH 1
390
+
391
+ /// Half rounding mode.
392
+ /// In correspondence with `FLT_ROUNDS` from `<cfloat>` this symbol expands to the rounding mode used for
393
+ /// half-precision operations. It is an alias for [HALF_ROUND_STYLE](\ref HALF_ROUND_STYLE).
394
+ ///
395
+ /// **See also:** Documentation for [FLT_ROUNDS](https://en.cppreference.com/w/cpp/types/climits/FLT_ROUNDS)
396
+ #define HLF_ROUNDS HALF_ROUND_STYLE
397
+
398
+ #ifndef FP_ILOGB0
399
+ #define FP_ILOGB0 INT_MIN
400
+ #endif
401
+ #ifndef FP_ILOGBNAN
402
+ #define FP_ILOGBNAN INT_MAX
403
+ #endif
404
+ #ifndef FP_SUBNORMAL
405
+ #define FP_SUBNORMAL 0
406
+ #endif
407
+ #ifndef FP_ZERO
408
+ #define FP_ZERO 1
409
+ #endif
410
+ #ifndef FP_NAN
411
+ #define FP_NAN 2
412
+ #endif
413
+ #ifndef FP_INFINITE
414
+ #define FP_INFINITE 3
415
+ #endif
416
+ #ifndef FP_NORMAL
417
+ #define FP_NORMAL 4
418
+ #endif
419
+
420
+ #if !HALF_ENABLE_CPP11_CFENV && !defined(FE_ALL_EXCEPT)
421
+ #define FE_INVALID 0x10
422
+ #define FE_DIVBYZERO 0x08
423
+ #define FE_OVERFLOW 0x04
424
+ #define FE_UNDERFLOW 0x02
425
+ #define FE_INEXACT 0x01
426
+ #define FE_ALL_EXCEPT (FE_INVALID|FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW|FE_INEXACT)
427
+ #endif
428
+
429
+
430
+ /// Main namespace for half-precision functionality.
431
+ /// This namespace contains all the functionality provided by the library.
432
+ namespace half_float
433
+ {
434
+ class half;
435
+
436
+ #if HALF_ENABLE_CPP11_USER_LITERALS
437
+ /// Library-defined half-precision literals.
438
+ /// Import this namespace to enable half-precision floating-point literals:
439
+ /// ~~~~{.cpp}
440
+ /// using namespace half_float::literal;
441
+ /// half_float::half = 4.2_h;
442
+ /// ~~~~
443
+ namespace literal
444
+ {
445
+ half operator "" _h(long double);
446
+ }
447
+ #endif
448
+
449
+ /// \internal
450
+ /// \brief Implementation details.
451
+ namespace detail
452
+ {
453
+ #if HALF_ENABLE_CPP11_TYPE_TRAITS
454
+ /// Conditional type.
455
+ template<bool B,typename T,typename F> struct conditional : std::conditional<B,T,F> {};
456
+
457
+ /// Helper for tag dispatching.
458
+ template<bool B> struct bool_type : std::integral_constant<bool,B> {};
459
+ using std::true_type;
460
+ using std::false_type;
461
+
462
+ /// Type traits for floating-point types.
463
+ template<typename T> struct is_float : std::is_floating_point<T> {};
464
+ #else
465
+ /// Conditional type.
466
+ template<bool,typename T,typename> struct conditional { typedef T type; };
467
+ template<typename T,typename F> struct conditional<false,T,F> { typedef F type; };
468
+
469
+ /// Helper for tag dispatching.
470
+ template<bool> struct bool_type {};
471
+ typedef bool_type<true> true_type;
472
+ typedef bool_type<false> false_type;
473
+
474
+ /// Type traits for floating-point types.
475
+ template<typename> struct is_float : false_type {};
476
+ template<typename T> struct is_float<const T> : is_float<T> {};
477
+ template<typename T> struct is_float<volatile T> : is_float<T> {};
478
+ template<typename T> struct is_float<const volatile T> : is_float<T> {};
479
+ template<> struct is_float<float> : true_type {};
480
+ template<> struct is_float<double> : true_type {};
481
+ template<> struct is_float<long double> : true_type {};
482
+ #endif
483
+
484
+ /// Type traits for floating-point bits.
485
+ template<typename T> struct bits { typedef unsigned char type; };
486
+ template<typename T> struct bits<const T> : bits<T> {};
487
+ template<typename T> struct bits<volatile T> : bits<T> {};
488
+ template<typename T> struct bits<const volatile T> : bits<T> {};
489
+
490
+ #if HALF_ENABLE_CPP11_CSTDINT
491
+ /// Unsigned integer of (at least) 16 bits width.
492
+ typedef std::uint_least16_t uint16;
493
+
494
+ /// Fastest unsigned integer of (at least) 32 bits width.
495
+ typedef std::uint_fast32_t uint32;
496
+
497
+ /// Fastest signed integer of (at least) 32 bits width.
498
+ typedef std::int_fast32_t int32;
499
+
500
+ /// Unsigned integer of (at least) 32 bits width.
501
+ template<> struct bits<float> { typedef std::uint_least32_t type; };
502
+
503
+ /// Unsigned integer of (at least) 64 bits width.
504
+ template<> struct bits<double> { typedef std::uint_least64_t type; };
505
+ #else
506
+ /// Unsigned integer of (at least) 16 bits width.
507
+ typedef unsigned short uint16;
508
+
509
+ /// Fastest unsigned integer of (at least) 32 bits width.
510
+ typedef unsigned long uint32;
511
+
512
+ /// Fastest unsigned integer of (at least) 32 bits width.
513
+ typedef long int32;
514
+
515
+ /// Unsigned integer of (at least) 32 bits width.
516
+ template<> struct bits<float> : conditional<std::numeric_limits<unsigned int>::digits>=32,unsigned int,unsigned long> {};
517
+
518
+ #if HALF_ENABLE_CPP11_LONG_LONG
519
+ /// Unsigned integer of (at least) 64 bits width.
520
+ template<> struct bits<double> : conditional<std::numeric_limits<unsigned long>::digits>=64,unsigned long,unsigned long long> {};
521
+ #else
522
+ /// Unsigned integer of (at least) 64 bits width.
523
+ template<> struct bits<double> { typedef unsigned long type; };
524
+ #endif
525
+ #endif
526
+
527
+ #ifdef HALF_ARITHMETIC_TYPE
528
+ /// Type to use for arithmetic computations and mathematic functions internally.
529
+ typedef HALF_ARITHMETIC_TYPE internal_t;
530
+ #endif
531
+
532
+ /// Tag type for binary construction.
533
+ struct binary_t {};
534
+
535
+ /// Tag for binary construction.
536
+ HALF_CONSTEXPR_CONST binary_t binary = binary_t();
537
+
538
+ /// \name Implementation defined classification and arithmetic
539
+ /// \{
540
+
541
+ /// Check for infinity.
542
+ /// \tparam T argument type (builtin floating-point type)
543
+ /// \param arg value to query
544
+ /// \retval true if infinity
545
+ /// \retval false else
546
+ template<typename T> bool builtin_isinf(T arg)
547
+ {
548
+ #if HALF_ENABLE_CPP11_CMATH
549
+ return std::isinf(arg);
550
+ #elif defined(_MSC_VER)
551
+ return !::_finite(static_cast<double>(arg)) && !::_isnan(static_cast<double>(arg));
552
+ #else
553
+ return arg == std::numeric_limits<T>::infinity() || arg == -std::numeric_limits<T>::infinity();
554
+ #endif
555
+ }
556
+
557
+ /// Check for NaN.
558
+ /// \tparam T argument type (builtin floating-point type)
559
+ /// \param arg value to query
560
+ /// \retval true if not a number
561
+ /// \retval false else
562
+ template<typename T> bool builtin_isnan(T arg)
563
+ {
564
+ #if HALF_ENABLE_CPP11_CMATH
565
+ return std::isnan(arg);
566
+ #elif defined(_MSC_VER)
567
+ return ::_isnan(static_cast<double>(arg)) != 0;
568
+ #else
569
+ return arg != arg;
570
+ #endif
571
+ }
572
+
573
+ /// Check sign.
574
+ /// \tparam T argument type (builtin floating-point type)
575
+ /// \param arg value to query
576
+ /// \retval true if signbit set
577
+ /// \retval false else
578
+ template<typename T> bool builtin_signbit(T arg)
579
+ {
580
+ #if HALF_ENABLE_CPP11_CMATH
581
+ return std::signbit(arg);
582
+ #else
583
+ return arg < T() || (arg == T() && T(1)/arg < T());
584
+ #endif
585
+ }
586
+
587
+ /// Platform-independent sign mask.
588
+ /// \param arg integer value in two's complement
589
+ /// \retval -1 if \a arg negative
590
+ /// \retval 0 if \a arg positive
591
+ inline uint32 sign_mask(uint32 arg)
592
+ {
593
+ static const int N = std::numeric_limits<uint32>::digits - 1;
594
+ #if HALF_TWOS_COMPLEMENT_INT
595
+ return static_cast<int32>(arg) >> N;
596
+ #else
597
+ return -((arg>>N)&1);
598
+ #endif
599
+ }
600
+
601
+ /// Platform-independent arithmetic right shift.
602
+ /// \param arg integer value in two's complement
603
+ /// \param i shift amount (at most 31)
604
+ /// \return \a arg right shifted for \a i bits with possible sign extension
605
+ inline uint32 arithmetic_shift(uint32 arg, int i)
606
+ {
607
+ #if HALF_TWOS_COMPLEMENT_INT
608
+ return static_cast<int32>(arg) >> i;
609
+ #else
610
+ return static_cast<int32>(arg)/(static_cast<int32>(1)<<i) - ((arg>>(std::numeric_limits<uint32>::digits-1))&1);
611
+ #endif
612
+ }
613
+
614
+ /// \}
615
+ /// \name Error handling
616
+ /// \{
617
+
618
+ /// Internal exception flags.
619
+ /// \return reference to global exception flags
620
+ inline int& errflags() { HALF_THREAD_LOCAL int flags = 0; return flags; }
621
+
622
+ /// Raise floating-point exception.
623
+ /// \param flags exceptions to raise
624
+ /// \param cond condition to raise exceptions for
625
+ inline void raise(int HALF_UNUSED_NOERR(flags), bool HALF_UNUSED_NOERR(cond) = true)
626
+ {
627
+ #if HALF_ERRHANDLING
628
+ if(!cond)
629
+ return;
630
+ #if HALF_ERRHANDLING_FLAGS
631
+ errflags() |= flags;
632
+ #endif
633
+ #if HALF_ERRHANDLING_ERRNO
634
+ if(flags & FE_INVALID)
635
+ errno = EDOM;
636
+ else if(flags & (FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW))
637
+ errno = ERANGE;
638
+ #endif
639
+ #if HALF_ERRHANDLING_FENV && HALF_ENABLE_CPP11_CFENV
640
+ std::feraiseexcept(flags);
641
+ #endif
642
+ #ifdef HALF_ERRHANDLING_THROW_INVALID
643
+ if(flags & FE_INVALID)
644
+ throw std::domain_error(HALF_ERRHANDLING_THROW_INVALID);
645
+ #endif
646
+ #ifdef HALF_ERRHANDLING_THROW_DIVBYZERO
647
+ if(flags & FE_DIVBYZERO)
648
+ throw std::domain_error(HALF_ERRHANDLING_THROW_DIVBYZERO);
649
+ #endif
650
+ #ifdef HALF_ERRHANDLING_THROW_OVERFLOW
651
+ if(flags & FE_OVERFLOW)
652
+ throw std::overflow_error(HALF_ERRHANDLING_THROW_OVERFLOW);
653
+ #endif
654
+ #ifdef HALF_ERRHANDLING_THROW_UNDERFLOW
655
+ if(flags & FE_UNDERFLOW)
656
+ throw std::underflow_error(HALF_ERRHANDLING_THROW_UNDERFLOW);
657
+ #endif
658
+ #ifdef HALF_ERRHANDLING_THROW_INEXACT
659
+ if(flags & FE_INEXACT)
660
+ throw std::range_error(HALF_ERRHANDLING_THROW_INEXACT);
661
+ #endif
662
+ #if HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
663
+ if((flags & FE_UNDERFLOW) && !(flags & FE_INEXACT))
664
+ raise(FE_INEXACT);
665
+ #endif
666
+ #if HALF_ERRHANDLING_OVERFLOW_TO_INEXACT
667
+ if((flags & FE_OVERFLOW) && !(flags & FE_INEXACT))
668
+ raise(FE_INEXACT);
669
+ #endif
670
+ #endif
671
+ }
672
+
673
+ /// Check and signal for any NaN.
674
+ /// \param x first half-precision value to check
675
+ /// \param y second half-precision value to check
676
+ /// \retval true if either \a x or \a y is NaN
677
+ /// \retval false else
678
+ /// \exception FE_INVALID if \a x or \a y is NaN
679
+ inline HALF_CONSTEXPR_NOERR bool compsignal(unsigned int x, unsigned int y)
680
+ {
681
+ #if HALF_ERRHANDLING
682
+ raise(FE_INVALID, (x&0x7FFF)>0x7C00 || (y&0x7FFF)>0x7C00);
683
+ #endif
684
+ return (x&0x7FFF) > 0x7C00 || (y&0x7FFF) > 0x7C00;
685
+ }
686
+
687
+ /// Signal and silence signaling NaN.
688
+ /// \param nan half-precision NaN value
689
+ /// \return quiet NaN
690
+ /// \exception FE_INVALID if \a nan is signaling NaN
691
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int nan)
692
+ {
693
+ #if HALF_ERRHANDLING
694
+ raise(FE_INVALID, !(nan&0x200));
695
+ #endif
696
+ return nan | 0x200;
697
+ }
698
+
699
+ /// Signal and silence signaling NaNs.
700
+ /// \param x first half-precision value to check
701
+ /// \param y second half-precision value to check
702
+ /// \return quiet NaN
703
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
704
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y)
705
+ {
706
+ #if HALF_ERRHANDLING
707
+ raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200)));
708
+ #endif
709
+ return ((x&0x7FFF)>0x7C00) ? (x|0x200) : (y|0x200);
710
+ }
711
+
712
+ /// Signal and silence signaling NaNs.
713
+ /// \param x first half-precision value to check
714
+ /// \param y second half-precision value to check
715
+ /// \param z third half-precision value to check
716
+ /// \return quiet NaN
717
+ /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN
718
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y, unsigned int z)
719
+ {
720
+ #if HALF_ERRHANDLING
721
+ raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200)) || ((z&0x7FFF)>0x7C00 && !(z&0x200)));
722
+ #endif
723
+ return ((x&0x7FFF)>0x7C00) ? (x|0x200) : ((y&0x7FFF)>0x7C00) ? (y|0x200) : (z|0x200);
724
+ }
725
+
726
+ /// Select value or signaling NaN.
727
+ /// \param x preferred half-precision value
728
+ /// \param y ignored half-precision value except for signaling NaN
729
+ /// \return \a y if signaling NaN, \a x otherwise
730
+ /// \exception FE_INVALID if \a y is signaling NaN
731
+ inline HALF_CONSTEXPR_NOERR unsigned int select(unsigned int x, unsigned int HALF_UNUSED_NOERR(y))
732
+ {
733
+ #if HALF_ERRHANDLING
734
+ return (((y&0x7FFF)>0x7C00) && !(y&0x200)) ? signal(y) : x;
735
+ #else
736
+ return x;
737
+ #endif
738
+ }
739
+
740
+ /// Raise domain error and return NaN.
741
+ /// return quiet NaN
742
+ /// \exception FE_INVALID
743
+ inline HALF_CONSTEXPR_NOERR unsigned int invalid()
744
+ {
745
+ #if HALF_ERRHANDLING
746
+ raise(FE_INVALID);
747
+ #endif
748
+ return 0x7FFF;
749
+ }
750
+
751
+ /// Raise pole error and return infinity.
752
+ /// \param sign half-precision value with sign bit only
753
+ /// \return half-precision infinity with sign of \a sign
754
+ /// \exception FE_DIVBYZERO
755
+ inline HALF_CONSTEXPR_NOERR unsigned int pole(unsigned int sign = 0)
756
+ {
757
+ #if HALF_ERRHANDLING
758
+ raise(FE_DIVBYZERO);
759
+ #endif
760
+ return sign | 0x7C00;
761
+ }
762
+
763
+ /// Check value for underflow.
764
+ /// \param arg non-zero half-precision value to check
765
+ /// \return \a arg
766
+ /// \exception FE_UNDERFLOW if arg is subnormal
767
+ inline HALF_CONSTEXPR_NOERR unsigned int check_underflow(unsigned int arg)
768
+ {
769
+ #if HALF_ERRHANDLING && !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
770
+ raise(FE_UNDERFLOW, !(arg&0x7C00));
771
+ #endif
772
+ return arg;
773
+ }
774
+
775
+ /// \}
776
+ /// \name Conversion and rounding
777
+ /// \{
778
+
779
+ /// Half-precision overflow.
780
+ /// \tparam R rounding mode to use
781
+ /// \param sign half-precision value with sign bit only
782
+ /// \return rounded overflowing half-precision value
783
+ /// \exception FE_OVERFLOW
784
+ template<std::float_round_style R> HALF_CONSTEXPR_NOERR unsigned int overflow(unsigned int sign = 0)
785
+ {
786
+ #if HALF_ERRHANDLING
787
+ raise(FE_OVERFLOW);
788
+ #endif
789
+ return (R==std::round_toward_infinity) ? (sign+0x7C00-(sign>>15)) :
790
+ (R==std::round_toward_neg_infinity) ? (sign+0x7BFF+(sign>>15)) :
791
+ (R==std::round_toward_zero) ? (sign|0x7BFF) :
792
+ (sign|0x7C00);
793
+ }
794
+
795
+ /// Half-precision underflow.
796
+ /// \tparam R rounding mode to use
797
+ /// \param sign half-precision value with sign bit only
798
+ /// \return rounded underflowing half-precision value
799
+ /// \exception FE_UNDERFLOW
800
+ template<std::float_round_style R> HALF_CONSTEXPR_NOERR unsigned int underflow(unsigned int sign = 0)
801
+ {
802
+ #if HALF_ERRHANDLING
803
+ raise(FE_UNDERFLOW);
804
+ #endif
805
+ return (R==std::round_toward_infinity) ? (sign+1-(sign>>15)) :
806
+ (R==std::round_toward_neg_infinity) ? (sign+(sign>>15)) :
807
+ sign;
808
+ }
809
+
810
+ /// Round half-precision number.
811
+ /// \tparam R rounding mode to use
812
+ /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results
813
+ /// \param value finite half-precision number to round
814
+ /// \param g guard bit (most significant discarded bit)
815
+ /// \param s sticky bit (or of all but the most significant discarded bits)
816
+ /// \return rounded half-precision value
817
+ /// \exception FE_OVERFLOW on overflows
818
+ /// \exception FE_UNDERFLOW on underflows
819
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
820
+ template<std::float_round_style R,bool I> HALF_CONSTEXPR_NOERR unsigned int rounded(unsigned int value, int g, int s)
821
+ {
822
+ #if HALF_ERRHANDLING
823
+ value += (R==std::round_to_nearest) ? (g&(s|value)) :
824
+ (R==std::round_toward_infinity) ? (~(value>>15)&(g|s)) :
825
+ (R==std::round_toward_neg_infinity) ? ((value>>15)&(g|s)) : 0;
826
+ if((value&0x7C00) == 0x7C00)
827
+ raise(FE_OVERFLOW);
828
+ else if(value & 0x7C00)
829
+ raise(FE_INEXACT, I || (g|s)!=0);
830
+ else
831
+ raise(FE_UNDERFLOW, !(HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT) || I || (g|s)!=0);
832
+ return value;
833
+ #else
834
+ return (R==std::round_to_nearest) ? (value+(g&(s|value))) :
835
+ (R==std::round_toward_infinity) ? (value+(~(value>>15)&(g|s))) :
836
+ (R==std::round_toward_neg_infinity) ? (value+((value>>15)&(g|s))) :
837
+ value;
838
+ #endif
839
+ }
840
+
841
+ /// Round half-precision number to nearest integer value.
842
+ /// \tparam R rounding mode to use
843
+ /// \tparam E `true` for round to even, `false` for round away from zero
844
+ /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it
845
+ /// \param value half-precision value to round
846
+ /// \return half-precision bits for nearest integral value
847
+ /// \exception FE_INVALID for signaling NaN
848
+ /// \exception FE_INEXACT if value had to be rounded and \a I is `true`
849
+ template<std::float_round_style R,bool E,bool I> unsigned int integral(unsigned int value)
850
+ {
851
+ unsigned int abs = value & 0x7FFF;
852
+ if(abs < 0x3C00)
853
+ {
854
+ raise(FE_INEXACT, I);
855
+ return ((R==std::round_to_nearest) ? (0x3C00&-static_cast<unsigned>(abs>=(0x3800+E))) :
856
+ (R==std::round_toward_infinity) ? (0x3C00&-(~(value>>15)&(abs!=0))) :
857
+ (R==std::round_toward_neg_infinity) ? (0x3C00&-static_cast<unsigned>(value>0x8000)) :
858
+ 0) | (value&0x8000);
859
+ }
860
+ if(abs >= 0x6400)
861
+ return (abs>0x7C00) ? signal(value) : value;
862
+ unsigned int exp = 25 - (abs>>10), mask = (1<<exp) - 1;
863
+ raise(FE_INEXACT, I && (value&mask));
864
+ return (( (R==std::round_to_nearest) ? ((1<<(exp-1))-(~(value>>exp)&E)) :
865
+ (R==std::round_toward_infinity) ? (mask&((value>>15)-1)) :
866
+ (R==std::round_toward_neg_infinity) ? (mask&-(value>>15)) :
867
+ 0) + value) & ~mask;
868
+ }
869
+
870
+ /// Convert fixed point to half-precision floating-point.
871
+ /// \tparam R rounding mode to use
872
+ /// \tparam F number of fractional bits (at least 11)
873
+ /// \tparam S `true` for signed, `false` for unsigned
874
+ /// \tparam N `true` for additional normalization step, `false` if already normalized to 1.F
875
+ /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results
876
+ /// \param m mantissa in Q1.F fixed point format
877
+ /// \param exp exponent
878
+ /// \param sign half-precision value with sign bit only
879
+ /// \param s sticky bit (or of all but the most significant already discarded bits)
880
+ /// \return value converted to half-precision
881
+ /// \exception FE_OVERFLOW on overflows
882
+ /// \exception FE_UNDERFLOW on underflows
883
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
884
+ template<std::float_round_style R,unsigned int F,bool S,bool N,bool I> unsigned int fixed2half(uint32 m, int exp = 14, unsigned int sign = 0, int s = 0)
885
+ {
886
+ if(S)
887
+ {
888
+ uint32 msign = sign_mask(m);
889
+ m = (m^msign) - msign;
890
+ sign = msign & 0x8000;
891
+ }
892
+ if(N)
893
+ for(; m<(static_cast<uint32>(1)<<F) && exp; m<<=1,--exp) ;
894
+ else if(exp < 0)
895
+ return rounded<R,I>(sign+(m>>(F-10-exp)), (m>>(F-11-exp))&1, s|((m&((static_cast<uint32>(1)<<(F-11-exp))-1))!=0));
896
+ return rounded<R,I>(sign+(exp<<10)+(m>>(F-10)), (m>>(F-11))&1, s|((m&((static_cast<uint32>(1)<<(F-11))-1))!=0));
897
+ }
898
+
899
+ /// Convert IEEE single-precision to half-precision.
900
+ /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf).
901
+ /// \tparam R rounding mode to use
902
+ /// \param value single-precision value to convert
903
+ /// \return rounded half-precision value
904
+ /// \exception FE_OVERFLOW on overflows
905
+ /// \exception FE_UNDERFLOW on underflows
906
+ /// \exception FE_INEXACT if value had to be rounded
907
+ template<std::float_round_style R> unsigned int float2half_impl(float value, true_type)
908
+ {
909
+ #if HALF_ENABLE_F16C_INTRINSICS
910
+ return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(value),
911
+ (R==std::round_to_nearest) ? _MM_FROUND_TO_NEAREST_INT :
912
+ (R==std::round_toward_zero) ? _MM_FROUND_TO_ZERO :
913
+ (R==std::round_toward_infinity) ? _MM_FROUND_TO_POS_INF :
914
+ (R==std::round_toward_neg_infinity) ? _MM_FROUND_TO_NEG_INF :
915
+ _MM_FROUND_CUR_DIRECTION));
916
+ #else
917
+ bits<float>::type fbits;
918
+ std::memcpy(&fbits, &value, sizeof(float));
919
+ #if 1
920
+ unsigned int sign = (fbits>>16) & 0x8000;
921
+ fbits &= 0x7FFFFFFF;
922
+ if(fbits >= 0x7F800000)
923
+ return sign | 0x7C00 | ((fbits>0x7F800000) ? (0x200|((fbits>>13)&0x3FF)) : 0);
924
+ if(fbits >= 0x47800000)
925
+ return overflow<R>(sign);
926
+ if(fbits >= 0x38800000)
927
+ return rounded<R,false>(sign|(((fbits>>23)-112)<<10)|((fbits>>13)&0x3FF), (fbits>>12)&1, (fbits&0xFFF)!=0);
928
+ if(fbits >= 0x33000000)
929
+ {
930
+ int i = 125 - (fbits>>23);
931
+ fbits = (fbits&0x7FFFFF) | 0x800000;
932
+ return rounded<R,false>(sign|(fbits>>(i+1)), (fbits>>i)&1, (fbits&((static_cast<uint32>(1)<<i)-1))!=0);
933
+ }
934
+ if(fbits != 0)
935
+ return underflow<R>(sign);
936
+ return sign;
937
+ #else
938
+ static const uint16 base_table[512] = {
939
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
940
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
941
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
942
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
943
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
944
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
945
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080, 0x0100,
946
+ 0x0200, 0x0400, 0x0800, 0x0C00, 0x1000, 0x1400, 0x1800, 0x1C00, 0x2000, 0x2400, 0x2800, 0x2C00, 0x3000, 0x3400, 0x3800, 0x3C00,
947
+ 0x4000, 0x4400, 0x4800, 0x4C00, 0x5000, 0x5400, 0x5800, 0x5C00, 0x6000, 0x6400, 0x6800, 0x6C00, 0x7000, 0x7400, 0x7800, 0x7BFF,
948
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
949
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
950
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
951
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
952
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
953
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
954
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7C00,
955
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
956
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
957
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
958
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
959
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
960
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
961
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008, 0x8010, 0x8020, 0x8040, 0x8080, 0x8100,
962
+ 0x8200, 0x8400, 0x8800, 0x8C00, 0x9000, 0x9400, 0x9800, 0x9C00, 0xA000, 0xA400, 0xA800, 0xAC00, 0xB000, 0xB400, 0xB800, 0xBC00,
963
+ 0xC000, 0xC400, 0xC800, 0xCC00, 0xD000, 0xD400, 0xD800, 0xDC00, 0xE000, 0xE400, 0xE800, 0xEC00, 0xF000, 0xF400, 0xF800, 0xFBFF,
964
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
965
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
966
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
967
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
968
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
969
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
970
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFC00 };
971
+ static const unsigned char shift_table[256] = {
972
+ 24, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
973
+ 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
974
+ 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
975
+ 25, 25, 25, 25, 25, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
976
+ 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
977
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
978
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
979
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 13 };
980
+ int sexp = fbits >> 23, exp = sexp & 0xFF, i = shift_table[exp];
981
+ fbits &= 0x7FFFFF;
982
+ uint32 m = (fbits|((exp!=0)<<23)) & -static_cast<uint32>(exp!=0xFF);
983
+ return rounded<R,false>(base_table[sexp]+(fbits>>i), (m>>(i-1))&1, (((static_cast<uint32>(1)<<(i-1))-1)&m)!=0);
984
+ #endif
985
+ #endif
986
+ }
987
+
988
+ /// Convert IEEE double-precision to half-precision.
989
+ /// \tparam R rounding mode to use
990
+ /// \param value double-precision value to convert
991
+ /// \return rounded half-precision value
992
+ /// \exception FE_OVERFLOW on overflows
993
+ /// \exception FE_UNDERFLOW on underflows
994
+ /// \exception FE_INEXACT if value had to be rounded
995
+ template<std::float_round_style R> unsigned int float2half_impl(double value, true_type)
996
+ {
997
+ #if HALF_ENABLE_F16C_INTRINSICS
998
+ if(R == std::round_indeterminate)
999
+ return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_cvtpd_ps(_mm_set_sd(value)), _MM_FROUND_CUR_DIRECTION));
1000
+ #endif
1001
+ bits<double>::type dbits;
1002
+ std::memcpy(&dbits, &value, sizeof(double));
1003
+ uint32 hi = dbits >> 32, lo = dbits & 0xFFFFFFFF;
1004
+ unsigned int sign = (hi>>16) & 0x8000;
1005
+ hi &= 0x7FFFFFFF;
1006
+ if(hi >= 0x7FF00000)
1007
+ return sign | 0x7C00 | ((dbits&0xFFFFFFFFFFFFF) ? (0x200|((hi>>10)&0x3FF)) : 0);
1008
+ if(hi >= 0x40F00000)
1009
+ return overflow<R>(sign);
1010
+ if(hi >= 0x3F100000)
1011
+ return rounded<R,false>(sign|(((hi>>20)-1008)<<10)|((hi>>10)&0x3FF), (hi>>9)&1, ((hi&0x1FF)|lo)!=0);
1012
+ if(hi >= 0x3E600000)
1013
+ {
1014
+ int i = 1018 - (hi>>20);
1015
+ hi = (hi&0xFFFFF) | 0x100000;
1016
+ return rounded<R,false>(sign|(hi>>(i+1)), (hi>>i)&1, ((hi&((static_cast<uint32>(1)<<i)-1))|lo)!=0);
1017
+ }
1018
+ if((hi|lo) != 0)
1019
+ return underflow<R>(sign);
1020
+ return sign;
1021
+ }
1022
+
1023
+ /// Convert non-IEEE floating-point to half-precision.
1024
+ /// \tparam R rounding mode to use
1025
+ /// \tparam T source type (builtin floating-point type)
1026
+ /// \param value floating-point value to convert
1027
+ /// \return rounded half-precision value
1028
+ /// \exception FE_OVERFLOW on overflows
1029
+ /// \exception FE_UNDERFLOW on underflows
1030
+ /// \exception FE_INEXACT if value had to be rounded
1031
+ template<std::float_round_style R,typename T> unsigned int float2half_impl(T value, ...)
1032
+ {
1033
+ unsigned int hbits = static_cast<unsigned>(builtin_signbit(value)) << 15;
1034
+ if(value == T())
1035
+ return hbits;
1036
+ if(builtin_isnan(value))
1037
+ return hbits | 0x7FFF;
1038
+ if(builtin_isinf(value))
1039
+ return hbits | 0x7C00;
1040
+ int exp;
1041
+ std::frexp(value, &exp);
1042
+ if(exp > 16)
1043
+ return overflow<R>(hbits);
1044
+ if(exp < -13)
1045
+ value = std::ldexp(value, 25);
1046
+ else
1047
+ {
1048
+ value = std::ldexp(value, 12-exp);
1049
+ hbits |= ((exp+13)<<10);
1050
+ }
1051
+ T ival, frac = std::modf(value, &ival);
1052
+ int m = std::abs(static_cast<int>(ival));
1053
+ return rounded<R,false>(hbits+(m>>1), m&1, frac!=T());
1054
+ }
1055
+
1056
+ /// Convert floating-point to half-precision.
1057
+ /// \tparam R rounding mode to use
1058
+ /// \tparam T source type (builtin floating-point type)
1059
+ /// \param value floating-point value to convert
1060
+ /// \return rounded half-precision value
1061
+ /// \exception FE_OVERFLOW on overflows
1062
+ /// \exception FE_UNDERFLOW on underflows
1063
+ /// \exception FE_INEXACT if value had to be rounded
1064
+ template<std::float_round_style R,typename T> unsigned int float2half(T value)
1065
+ {
1066
+ return float2half_impl<R>(value, bool_type<std::numeric_limits<T>::is_iec559&&sizeof(typename bits<T>::type)==sizeof(T)>());
1067
+ }
1068
+
1069
+ /// Convert integer to half-precision floating-point.
1070
+ /// \tparam R rounding mode to use
1071
+ /// \tparam T type to convert (builtin integer type)
1072
+ /// \param value integral value to convert
1073
+ /// \return rounded half-precision value
1074
+ /// \exception FE_OVERFLOW on overflows
1075
+ /// \exception FE_INEXACT if value had to be rounded
1076
+ template<std::float_round_style R,typename T> unsigned int int2half(T value)
1077
+ {
1078
+ unsigned int bits = static_cast<unsigned>(value<0) << 15;
1079
+ if(!value)
1080
+ return bits;
1081
+ if(bits)
1082
+ value = -value;
1083
+ if(value > 0xFFFF)
1084
+ return overflow<R>(bits);
1085
+ unsigned int m = static_cast<unsigned int>(value), exp = 24;
1086
+ for(; m<0x400; m<<=1,--exp) ;
1087
+ for(; m>0x7FF; m>>=1,++exp) ;
1088
+ bits |= (exp<<10) + m;
1089
+ return (exp>24) ? rounded<R,false>(bits, (value>>(exp-25))&1, (((1<<(exp-25))-1)&value)!=0) : bits;
1090
+ }
1091
+
1092
+ /// Convert half-precision to IEEE single-precision.
1093
+ /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf).
1094
+ /// \param value half-precision value to convert
1095
+ /// \return single-precision value
1096
+ inline float half2float_impl(unsigned int value, float, true_type)
1097
+ {
1098
+ #if HALF_ENABLE_F16C_INTRINSICS
1099
+ return _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(value)));
1100
+ #else
1101
+ #if 0
1102
+ bits<float>::type fbits = static_cast<bits<float>::type>(value&0x8000) << 16;
1103
+ int abs = value & 0x7FFF;
1104
+ if(abs)
1105
+ {
1106
+ fbits |= 0x38000000 << static_cast<unsigned>(abs>=0x7C00);
1107
+ for(; abs<0x400; abs<<=1,fbits-=0x800000) ;
1108
+ fbits += static_cast<bits<float>::type>(abs) << 13;
1109
+ }
1110
+ #else
1111
+ static const bits<float>::type mantissa_table[2048] = {
1112
+ 0x00000000, 0x33800000, 0x34000000, 0x34400000, 0x34800000, 0x34A00000, 0x34C00000, 0x34E00000, 0x35000000, 0x35100000, 0x35200000, 0x35300000, 0x35400000, 0x35500000, 0x35600000, 0x35700000,
1113
+ 0x35800000, 0x35880000, 0x35900000, 0x35980000, 0x35A00000, 0x35A80000, 0x35B00000, 0x35B80000, 0x35C00000, 0x35C80000, 0x35D00000, 0x35D80000, 0x35E00000, 0x35E80000, 0x35F00000, 0x35F80000,
1114
+ 0x36000000, 0x36040000, 0x36080000, 0x360C0000, 0x36100000, 0x36140000, 0x36180000, 0x361C0000, 0x36200000, 0x36240000, 0x36280000, 0x362C0000, 0x36300000, 0x36340000, 0x36380000, 0x363C0000,
1115
+ 0x36400000, 0x36440000, 0x36480000, 0x364C0000, 0x36500000, 0x36540000, 0x36580000, 0x365C0000, 0x36600000, 0x36640000, 0x36680000, 0x366C0000, 0x36700000, 0x36740000, 0x36780000, 0x367C0000,
1116
+ 0x36800000, 0x36820000, 0x36840000, 0x36860000, 0x36880000, 0x368A0000, 0x368C0000, 0x368E0000, 0x36900000, 0x36920000, 0x36940000, 0x36960000, 0x36980000, 0x369A0000, 0x369C0000, 0x369E0000,
1117
+ 0x36A00000, 0x36A20000, 0x36A40000, 0x36A60000, 0x36A80000, 0x36AA0000, 0x36AC0000, 0x36AE0000, 0x36B00000, 0x36B20000, 0x36B40000, 0x36B60000, 0x36B80000, 0x36BA0000, 0x36BC0000, 0x36BE0000,
1118
+ 0x36C00000, 0x36C20000, 0x36C40000, 0x36C60000, 0x36C80000, 0x36CA0000, 0x36CC0000, 0x36CE0000, 0x36D00000, 0x36D20000, 0x36D40000, 0x36D60000, 0x36D80000, 0x36DA0000, 0x36DC0000, 0x36DE0000,
1119
+ 0x36E00000, 0x36E20000, 0x36E40000, 0x36E60000, 0x36E80000, 0x36EA0000, 0x36EC0000, 0x36EE0000, 0x36F00000, 0x36F20000, 0x36F40000, 0x36F60000, 0x36F80000, 0x36FA0000, 0x36FC0000, 0x36FE0000,
1120
+ 0x37000000, 0x37010000, 0x37020000, 0x37030000, 0x37040000, 0x37050000, 0x37060000, 0x37070000, 0x37080000, 0x37090000, 0x370A0000, 0x370B0000, 0x370C0000, 0x370D0000, 0x370E0000, 0x370F0000,
1121
+ 0x37100000, 0x37110000, 0x37120000, 0x37130000, 0x37140000, 0x37150000, 0x37160000, 0x37170000, 0x37180000, 0x37190000, 0x371A0000, 0x371B0000, 0x371C0000, 0x371D0000, 0x371E0000, 0x371F0000,
1122
+ 0x37200000, 0x37210000, 0x37220000, 0x37230000, 0x37240000, 0x37250000, 0x37260000, 0x37270000, 0x37280000, 0x37290000, 0x372A0000, 0x372B0000, 0x372C0000, 0x372D0000, 0x372E0000, 0x372F0000,
1123
+ 0x37300000, 0x37310000, 0x37320000, 0x37330000, 0x37340000, 0x37350000, 0x37360000, 0x37370000, 0x37380000, 0x37390000, 0x373A0000, 0x373B0000, 0x373C0000, 0x373D0000, 0x373E0000, 0x373F0000,
1124
+ 0x37400000, 0x37410000, 0x37420000, 0x37430000, 0x37440000, 0x37450000, 0x37460000, 0x37470000, 0x37480000, 0x37490000, 0x374A0000, 0x374B0000, 0x374C0000, 0x374D0000, 0x374E0000, 0x374F0000,
1125
+ 0x37500000, 0x37510000, 0x37520000, 0x37530000, 0x37540000, 0x37550000, 0x37560000, 0x37570000, 0x37580000, 0x37590000, 0x375A0000, 0x375B0000, 0x375C0000, 0x375D0000, 0x375E0000, 0x375F0000,
1126
+ 0x37600000, 0x37610000, 0x37620000, 0x37630000, 0x37640000, 0x37650000, 0x37660000, 0x37670000, 0x37680000, 0x37690000, 0x376A0000, 0x376B0000, 0x376C0000, 0x376D0000, 0x376E0000, 0x376F0000,
1127
+ 0x37700000, 0x37710000, 0x37720000, 0x37730000, 0x37740000, 0x37750000, 0x37760000, 0x37770000, 0x37780000, 0x37790000, 0x377A0000, 0x377B0000, 0x377C0000, 0x377D0000, 0x377E0000, 0x377F0000,
1128
+ 0x37800000, 0x37808000, 0x37810000, 0x37818000, 0x37820000, 0x37828000, 0x37830000, 0x37838000, 0x37840000, 0x37848000, 0x37850000, 0x37858000, 0x37860000, 0x37868000, 0x37870000, 0x37878000,
1129
+ 0x37880000, 0x37888000, 0x37890000, 0x37898000, 0x378A0000, 0x378A8000, 0x378B0000, 0x378B8000, 0x378C0000, 0x378C8000, 0x378D0000, 0x378D8000, 0x378E0000, 0x378E8000, 0x378F0000, 0x378F8000,
1130
+ 0x37900000, 0x37908000, 0x37910000, 0x37918000, 0x37920000, 0x37928000, 0x37930000, 0x37938000, 0x37940000, 0x37948000, 0x37950000, 0x37958000, 0x37960000, 0x37968000, 0x37970000, 0x37978000,
1131
+ 0x37980000, 0x37988000, 0x37990000, 0x37998000, 0x379A0000, 0x379A8000, 0x379B0000, 0x379B8000, 0x379C0000, 0x379C8000, 0x379D0000, 0x379D8000, 0x379E0000, 0x379E8000, 0x379F0000, 0x379F8000,
1132
+ 0x37A00000, 0x37A08000, 0x37A10000, 0x37A18000, 0x37A20000, 0x37A28000, 0x37A30000, 0x37A38000, 0x37A40000, 0x37A48000, 0x37A50000, 0x37A58000, 0x37A60000, 0x37A68000, 0x37A70000, 0x37A78000,
1133
+ 0x37A80000, 0x37A88000, 0x37A90000, 0x37A98000, 0x37AA0000, 0x37AA8000, 0x37AB0000, 0x37AB8000, 0x37AC0000, 0x37AC8000, 0x37AD0000, 0x37AD8000, 0x37AE0000, 0x37AE8000, 0x37AF0000, 0x37AF8000,
1134
+ 0x37B00000, 0x37B08000, 0x37B10000, 0x37B18000, 0x37B20000, 0x37B28000, 0x37B30000, 0x37B38000, 0x37B40000, 0x37B48000, 0x37B50000, 0x37B58000, 0x37B60000, 0x37B68000, 0x37B70000, 0x37B78000,
1135
+ 0x37B80000, 0x37B88000, 0x37B90000, 0x37B98000, 0x37BA0000, 0x37BA8000, 0x37BB0000, 0x37BB8000, 0x37BC0000, 0x37BC8000, 0x37BD0000, 0x37BD8000, 0x37BE0000, 0x37BE8000, 0x37BF0000, 0x37BF8000,
1136
+ 0x37C00000, 0x37C08000, 0x37C10000, 0x37C18000, 0x37C20000, 0x37C28000, 0x37C30000, 0x37C38000, 0x37C40000, 0x37C48000, 0x37C50000, 0x37C58000, 0x37C60000, 0x37C68000, 0x37C70000, 0x37C78000,
1137
+ 0x37C80000, 0x37C88000, 0x37C90000, 0x37C98000, 0x37CA0000, 0x37CA8000, 0x37CB0000, 0x37CB8000, 0x37CC0000, 0x37CC8000, 0x37CD0000, 0x37CD8000, 0x37CE0000, 0x37CE8000, 0x37CF0000, 0x37CF8000,
1138
+ 0x37D00000, 0x37D08000, 0x37D10000, 0x37D18000, 0x37D20000, 0x37D28000, 0x37D30000, 0x37D38000, 0x37D40000, 0x37D48000, 0x37D50000, 0x37D58000, 0x37D60000, 0x37D68000, 0x37D70000, 0x37D78000,
1139
+ 0x37D80000, 0x37D88000, 0x37D90000, 0x37D98000, 0x37DA0000, 0x37DA8000, 0x37DB0000, 0x37DB8000, 0x37DC0000, 0x37DC8000, 0x37DD0000, 0x37DD8000, 0x37DE0000, 0x37DE8000, 0x37DF0000, 0x37DF8000,
1140
+ 0x37E00000, 0x37E08000, 0x37E10000, 0x37E18000, 0x37E20000, 0x37E28000, 0x37E30000, 0x37E38000, 0x37E40000, 0x37E48000, 0x37E50000, 0x37E58000, 0x37E60000, 0x37E68000, 0x37E70000, 0x37E78000,
1141
+ 0x37E80000, 0x37E88000, 0x37E90000, 0x37E98000, 0x37EA0000, 0x37EA8000, 0x37EB0000, 0x37EB8000, 0x37EC0000, 0x37EC8000, 0x37ED0000, 0x37ED8000, 0x37EE0000, 0x37EE8000, 0x37EF0000, 0x37EF8000,
1142
+ 0x37F00000, 0x37F08000, 0x37F10000, 0x37F18000, 0x37F20000, 0x37F28000, 0x37F30000, 0x37F38000, 0x37F40000, 0x37F48000, 0x37F50000, 0x37F58000, 0x37F60000, 0x37F68000, 0x37F70000, 0x37F78000,
1143
+ 0x37F80000, 0x37F88000, 0x37F90000, 0x37F98000, 0x37FA0000, 0x37FA8000, 0x37FB0000, 0x37FB8000, 0x37FC0000, 0x37FC8000, 0x37FD0000, 0x37FD8000, 0x37FE0000, 0x37FE8000, 0x37FF0000, 0x37FF8000,
1144
+ 0x38000000, 0x38004000, 0x38008000, 0x3800C000, 0x38010000, 0x38014000, 0x38018000, 0x3801C000, 0x38020000, 0x38024000, 0x38028000, 0x3802C000, 0x38030000, 0x38034000, 0x38038000, 0x3803C000,
1145
+ 0x38040000, 0x38044000, 0x38048000, 0x3804C000, 0x38050000, 0x38054000, 0x38058000, 0x3805C000, 0x38060000, 0x38064000, 0x38068000, 0x3806C000, 0x38070000, 0x38074000, 0x38078000, 0x3807C000,
1146
+ 0x38080000, 0x38084000, 0x38088000, 0x3808C000, 0x38090000, 0x38094000, 0x38098000, 0x3809C000, 0x380A0000, 0x380A4000, 0x380A8000, 0x380AC000, 0x380B0000, 0x380B4000, 0x380B8000, 0x380BC000,
1147
+ 0x380C0000, 0x380C4000, 0x380C8000, 0x380CC000, 0x380D0000, 0x380D4000, 0x380D8000, 0x380DC000, 0x380E0000, 0x380E4000, 0x380E8000, 0x380EC000, 0x380F0000, 0x380F4000, 0x380F8000, 0x380FC000,
1148
+ 0x38100000, 0x38104000, 0x38108000, 0x3810C000, 0x38110000, 0x38114000, 0x38118000, 0x3811C000, 0x38120000, 0x38124000, 0x38128000, 0x3812C000, 0x38130000, 0x38134000, 0x38138000, 0x3813C000,
1149
+ 0x38140000, 0x38144000, 0x38148000, 0x3814C000, 0x38150000, 0x38154000, 0x38158000, 0x3815C000, 0x38160000, 0x38164000, 0x38168000, 0x3816C000, 0x38170000, 0x38174000, 0x38178000, 0x3817C000,
1150
+ 0x38180000, 0x38184000, 0x38188000, 0x3818C000, 0x38190000, 0x38194000, 0x38198000, 0x3819C000, 0x381A0000, 0x381A4000, 0x381A8000, 0x381AC000, 0x381B0000, 0x381B4000, 0x381B8000, 0x381BC000,
1151
+ 0x381C0000, 0x381C4000, 0x381C8000, 0x381CC000, 0x381D0000, 0x381D4000, 0x381D8000, 0x381DC000, 0x381E0000, 0x381E4000, 0x381E8000, 0x381EC000, 0x381F0000, 0x381F4000, 0x381F8000, 0x381FC000,
1152
+ 0x38200000, 0x38204000, 0x38208000, 0x3820C000, 0x38210000, 0x38214000, 0x38218000, 0x3821C000, 0x38220000, 0x38224000, 0x38228000, 0x3822C000, 0x38230000, 0x38234000, 0x38238000, 0x3823C000,
1153
+ 0x38240000, 0x38244000, 0x38248000, 0x3824C000, 0x38250000, 0x38254000, 0x38258000, 0x3825C000, 0x38260000, 0x38264000, 0x38268000, 0x3826C000, 0x38270000, 0x38274000, 0x38278000, 0x3827C000,
1154
+ 0x38280000, 0x38284000, 0x38288000, 0x3828C000, 0x38290000, 0x38294000, 0x38298000, 0x3829C000, 0x382A0000, 0x382A4000, 0x382A8000, 0x382AC000, 0x382B0000, 0x382B4000, 0x382B8000, 0x382BC000,
1155
+ 0x382C0000, 0x382C4000, 0x382C8000, 0x382CC000, 0x382D0000, 0x382D4000, 0x382D8000, 0x382DC000, 0x382E0000, 0x382E4000, 0x382E8000, 0x382EC000, 0x382F0000, 0x382F4000, 0x382F8000, 0x382FC000,
1156
+ 0x38300000, 0x38304000, 0x38308000, 0x3830C000, 0x38310000, 0x38314000, 0x38318000, 0x3831C000, 0x38320000, 0x38324000, 0x38328000, 0x3832C000, 0x38330000, 0x38334000, 0x38338000, 0x3833C000,
1157
+ 0x38340000, 0x38344000, 0x38348000, 0x3834C000, 0x38350000, 0x38354000, 0x38358000, 0x3835C000, 0x38360000, 0x38364000, 0x38368000, 0x3836C000, 0x38370000, 0x38374000, 0x38378000, 0x3837C000,
1158
+ 0x38380000, 0x38384000, 0x38388000, 0x3838C000, 0x38390000, 0x38394000, 0x38398000, 0x3839C000, 0x383A0000, 0x383A4000, 0x383A8000, 0x383AC000, 0x383B0000, 0x383B4000, 0x383B8000, 0x383BC000,
1159
+ 0x383C0000, 0x383C4000, 0x383C8000, 0x383CC000, 0x383D0000, 0x383D4000, 0x383D8000, 0x383DC000, 0x383E0000, 0x383E4000, 0x383E8000, 0x383EC000, 0x383F0000, 0x383F4000, 0x383F8000, 0x383FC000,
1160
+ 0x38400000, 0x38404000, 0x38408000, 0x3840C000, 0x38410000, 0x38414000, 0x38418000, 0x3841C000, 0x38420000, 0x38424000, 0x38428000, 0x3842C000, 0x38430000, 0x38434000, 0x38438000, 0x3843C000,
1161
+ 0x38440000, 0x38444000, 0x38448000, 0x3844C000, 0x38450000, 0x38454000, 0x38458000, 0x3845C000, 0x38460000, 0x38464000, 0x38468000, 0x3846C000, 0x38470000, 0x38474000, 0x38478000, 0x3847C000,
1162
+ 0x38480000, 0x38484000, 0x38488000, 0x3848C000, 0x38490000, 0x38494000, 0x38498000, 0x3849C000, 0x384A0000, 0x384A4000, 0x384A8000, 0x384AC000, 0x384B0000, 0x384B4000, 0x384B8000, 0x384BC000,
1163
+ 0x384C0000, 0x384C4000, 0x384C8000, 0x384CC000, 0x384D0000, 0x384D4000, 0x384D8000, 0x384DC000, 0x384E0000, 0x384E4000, 0x384E8000, 0x384EC000, 0x384F0000, 0x384F4000, 0x384F8000, 0x384FC000,
1164
+ 0x38500000, 0x38504000, 0x38508000, 0x3850C000, 0x38510000, 0x38514000, 0x38518000, 0x3851C000, 0x38520000, 0x38524000, 0x38528000, 0x3852C000, 0x38530000, 0x38534000, 0x38538000, 0x3853C000,
1165
+ 0x38540000, 0x38544000, 0x38548000, 0x3854C000, 0x38550000, 0x38554000, 0x38558000, 0x3855C000, 0x38560000, 0x38564000, 0x38568000, 0x3856C000, 0x38570000, 0x38574000, 0x38578000, 0x3857C000,
1166
+ 0x38580000, 0x38584000, 0x38588000, 0x3858C000, 0x38590000, 0x38594000, 0x38598000, 0x3859C000, 0x385A0000, 0x385A4000, 0x385A8000, 0x385AC000, 0x385B0000, 0x385B4000, 0x385B8000, 0x385BC000,
1167
+ 0x385C0000, 0x385C4000, 0x385C8000, 0x385CC000, 0x385D0000, 0x385D4000, 0x385D8000, 0x385DC000, 0x385E0000, 0x385E4000, 0x385E8000, 0x385EC000, 0x385F0000, 0x385F4000, 0x385F8000, 0x385FC000,
1168
+ 0x38600000, 0x38604000, 0x38608000, 0x3860C000, 0x38610000, 0x38614000, 0x38618000, 0x3861C000, 0x38620000, 0x38624000, 0x38628000, 0x3862C000, 0x38630000, 0x38634000, 0x38638000, 0x3863C000,
1169
+ 0x38640000, 0x38644000, 0x38648000, 0x3864C000, 0x38650000, 0x38654000, 0x38658000, 0x3865C000, 0x38660000, 0x38664000, 0x38668000, 0x3866C000, 0x38670000, 0x38674000, 0x38678000, 0x3867C000,
1170
+ 0x38680000, 0x38684000, 0x38688000, 0x3868C000, 0x38690000, 0x38694000, 0x38698000, 0x3869C000, 0x386A0000, 0x386A4000, 0x386A8000, 0x386AC000, 0x386B0000, 0x386B4000, 0x386B8000, 0x386BC000,
1171
+ 0x386C0000, 0x386C4000, 0x386C8000, 0x386CC000, 0x386D0000, 0x386D4000, 0x386D8000, 0x386DC000, 0x386E0000, 0x386E4000, 0x386E8000, 0x386EC000, 0x386F0000, 0x386F4000, 0x386F8000, 0x386FC000,
1172
+ 0x38700000, 0x38704000, 0x38708000, 0x3870C000, 0x38710000, 0x38714000, 0x38718000, 0x3871C000, 0x38720000, 0x38724000, 0x38728000, 0x3872C000, 0x38730000, 0x38734000, 0x38738000, 0x3873C000,
1173
+ 0x38740000, 0x38744000, 0x38748000, 0x3874C000, 0x38750000, 0x38754000, 0x38758000, 0x3875C000, 0x38760000, 0x38764000, 0x38768000, 0x3876C000, 0x38770000, 0x38774000, 0x38778000, 0x3877C000,
1174
+ 0x38780000, 0x38784000, 0x38788000, 0x3878C000, 0x38790000, 0x38794000, 0x38798000, 0x3879C000, 0x387A0000, 0x387A4000, 0x387A8000, 0x387AC000, 0x387B0000, 0x387B4000, 0x387B8000, 0x387BC000,
1175
+ 0x387C0000, 0x387C4000, 0x387C8000, 0x387CC000, 0x387D0000, 0x387D4000, 0x387D8000, 0x387DC000, 0x387E0000, 0x387E4000, 0x387E8000, 0x387EC000, 0x387F0000, 0x387F4000, 0x387F8000, 0x387FC000,
1176
+ 0x38000000, 0x38002000, 0x38004000, 0x38006000, 0x38008000, 0x3800A000, 0x3800C000, 0x3800E000, 0x38010000, 0x38012000, 0x38014000, 0x38016000, 0x38018000, 0x3801A000, 0x3801C000, 0x3801E000,
1177
+ 0x38020000, 0x38022000, 0x38024000, 0x38026000, 0x38028000, 0x3802A000, 0x3802C000, 0x3802E000, 0x38030000, 0x38032000, 0x38034000, 0x38036000, 0x38038000, 0x3803A000, 0x3803C000, 0x3803E000,
1178
+ 0x38040000, 0x38042000, 0x38044000, 0x38046000, 0x38048000, 0x3804A000, 0x3804C000, 0x3804E000, 0x38050000, 0x38052000, 0x38054000, 0x38056000, 0x38058000, 0x3805A000, 0x3805C000, 0x3805E000,
1179
+ 0x38060000, 0x38062000, 0x38064000, 0x38066000, 0x38068000, 0x3806A000, 0x3806C000, 0x3806E000, 0x38070000, 0x38072000, 0x38074000, 0x38076000, 0x38078000, 0x3807A000, 0x3807C000, 0x3807E000,
1180
+ 0x38080000, 0x38082000, 0x38084000, 0x38086000, 0x38088000, 0x3808A000, 0x3808C000, 0x3808E000, 0x38090000, 0x38092000, 0x38094000, 0x38096000, 0x38098000, 0x3809A000, 0x3809C000, 0x3809E000,
1181
+ 0x380A0000, 0x380A2000, 0x380A4000, 0x380A6000, 0x380A8000, 0x380AA000, 0x380AC000, 0x380AE000, 0x380B0000, 0x380B2000, 0x380B4000, 0x380B6000, 0x380B8000, 0x380BA000, 0x380BC000, 0x380BE000,
1182
+ 0x380C0000, 0x380C2000, 0x380C4000, 0x380C6000, 0x380C8000, 0x380CA000, 0x380CC000, 0x380CE000, 0x380D0000, 0x380D2000, 0x380D4000, 0x380D6000, 0x380D8000, 0x380DA000, 0x380DC000, 0x380DE000,
1183
+ 0x380E0000, 0x380E2000, 0x380E4000, 0x380E6000, 0x380E8000, 0x380EA000, 0x380EC000, 0x380EE000, 0x380F0000, 0x380F2000, 0x380F4000, 0x380F6000, 0x380F8000, 0x380FA000, 0x380FC000, 0x380FE000,
1184
+ 0x38100000, 0x38102000, 0x38104000, 0x38106000, 0x38108000, 0x3810A000, 0x3810C000, 0x3810E000, 0x38110000, 0x38112000, 0x38114000, 0x38116000, 0x38118000, 0x3811A000, 0x3811C000, 0x3811E000,
1185
+ 0x38120000, 0x38122000, 0x38124000, 0x38126000, 0x38128000, 0x3812A000, 0x3812C000, 0x3812E000, 0x38130000, 0x38132000, 0x38134000, 0x38136000, 0x38138000, 0x3813A000, 0x3813C000, 0x3813E000,
1186
+ 0x38140000, 0x38142000, 0x38144000, 0x38146000, 0x38148000, 0x3814A000, 0x3814C000, 0x3814E000, 0x38150000, 0x38152000, 0x38154000, 0x38156000, 0x38158000, 0x3815A000, 0x3815C000, 0x3815E000,
1187
+ 0x38160000, 0x38162000, 0x38164000, 0x38166000, 0x38168000, 0x3816A000, 0x3816C000, 0x3816E000, 0x38170000, 0x38172000, 0x38174000, 0x38176000, 0x38178000, 0x3817A000, 0x3817C000, 0x3817E000,
1188
+ 0x38180000, 0x38182000, 0x38184000, 0x38186000, 0x38188000, 0x3818A000, 0x3818C000, 0x3818E000, 0x38190000, 0x38192000, 0x38194000, 0x38196000, 0x38198000, 0x3819A000, 0x3819C000, 0x3819E000,
1189
+ 0x381A0000, 0x381A2000, 0x381A4000, 0x381A6000, 0x381A8000, 0x381AA000, 0x381AC000, 0x381AE000, 0x381B0000, 0x381B2000, 0x381B4000, 0x381B6000, 0x381B8000, 0x381BA000, 0x381BC000, 0x381BE000,
1190
+ 0x381C0000, 0x381C2000, 0x381C4000, 0x381C6000, 0x381C8000, 0x381CA000, 0x381CC000, 0x381CE000, 0x381D0000, 0x381D2000, 0x381D4000, 0x381D6000, 0x381D8000, 0x381DA000, 0x381DC000, 0x381DE000,
1191
+ 0x381E0000, 0x381E2000, 0x381E4000, 0x381E6000, 0x381E8000, 0x381EA000, 0x381EC000, 0x381EE000, 0x381F0000, 0x381F2000, 0x381F4000, 0x381F6000, 0x381F8000, 0x381FA000, 0x381FC000, 0x381FE000,
1192
+ 0x38200000, 0x38202000, 0x38204000, 0x38206000, 0x38208000, 0x3820A000, 0x3820C000, 0x3820E000, 0x38210000, 0x38212000, 0x38214000, 0x38216000, 0x38218000, 0x3821A000, 0x3821C000, 0x3821E000,
1193
+ 0x38220000, 0x38222000, 0x38224000, 0x38226000, 0x38228000, 0x3822A000, 0x3822C000, 0x3822E000, 0x38230000, 0x38232000, 0x38234000, 0x38236000, 0x38238000, 0x3823A000, 0x3823C000, 0x3823E000,
1194
+ 0x38240000, 0x38242000, 0x38244000, 0x38246000, 0x38248000, 0x3824A000, 0x3824C000, 0x3824E000, 0x38250000, 0x38252000, 0x38254000, 0x38256000, 0x38258000, 0x3825A000, 0x3825C000, 0x3825E000,
1195
+ 0x38260000, 0x38262000, 0x38264000, 0x38266000, 0x38268000, 0x3826A000, 0x3826C000, 0x3826E000, 0x38270000, 0x38272000, 0x38274000, 0x38276000, 0x38278000, 0x3827A000, 0x3827C000, 0x3827E000,
1196
+ 0x38280000, 0x38282000, 0x38284000, 0x38286000, 0x38288000, 0x3828A000, 0x3828C000, 0x3828E000, 0x38290000, 0x38292000, 0x38294000, 0x38296000, 0x38298000, 0x3829A000, 0x3829C000, 0x3829E000,
1197
+ 0x382A0000, 0x382A2000, 0x382A4000, 0x382A6000, 0x382A8000, 0x382AA000, 0x382AC000, 0x382AE000, 0x382B0000, 0x382B2000, 0x382B4000, 0x382B6000, 0x382B8000, 0x382BA000, 0x382BC000, 0x382BE000,
1198
+ 0x382C0000, 0x382C2000, 0x382C4000, 0x382C6000, 0x382C8000, 0x382CA000, 0x382CC000, 0x382CE000, 0x382D0000, 0x382D2000, 0x382D4000, 0x382D6000, 0x382D8000, 0x382DA000, 0x382DC000, 0x382DE000,
1199
+ 0x382E0000, 0x382E2000, 0x382E4000, 0x382E6000, 0x382E8000, 0x382EA000, 0x382EC000, 0x382EE000, 0x382F0000, 0x382F2000, 0x382F4000, 0x382F6000, 0x382F8000, 0x382FA000, 0x382FC000, 0x382FE000,
1200
+ 0x38300000, 0x38302000, 0x38304000, 0x38306000, 0x38308000, 0x3830A000, 0x3830C000, 0x3830E000, 0x38310000, 0x38312000, 0x38314000, 0x38316000, 0x38318000, 0x3831A000, 0x3831C000, 0x3831E000,
1201
+ 0x38320000, 0x38322000, 0x38324000, 0x38326000, 0x38328000, 0x3832A000, 0x3832C000, 0x3832E000, 0x38330000, 0x38332000, 0x38334000, 0x38336000, 0x38338000, 0x3833A000, 0x3833C000, 0x3833E000,
1202
+ 0x38340000, 0x38342000, 0x38344000, 0x38346000, 0x38348000, 0x3834A000, 0x3834C000, 0x3834E000, 0x38350000, 0x38352000, 0x38354000, 0x38356000, 0x38358000, 0x3835A000, 0x3835C000, 0x3835E000,
1203
+ 0x38360000, 0x38362000, 0x38364000, 0x38366000, 0x38368000, 0x3836A000, 0x3836C000, 0x3836E000, 0x38370000, 0x38372000, 0x38374000, 0x38376000, 0x38378000, 0x3837A000, 0x3837C000, 0x3837E000,
1204
+ 0x38380000, 0x38382000, 0x38384000, 0x38386000, 0x38388000, 0x3838A000, 0x3838C000, 0x3838E000, 0x38390000, 0x38392000, 0x38394000, 0x38396000, 0x38398000, 0x3839A000, 0x3839C000, 0x3839E000,
1205
+ 0x383A0000, 0x383A2000, 0x383A4000, 0x383A6000, 0x383A8000, 0x383AA000, 0x383AC000, 0x383AE000, 0x383B0000, 0x383B2000, 0x383B4000, 0x383B6000, 0x383B8000, 0x383BA000, 0x383BC000, 0x383BE000,
1206
+ 0x383C0000, 0x383C2000, 0x383C4000, 0x383C6000, 0x383C8000, 0x383CA000, 0x383CC000, 0x383CE000, 0x383D0000, 0x383D2000, 0x383D4000, 0x383D6000, 0x383D8000, 0x383DA000, 0x383DC000, 0x383DE000,
1207
+ 0x383E0000, 0x383E2000, 0x383E4000, 0x383E6000, 0x383E8000, 0x383EA000, 0x383EC000, 0x383EE000, 0x383F0000, 0x383F2000, 0x383F4000, 0x383F6000, 0x383F8000, 0x383FA000, 0x383FC000, 0x383FE000,
1208
+ 0x38400000, 0x38402000, 0x38404000, 0x38406000, 0x38408000, 0x3840A000, 0x3840C000, 0x3840E000, 0x38410000, 0x38412000, 0x38414000, 0x38416000, 0x38418000, 0x3841A000, 0x3841C000, 0x3841E000,
1209
+ 0x38420000, 0x38422000, 0x38424000, 0x38426000, 0x38428000, 0x3842A000, 0x3842C000, 0x3842E000, 0x38430000, 0x38432000, 0x38434000, 0x38436000, 0x38438000, 0x3843A000, 0x3843C000, 0x3843E000,
1210
+ 0x38440000, 0x38442000, 0x38444000, 0x38446000, 0x38448000, 0x3844A000, 0x3844C000, 0x3844E000, 0x38450000, 0x38452000, 0x38454000, 0x38456000, 0x38458000, 0x3845A000, 0x3845C000, 0x3845E000,
1211
+ 0x38460000, 0x38462000, 0x38464000, 0x38466000, 0x38468000, 0x3846A000, 0x3846C000, 0x3846E000, 0x38470000, 0x38472000, 0x38474000, 0x38476000, 0x38478000, 0x3847A000, 0x3847C000, 0x3847E000,
1212
+ 0x38480000, 0x38482000, 0x38484000, 0x38486000, 0x38488000, 0x3848A000, 0x3848C000, 0x3848E000, 0x38490000, 0x38492000, 0x38494000, 0x38496000, 0x38498000, 0x3849A000, 0x3849C000, 0x3849E000,
1213
+ 0x384A0000, 0x384A2000, 0x384A4000, 0x384A6000, 0x384A8000, 0x384AA000, 0x384AC000, 0x384AE000, 0x384B0000, 0x384B2000, 0x384B4000, 0x384B6000, 0x384B8000, 0x384BA000, 0x384BC000, 0x384BE000,
1214
+ 0x384C0000, 0x384C2000, 0x384C4000, 0x384C6000, 0x384C8000, 0x384CA000, 0x384CC000, 0x384CE000, 0x384D0000, 0x384D2000, 0x384D4000, 0x384D6000, 0x384D8000, 0x384DA000, 0x384DC000, 0x384DE000,
1215
+ 0x384E0000, 0x384E2000, 0x384E4000, 0x384E6000, 0x384E8000, 0x384EA000, 0x384EC000, 0x384EE000, 0x384F0000, 0x384F2000, 0x384F4000, 0x384F6000, 0x384F8000, 0x384FA000, 0x384FC000, 0x384FE000,
1216
+ 0x38500000, 0x38502000, 0x38504000, 0x38506000, 0x38508000, 0x3850A000, 0x3850C000, 0x3850E000, 0x38510000, 0x38512000, 0x38514000, 0x38516000, 0x38518000, 0x3851A000, 0x3851C000, 0x3851E000,
1217
+ 0x38520000, 0x38522000, 0x38524000, 0x38526000, 0x38528000, 0x3852A000, 0x3852C000, 0x3852E000, 0x38530000, 0x38532000, 0x38534000, 0x38536000, 0x38538000, 0x3853A000, 0x3853C000, 0x3853E000,
1218
+ 0x38540000, 0x38542000, 0x38544000, 0x38546000, 0x38548000, 0x3854A000, 0x3854C000, 0x3854E000, 0x38550000, 0x38552000, 0x38554000, 0x38556000, 0x38558000, 0x3855A000, 0x3855C000, 0x3855E000,
1219
+ 0x38560000, 0x38562000, 0x38564000, 0x38566000, 0x38568000, 0x3856A000, 0x3856C000, 0x3856E000, 0x38570000, 0x38572000, 0x38574000, 0x38576000, 0x38578000, 0x3857A000, 0x3857C000, 0x3857E000,
1220
+ 0x38580000, 0x38582000, 0x38584000, 0x38586000, 0x38588000, 0x3858A000, 0x3858C000, 0x3858E000, 0x38590000, 0x38592000, 0x38594000, 0x38596000, 0x38598000, 0x3859A000, 0x3859C000, 0x3859E000,
1221
+ 0x385A0000, 0x385A2000, 0x385A4000, 0x385A6000, 0x385A8000, 0x385AA000, 0x385AC000, 0x385AE000, 0x385B0000, 0x385B2000, 0x385B4000, 0x385B6000, 0x385B8000, 0x385BA000, 0x385BC000, 0x385BE000,
1222
+ 0x385C0000, 0x385C2000, 0x385C4000, 0x385C6000, 0x385C8000, 0x385CA000, 0x385CC000, 0x385CE000, 0x385D0000, 0x385D2000, 0x385D4000, 0x385D6000, 0x385D8000, 0x385DA000, 0x385DC000, 0x385DE000,
1223
+ 0x385E0000, 0x385E2000, 0x385E4000, 0x385E6000, 0x385E8000, 0x385EA000, 0x385EC000, 0x385EE000, 0x385F0000, 0x385F2000, 0x385F4000, 0x385F6000, 0x385F8000, 0x385FA000, 0x385FC000, 0x385FE000,
1224
+ 0x38600000, 0x38602000, 0x38604000, 0x38606000, 0x38608000, 0x3860A000, 0x3860C000, 0x3860E000, 0x38610000, 0x38612000, 0x38614000, 0x38616000, 0x38618000, 0x3861A000, 0x3861C000, 0x3861E000,
1225
+ 0x38620000, 0x38622000, 0x38624000, 0x38626000, 0x38628000, 0x3862A000, 0x3862C000, 0x3862E000, 0x38630000, 0x38632000, 0x38634000, 0x38636000, 0x38638000, 0x3863A000, 0x3863C000, 0x3863E000,
1226
+ 0x38640000, 0x38642000, 0x38644000, 0x38646000, 0x38648000, 0x3864A000, 0x3864C000, 0x3864E000, 0x38650000, 0x38652000, 0x38654000, 0x38656000, 0x38658000, 0x3865A000, 0x3865C000, 0x3865E000,
1227
+ 0x38660000, 0x38662000, 0x38664000, 0x38666000, 0x38668000, 0x3866A000, 0x3866C000, 0x3866E000, 0x38670000, 0x38672000, 0x38674000, 0x38676000, 0x38678000, 0x3867A000, 0x3867C000, 0x3867E000,
1228
+ 0x38680000, 0x38682000, 0x38684000, 0x38686000, 0x38688000, 0x3868A000, 0x3868C000, 0x3868E000, 0x38690000, 0x38692000, 0x38694000, 0x38696000, 0x38698000, 0x3869A000, 0x3869C000, 0x3869E000,
1229
+ 0x386A0000, 0x386A2000, 0x386A4000, 0x386A6000, 0x386A8000, 0x386AA000, 0x386AC000, 0x386AE000, 0x386B0000, 0x386B2000, 0x386B4000, 0x386B6000, 0x386B8000, 0x386BA000, 0x386BC000, 0x386BE000,
1230
+ 0x386C0000, 0x386C2000, 0x386C4000, 0x386C6000, 0x386C8000, 0x386CA000, 0x386CC000, 0x386CE000, 0x386D0000, 0x386D2000, 0x386D4000, 0x386D6000, 0x386D8000, 0x386DA000, 0x386DC000, 0x386DE000,
1231
+ 0x386E0000, 0x386E2000, 0x386E4000, 0x386E6000, 0x386E8000, 0x386EA000, 0x386EC000, 0x386EE000, 0x386F0000, 0x386F2000, 0x386F4000, 0x386F6000, 0x386F8000, 0x386FA000, 0x386FC000, 0x386FE000,
1232
+ 0x38700000, 0x38702000, 0x38704000, 0x38706000, 0x38708000, 0x3870A000, 0x3870C000, 0x3870E000, 0x38710000, 0x38712000, 0x38714000, 0x38716000, 0x38718000, 0x3871A000, 0x3871C000, 0x3871E000,
1233
+ 0x38720000, 0x38722000, 0x38724000, 0x38726000, 0x38728000, 0x3872A000, 0x3872C000, 0x3872E000, 0x38730000, 0x38732000, 0x38734000, 0x38736000, 0x38738000, 0x3873A000, 0x3873C000, 0x3873E000,
1234
+ 0x38740000, 0x38742000, 0x38744000, 0x38746000, 0x38748000, 0x3874A000, 0x3874C000, 0x3874E000, 0x38750000, 0x38752000, 0x38754000, 0x38756000, 0x38758000, 0x3875A000, 0x3875C000, 0x3875E000,
1235
+ 0x38760000, 0x38762000, 0x38764000, 0x38766000, 0x38768000, 0x3876A000, 0x3876C000, 0x3876E000, 0x38770000, 0x38772000, 0x38774000, 0x38776000, 0x38778000, 0x3877A000, 0x3877C000, 0x3877E000,
1236
+ 0x38780000, 0x38782000, 0x38784000, 0x38786000, 0x38788000, 0x3878A000, 0x3878C000, 0x3878E000, 0x38790000, 0x38792000, 0x38794000, 0x38796000, 0x38798000, 0x3879A000, 0x3879C000, 0x3879E000,
1237
+ 0x387A0000, 0x387A2000, 0x387A4000, 0x387A6000, 0x387A8000, 0x387AA000, 0x387AC000, 0x387AE000, 0x387B0000, 0x387B2000, 0x387B4000, 0x387B6000, 0x387B8000, 0x387BA000, 0x387BC000, 0x387BE000,
1238
+ 0x387C0000, 0x387C2000, 0x387C4000, 0x387C6000, 0x387C8000, 0x387CA000, 0x387CC000, 0x387CE000, 0x387D0000, 0x387D2000, 0x387D4000, 0x387D6000, 0x387D8000, 0x387DA000, 0x387DC000, 0x387DE000,
1239
+ 0x387E0000, 0x387E2000, 0x387E4000, 0x387E6000, 0x387E8000, 0x387EA000, 0x387EC000, 0x387EE000, 0x387F0000, 0x387F2000, 0x387F4000, 0x387F6000, 0x387F8000, 0x387FA000, 0x387FC000, 0x387FE000 };
1240
+ static const bits<float>::type exponent_table[64] = {
1241
+ 0x00000000, 0x00800000, 0x01000000, 0x01800000, 0x02000000, 0x02800000, 0x03000000, 0x03800000, 0x04000000, 0x04800000, 0x05000000, 0x05800000, 0x06000000, 0x06800000, 0x07000000, 0x07800000,
1242
+ 0x08000000, 0x08800000, 0x09000000, 0x09800000, 0x0A000000, 0x0A800000, 0x0B000000, 0x0B800000, 0x0C000000, 0x0C800000, 0x0D000000, 0x0D800000, 0x0E000000, 0x0E800000, 0x0F000000, 0x47800000,
1243
+ 0x80000000, 0x80800000, 0x81000000, 0x81800000, 0x82000000, 0x82800000, 0x83000000, 0x83800000, 0x84000000, 0x84800000, 0x85000000, 0x85800000, 0x86000000, 0x86800000, 0x87000000, 0x87800000,
1244
+ 0x88000000, 0x88800000, 0x89000000, 0x89800000, 0x8A000000, 0x8A800000, 0x8B000000, 0x8B800000, 0x8C000000, 0x8C800000, 0x8D000000, 0x8D800000, 0x8E000000, 0x8E800000, 0x8F000000, 0xC7800000 };
1245
+ static const unsigned short offset_table[64] = {
1246
+ 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024,
1247
+ 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024 };
1248
+ bits<float>::type fbits = mantissa_table[offset_table[value>>10]+(value&0x3FF)] + exponent_table[value>>10];
1249
+ #endif
1250
+ float out;
1251
+ std::memcpy(&out, &fbits, sizeof(float));
1252
+ return out;
1253
+ #endif
1254
+ }
1255
+
1256
+ /// Convert half-precision to IEEE double-precision.
1257
+ /// \param value half-precision value to convert
1258
+ /// \return double-precision value
1259
+ inline double half2float_impl(unsigned int value, double, true_type)
1260
+ {
1261
+ #if HALF_ENABLE_F16C_INTRINSICS
1262
+ return _mm_cvtsd_f64(_mm_cvtps_pd(_mm_cvtph_ps(_mm_cvtsi32_si128(value))));
1263
+ #else
1264
+ uint32 hi = static_cast<uint32>(value&0x8000) << 16;
1265
+ unsigned int abs = value & 0x7FFF;
1266
+ if(abs)
1267
+ {
1268
+ hi |= 0x3F000000 << static_cast<unsigned>(abs>=0x7C00);
1269
+ for(; abs<0x400; abs<<=1,hi-=0x100000) ;
1270
+ hi += static_cast<uint32>(abs) << 10;
1271
+ }
1272
+ bits<double>::type dbits = static_cast<bits<double>::type>(hi) << 32;
1273
+ double out;
1274
+ std::memcpy(&out, &dbits, sizeof(double));
1275
+ return out;
1276
+ #endif
1277
+ }
1278
+
1279
+ /// Convert half-precision to non-IEEE floating-point.
1280
+ /// \tparam T type to convert to (builtin integer type)
1281
+ /// \param value half-precision value to convert
1282
+ /// \return floating-point value
1283
+ template<typename T> T half2float_impl(unsigned int value, T, ...)
1284
+ {
1285
+ T out;
1286
+ unsigned int abs = value & 0x7FFF;
1287
+ if(abs > 0x7C00)
1288
+ out = (std::numeric_limits<T>::has_signaling_NaN && !(abs&0x200)) ? std::numeric_limits<T>::signaling_NaN() :
1289
+ std::numeric_limits<T>::has_quiet_NaN ? std::numeric_limits<T>::quiet_NaN() : T();
1290
+ else if(abs == 0x7C00)
1291
+ out = std::numeric_limits<T>::has_infinity ? std::numeric_limits<T>::infinity() : std::numeric_limits<T>::max();
1292
+ else if(abs > 0x3FF)
1293
+ out = std::ldexp(static_cast<T>((abs&0x3FF)|0x400), (abs>>10)-25);
1294
+ else
1295
+ out = std::ldexp(static_cast<T>(abs), -24);
1296
+ return (value&0x8000) ? -out : out;
1297
+ }
1298
+
1299
+ /// Convert half-precision to floating-point.
1300
+ /// \tparam T type to convert to (builtin integer type)
1301
+ /// \param value half-precision value to convert
1302
+ /// \return floating-point value
1303
+ template<typename T> T half2float(unsigned int value)
1304
+ {
1305
+ return half2float_impl(value, T(), bool_type<std::numeric_limits<T>::is_iec559&&sizeof(typename bits<T>::type)==sizeof(T)>());
1306
+ }
1307
+
1308
+ /// Convert half-precision floating-point to integer.
1309
+ /// \tparam R rounding mode to use
1310
+ /// \tparam E `true` for round to even, `false` for round away from zero
1311
+ /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it
1312
+ /// \tparam T type to convert to (buitlin integer type with at least 16 bits precision, excluding any implicit sign bits)
1313
+ /// \param value half-precision value to convert
1314
+ /// \return rounded integer value
1315
+ /// \exception FE_INVALID if value is not representable in type \a T
1316
+ /// \exception FE_INEXACT if value had to be rounded and \a I is `true`
1317
+ template<std::float_round_style R,bool E,bool I,typename T> T half2int(unsigned int value)
1318
+ {
1319
+ unsigned int abs = value & 0x7FFF;
1320
+ if(abs >= 0x7C00)
1321
+ {
1322
+ raise(FE_INVALID);
1323
+ return (value&0x8000) ? std::numeric_limits<T>::min() : std::numeric_limits<T>::max();
1324
+ }
1325
+ if(abs < 0x3800)
1326
+ {
1327
+ raise(FE_INEXACT, I);
1328
+ return (R==std::round_toward_infinity) ? T(~(value>>15)&(abs!=0)) :
1329
+ (R==std::round_toward_neg_infinity) ? -T(value>0x8000) :
1330
+ T();
1331
+ }
1332
+ int exp = 25 - (abs>>10);
1333
+ unsigned int m = (value&0x3FF) | 0x400;
1334
+ int32 i = static_cast<int32>((exp<=0) ? (m<<-exp) : ((m+(
1335
+ (R==std::round_to_nearest) ? ((1<<(exp-1))-(~(m>>exp)&E)) :
1336
+ (R==std::round_toward_infinity) ? (((1<<exp)-1)&((value>>15)-1)) :
1337
+ (R==std::round_toward_neg_infinity) ? (((1<<exp)-1)&-(value>>15)) : 0))>>exp));
1338
+ if((!std::numeric_limits<T>::is_signed && (value&0x8000)) || (std::numeric_limits<T>::digits<16 &&
1339
+ ((value&0x8000) ? (-i<std::numeric_limits<T>::min()) : (i>std::numeric_limits<T>::max()))))
1340
+ raise(FE_INVALID);
1341
+ else if(I && exp > 0 && (m&((1<<exp)-1)))
1342
+ raise(FE_INEXACT);
1343
+ return static_cast<T>((value&0x8000) ? -i : i);
1344
+ }
1345
+
1346
+ /// \}
1347
+ /// \name Mathematics
1348
+ /// \{
1349
+
1350
+ /// upper part of 64-bit multiplication.
1351
+ /// \tparam R rounding mode to use
1352
+ /// \param x first factor
1353
+ /// \param y second factor
1354
+ /// \return upper 32 bit of \a x * \a y
1355
+ template<std::float_round_style R> uint32 mulhi(uint32 x, uint32 y)
1356
+ {
1357
+ uint32 xy = (x>>16) * (y&0xFFFF), yx = (x&0xFFFF) * (y>>16), c = (xy&0xFFFF) + (yx&0xFFFF) + (((x&0xFFFF)*(y&0xFFFF))>>16);
1358
+ return (x>>16)*(y>>16) + (xy>>16) + (yx>>16) + (c>>16) +
1359
+ ((R==std::round_to_nearest) ? ((c>>15)&1) : (R==std::round_toward_infinity) ? ((c&0xFFFF)!=0) : 0);
1360
+ }
1361
+
1362
+ /// 64-bit multiplication.
1363
+ /// \param x first factor
1364
+ /// \param y second factor
1365
+ /// \return upper 32 bit of \a x * \a y rounded to nearest
1366
+ inline uint32 multiply64(uint32 x, uint32 y)
1367
+ {
1368
+ #if HALF_ENABLE_CPP11_LONG_LONG
1369
+ return static_cast<uint32>((static_cast<unsigned long long>(x)*static_cast<unsigned long long>(y)+0x80000000)>>32);
1370
+ #else
1371
+ return mulhi<std::round_to_nearest>(x, y);
1372
+ #endif
1373
+ }
1374
+
1375
+ /// 64-bit division.
1376
+ /// \param x upper 32 bit of dividend
1377
+ /// \param y divisor
1378
+ /// \param s variable to store sticky bit for rounding
1379
+ /// \return (\a x << 32) / \a y
1380
+ inline uint32 divide64(uint32 x, uint32 y, int &s)
1381
+ {
1382
+ #if HALF_ENABLE_CPP11_LONG_LONG
1383
+ unsigned long long xx = static_cast<unsigned long long>(x) << 32;
1384
+ return s = (xx%y!=0), static_cast<uint32>(xx/y);
1385
+ #else
1386
+ y >>= 1;
1387
+ uint32 rem = x, div = 0;
1388
+ for(unsigned int i=0; i<32; ++i)
1389
+ {
1390
+ div <<= 1;
1391
+ if(rem >= y)
1392
+ {
1393
+ rem -= y;
1394
+ div |= 1;
1395
+ }
1396
+ rem <<= 1;
1397
+ }
1398
+ return s = rem > 1, div;
1399
+ #endif
1400
+ }
1401
+
1402
+ /// Half precision positive modulus.
1403
+ /// \tparam Q `true` to compute full quotient, `false` else
1404
+ /// \tparam R `true` to compute signed remainder, `false` for positive remainder
1405
+ /// \param x first operand as positive finite half-precision value
1406
+ /// \param y second operand as positive finite half-precision value
1407
+ /// \param quo adress to store quotient at, `nullptr` if \a Q `false`
1408
+ /// \return modulus of \a x / \a y
1409
+ template<bool Q,bool R> unsigned int mod(unsigned int x, unsigned int y, int *quo = NULL)
1410
+ {
1411
+ unsigned int q = 0;
1412
+ if(x > y)
1413
+ {
1414
+ int absx = x, absy = y, expx = 0, expy = 0;
1415
+ for(; absx<0x400; absx<<=1,--expx) ;
1416
+ for(; absy<0x400; absy<<=1,--expy) ;
1417
+ expx += absx >> 10;
1418
+ expy += absy >> 10;
1419
+ int mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
1420
+ for(int d=expx-expy; d; --d)
1421
+ {
1422
+ if(!Q && mx == my)
1423
+ return 0;
1424
+ if(mx >= my)
1425
+ {
1426
+ mx -= my;
1427
+ q += Q;
1428
+ }
1429
+ mx <<= 1;
1430
+ q <<= static_cast<int>(Q);
1431
+ }
1432
+ if(!Q && mx == my)
1433
+ return 0;
1434
+ if(mx >= my)
1435
+ {
1436
+ mx -= my;
1437
+ ++q;
1438
+ }
1439
+ if(Q)
1440
+ {
1441
+ q &= (1<<(std::numeric_limits<int>::digits-1)) - 1;
1442
+ if(!mx)
1443
+ return *quo = q, 0;
1444
+ }
1445
+ for(; mx<0x400; mx<<=1,--expy) ;
1446
+ x = (expy>0) ? ((expy<<10)|(mx&0x3FF)) : (mx>>(1-expy));
1447
+ }
1448
+ if(R)
1449
+ {
1450
+ unsigned int a, b;
1451
+ if(y < 0x800)
1452
+ {
1453
+ a = (x<0x400) ? (x<<1) : (x+0x400);
1454
+ b = y;
1455
+ }
1456
+ else
1457
+ {
1458
+ a = x;
1459
+ b = y - 0x400;
1460
+ }
1461
+ if(a > b || (a == b && (q&1)))
1462
+ {
1463
+ int exp = (y>>10) + (y<=0x3FF), d = exp - (x>>10) - (x<=0x3FF);
1464
+ int m = (((y&0x3FF)|((y>0x3FF)<<10))<<1) - (((x&0x3FF)|((x>0x3FF)<<10))<<(1-d));
1465
+ for(; m<0x800 && exp>1; m<<=1,--exp) ;
1466
+ x = 0x8000 + ((exp-1)<<10) + (m>>1);
1467
+ q += Q;
1468
+ }
1469
+ }
1470
+ if(Q)
1471
+ *quo = q;
1472
+ return x;
1473
+ }
1474
+
1475
+ /// Fixed point square root.
1476
+ /// \tparam F number of fractional bits
1477
+ /// \param r radicand in Q1.F fixed point format
1478
+ /// \param exp exponent
1479
+ /// \return square root as Q1.F/2
1480
+ template<unsigned int F> uint32 sqrt(uint32 &r, int &exp)
1481
+ {
1482
+ int i = exp & 1;
1483
+ r <<= i;
1484
+ exp = (exp-i) / 2;
1485
+ uint32 m = 0;
1486
+ for(uint32 bit=static_cast<uint32>(1)<<F; bit; bit>>=2)
1487
+ {
1488
+ if(r < m+bit)
1489
+ m >>= 1;
1490
+ else
1491
+ {
1492
+ r -= m + bit;
1493
+ m = (m>>1) + bit;
1494
+ }
1495
+ }
1496
+ return m;
1497
+ }
1498
+
1499
+ /// Fixed point binary exponential.
1500
+ /// This uses the BKM algorithm in E-mode.
1501
+ /// \param m exponent in [0,1) as Q0.31
1502
+ /// \param n number of iterations (at most 32)
1503
+ /// \return 2 ^ \a m as Q1.31
1504
+ inline uint32 exp2(uint32 m, unsigned int n = 32)
1505
+ {
1506
+ static const uint32 logs[] = {
1507
+ 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B,
1508
+ 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153,
1509
+ 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171,
1510
+ 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 };
1511
+ if(!m)
1512
+ return 0x80000000;
1513
+ uint32 mx = 0x80000000, my = 0;
1514
+ for(unsigned int i=1; i<n; ++i)
1515
+ {
1516
+ uint32 mz = my + logs[i];
1517
+ if(mz <= m)
1518
+ {
1519
+ my = mz;
1520
+ mx += mx >> i;
1521
+ }
1522
+ }
1523
+ return mx;
1524
+ }
1525
+
1526
+ /// Fixed point binary logarithm.
1527
+ /// This uses the BKM algorithm in L-mode.
1528
+ /// \param m mantissa in [1,2) as Q1.30
1529
+ /// \param n number of iterations (at most 32)
1530
+ /// \return log2(\a m) as Q0.31
1531
+ inline uint32 log2(uint32 m, unsigned int n = 32)
1532
+ {
1533
+ static const uint32 logs[] = {
1534
+ 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B,
1535
+ 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153,
1536
+ 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171,
1537
+ 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 };
1538
+ if(m == 0x40000000)
1539
+ return 0;
1540
+ uint32 mx = 0x40000000, my = 0;
1541
+ for(unsigned int i=1; i<n; ++i)
1542
+ {
1543
+ uint32 mz = mx + (mx>>i);
1544
+ if(mz <= m)
1545
+ {
1546
+ mx = mz;
1547
+ my += logs[i];
1548
+ }
1549
+ }
1550
+ return my;
1551
+ }
1552
+
1553
+ /// Fixed point sine and cosine.
1554
+ /// This uses the CORDIC algorithm in rotation mode.
1555
+ /// \param mz angle in [-pi/2,pi/2] as Q1.30
1556
+ /// \param n number of iterations (at most 31)
1557
+ /// \return sine and cosine of \a mz as Q1.30
1558
+ inline std::pair<uint32,uint32> sincos(uint32 mz, unsigned int n = 31)
1559
+ {
1560
+ static const uint32 angles[] = {
1561
+ 0x3243F6A9, 0x1DAC6705, 0x0FADBAFD, 0x07F56EA7, 0x03FEAB77, 0x01FFD55C, 0x00FFFAAB, 0x007FFF55,
1562
+ 0x003FFFEB, 0x001FFFFD, 0x00100000, 0x00080000, 0x00040000, 0x00020000, 0x00010000, 0x00008000,
1563
+ 0x00004000, 0x00002000, 0x00001000, 0x00000800, 0x00000400, 0x00000200, 0x00000100, 0x00000080,
1564
+ 0x00000040, 0x00000020, 0x00000010, 0x00000008, 0x00000004, 0x00000002, 0x00000001 };
1565
+ uint32 mx = 0x26DD3B6A, my = 0;
1566
+ for(unsigned int i=0; i<n; ++i)
1567
+ {
1568
+ uint32 sign = sign_mask(mz);
1569
+ uint32 tx = mx - (arithmetic_shift(my, i)^sign) + sign;
1570
+ uint32 ty = my + (arithmetic_shift(mx, i)^sign) - sign;
1571
+ mx = tx; my = ty; mz -= (angles[i]^sign) - sign;
1572
+ }
1573
+ return std::make_pair(my, mx);
1574
+ }
1575
+
1576
+ /// Fixed point arc tangent.
1577
+ /// This uses the CORDIC algorithm in vectoring mode.
1578
+ /// \param my y coordinate as Q0.30
1579
+ /// \param mx x coordinate as Q0.30
1580
+ /// \param n number of iterations (at most 31)
1581
+ /// \return arc tangent of \a my / \a mx as Q1.30
1582
+ inline uint32 atan2(uint32 my, uint32 mx, unsigned int n = 31)
1583
+ {
1584
+ static const uint32 angles[] = {
1585
+ 0x3243F6A9, 0x1DAC6705, 0x0FADBAFD, 0x07F56EA7, 0x03FEAB77, 0x01FFD55C, 0x00FFFAAB, 0x007FFF55,
1586
+ 0x003FFFEB, 0x001FFFFD, 0x00100000, 0x00080000, 0x00040000, 0x00020000, 0x00010000, 0x00008000,
1587
+ 0x00004000, 0x00002000, 0x00001000, 0x00000800, 0x00000400, 0x00000200, 0x00000100, 0x00000080,
1588
+ 0x00000040, 0x00000020, 0x00000010, 0x00000008, 0x00000004, 0x00000002, 0x00000001 };
1589
+ uint32 mz = 0;
1590
+ for(unsigned int i=0; i<n; ++i)
1591
+ {
1592
+ uint32 sign = sign_mask(my);
1593
+ uint32 tx = mx + (arithmetic_shift(my, i)^sign) - sign;
1594
+ uint32 ty = my - (arithmetic_shift(mx, i)^sign) + sign;
1595
+ mx = tx; my = ty; mz += (angles[i]^sign) - sign;
1596
+ }
1597
+ return mz;
1598
+ }
1599
+
1600
+ /// Reduce argument for trigonometric functions.
1601
+ /// \param abs half-precision floating-point value
1602
+ /// \param k value to take quarter period
1603
+ /// \return \a abs reduced to [-pi/4,pi/4] as Q0.30
1604
+ inline uint32 angle_arg(unsigned int abs, int &k)
1605
+ {
1606
+ uint32 m = (abs&0x3FF) | ((abs>0x3FF)<<10);
1607
+ int exp = (abs>>10) + (abs<=0x3FF) - 15;
1608
+ if(abs < 0x3A48)
1609
+ return k = 0, m << (exp+20);
1610
+ #if HALF_ENABLE_CPP11_LONG_LONG
1611
+ unsigned long long y = m * 0xA2F9836E4E442, mask = (1ULL<<(62-exp)) - 1, yi = (y+(mask>>1)) & ~mask, f = y - yi;
1612
+ uint32 sign = -static_cast<uint32>(f>>63);
1613
+ k = static_cast<int>(yi>>(62-exp));
1614
+ return (multiply64(static_cast<uint32>((sign ? -f : f)>>(31-exp)), 0xC90FDAA2)^sign) - sign;
1615
+ #else
1616
+ uint32 yh = m*0xA2F98 + mulhi<std::round_toward_zero>(m, 0x36E4E442), yl = (m*0x36E4E442) & 0xFFFFFFFF;
1617
+ uint32 mask = (static_cast<uint32>(1)<<(30-exp)) - 1, yi = (yh+(mask>>1)) & ~mask, sign = -static_cast<uint32>(yi>yh);
1618
+ k = static_cast<int>(yi>>(30-exp));
1619
+ uint32 fh = (yh^sign) + (yi^~sign) - ~sign, fl = (yl^sign) - sign;
1620
+ return (multiply64((exp>-1) ? (((fh<<(1+exp))&0xFFFFFFFF)|((fl&0xFFFFFFFF)>>(31-exp))) : fh, 0xC90FDAA2)^sign) - sign;
1621
+ #endif
1622
+ }
1623
+
1624
+ /// Get arguments for atan2 function.
1625
+ /// \param abs half-precision floating-point value
1626
+ /// \return \a abs and sqrt(1 - \a abs^2) as Q0.30
1627
+ inline std::pair<uint32,uint32> atan2_args(unsigned int abs)
1628
+ {
1629
+ int exp = -15;
1630
+ for(; abs<0x400; abs<<=1,--exp) ;
1631
+ exp += abs >> 10;
1632
+ uint32 my = ((abs&0x3FF)|0x400) << 5, r = my * my;
1633
+ int rexp = 2 * exp;
1634
+ r = 0x40000000 - ((rexp>-31) ? ((r>>-rexp)|((r&((static_cast<uint32>(1)<<-rexp)-1))!=0)) : 1);
1635
+ for(rexp=0; r<0x40000000; r<<=1,--rexp) ;
1636
+ uint32 mx = sqrt<30>(r, rexp);
1637
+ int d = exp - rexp;
1638
+ if(d < 0)
1639
+ return std::make_pair((d<-14) ? ((my>>(-d-14))+((my>>(-d-15))&1)) : (my<<(14+d)), (mx<<14)+(r<<13)/mx);
1640
+ if(d > 0)
1641
+ return std::make_pair(my<<14, (d>14) ? ((mx>>(d-14))+((mx>>(d-15))&1)) : ((d==14) ? mx : ((mx<<(14-d))+(r<<(13-d))/mx)));
1642
+ return std::make_pair(my<<13, (mx<<13)+(r<<12)/mx);
1643
+ }
1644
+
1645
+ /// Get exponentials for hyperbolic computation
1646
+ /// \param abs half-precision floating-point value
1647
+ /// \param exp variable to take unbiased exponent of larger result
1648
+ /// \param n number of BKM iterations (at most 32)
1649
+ /// \return exp(abs) and exp(-\a abs) as Q1.31 with same exponent
1650
+ inline std::pair<uint32,uint32> hyperbolic_args(unsigned int abs, int &exp, unsigned int n = 32)
1651
+ {
1652
+ uint32 mx = detail::multiply64(static_cast<uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29), my;
1653
+ int e = (abs>>10) + (abs<=0x3FF);
1654
+ if(e < 14)
1655
+ {
1656
+ exp = 0;
1657
+ mx >>= 14 - e;
1658
+ }
1659
+ else
1660
+ {
1661
+ exp = mx >> (45-e);
1662
+ mx = (mx<<(e-14)) & 0x7FFFFFFF;
1663
+ }
1664
+ mx = exp2(mx, n);
1665
+ int d = exp << 1, s;
1666
+ if(mx > 0x80000000)
1667
+ {
1668
+ my = divide64(0x80000000, mx, s);
1669
+ my |= s;
1670
+ ++d;
1671
+ }
1672
+ else
1673
+ my = mx;
1674
+ return std::make_pair(mx, (d<31) ? ((my>>d)|((my&((static_cast<uint32>(1)<<d)-1))!=0)) : 1);
1675
+ }
1676
+
1677
+ /// Postprocessing for binary exponential.
1678
+ /// \tparam R rounding mode to use
1679
+ /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results
1680
+ /// \param m mantissa as Q1.31
1681
+ /// \param exp absolute value of unbiased exponent
1682
+ /// \param esign sign of actual exponent
1683
+ /// \param sign sign bit of result
1684
+ /// \return value converted to half-precision
1685
+ /// \exception FE_OVERFLOW on overflows
1686
+ /// \exception FE_UNDERFLOW on underflows
1687
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
1688
+ template<std::float_round_style R,bool I> unsigned int exp2_post(uint32 m, int exp, bool esign, unsigned int sign = 0)
1689
+ {
1690
+ int s = 0;
1691
+ if(esign)
1692
+ {
1693
+ if(m > 0x80000000)
1694
+ {
1695
+ m = divide64(0x80000000, m, s);
1696
+ ++exp;
1697
+ }
1698
+ if(exp > 25)
1699
+ return underflow<R>(sign);
1700
+ else if(exp == 25)
1701
+ return rounded<R,I>(sign, 1, (m&0x7FFFFFFF)!=0);
1702
+ exp = -exp;
1703
+ }
1704
+ else if(exp > 15)
1705
+ return overflow<R>(sign);
1706
+ return fixed2half<R,31,false,false,I>(m, exp+14, sign, s);
1707
+ }
1708
+
1709
+ /// Postprocessing for binary logarithm.
1710
+ /// \tparam R rounding mode to use
1711
+ /// \tparam L logarithm for base transformation as Q1.31
1712
+ /// \param m fractional part of logarithm as Q0.31
1713
+ /// \param ilog signed integer part of logarithm
1714
+ /// \param exp biased exponent of result
1715
+ /// \param sign sign bit of result
1716
+ /// \return value base-transformed and converted to half-precision
1717
+ /// \exception FE_OVERFLOW on overflows
1718
+ /// \exception FE_UNDERFLOW on underflows
1719
+ /// \exception FE_INEXACT if no other exception occurred
1720
+ template<std::float_round_style R,uint32 L> unsigned int log2_post(uint32 m, int ilog, int exp, unsigned int sign = 0)
1721
+ {
1722
+ uint32 msign = sign_mask(ilog);
1723
+ m = (((static_cast<uint32>(ilog)<<27)+(m>>4))^msign) - msign;
1724
+ if(!m)
1725
+ return 0;
1726
+ for(; m<0x80000000; m<<=1,--exp) ;
1727
+ int i = m >= L, s;
1728
+ exp += i;
1729
+ m >>= 1 + i;
1730
+ sign ^= msign & 0x8000;
1731
+ if(exp < -11)
1732
+ return underflow<R>(sign);
1733
+ m = divide64(m, L, s);
1734
+ return fixed2half<R,30,false,false,true>(m, exp, sign, 1);
1735
+ }
1736
+
1737
+ /// Hypotenuse square root and postprocessing.
1738
+ /// \tparam R rounding mode to use
1739
+ /// \param r mantissa as Q2.30
1740
+ /// \param exp unbiased exponent
1741
+ /// \return square root converted to half-precision
1742
+ /// \exception FE_OVERFLOW on overflows
1743
+ /// \exception FE_UNDERFLOW on underflows
1744
+ /// \exception FE_INEXACT if value had to be rounded
1745
+ template<std::float_round_style R> unsigned int hypot_post(uint32 r, int exp)
1746
+ {
1747
+ int i = r >> 31;
1748
+ if((exp+=i) > 46)
1749
+ return overflow<R>();
1750
+ if(exp < -34)
1751
+ return underflow<R>();
1752
+ r = (r>>i) | (r&i);
1753
+ uint32 m = sqrt<30>(r, exp+=15);
1754
+ return fixed2half<R,15,false,false,false>(m, exp-1, 0, r!=0);
1755
+ }
1756
+
1757
+ /// Division and postprocessing for tangents.
1758
+ /// \tparam R rounding mode to use
1759
+ /// \param my dividend as Q1.31
1760
+ /// \param mx divisor as Q1.31
1761
+ /// \param exp biased exponent of result
1762
+ /// \param sign sign bit of result
1763
+ /// \return quotient converted to half-precision
1764
+ /// \exception FE_OVERFLOW on overflows
1765
+ /// \exception FE_UNDERFLOW on underflows
1766
+ /// \exception FE_INEXACT if no other exception occurred
1767
+ template<std::float_round_style R> unsigned int tangent_post(uint32 my, uint32 mx, int exp, unsigned int sign = 0)
1768
+ {
1769
+ int i = my >= mx, s;
1770
+ exp += i;
1771
+ if(exp > 29)
1772
+ return overflow<R>(sign);
1773
+ if(exp < -11)
1774
+ return underflow<R>(sign);
1775
+ uint32 m = divide64(my>>(i+1), mx, s);
1776
+ return fixed2half<R,30,false,false,true>(m, exp, sign, s);
1777
+ }
1778
+
1779
+ /// Area function and postprocessing.
1780
+ /// This computes the value directly in Q2.30 using the representation `asinh|acosh(x) = log(x+sqrt(x^2+|-1))`.
1781
+ /// \tparam R rounding mode to use
1782
+ /// \tparam S `true` for asinh, `false` for acosh
1783
+ /// \param arg half-precision argument
1784
+ /// \return asinh|acosh(\a arg) converted to half-precision
1785
+ /// \exception FE_OVERFLOW on overflows
1786
+ /// \exception FE_UNDERFLOW on underflows
1787
+ /// \exception FE_INEXACT if no other exception occurred
1788
+ template<std::float_round_style R,bool S> unsigned int area(unsigned int arg)
1789
+ {
1790
+ int abs = arg & 0x7FFF, expx = (abs>>10) + (abs<=0x3FF) - 15, expy = -15, ilog, i;
1791
+ uint32 mx = static_cast<uint32>((abs&0x3FF)|((abs>0x3FF)<<10)) << 20, my, r;
1792
+ for(; abs<0x400; abs<<=1,--expy) ;
1793
+ expy += abs >> 10;
1794
+ r = ((abs&0x3FF)|0x400) << 5;
1795
+ r *= r;
1796
+ i = r >> 31;
1797
+ expy = 2*expy + i;
1798
+ r >>= i;
1799
+ if(S)
1800
+ {
1801
+ if(expy < 0)
1802
+ {
1803
+ r = 0x40000000 + ((expy>-30) ? ((r>>-expy)|((r&((static_cast<uint32>(1)<<-expy)-1))!=0)) : 1);
1804
+ expy = 0;
1805
+ }
1806
+ else
1807
+ {
1808
+ r += 0x40000000 >> expy;
1809
+ i = r >> 31;
1810
+ r = (r>>i) | (r&i);
1811
+ expy += i;
1812
+ }
1813
+ }
1814
+ else
1815
+ {
1816
+ r -= 0x40000000 >> expy;
1817
+ for(; r<0x40000000; r<<=1,--expy) ;
1818
+ }
1819
+ my = sqrt<30>(r, expy);
1820
+ my = (my<<15) + (r<<14)/my;
1821
+ if(S)
1822
+ {
1823
+ mx >>= expy - expx;
1824
+ ilog = expy;
1825
+ }
1826
+ else
1827
+ {
1828
+ my >>= expx - expy;
1829
+ ilog = expx;
1830
+ }
1831
+ my += mx;
1832
+ i = my >> 31;
1833
+ static const int G = S && (R==std::round_to_nearest);
1834
+ return log2_post<R,0xB8AA3B2A>(log2(my>>i, 26+S+G)+(G<<3), ilog+i, 17, arg&(static_cast<unsigned>(S)<<15));
1835
+ }
1836
+
1837
+ /// Class for 1.31 unsigned floating-point computation
1838
+ struct f31
1839
+ {
1840
+ /// Constructor.
1841
+ /// \param mant mantissa as 1.31
1842
+ /// \param e exponent
1843
+ HALF_CONSTEXPR f31(uint32 mant, int e) : m(mant), exp(e) {}
1844
+
1845
+ /// Constructor.
1846
+ /// \param abs unsigned half-precision value
1847
+ f31(unsigned int abs) : exp(-15)
1848
+ {
1849
+ for(; abs<0x400; abs<<=1,--exp) ;
1850
+ m = static_cast<uint32>((abs&0x3FF)|0x400) << 21;
1851
+ exp += (abs>>10);
1852
+ }
1853
+
1854
+ /// Addition operator.
1855
+ /// \param a first operand
1856
+ /// \param b second operand
1857
+ /// \return \a a + \a b
1858
+ friend f31 operator+(f31 a, f31 b)
1859
+ {
1860
+ if(b.exp > a.exp)
1861
+ std::swap(a, b);
1862
+ int d = a.exp - b.exp;
1863
+ uint32 m = a.m + ((d<32) ? (b.m>>d) : 0);
1864
+ int i = (m&0xFFFFFFFF) < a.m;
1865
+ return f31(((m+i)>>i)|0x80000000, a.exp+i);
1866
+ }
1867
+
1868
+ /// Subtraction operator.
1869
+ /// \param a first operand
1870
+ /// \param b second operand
1871
+ /// \return \a a - \a b
1872
+ friend f31 operator-(f31 a, f31 b)
1873
+ {
1874
+ int d = a.exp - b.exp, exp = a.exp;
1875
+ uint32 m = a.m - ((d<32) ? (b.m>>d) : 0);
1876
+ if(!m)
1877
+ return f31(0, -32);
1878
+ for(; m<0x80000000; m<<=1,--exp) ;
1879
+ return f31(m, exp);
1880
+ }
1881
+
1882
+ /// Multiplication operator.
1883
+ /// \param a first operand
1884
+ /// \param b second operand
1885
+ /// \return \a a * \a b
1886
+ friend f31 operator*(f31 a, f31 b)
1887
+ {
1888
+ uint32 m = multiply64(a.m, b.m);
1889
+ int i = m >> 31;
1890
+ return f31(m<<(1-i), a.exp + b.exp + i);
1891
+ }
1892
+
1893
+ /// Division operator.
1894
+ /// \param a first operand
1895
+ /// \param b second operand
1896
+ /// \return \a a / \a b
1897
+ friend f31 operator/(f31 a, f31 b)
1898
+ {
1899
+ int i = a.m >= b.m, s;
1900
+ uint32 m = divide64((a.m+i)>>i, b.m, s);
1901
+ return f31(m, a.exp - b.exp + i - 1);
1902
+ }
1903
+
1904
+ uint32 m; ///< mantissa as 1.31.
1905
+ int exp; ///< exponent.
1906
+ };
1907
+
1908
+ /// Error function and postprocessing.
1909
+ /// This computes the value directly in Q1.31 using the approximations given
1910
+ /// [here](https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions).
1911
+ /// \tparam R rounding mode to use
1912
+ /// \tparam C `true` for comlementary error function, `false` else
1913
+ /// \param arg half-precision function argument
1914
+ /// \return approximated value of error function in half-precision
1915
+ /// \exception FE_OVERFLOW on overflows
1916
+ /// \exception FE_UNDERFLOW on underflows
1917
+ /// \exception FE_INEXACT if no other exception occurred
1918
+ template<std::float_round_style R,bool C> unsigned int erf(unsigned int arg)
1919
+ {
1920
+ unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000;
1921
+ f31 x(abs), x2 = x * x * f31(0xB8AA3B29, 0), t = f31(0x80000000, 0) / (f31(0x80000000, 0)+f31(0xA7BA054A, -2)*x), t2 = t * t;
1922
+ f31 e = ((f31(0x87DC2213, 0)*t2+f31(0xB5F0E2AE, 0))*t2+f31(0x82790637, -2)-(f31(0xBA00E2B8, 0)*t2+f31(0x91A98E62, -2))*t) * t /
1923
+ ((x2.exp<0) ? f31(exp2((x2.exp>-32) ? (x2.m>>-x2.exp) : 0, 30), 0) : f31(exp2((x2.m<<x2.exp)&0x7FFFFFFF, 22), x2.m>>(31-x2.exp)));
1924
+ return (!C || sign) ? fixed2half<R,31,false,true,true>(0x80000000-(e.m>>(C-e.exp)), 14+C, sign&(C-1U)) :
1925
+ (e.exp<-25) ? underflow<R>() : fixed2half<R,30,false,false,true>(e.m>>1, e.exp+14, 0, e.m&1);
1926
+ }
1927
+
1928
+ /// Gamma function and postprocessing.
1929
+ /// This approximates the value of either the gamma function or its logarithm directly in Q1.31.
1930
+ /// \tparam R rounding mode to use
1931
+ /// \tparam L `true` for lograithm of gamma function, `false` for gamma function
1932
+ /// \param arg half-precision floating-point value
1933
+ /// \return lgamma/tgamma(\a arg) in half-precision
1934
+ /// \exception FE_OVERFLOW on overflows
1935
+ /// \exception FE_UNDERFLOW on underflows
1936
+ /// \exception FE_INEXACT if \a arg is not a positive integer
1937
+ template<std::float_round_style R,bool L> unsigned int gamma(unsigned int arg)
1938
+ {
1939
+ /* static const double p[] ={ 2.50662827563479526904, 225.525584619175212544, -268.295973841304927459, 80.9030806934622512966, -5.00757863970517583837, 0.0114684895434781459556 };
1940
+ double t = arg + 4.65, s = p[0];
1941
+ for(unsigned int i=0; i<5; ++i)
1942
+ s += p[i+1] / (arg+i);
1943
+ return std::log(s) + (arg-0.5)*std::log(t) - t;
1944
+ */ static const f31 pi(0xC90FDAA2, 1), lbe(0xB8AA3B29, 0);
1945
+ unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000;
1946
+ bool bsign = sign != 0;
1947
+ f31 z(abs), x = sign ? (z+f31(0x80000000, 0)) : z, t = x + f31(0x94CCCCCD, 2), s =
1948
+ f31(0xA06C9901, 1) + f31(0xBBE654E2, -7)/(x+f31(0x80000000, 2)) + f31(0xA1CE6098, 6)/(x+f31(0x80000000, 1))
1949
+ + f31(0xE1868CB7, 7)/x - f31(0x8625E279, 8)/(x+f31(0x80000000, 0)) - f31(0xA03E158F, 2)/(x+f31(0xC0000000, 1));
1950
+ int i = (s.exp>=2) + (s.exp>=4) + (s.exp>=8) + (s.exp>=16);
1951
+ s = f31((static_cast<uint32>(s.exp)<<(31-i))+(log2(s.m>>1, 28)>>i), i) / lbe;
1952
+ if(x.exp != -1 || x.m != 0x80000000)
1953
+ {
1954
+ i = (t.exp>=2) + (t.exp>=4) + (t.exp>=8);
1955
+ f31 l = f31((static_cast<uint32>(t.exp)<<(31-i))+(log2(t.m>>1, 30)>>i), i) / lbe;
1956
+ s = (x.exp<-1) ? (s-(f31(0x80000000, -1)-x)*l) : (s+(x-f31(0x80000000, -1))*l);
1957
+ }
1958
+ s = x.exp ? (s-t) : (t-s);
1959
+ if(bsign)
1960
+ {
1961
+ if(z.exp >= 0)
1962
+ {
1963
+ sign &= (L|((z.m>>(31-z.exp))&1)) - 1;
1964
+ for(z=f31((z.m<<(1+z.exp))&0xFFFFFFFF, -1); z.m<0x80000000; z.m<<=1,--z.exp) ;
1965
+ }
1966
+ if(z.exp == -1)
1967
+ z = f31(0x80000000, 0) - z;
1968
+ if(z.exp < -1)
1969
+ {
1970
+ z = z * pi;
1971
+ z.m = sincos(z.m>>(1-z.exp), 30).first;
1972
+ for(z.exp=1; z.m<0x80000000; z.m<<=1,--z.exp) ;
1973
+ }
1974
+ else
1975
+ z = f31(0x80000000, 0);
1976
+ }
1977
+ if(L)
1978
+ {
1979
+ if(bsign)
1980
+ {
1981
+ f31 l(0x92868247, 0);
1982
+ if(z.exp < 0)
1983
+ {
1984
+ uint32 m = log2((z.m+1)>>1, 27);
1985
+ z = f31(-((static_cast<uint32>(z.exp)<<26)+(m>>5)), 5);
1986
+ for(; z.m<0x80000000; z.m<<=1,--z.exp) ;
1987
+ l = l + z / lbe;
1988
+ }
1989
+ sign = static_cast<unsigned>(x.exp&&(l.exp<s.exp||(l.exp==s.exp&&l.m<s.m))) << 15;
1990
+ s = sign ? (s-l) : x.exp ? (l-s) : (l+s);
1991
+ }
1992
+ else
1993
+ {
1994
+ sign = static_cast<unsigned>(x.exp==0) << 15;
1995
+ if(s.exp < -24)
1996
+ return underflow<R>(sign);
1997
+ if(s.exp > 15)
1998
+ return overflow<R>(sign);
1999
+ }
2000
+ }
2001
+ else
2002
+ {
2003
+ s = s * lbe;
2004
+ uint32 m;
2005
+ if(s.exp < 0)
2006
+ {
2007
+ m = s.m >> -s.exp;
2008
+ s.exp = 0;
2009
+ }
2010
+ else
2011
+ {
2012
+ m = (s.m<<s.exp) & 0x7FFFFFFF;
2013
+ s.exp = (s.m>>(31-s.exp));
2014
+ }
2015
+ s.m = exp2(m, 27);
2016
+ if(!x.exp)
2017
+ s = f31(0x80000000, 0) / s;
2018
+ if(bsign)
2019
+ {
2020
+ if(z.exp < 0)
2021
+ s = s * z;
2022
+ s = pi / s;
2023
+ if(s.exp < -24)
2024
+ return underflow<R>(sign);
2025
+ }
2026
+ else if(z.exp > 0 && !(z.m&((1<<(31-z.exp))-1)))
2027
+ return ((s.exp+14)<<10) + (s.m>>21);
2028
+ if(s.exp > 15)
2029
+ return overflow<R>(sign);
2030
+ }
2031
+ return fixed2half<R,31,false,false,true>(s.m, s.exp+14, sign);
2032
+ }
2033
+ /// \}
2034
+
2035
+ template<typename,typename,std::float_round_style> struct half_caster;
2036
+ }
2037
+
2038
+ /// Half-precision floating-point type.
2039
+ /// This class implements an IEEE-conformant half-precision floating-point type with the usual arithmetic
2040
+ /// operators and conversions. It is implicitly convertible to single-precision floating-point, which makes artihmetic
2041
+ /// expressions and functions with mixed-type operands to be of the most precise operand type.
2042
+ ///
2043
+ /// According to the C++98/03 definition, the half type is not a POD type. But according to C++11's less strict and
2044
+ /// extended definitions it is both a standard layout type and a trivially copyable type (even if not a POD type), which
2045
+ /// means it can be standard-conformantly copied using raw binary copies. But in this context some more words about the
2046
+ /// actual size of the type. Although the half is representing an IEEE 16-bit type, it does not neccessarily have to be of
2047
+ /// exactly 16-bits size. But on any reasonable implementation the actual binary representation of this type will most
2048
+ /// probably not ivolve any additional "magic" or padding beyond the simple binary representation of the underlying 16-bit
2049
+ /// IEEE number, even if not strictly guaranteed by the standard. But even then it only has an actual size of 16 bits if
2050
+ /// your C++ implementation supports an unsigned integer type of exactly 16 bits width. But this should be the case on
2051
+ /// nearly any reasonable platform.
2052
+ ///
2053
+ /// So if your C++ implementation is not totally exotic or imposes special alignment requirements, it is a reasonable
2054
+ /// assumption that the data of a half is just comprised of the 2 bytes of the underlying IEEE representation.
2055
+ class half
2056
+ {
2057
+ public:
2058
+ /// \name Construction and assignment
2059
+ /// \{
2060
+
2061
+ /// Default constructor.
2062
+ /// This initializes the half to 0. Although this does not match the builtin types' default-initialization semantics
2063
+ /// and may be less efficient than no initialization, it is needed to provide proper value-initialization semantics.
2064
+ HALF_CONSTEXPR half() HALF_NOEXCEPT : data_() {}
2065
+
2066
+ /// Conversion constructor.
2067
+ /// \param rhs float to convert
2068
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2069
+ explicit half(float rhs) : data_(static_cast<detail::uint16>(detail::float2half<round_style>(rhs))) {}
2070
+
2071
+ /// Conversion to single-precision.
2072
+ /// \return single precision value representing expression value
2073
+ operator float() const { return detail::half2float<float>(data_); }
2074
+
2075
+ /// Assignment operator.
2076
+ /// \param rhs single-precision value to copy from
2077
+ /// \return reference to this half
2078
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2079
+ half& operator=(float rhs) { data_ = static_cast<detail::uint16>(detail::float2half<round_style>(rhs)); return *this; }
2080
+
2081
+ /// \}
2082
+ /// \name Arithmetic updates
2083
+ /// \{
2084
+
2085
+ /// Arithmetic assignment.
2086
+ /// \tparam T type of concrete half expression
2087
+ /// \param rhs half expression to add
2088
+ /// \return reference to this half
2089
+ /// \exception FE_... according to operator+(half,half)
2090
+ half& operator+=(half rhs) { return *this = *this + rhs; }
2091
+
2092
+ /// Arithmetic assignment.
2093
+ /// \tparam T type of concrete half expression
2094
+ /// \param rhs half expression to subtract
2095
+ /// \return reference to this half
2096
+ /// \exception FE_... according to operator-(half,half)
2097
+ half& operator-=(half rhs) { return *this = *this - rhs; }
2098
+
2099
+ /// Arithmetic assignment.
2100
+ /// \tparam T type of concrete half expression
2101
+ /// \param rhs half expression to multiply with
2102
+ /// \return reference to this half
2103
+ /// \exception FE_... according to operator*(half,half)
2104
+ half& operator*=(half rhs) { return *this = *this * rhs; }
2105
+
2106
+ /// Arithmetic assignment.
2107
+ /// \tparam T type of concrete half expression
2108
+ /// \param rhs half expression to divide by
2109
+ /// \return reference to this half
2110
+ /// \exception FE_... according to operator/(half,half)
2111
+ half& operator/=(half rhs) { return *this = *this / rhs; }
2112
+
2113
+ /// Arithmetic assignment.
2114
+ /// \param rhs single-precision value to add
2115
+ /// \return reference to this half
2116
+ /// \exception FE_... according to operator=()
2117
+ half& operator+=(float rhs) { return *this = *this + rhs; }
2118
+
2119
+ /// Arithmetic assignment.
2120
+ /// \param rhs single-precision value to subtract
2121
+ /// \return reference to this half
2122
+ /// \exception FE_... according to operator=()
2123
+ half& operator-=(float rhs) { return *this = *this - rhs; }
2124
+
2125
+ /// Arithmetic assignment.
2126
+ /// \param rhs single-precision value to multiply with
2127
+ /// \return reference to this half
2128
+ /// \exception FE_... according to operator=()
2129
+ half& operator*=(float rhs) { return *this = *this * rhs; }
2130
+
2131
+ /// Arithmetic assignment.
2132
+ /// \param rhs single-precision value to divide by
2133
+ /// \return reference to this half
2134
+ /// \exception FE_... according to operator=()
2135
+ half& operator/=(float rhs) { return *this = *this / rhs; }
2136
+
2137
+ /// \}
2138
+ /// \name Increment and decrement
2139
+ /// \{
2140
+
2141
+ /// Prefix increment.
2142
+ /// \return incremented half value
2143
+ /// \exception FE_... according to operator+(half,half)
2144
+ half& operator++() { return *this = *this + half(detail::binary, 0x3C00); }
2145
+
2146
+ /// Prefix decrement.
2147
+ /// \return decremented half value
2148
+ /// \exception FE_... according to operator-(half,half)
2149
+ half& operator--() { return *this = *this + half(detail::binary, 0xBC00); }
2150
+
2151
+ /// Postfix increment.
2152
+ /// \return non-incremented half value
2153
+ /// \exception FE_... according to operator+(half,half)
2154
+ half operator++(int) { half out(*this); ++*this; return out; }
2155
+
2156
+ /// Postfix decrement.
2157
+ /// \return non-decremented half value
2158
+ /// \exception FE_... according to operator-(half,half)
2159
+ half operator--(int) { half out(*this); --*this; return out; }
2160
+ /// \}
2161
+
2162
+ private:
2163
+ /// Rounding mode to use
2164
+ static const std::float_round_style round_style = (std::float_round_style)(HALF_ROUND_STYLE);
2165
+
2166
+ /// Constructor.
2167
+ /// \param bits binary representation to set half to
2168
+ HALF_CONSTEXPR half(detail::binary_t, unsigned int bits) HALF_NOEXCEPT : data_(static_cast<detail::uint16>(bits)) {}
2169
+
2170
+ /// Internal binary representation
2171
+ detail::uint16 data_;
2172
+
2173
+ #ifndef HALF_DOXYGEN_ONLY
2174
+ friend HALF_CONSTEXPR_NOERR bool operator==(half, half);
2175
+ friend HALF_CONSTEXPR_NOERR bool operator!=(half, half);
2176
+ friend HALF_CONSTEXPR_NOERR bool operator<(half, half);
2177
+ friend HALF_CONSTEXPR_NOERR bool operator>(half, half);
2178
+ friend HALF_CONSTEXPR_NOERR bool operator<=(half, half);
2179
+ friend HALF_CONSTEXPR_NOERR bool operator>=(half, half);
2180
+ friend HALF_CONSTEXPR half operator-(half);
2181
+ friend half operator+(half, half);
2182
+ friend half operator-(half, half);
2183
+ friend half operator*(half, half);
2184
+ friend half operator/(half, half);
2185
+ template<typename charT,typename traits> friend std::basic_ostream<charT,traits>& operator<<(std::basic_ostream<charT,traits>&, half);
2186
+ template<typename charT,typename traits> friend std::basic_istream<charT,traits>& operator>>(std::basic_istream<charT,traits>&, half&);
2187
+ friend HALF_CONSTEXPR half fabs(half);
2188
+ friend half fmod(half, half);
2189
+ friend half remainder(half, half);
2190
+ friend half remquo(half, half, int*);
2191
+ friend half fma(half, half, half);
2192
+ friend HALF_CONSTEXPR_NOERR half fmax(half, half);
2193
+ friend HALF_CONSTEXPR_NOERR half fmin(half, half);
2194
+ friend half fdim(half, half);
2195
+ friend half nanh(const char*);
2196
+ friend half exp(half);
2197
+ friend half exp2(half);
2198
+ friend half expm1(half);
2199
+ friend half log(half);
2200
+ friend half log10(half);
2201
+ friend half log2(half);
2202
+ friend half log1p(half);
2203
+ friend half sqrt(half);
2204
+ friend half cbrt(half);
2205
+ friend half hypot(half, half);
2206
+ friend half hypot(half, half, half);
2207
+ friend half pow(half, half);
2208
+ friend void sincos(half, half*, half*);
2209
+ friend half sin(half);
2210
+ friend half cos(half);
2211
+ friend half tan(half);
2212
+ friend half asin(half);
2213
+ friend half acos(half);
2214
+ friend half atan(half);
2215
+ friend half atan2(half, half);
2216
+ friend half sinh(half);
2217
+ friend half cosh(half);
2218
+ friend half tanh(half);
2219
+ friend half asinh(half);
2220
+ friend half acosh(half);
2221
+ friend half atanh(half);
2222
+ friend half erf(half);
2223
+ friend half erfc(half);
2224
+ friend half lgamma(half);
2225
+ friend half tgamma(half);
2226
+ friend half ceil(half);
2227
+ friend half floor(half);
2228
+ friend half trunc(half);
2229
+ friend half round(half);
2230
+ friend long lround(half);
2231
+ friend half rint(half);
2232
+ friend long lrint(half);
2233
+ friend half nearbyint(half);
2234
+ #ifdef HALF_ENABLE_CPP11_LONG_LONG
2235
+ friend long long llround(half);
2236
+ friend long long llrint(half);
2237
+ #endif
2238
+ friend half frexp(half, int*);
2239
+ friend half scalbln(half, long);
2240
+ friend half modf(half, half*);
2241
+ friend int ilogb(half);
2242
+ friend half logb(half);
2243
+ friend half nextafter(half, half);
2244
+ friend half nexttoward(half, long double);
2245
+ friend HALF_CONSTEXPR half copysign(half, half);
2246
+ friend HALF_CONSTEXPR int fpclassify(half);
2247
+ friend HALF_CONSTEXPR bool isfinite(half);
2248
+ friend HALF_CONSTEXPR bool isinf(half);
2249
+ friend HALF_CONSTEXPR bool isnan(half);
2250
+ friend HALF_CONSTEXPR bool isnormal(half);
2251
+ friend HALF_CONSTEXPR bool signbit(half);
2252
+ friend HALF_CONSTEXPR bool isgreater(half, half);
2253
+ friend HALF_CONSTEXPR bool isgreaterequal(half, half);
2254
+ friend HALF_CONSTEXPR bool isless(half, half);
2255
+ friend HALF_CONSTEXPR bool islessequal(half, half);
2256
+ friend HALF_CONSTEXPR bool islessgreater(half, half);
2257
+ template<typename,typename,std::float_round_style> friend struct detail::half_caster;
2258
+ friend class std::numeric_limits<half>;
2259
+ #if HALF_ENABLE_CPP11_HASH
2260
+ friend struct std::hash<half>;
2261
+ #endif
2262
+ #if HALF_ENABLE_CPP11_USER_LITERALS
2263
+ friend half literal::operator "" _h(long double);
2264
+ #endif
2265
+ #endif
2266
+ };
2267
+
2268
+ #if HALF_ENABLE_CPP11_USER_LITERALS
2269
+ namespace literal
2270
+ {
2271
+ /// Half literal.
2272
+ /// While this returns a properly rounded half-precision value, half literals can unfortunately not be constant
2273
+ /// expressions due to rather involved conversions. So don't expect this to be a literal literal without involving
2274
+ /// conversion operations at runtime. It is a convenience feature, not a performance optimization.
2275
+ /// \param value literal value
2276
+ /// \return half with of given value (possibly rounded)
2277
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2278
+ inline half operator "" _h(long double value) { return half(detail::binary, detail::float2half<half::round_style>(value)); }
2279
+ }
2280
+ #endif
2281
+
2282
+ namespace detail
2283
+ {
2284
+ /// Helper class for half casts.
2285
+ /// This class template has to be specialized for all valid cast arguments to define an appropriate static
2286
+ /// `cast` member function and a corresponding `type` member denoting its return type.
2287
+ /// \tparam T destination type
2288
+ /// \tparam U source type
2289
+ /// \tparam R rounding mode to use
2290
+ template<typename T,typename U,std::float_round_style R=(std::float_round_style)(HALF_ROUND_STYLE)> struct half_caster {};
2291
+ template<typename U,std::float_round_style R> struct half_caster<half,U,R>
2292
+ {
2293
+ #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS
2294
+ static_assert(std::is_arithmetic<U>::value, "half_cast from non-arithmetic type unsupported");
2295
+ #endif
2296
+
2297
+ static half cast(U arg) { return cast_impl(arg, is_float<U>()); };
2298
+
2299
+ private:
2300
+ static half cast_impl(U arg, true_type) { return half(binary, float2half<R>(arg)); }
2301
+ static half cast_impl(U arg, false_type) { return half(binary, int2half<R>(arg)); }
2302
+ };
2303
+ template<typename T,std::float_round_style R> struct half_caster<T,half,R>
2304
+ {
2305
+ #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS
2306
+ static_assert(std::is_arithmetic<T>::value, "half_cast to non-arithmetic type unsupported");
2307
+ #endif
2308
+
2309
+ static T cast(half arg) { return cast_impl(arg, is_float<T>()); }
2310
+
2311
+ private:
2312
+ static T cast_impl(half arg, true_type) { return half2float<T>(arg.data_); }
2313
+ static T cast_impl(half arg, false_type) { return half2int<R,true,true,T>(arg.data_); }
2314
+ };
2315
+ template<std::float_round_style R> struct half_caster<half,half,R>
2316
+ {
2317
+ static half cast(half arg) { return arg; }
2318
+ };
2319
+ }
2320
+ }
2321
+
2322
+ /// Extensions to the C++ standard library.
2323
+ namespace std
2324
+ {
2325
+ /// Numeric limits for half-precision floats.
2326
+ /// **See also:** Documentation for [std::numeric_limits](https://en.cppreference.com/w/cpp/types/numeric_limits)
2327
+ template<> class numeric_limits<half_float::half>
2328
+ {
2329
+ public:
2330
+ /// Is template specialization.
2331
+ static HALF_CONSTEXPR_CONST bool is_specialized = true;
2332
+
2333
+ /// Supports signed values.
2334
+ static HALF_CONSTEXPR_CONST bool is_signed = true;
2335
+
2336
+ /// Is not an integer type.
2337
+ static HALF_CONSTEXPR_CONST bool is_integer = false;
2338
+
2339
+ /// Is not exact.
2340
+ static HALF_CONSTEXPR_CONST bool is_exact = false;
2341
+
2342
+ /// Doesn't provide modulo arithmetic.
2343
+ static HALF_CONSTEXPR_CONST bool is_modulo = false;
2344
+
2345
+ /// Has a finite set of values.
2346
+ static HALF_CONSTEXPR_CONST bool is_bounded = true;
2347
+
2348
+ /// IEEE conformant.
2349
+ static HALF_CONSTEXPR_CONST bool is_iec559 = true;
2350
+
2351
+ /// Supports infinity.
2352
+ static HALF_CONSTEXPR_CONST bool has_infinity = true;
2353
+
2354
+ /// Supports quiet NaNs.
2355
+ static HALF_CONSTEXPR_CONST bool has_quiet_NaN = true;
2356
+
2357
+ /// Supports signaling NaNs.
2358
+ static HALF_CONSTEXPR_CONST bool has_signaling_NaN = true;
2359
+
2360
+ /// Supports subnormal values.
2361
+ static HALF_CONSTEXPR_CONST float_denorm_style has_denorm = denorm_present;
2362
+
2363
+ /// Supports no denormalization detection.
2364
+ static HALF_CONSTEXPR_CONST bool has_denorm_loss = false;
2365
+
2366
+ #if HALF_ERRHANDLING_THROWS
2367
+ static HALF_CONSTEXPR_CONST bool traps = true;
2368
+ #else
2369
+ /// Traps only if [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID) is acitvated.
2370
+ static HALF_CONSTEXPR_CONST bool traps = false;
2371
+ #endif
2372
+
2373
+ /// Does not support no pre-rounding underflow detection.
2374
+ static HALF_CONSTEXPR_CONST bool tinyness_before = false;
2375
+
2376
+ /// Rounding mode.
2377
+ static HALF_CONSTEXPR_CONST float_round_style round_style = half_float::half::round_style;
2378
+
2379
+ /// Significant digits.
2380
+ static HALF_CONSTEXPR_CONST int digits = 11;
2381
+
2382
+ /// Significant decimal digits.
2383
+ static HALF_CONSTEXPR_CONST int digits10 = 3;
2384
+
2385
+ /// Required decimal digits to represent all possible values.
2386
+ static HALF_CONSTEXPR_CONST int max_digits10 = 5;
2387
+
2388
+ /// Number base.
2389
+ static HALF_CONSTEXPR_CONST int radix = 2;
2390
+
2391
+ /// One more than smallest exponent.
2392
+ static HALF_CONSTEXPR_CONST int min_exponent = -13;
2393
+
2394
+ /// Smallest normalized representable power of 10.
2395
+ static HALF_CONSTEXPR_CONST int min_exponent10 = -4;
2396
+
2397
+ /// One more than largest exponent
2398
+ static HALF_CONSTEXPR_CONST int max_exponent = 16;
2399
+
2400
+ /// Largest finitely representable power of 10.
2401
+ static HALF_CONSTEXPR_CONST int max_exponent10 = 4;
2402
+
2403
+ /// Smallest positive normal value.
2404
+ static HALF_CONSTEXPR half_float::half min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0400); }
2405
+
2406
+ /// Smallest finite value.
2407
+ static HALF_CONSTEXPR half_float::half lowest() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0xFBFF); }
2408
+
2409
+ /// Largest finite value.
2410
+ static HALF_CONSTEXPR half_float::half max() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7BFF); }
2411
+
2412
+ /// Difference between 1 and next representable value.
2413
+ static HALF_CONSTEXPR half_float::half epsilon() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x1400); }
2414
+
2415
+ /// Maximum rounding error in ULP (units in the last place).
2416
+ static HALF_CONSTEXPR half_float::half round_error() HALF_NOTHROW
2417
+ { return half_float::half(half_float::detail::binary, (round_style==std::round_to_nearest) ? 0x3800 : 0x3C00); }
2418
+
2419
+ /// Positive infinity.
2420
+ static HALF_CONSTEXPR half_float::half infinity() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7C00); }
2421
+
2422
+ /// Quiet NaN.
2423
+ static HALF_CONSTEXPR half_float::half quiet_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7FFF); }
2424
+
2425
+ /// Signaling NaN.
2426
+ static HALF_CONSTEXPR half_float::half signaling_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7DFF); }
2427
+
2428
+ /// Smallest positive subnormal value.
2429
+ static HALF_CONSTEXPR half_float::half denorm_min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0001); }
2430
+ };
2431
+
2432
+ #if HALF_ENABLE_CPP11_HASH
2433
+ /// Hash function for half-precision floats.
2434
+ /// This is only defined if C++11 `std::hash` is supported and enabled.
2435
+ ///
2436
+ /// **See also:** Documentation for [std::hash](https://en.cppreference.com/w/cpp/utility/hash)
2437
+ template<> struct hash<half_float::half>
2438
+ {
2439
+ /// Type of function argument.
2440
+ typedef half_float::half argument_type;
2441
+
2442
+ /// Function return type.
2443
+ typedef size_t result_type;
2444
+
2445
+ /// Compute hash function.
2446
+ /// \param arg half to hash
2447
+ /// \return hash value
2448
+ result_type operator()(argument_type arg) const { return hash<half_float::detail::uint16>()(arg.data_&-static_cast<unsigned>(arg.data_!=0x8000)); }
2449
+ };
2450
+ #endif
2451
+ }
2452
+
2453
+ namespace half_float
2454
+ {
2455
+ /// \anchor compop
2456
+ /// \name Comparison operators
2457
+ /// \{
2458
+
2459
+ /// Comparison for equality.
2460
+ /// \param x first operand
2461
+ /// \param y second operand
2462
+ /// \retval true if operands equal
2463
+ /// \retval false else
2464
+ /// \exception FE_INVALID if \a x or \a y is NaN
2465
+ inline HALF_CONSTEXPR_NOERR bool operator==(half x, half y)
2466
+ {
2467
+ return !detail::compsignal(x.data_, y.data_) && (x.data_==y.data_ || !((x.data_|y.data_)&0x7FFF));
2468
+ }
2469
+
2470
+ /// Comparison for inequality.
2471
+ /// \param x first operand
2472
+ /// \param y second operand
2473
+ /// \retval true if operands not equal
2474
+ /// \retval false else
2475
+ /// \exception FE_INVALID if \a x or \a y is NaN
2476
+ inline HALF_CONSTEXPR_NOERR bool operator!=(half x, half y)
2477
+ {
2478
+ return detail::compsignal(x.data_, y.data_) || (x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF));
2479
+ }
2480
+
2481
+ /// Comparison for less than.
2482
+ /// \param x first operand
2483
+ /// \param y second operand
2484
+ /// \retval true if \a x less than \a y
2485
+ /// \retval false else
2486
+ /// \exception FE_INVALID if \a x or \a y is NaN
2487
+ inline HALF_CONSTEXPR_NOERR bool operator<(half x, half y)
2488
+ {
2489
+ return !detail::compsignal(x.data_, y.data_) &&
2490
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
2491
+ }
2492
+
2493
+ /// Comparison for greater than.
2494
+ /// \param x first operand
2495
+ /// \param y second operand
2496
+ /// \retval true if \a x greater than \a y
2497
+ /// \retval false else
2498
+ /// \exception FE_INVALID if \a x or \a y is NaN
2499
+ inline HALF_CONSTEXPR_NOERR bool operator>(half x, half y)
2500
+ {
2501
+ return !detail::compsignal(x.data_, y.data_) &&
2502
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
2503
+ }
2504
+
2505
+ /// Comparison for less equal.
2506
+ /// \param x first operand
2507
+ /// \param y second operand
2508
+ /// \retval true if \a x less equal \a y
2509
+ /// \retval false else
2510
+ /// \exception FE_INVALID if \a x or \a y is NaN
2511
+ inline HALF_CONSTEXPR_NOERR bool operator<=(half x, half y)
2512
+ {
2513
+ return !detail::compsignal(x.data_, y.data_) &&
2514
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
2515
+ }
2516
+
2517
+ /// Comparison for greater equal.
2518
+ /// \param x first operand
2519
+ /// \param y second operand
2520
+ /// \retval true if \a x greater equal \a y
2521
+ /// \retval false else
2522
+ /// \exception FE_INVALID if \a x or \a y is NaN
2523
+ inline HALF_CONSTEXPR_NOERR bool operator>=(half x, half y)
2524
+ {
2525
+ return !detail::compsignal(x.data_, y.data_) &&
2526
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
2527
+ }
2528
+
2529
+ /// \}
2530
+ /// \anchor arithmetics
2531
+ /// \name Arithmetic operators
2532
+ /// \{
2533
+
2534
+ /// Identity.
2535
+ /// \param arg operand
2536
+ /// \return unchanged operand
2537
+ inline HALF_CONSTEXPR half operator+(half arg) { return arg; }
2538
+
2539
+ /// Negation.
2540
+ /// \param arg operand
2541
+ /// \return negated operand
2542
+ inline HALF_CONSTEXPR half operator-(half arg) { return half(detail::binary, arg.data_^0x8000); }
2543
+
2544
+ /// Addition.
2545
+ /// This operation is exact to rounding for all rounding modes.
2546
+ /// \param x left operand
2547
+ /// \param y right operand
2548
+ /// \return sum of half expressions
2549
+ /// \exception FE_INVALID if \a x and \a y are infinities with different signs or signaling NaNs
2550
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2551
+ inline half operator+(half x, half y)
2552
+ {
2553
+ #ifdef HALF_ARITHMETIC_TYPE
2554
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)+detail::half2float<detail::internal_t>(y.data_)));
2555
+ #else
2556
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF;
2557
+ bool sub = ((x.data_^y.data_)&0x8000) != 0;
2558
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2559
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : (absy!=0x7C00) ? x.data_ :
2560
+ (sub && absx==0x7C00) ? detail::invalid() : y.data_);
2561
+ if(!absx)
2562
+ return absy ? y : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (x.data_|y.data_) : (x.data_&y.data_));
2563
+ if(!absy)
2564
+ return x;
2565
+ unsigned int sign = ((sub && absy>absx) ? y.data_ : x.data_) & 0x8000;
2566
+ if(absy > absx)
2567
+ std::swap(absx, absy);
2568
+ int exp = (absx>>10) + (absx<=0x3FF), d = exp - (absy>>10) - (absy<=0x3FF), mx = ((absx&0x3FF)|((absx>0x3FF)<<10)) << 3, my;
2569
+ if(d < 13)
2570
+ {
2571
+ my = ((absy&0x3FF)|((absy>0x3FF)<<10)) << 3;
2572
+ my = (my>>d) | ((my&((1<<d)-1))!=0);
2573
+ }
2574
+ else
2575
+ my = 1;
2576
+ if(sub)
2577
+ {
2578
+ if(!(mx-=my))
2579
+ return half(detail::binary, static_cast<unsigned>(half::round_style==std::round_toward_neg_infinity)<<15);
2580
+ for(; mx<0x2000 && exp>1; mx<<=1,--exp) ;
2581
+ }
2582
+ else
2583
+ {
2584
+ mx += my;
2585
+ int i = mx >> 14;
2586
+ if((exp+=i) > 30)
2587
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
2588
+ mx = (mx>>i) | (mx&i);
2589
+ }
2590
+ return half(detail::binary, detail::rounded<half::round_style,false>(sign+((exp-1)<<10)+(mx>>3), (mx>>2)&1, (mx&0x3)!=0));
2591
+ #endif
2592
+ }
2593
+
2594
+ /// Subtraction.
2595
+ /// This operation is exact to rounding for all rounding modes.
2596
+ /// \param x left operand
2597
+ /// \param y right operand
2598
+ /// \return difference of half expressions
2599
+ /// \exception FE_INVALID if \a x and \a y are infinities with equal signs or signaling NaNs
2600
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2601
+ inline half operator-(half x, half y)
2602
+ {
2603
+ #ifdef HALF_ARITHMETIC_TYPE
2604
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)-detail::half2float<detail::internal_t>(y.data_)));
2605
+ #else
2606
+ return x + -y;
2607
+ #endif
2608
+ }
2609
+
2610
+ /// Multiplication.
2611
+ /// This operation is exact to rounding for all rounding modes.
2612
+ /// \param x left operand
2613
+ /// \param y right operand
2614
+ /// \return product of half expressions
2615
+ /// \exception FE_INVALID if multiplying 0 with infinity or if \a x or \a y is signaling NaN
2616
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2617
+ inline half operator*(half x, half y)
2618
+ {
2619
+ #ifdef HALF_ARITHMETIC_TYPE
2620
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)*detail::half2float<detail::internal_t>(y.data_)));
2621
+ #else
2622
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -16;
2623
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
2624
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2625
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
2626
+ ((absx==0x7C00 && !absy)||(absy==0x7C00 && !absx)) ? detail::invalid() : (sign|0x7C00));
2627
+ if(!absx || !absy)
2628
+ return half(detail::binary, sign);
2629
+ for(; absx<0x400; absx<<=1,--exp) ;
2630
+ for(; absy<0x400; absy<<=1,--exp) ;
2631
+ detail::uint32 m = static_cast<detail::uint32>((absx&0x3FF)|0x400) * static_cast<detail::uint32>((absy&0x3FF)|0x400);
2632
+ int i = m >> 21, s = m & i;
2633
+ exp += (absx>>10) + (absy>>10) + i;
2634
+ if(exp > 29)
2635
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
2636
+ else if(exp < -11)
2637
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
2638
+ return half(detail::binary, detail::fixed2half<half::round_style,20,false,false,false>(m>>i, exp, sign, s));
2639
+ #endif
2640
+ }
2641
+
2642
+ /// Division.
2643
+ /// This operation is exact to rounding for all rounding modes.
2644
+ /// \param x left operand
2645
+ /// \param y right operand
2646
+ /// \return quotient of half expressions
2647
+ /// \exception FE_INVALID if dividing 0s or infinities with each other or if \a x or \a y is signaling NaN
2648
+ /// \exception FE_DIVBYZERO if dividing finite value by 0
2649
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2650
+ inline half operator/(half x, half y)
2651
+ {
2652
+ #ifdef HALF_ARITHMETIC_TYPE
2653
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)/detail::half2float<detail::internal_t>(y.data_)));
2654
+ #else
2655
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = 14;
2656
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
2657
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2658
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
2659
+ (absx==absy) ? detail::invalid() : (sign|((absx==0x7C00) ? 0x7C00 : 0)));
2660
+ if(!absx)
2661
+ return half(detail::binary, absy ? sign : detail::invalid());
2662
+ if(!absy)
2663
+ return half(detail::binary, detail::pole(sign));
2664
+ for(; absx<0x400; absx<<=1,--exp) ;
2665
+ for(; absy<0x400; absy<<=1,++exp) ;
2666
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
2667
+ int i = mx < my;
2668
+ exp += (absx>>10) - (absy>>10) - i;
2669
+ if(exp > 29)
2670
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
2671
+ else if(exp < -11)
2672
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
2673
+ mx <<= 12 + i;
2674
+ my <<= 1;
2675
+ return half(detail::binary, detail::fixed2half<half::round_style,11,false,false,false>(mx/my, exp, sign, mx%my!=0));
2676
+ #endif
2677
+ }
2678
+
2679
+ /// \}
2680
+ /// \anchor streaming
2681
+ /// \name Input and output
2682
+ /// \{
2683
+
2684
+ /// Output operator.
2685
+ /// This uses the built-in functionality for streaming out floating-point numbers.
2686
+ /// \param out output stream to write into
2687
+ /// \param arg half expression to write
2688
+ /// \return reference to output stream
2689
+ template<typename charT,typename traits> std::basic_ostream<charT,traits>& operator<<(std::basic_ostream<charT,traits> &out, half arg)
2690
+ {
2691
+ #ifdef HALF_ARITHMETIC_TYPE
2692
+ return out << detail::half2float<detail::internal_t>(arg.data_);
2693
+ #else
2694
+ return out << detail::half2float<float>(arg.data_);
2695
+ #endif
2696
+ }
2697
+
2698
+ /// Input operator.
2699
+ /// This uses the built-in functionality for streaming in floating-point numbers, specifically double precision floating
2700
+ /// point numbers (unless overridden with [HALF_ARITHMETIC_TYPE](\ref HALF_ARITHMETIC_TYPE)). So the input string is first
2701
+ /// rounded to double precision using the underlying platform's current floating-point rounding mode before being rounded
2702
+ /// to half-precision using the library's half-precision rounding mode.
2703
+ /// \param in input stream to read from
2704
+ /// \param arg half to read into
2705
+ /// \return reference to input stream
2706
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2707
+ template<typename charT,typename traits> std::basic_istream<charT,traits>& operator>>(std::basic_istream<charT,traits> &in, half &arg)
2708
+ {
2709
+ #ifdef HALF_ARITHMETIC_TYPE
2710
+ detail::internal_t f;
2711
+ #else
2712
+ double f;
2713
+ #endif
2714
+ if(in >> f)
2715
+ arg.data_ = detail::float2half<half::round_style>(f);
2716
+ return in;
2717
+ }
2718
+
2719
+ /// \}
2720
+ /// \anchor basic
2721
+ /// \name Basic mathematical operations
2722
+ /// \{
2723
+
2724
+ /// Absolute value.
2725
+ /// **See also:** Documentation for [std::fabs](https://en.cppreference.com/w/cpp/numeric/math/fabs).
2726
+ /// \param arg operand
2727
+ /// \return absolute value of \a arg
2728
+ inline HALF_CONSTEXPR half fabs(half arg) { return half(detail::binary, arg.data_&0x7FFF); }
2729
+
2730
+ /// Absolute value.
2731
+ /// **See also:** Documentation for [std::abs](https://en.cppreference.com/w/cpp/numeric/math/fabs).
2732
+ /// \param arg operand
2733
+ /// \return absolute value of \a arg
2734
+ inline HALF_CONSTEXPR half abs(half arg) { return fabs(arg); }
2735
+
2736
+ /// Remainder of division.
2737
+ /// **See also:** Documentation for [std::fmod](https://en.cppreference.com/w/cpp/numeric/math/fmod).
2738
+ /// \param x first operand
2739
+ /// \param y second operand
2740
+ /// \return remainder of floating-point division.
2741
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
2742
+ inline half fmod(half x, half y)
2743
+ {
2744
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000;
2745
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2746
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
2747
+ (absx==0x7C00) ? detail::invalid() : x.data_);
2748
+ if(!absy)
2749
+ return half(detail::binary, detail::invalid());
2750
+ if(!absx)
2751
+ return x;
2752
+ if(absx == absy)
2753
+ return half(detail::binary, sign);
2754
+ return half(detail::binary, sign|detail::mod<false,false>(absx, absy));
2755
+ }
2756
+
2757
+ /// Remainder of division.
2758
+ /// **See also:** Documentation for [std::remainder](https://en.cppreference.com/w/cpp/numeric/math/remainder).
2759
+ /// \param x first operand
2760
+ /// \param y second operand
2761
+ /// \return remainder of floating-point division.
2762
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
2763
+ inline half remainder(half x, half y)
2764
+ {
2765
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000;
2766
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2767
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
2768
+ (absx==0x7C00) ? detail::invalid() : x.data_);
2769
+ if(!absy)
2770
+ return half(detail::binary, detail::invalid());
2771
+ if(absx == absy)
2772
+ return half(detail::binary, sign);
2773
+ return half(detail::binary, sign^detail::mod<false,true>(absx, absy));
2774
+ }
2775
+
2776
+ /// Remainder of division.
2777
+ /// **See also:** Documentation for [std::remquo](https://en.cppreference.com/w/cpp/numeric/math/remquo).
2778
+ /// \param x first operand
2779
+ /// \param y second operand
2780
+ /// \param quo address to store some bits of quotient at
2781
+ /// \return remainder of floating-point division.
2782
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
2783
+ inline half remquo(half x, half y, int *quo)
2784
+ {
2785
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, value = x.data_ & 0x8000;
2786
+ if(absx >= 0x7C00 || absy >= 0x7C00)
2787
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
2788
+ (absx==0x7C00) ? detail::invalid() : (*quo = 0, x.data_));
2789
+ if(!absy)
2790
+ return half(detail::binary, detail::invalid());
2791
+ bool qsign = ((value^y.data_)&0x8000) != 0;
2792
+ int q = 1;
2793
+ if(absx != absy)
2794
+ value ^= detail::mod<true, true>(absx, absy, &q);
2795
+ return *quo = qsign ? -q : q, half(detail::binary, value);
2796
+ }
2797
+
2798
+ /// Fused multiply add.
2799
+ /// This function is exact to rounding for all rounding modes.
2800
+ ///
2801
+ /// **See also:** Documentation for [std::fma](https://en.cppreference.com/w/cpp/numeric/math/fma).
2802
+ /// \param x first operand
2803
+ /// \param y second operand
2804
+ /// \param z third operand
2805
+ /// \return ( \a x * \a y ) + \a z rounded as one operation.
2806
+ /// \exception FE_INVALID according to operator*() and operator+() unless any argument is a quiet NaN and no argument is a signaling NaN
2807
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding the final addition
2808
+ inline half fma(half x, half y, half z)
2809
+ {
2810
+ #ifdef HALF_ARITHMETIC_TYPE
2811
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_), fz = detail::half2float<detail::internal_t>(z.data_);
2812
+ #if HALF_ENABLE_CPP11_CMATH && FP_FAST_FMA
2813
+ return half(detail::binary, detail::float2half<half::round_style>(std::fma(fx, fy, fz)));
2814
+ #else
2815
+ return half(detail::binary, detail::float2half<half::round_style>(fx*fy+fz));
2816
+ #endif
2817
+ #else
2818
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, exp = -15;
2819
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
2820
+ bool sub = ((sign^z.data_)&0x8000) != 0;
2821
+ if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00)
2822
+ return (absx>0x7C00 || absy>0x7C00 || absz>0x7C00) ? half(detail::binary, detail::signal(x.data_, y.data_, z.data_)) :
2823
+ (absx==0x7C00) ? half(detail::binary, (!absy || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) :
2824
+ (absy==0x7C00) ? half(detail::binary, (!absx || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) : z;
2825
+ if(!absx || !absy)
2826
+ return absz ? z : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (z.data_|sign) : (z.data_&sign));
2827
+ for(; absx<0x400; absx<<=1,--exp) ;
2828
+ for(; absy<0x400; absy<<=1,--exp) ;
2829
+ detail::uint32 m = static_cast<detail::uint32>((absx&0x3FF)|0x400) * static_cast<detail::uint32>((absy&0x3FF)|0x400);
2830
+ int i = m >> 21;
2831
+ exp += (absx>>10) + (absy>>10) + i;
2832
+ m <<= 3 - i;
2833
+ if(absz)
2834
+ {
2835
+ int expz = 0;
2836
+ for(; absz<0x400; absz<<=1,--expz) ;
2837
+ expz += absz >> 10;
2838
+ detail::uint32 mz = static_cast<detail::uint32>((absz&0x3FF)|0x400) << 13;
2839
+ if(expz > exp || (expz == exp && mz > m))
2840
+ {
2841
+ std::swap(m, mz);
2842
+ std::swap(exp, expz);
2843
+ if(sub)
2844
+ sign = z.data_ & 0x8000;
2845
+ }
2846
+ int d = exp - expz;
2847
+ mz = (d<23) ? ((mz>>d)|((mz&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
2848
+ if(sub)
2849
+ {
2850
+ m = m - mz;
2851
+ if(!m)
2852
+ return half(detail::binary, static_cast<unsigned>(half::round_style==std::round_toward_neg_infinity)<<15);
2853
+ for(; m<0x800000; m<<=1,--exp) ;
2854
+ }
2855
+ else
2856
+ {
2857
+ m += mz;
2858
+ i = m >> 24;
2859
+ m = (m>>i) | (m&i);
2860
+ exp += i;
2861
+ }
2862
+ }
2863
+ if(exp > 30)
2864
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
2865
+ else if(exp < -10)
2866
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
2867
+ return half(detail::binary, detail::fixed2half<half::round_style,23,false,false,false>(m, exp-1, sign));
2868
+ #endif
2869
+ }
2870
+
2871
+ /// Maximum of half expressions.
2872
+ /// **See also:** Documentation for [std::fmax](https://en.cppreference.com/w/cpp/numeric/math/fmax).
2873
+ /// \param x first operand
2874
+ /// \param y second operand
2875
+ /// \return maximum of operands, ignoring quiet NaNs
2876
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
2877
+ inline HALF_CONSTEXPR_NOERR half fmax(half x, half y)
2878
+ {
2879
+ return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) <
2880
+ (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_));
2881
+ }
2882
+
2883
+ /// Minimum of half expressions.
2884
+ /// **See also:** Documentation for [std::fmin](https://en.cppreference.com/w/cpp/numeric/math/fmin).
2885
+ /// \param x first operand
2886
+ /// \param y second operand
2887
+ /// \return minimum of operands, ignoring quiet NaNs
2888
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
2889
+ inline HALF_CONSTEXPR_NOERR half fmin(half x, half y)
2890
+ {
2891
+ return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) >
2892
+ (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_));
2893
+ }
2894
+
2895
+ /// Positive difference.
2896
+ /// This function is exact to rounding for all rounding modes.
2897
+ ///
2898
+ /// **See also:** Documentation for [std::fdim](https://en.cppreference.com/w/cpp/numeric/math/fdim).
2899
+ /// \param x first operand
2900
+ /// \param y second operand
2901
+ /// \return \a x - \a y or 0 if difference negative
2902
+ /// \exception FE_... according to operator-(half,half)
2903
+ inline half fdim(half x, half y)
2904
+ {
2905
+ if(isnan(x) || isnan(y))
2906
+ return half(detail::binary, detail::signal(x.data_, y.data_));
2907
+ return (x.data_^(0x8000|(0x8000-(x.data_>>15)))) <= (y.data_^(0x8000|(0x8000-(y.data_>>15)))) ? half(detail::binary, 0) : (x-y);
2908
+ }
2909
+
2910
+ /// Get NaN value.
2911
+ /// **See also:** Documentation for [std::nan](https://en.cppreference.com/w/cpp/numeric/math/nan).
2912
+ /// \param arg string code
2913
+ /// \return quiet NaN
2914
+ inline half nanh(const char *arg)
2915
+ {
2916
+ unsigned int value = 0x7FFF;
2917
+ while(*arg)
2918
+ value ^= static_cast<unsigned>(*arg++) & 0xFF;
2919
+ return half(detail::binary, value);
2920
+ }
2921
+
2922
+ /// \}
2923
+ /// \anchor exponential
2924
+ /// \name Exponential functions
2925
+ /// \{
2926
+
2927
+ /// Exponential function.
2928
+ /// This function is exact to rounding for all rounding modes.
2929
+ ///
2930
+ /// **See also:** Documentation for [std::exp](https://en.cppreference.com/w/cpp/numeric/math/exp).
2931
+ /// \param arg function argument
2932
+ /// \return e raised to \a arg
2933
+ /// \exception FE_INVALID for signaling NaN
2934
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2935
+ inline half exp(half arg)
2936
+ {
2937
+ #ifdef HALF_ARITHMETIC_TYPE
2938
+ return half(detail::binary, detail::float2half<half::round_style>(std::exp(detail::half2float<detail::internal_t>(arg.data_))));
2939
+ #else
2940
+ int abs = arg.data_ & 0x7FFF;
2941
+ if(!abs)
2942
+ return half(detail::binary, 0x3C00);
2943
+ if(abs >= 0x7C00)
2944
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_));
2945
+ if(abs >= 0x4C80)
2946
+ return half(detail::binary, (arg.data_&0x8000) ? detail::underflow<half::round_style>() : detail::overflow<half::round_style>());
2947
+ detail::uint32 m = detail::multiply64(static_cast<detail::uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29);
2948
+ int e = (abs>>10) + (abs<=0x3FF), exp;
2949
+ if(e < 14)
2950
+ {
2951
+ exp = 0;
2952
+ m >>= 14 - e;
2953
+ }
2954
+ else
2955
+ {
2956
+ exp = m >> (45-e);
2957
+ m = (m<<(e-14)) & 0x7FFFFFFF;
2958
+ }
2959
+ return half(detail::binary, detail::exp2_post<half::round_style,true>(detail::exp2(m, 26), exp, (arg.data_&0x8000)!=0));
2960
+ #endif
2961
+ }
2962
+
2963
+ /// Binary exponential.
2964
+ /// This function is exact to rounding for all rounding modes.
2965
+ ///
2966
+ /// **See also:** Documentation for [std::exp2](https://en.cppreference.com/w/cpp/numeric/math/exp2).
2967
+ /// \param arg function argument
2968
+ /// \return 2 raised to \a arg
2969
+ /// \exception FE_INVALID for signaling NaN
2970
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
2971
+ inline half exp2(half arg)
2972
+ {
2973
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
2974
+ return half(detail::binary, detail::float2half<half::round_style>(std::exp2(detail::half2float<detail::internal_t>(arg.data_))));
2975
+ #else
2976
+ int abs = arg.data_ & 0x7FFF;
2977
+ if(!abs)
2978
+ return half(detail::binary, 0x3C00);
2979
+ if(abs >= 0x7C00)
2980
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_));
2981
+ if(abs >= 0x4E40)
2982
+ return half(detail::binary, (arg.data_&0x8000) ? detail::underflow<half::round_style>() : detail::overflow<half::round_style>());
2983
+ int e = (abs>>10) + (abs<=0x3FF), exp = (abs&0x3FF) + ((abs>0x3FF)<<10);
2984
+ detail::uint32 m = detail::exp2((static_cast<detail::uint32>(exp)<<(6+e))&0x7FFFFFFF, 28);
2985
+ exp >>= 25 - e;
2986
+ if(m == 0x80000000)
2987
+ {
2988
+ if(arg.data_&0x8000)
2989
+ exp = -exp;
2990
+ else if(exp > 15)
2991
+ return half(detail::binary, detail::overflow<half::round_style>());
2992
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,false,false>(m, exp+14));
2993
+ }
2994
+ return half(detail::binary, detail::exp2_post<half::round_style,true>(m, exp, (arg.data_&0x8000)!=0));
2995
+ #endif
2996
+ }
2997
+
2998
+ /// Exponential minus one.
2999
+ /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest`
3000
+ /// and in <1% of inputs for any other rounding mode.
3001
+ ///
3002
+ /// **See also:** Documentation for [std::expm1](https://en.cppreference.com/w/cpp/numeric/math/expm1).
3003
+ /// \param arg function argument
3004
+ /// \return e raised to \a arg and subtracted by 1
3005
+ /// \exception FE_INVALID for signaling NaN
3006
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3007
+ inline half expm1(half arg)
3008
+ {
3009
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3010
+ return half(detail::binary, detail::float2half<half::round_style>(std::expm1(detail::half2float<detail::internal_t>(arg.data_))));
3011
+ #else
3012
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
3013
+ if(!abs)
3014
+ return arg;
3015
+ if(abs >= 0x7C00)
3016
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00+(sign>>1)) : detail::signal(arg.data_));
3017
+ if(abs >= 0x4A00)
3018
+ return half(detail::binary, (arg.data_&0x8000) ? detail::rounded<half::round_style,true>(0xBBFF, 1, 1) : detail::overflow<half::round_style>());
3019
+ detail::uint32 m = detail::multiply64(static_cast<detail::uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29);
3020
+ int e = (abs>>10) + (abs<=0x3FF), exp;
3021
+ if(e < 14)
3022
+ {
3023
+ exp = 0;
3024
+ m >>= 14 - e;
3025
+ }
3026
+ else
3027
+ {
3028
+ exp = m >> (45-e);
3029
+ m = (m<<(e-14)) & 0x7FFFFFFF;
3030
+ }
3031
+ m = detail::exp2(m);
3032
+ if(sign)
3033
+ {
3034
+ int s = 0;
3035
+ if(m > 0x80000000)
3036
+ {
3037
+ ++exp;
3038
+ m = detail::divide64(0x80000000, m, s);
3039
+ }
3040
+ m = 0x80000000 - ((m>>exp)|((m&((static_cast<detail::uint32>(1)<<exp)-1))!=0)|s);
3041
+ exp = 0;
3042
+ }
3043
+ else
3044
+ m -= (exp<31) ? (0x80000000>>exp) : 1;
3045
+ for(exp+=14; m<0x80000000 && exp; m<<=1,--exp) ;
3046
+ if(exp > 29)
3047
+ return half(detail::binary, detail::overflow<half::round_style>());
3048
+ return half(detail::binary, detail::rounded<half::round_style,true>(sign+(exp<<10)+(m>>21), (m>>20)&1, (m&0xFFFFF)!=0));
3049
+ #endif
3050
+ }
3051
+
3052
+ /// Natural logarithm.
3053
+ /// This function is exact to rounding for all rounding modes.
3054
+ ///
3055
+ /// **See also:** Documentation for [std::log](https://en.cppreference.com/w/cpp/numeric/math/log).
3056
+ /// \param arg function argument
3057
+ /// \return logarithm of \a arg to base e
3058
+ /// \exception FE_INVALID for signaling NaN or negative argument
3059
+ /// \exception FE_DIVBYZERO for 0
3060
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3061
+ inline half log(half arg)
3062
+ {
3063
+ #ifdef HALF_ARITHMETIC_TYPE
3064
+ return half(detail::binary, detail::float2half<half::round_style>(std::log(detail::half2float<detail::internal_t>(arg.data_))));
3065
+ #else
3066
+ int abs = arg.data_ & 0x7FFF, exp = -15;
3067
+ if(!abs)
3068
+ return half(detail::binary, detail::pole(0x8000));
3069
+ if(arg.data_ & 0x8000)
3070
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
3071
+ if(abs >= 0x7C00)
3072
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
3073
+ for(; abs<0x400; abs<<=1,--exp) ;
3074
+ exp += abs >> 10;
3075
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(
3076
+ detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 27)+8, exp, 17));
3077
+ #endif
3078
+ }
3079
+
3080
+ /// Common logarithm.
3081
+ /// This function is exact to rounding for all rounding modes.
3082
+ ///
3083
+ /// **See also:** Documentation for [std::log10](https://en.cppreference.com/w/cpp/numeric/math/log10).
3084
+ /// \param arg function argument
3085
+ /// \return logarithm of \a arg to base 10
3086
+ /// \exception FE_INVALID for signaling NaN or negative argument
3087
+ /// \exception FE_DIVBYZERO for 0
3088
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3089
+ inline half log10(half arg)
3090
+ {
3091
+ #ifdef HALF_ARITHMETIC_TYPE
3092
+ return half(detail::binary, detail::float2half<half::round_style>(std::log10(detail::half2float<detail::internal_t>(arg.data_))));
3093
+ #else
3094
+ int abs = arg.data_ & 0x7FFF, exp = -15;
3095
+ if(!abs)
3096
+ return half(detail::binary, detail::pole(0x8000));
3097
+ if(arg.data_ & 0x8000)
3098
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
3099
+ if(abs >= 0x7C00)
3100
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
3101
+ switch(abs)
3102
+ {
3103
+ case 0x4900: return half(detail::binary, 0x3C00);
3104
+ case 0x5640: return half(detail::binary, 0x4000);
3105
+ case 0x63D0: return half(detail::binary, 0x4200);
3106
+ case 0x70E2: return half(detail::binary, 0x4400);
3107
+ }
3108
+ for(; abs<0x400; abs<<=1,--exp) ;
3109
+ exp += abs >> 10;
3110
+ return half(detail::binary, detail::log2_post<half::round_style,0xD49A784C>(
3111
+ detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 27)+8, exp, 16));
3112
+ #endif
3113
+ }
3114
+
3115
+ /// Binary logarithm.
3116
+ /// This function is exact to rounding for all rounding modes.
3117
+ ///
3118
+ /// **See also:** Documentation for [std::log2](https://en.cppreference.com/w/cpp/numeric/math/log2).
3119
+ /// \param arg function argument
3120
+ /// \return logarithm of \a arg to base 2
3121
+ /// \exception FE_INVALID for signaling NaN or negative argument
3122
+ /// \exception FE_DIVBYZERO for 0
3123
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3124
+ inline half log2(half arg)
3125
+ {
3126
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3127
+ return half(detail::binary, detail::float2half<half::round_style>(std::log2(detail::half2float<detail::internal_t>(arg.data_))));
3128
+ #else
3129
+ int abs = arg.data_ & 0x7FFF, exp = -15, s = 0;
3130
+ if(!abs)
3131
+ return half(detail::binary, detail::pole(0x8000));
3132
+ if(arg.data_ & 0x8000)
3133
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
3134
+ if(abs >= 0x7C00)
3135
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
3136
+ if(abs == 0x3C00)
3137
+ return half(detail::binary, 0);
3138
+ for(; abs<0x400; abs<<=1,--exp) ;
3139
+ exp += (abs>>10);
3140
+ if(!(abs&0x3FF))
3141
+ {
3142
+ unsigned int value = static_cast<unsigned>(exp<0) << 15, m = std::abs(exp) << 6;
3143
+ for(exp=18; m<0x400; m<<=1,--exp) ;
3144
+ return half(detail::binary, value+(exp<<10)+m);
3145
+ }
3146
+ detail::uint32 ilog = exp, sign = detail::sign_mask(ilog), m =
3147
+ (((ilog<<27)+(detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 28)>>4))^sign) - sign;
3148
+ if(!m)
3149
+ return half(detail::binary, 0);
3150
+ for(exp=14; m<0x8000000 && exp; m<<=1,--exp) ;
3151
+ for(; m>0xFFFFFFF; m>>=1,++exp)
3152
+ s |= m & 1;
3153
+ return half(detail::binary, detail::fixed2half<half::round_style,27,false,false,true>(m, exp, sign&0x8000, s));
3154
+ #endif
3155
+ }
3156
+
3157
+ /// Natural logarithm plus one.
3158
+ /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest`
3159
+ /// and in ~1% of inputs for any other rounding mode.
3160
+ ///
3161
+ /// **See also:** Documentation for [std::log1p](https://en.cppreference.com/w/cpp/numeric/math/log1p).
3162
+ /// \param arg function argument
3163
+ /// \return logarithm of \a arg plus 1 to base e
3164
+ /// \exception FE_INVALID for signaling NaN or argument <-1
3165
+ /// \exception FE_DIVBYZERO for -1
3166
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3167
+ inline half log1p(half arg)
3168
+ {
3169
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3170
+ return half(detail::binary, detail::float2half<half::round_style>(std::log1p(detail::half2float<detail::internal_t>(arg.data_))));
3171
+ #else
3172
+ if(arg.data_ >= 0xBC00)
3173
+ return half(detail::binary, (arg.data_==0xBC00) ? detail::pole(0x8000) : (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
3174
+ int abs = arg.data_ & 0x7FFF, exp = -15;
3175
+ if(!abs || abs >= 0x7C00)
3176
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
3177
+ for(; abs<0x400; abs<<=1,--exp) ;
3178
+ exp += abs >> 10;
3179
+ detail::uint32 m = static_cast<detail::uint32>((abs&0x3FF)|0x400) << 20;
3180
+ if(arg.data_ & 0x8000)
3181
+ {
3182
+ m = 0x40000000 - (m>>-exp);
3183
+ for(exp=0; m<0x40000000; m<<=1,--exp) ;
3184
+ }
3185
+ else
3186
+ {
3187
+ if(exp < 0)
3188
+ {
3189
+ m = 0x40000000 + (m>>-exp);
3190
+ exp = 0;
3191
+ }
3192
+ else
3193
+ {
3194
+ m += 0x40000000 >> exp;
3195
+ int i = m >> 31;
3196
+ m >>= i;
3197
+ exp += i;
3198
+ }
3199
+ }
3200
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(detail::log2(m), exp, 17));
3201
+ #endif
3202
+ }
3203
+
3204
+ /// \}
3205
+ /// \anchor power
3206
+ /// \name Power functions
3207
+ /// \{
3208
+
3209
+ /// Square root.
3210
+ /// This function is exact to rounding for all rounding modes.
3211
+ ///
3212
+ /// **See also:** Documentation for [std::sqrt](https://en.cppreference.com/w/cpp/numeric/math/sqrt).
3213
+ /// \param arg function argument
3214
+ /// \return square root of \a arg
3215
+ /// \exception FE_INVALID for signaling NaN and negative arguments
3216
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3217
+ inline half sqrt(half arg)
3218
+ {
3219
+ #ifdef HALF_ARITHMETIC_TYPE
3220
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(detail::half2float<detail::internal_t>(arg.data_))));
3221
+ #else
3222
+ int abs = arg.data_ & 0x7FFF, exp = 15;
3223
+ if(!abs || arg.data_ >= 0x7C00)
3224
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_>0x8000) ? detail::invalid() : arg.data_);
3225
+ for(; abs<0x400; abs<<=1,--exp) ;
3226
+ detail::uint32 r = static_cast<detail::uint32>((abs&0x3FF)|0x400) << 10, m = detail::sqrt<20>(r, exp+=abs>>10);
3227
+ return half(detail::binary, detail::rounded<half::round_style,false>((exp<<10)+(m&0x3FF), r>m, r!=0));
3228
+ #endif
3229
+ }
3230
+
3231
+ /// Cubic root.
3232
+ /// This function is exact to rounding for all rounding modes.
3233
+ ///
3234
+ /// **See also:** Documentation for [std::cbrt](https://en.cppreference.com/w/cpp/numeric/math/cbrt).
3235
+ /// \param arg function argument
3236
+ /// \return cubic root of \a arg
3237
+ /// \exception FE_INVALID for signaling NaN
3238
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3239
+ inline half cbrt(half arg)
3240
+ {
3241
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3242
+ return half(detail::binary, detail::float2half<half::round_style>(std::cbrt(detail::half2float<detail::internal_t>(arg.data_))));
3243
+ #else
3244
+ int abs = arg.data_ & 0x7FFF, exp = -15;
3245
+ if(!abs || abs == 0x3C00 || abs >= 0x7C00)
3246
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
3247
+ for(; abs<0x400; abs<<=1, --exp);
3248
+ detail::uint32 ilog = exp + (abs>>10), sign = detail::sign_mask(ilog), f, m =
3249
+ (((ilog<<27)+(detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 24)>>4))^sign) - sign;
3250
+ for(exp=2; m<0x80000000; m<<=1,--exp) ;
3251
+ m = detail::multiply64(m, 0xAAAAAAAB);
3252
+ int i = m >> 31, s;
3253
+ exp += i;
3254
+ m <<= 1 - i;
3255
+ if(exp < 0)
3256
+ {
3257
+ f = m >> -exp;
3258
+ exp = 0;
3259
+ }
3260
+ else
3261
+ {
3262
+ f = (m<<exp) & 0x7FFFFFFF;
3263
+ exp = m >> (31-exp);
3264
+ }
3265
+ m = detail::exp2(f, (half::round_style==std::round_to_nearest) ? 29 : 26);
3266
+ if(sign)
3267
+ {
3268
+ if(m > 0x80000000)
3269
+ {
3270
+ m = detail::divide64(0x80000000, m, s);
3271
+ ++exp;
3272
+ }
3273
+ exp = -exp;
3274
+ }
3275
+ return half(detail::binary, (half::round_style==std::round_to_nearest) ?
3276
+ detail::fixed2half<half::round_style,31,false,false,false>(m, exp+14, arg.data_&0x8000) :
3277
+ detail::fixed2half<half::round_style,23,false,false,false>((m+0x80)>>8, exp+14, arg.data_&0x8000));
3278
+ #endif
3279
+ }
3280
+
3281
+ /// Hypotenuse function.
3282
+ /// This function is exact to rounding for all rounding modes.
3283
+ ///
3284
+ /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot).
3285
+ /// \param x first argument
3286
+ /// \param y second argument
3287
+ /// \return square root of sum of squares without internal over- or underflows
3288
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
3289
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root
3290
+ inline half hypot(half x, half y)
3291
+ {
3292
+ #ifdef HALF_ARITHMETIC_TYPE
3293
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_);
3294
+ #if HALF_ENABLE_CPP11_CMATH
3295
+ return half(detail::binary, detail::float2half<half::round_style>(std::hypot(fx, fy)));
3296
+ #else
3297
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(fx*fx+fy*fy)));
3298
+ #endif
3299
+ #else
3300
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, expx = 0, expy = 0;
3301
+ if(absx >= 0x7C00 || absy >= 0x7C00)
3302
+ return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, y.data_) :
3303
+ (absy==0x7C00) ? detail::select(0x7C00, x.data_) : detail::signal(x.data_, y.data_));
3304
+ if(!absx)
3305
+ return half(detail::binary, absy ? detail::check_underflow(absy) : 0);
3306
+ if(!absy)
3307
+ return half(detail::binary, detail::check_underflow(absx));
3308
+ if(absy > absx)
3309
+ std::swap(absx, absy);
3310
+ for(; absx<0x400; absx<<=1,--expx) ;
3311
+ for(; absy<0x400; absy<<=1,--expy) ;
3312
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
3313
+ mx *= mx;
3314
+ my *= my;
3315
+ int ix = mx >> 21, iy = my >> 21;
3316
+ expx = 2*(expx+(absx>>10)) - 15 + ix;
3317
+ expy = 2*(expy+(absy>>10)) - 15 + iy;
3318
+ mx <<= 10 - ix;
3319
+ my <<= 10 - iy;
3320
+ int d = expx - expy;
3321
+ my = (d<30) ? ((my>>d)|((my&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
3322
+ return half(detail::binary, detail::hypot_post<half::round_style>(mx+my, expx));
3323
+ #endif
3324
+ }
3325
+
3326
+ /// Hypotenuse function.
3327
+ /// This function is exact to rounding for all rounding modes.
3328
+ ///
3329
+ /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot).
3330
+ /// \param x first argument
3331
+ /// \param y second argument
3332
+ /// \param z third argument
3333
+ /// \return square root of sum of squares without internal over- or underflows
3334
+ /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN
3335
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root
3336
+ inline half hypot(half x, half y, half z)
3337
+ {
3338
+ #ifdef HALF_ARITHMETIC_TYPE
3339
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_), fz = detail::half2float<detail::internal_t>(z.data_);
3340
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(fx*fx+fy*fy+fz*fz)));
3341
+ #else
3342
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, expx = 0, expy = 0, expz = 0;
3343
+ if(!absx)
3344
+ return hypot(y, z);
3345
+ if(!absy)
3346
+ return hypot(x, z);
3347
+ if(!absz)
3348
+ return hypot(x, y);
3349
+ if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00)
3350
+ return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, detail::select(y.data_, z.data_)) :
3351
+ (absy==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, z.data_)) :
3352
+ (absz==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, y.data_)) :
3353
+ detail::signal(x.data_, y.data_, z.data_));
3354
+ if(absz > absy)
3355
+ std::swap(absy, absz);
3356
+ if(absy > absx)
3357
+ std::swap(absx, absy);
3358
+ if(absz > absy)
3359
+ std::swap(absy, absz);
3360
+ for(; absx<0x400; absx<<=1,--expx) ;
3361
+ for(; absy<0x400; absy<<=1,--expy) ;
3362
+ for(; absz<0x400; absz<<=1,--expz) ;
3363
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400, mz = (absz&0x3FF) | 0x400;
3364
+ mx *= mx;
3365
+ my *= my;
3366
+ mz *= mz;
3367
+ int ix = mx >> 21, iy = my >> 21, iz = mz >> 21;
3368
+ expx = 2*(expx+(absx>>10)) - 15 + ix;
3369
+ expy = 2*(expy+(absy>>10)) - 15 + iy;
3370
+ expz = 2*(expz+(absz>>10)) - 15 + iz;
3371
+ mx <<= 10 - ix;
3372
+ my <<= 10 - iy;
3373
+ mz <<= 10 - iz;
3374
+ int d = expy - expz;
3375
+ mz = (d<30) ? ((mz>>d)|((mz&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
3376
+ my += mz;
3377
+ if(my & 0x80000000)
3378
+ {
3379
+ my = (my>>1) | (my&1);
3380
+ if(++expy > expx)
3381
+ {
3382
+ std::swap(mx, my);
3383
+ std::swap(expx, expy);
3384
+ }
3385
+ }
3386
+ d = expx - expy;
3387
+ my = (d<30) ? ((my>>d)|((my&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
3388
+ return half(detail::binary, detail::hypot_post<half::round_style>(mx+my, expx));
3389
+ #endif
3390
+ }
3391
+
3392
+ /// Power function.
3393
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.00025% of inputs.
3394
+ ///
3395
+ /// **See also:** Documentation for [std::pow](https://en.cppreference.com/w/cpp/numeric/math/pow).
3396
+ /// \param x base
3397
+ /// \param y exponent
3398
+ /// \return \a x raised to \a y
3399
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN or if \a x is finite an negative and \a y is finite and not integral
3400
+ /// \exception FE_DIVBYZERO if \a x is 0 and \a y is negative
3401
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3402
+ inline half pow(half x, half y)
3403
+ {
3404
+ #ifdef HALF_ARITHMETIC_TYPE
3405
+ return half(detail::binary, detail::float2half<half::round_style>(std::pow(detail::half2float<detail::internal_t>(x.data_), detail::half2float<detail::internal_t>(y.data_))));
3406
+ #else
3407
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -15;
3408
+ if(!absy || x.data_ == 0x3C00)
3409
+ return half(detail::binary, detail::select(0x3C00, (x.data_==0x3C00) ? y.data_ : x.data_));
3410
+ bool is_int = absy >= 0x6400 || (absy>=0x3C00 && !(absy&((1<<(25-(absy>>10)))-1)));
3411
+ unsigned int sign = x.data_ & (static_cast<unsigned>((absy<0x6800)&&is_int&&((absy>>(25-(absy>>10)))&1))<<15);
3412
+ if(absx >= 0x7C00 || absy >= 0x7C00)
3413
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
3414
+ (absy==0x7C00) ? ((absx==0x3C00) ? 0x3C00 : (!absx && y.data_==0xFC00) ? detail::pole() :
3415
+ (0x7C00&-((y.data_>>15)^(absx>0x3C00)))) : (sign|(0x7C00&((y.data_>>15)-1U))));
3416
+ if(!absx)
3417
+ return half(detail::binary, (y.data_&0x8000) ? detail::pole(sign) : sign);
3418
+ if((x.data_&0x8000) && !is_int)
3419
+ return half(detail::binary, detail::invalid());
3420
+ if(x.data_ == 0xBC00)
3421
+ return half(detail::binary, sign|0x3C00);
3422
+ if(y.data_ == 0x3800)
3423
+ return sqrt(x);
3424
+ if(y.data_ == 0x3C00)
3425
+ return half(detail::binary, detail::check_underflow(x.data_));
3426
+ if(y.data_ == 0x4000)
3427
+ return x * x;
3428
+ for(; absx<0x400; absx<<=1,--exp) ;
3429
+ detail::uint32 ilog = exp + (absx>>10), msign = detail::sign_mask(ilog), f, m =
3430
+ (((ilog<<27)+((detail::log2(static_cast<detail::uint32>((absx&0x3FF)|0x400)<<20)+8)>>4))^msign) - msign;
3431
+ for(exp=-11; m<0x80000000; m<<=1,--exp) ;
3432
+ for(; absy<0x400; absy<<=1,--exp) ;
3433
+ m = detail::multiply64(m, static_cast<detail::uint32>((absy&0x3FF)|0x400)<<21);
3434
+ int i = m >> 31;
3435
+ exp += (absy>>10) + i;
3436
+ m <<= 1 - i;
3437
+ if(exp < 0)
3438
+ {
3439
+ f = m >> -exp;
3440
+ exp = 0;
3441
+ }
3442
+ else
3443
+ {
3444
+ f = (m<<exp) & 0x7FFFFFFF;
3445
+ exp = m >> (31-exp);
3446
+ }
3447
+ return half(detail::binary, detail::exp2_post<half::round_style,false>(detail::exp2(f), exp, ((msign&1)^(y.data_>>15))!=0, sign));
3448
+ #endif
3449
+ }
3450
+
3451
+ /// \}
3452
+ /// \anchor trigonometric
3453
+ /// \name Trigonometric functions
3454
+ /// \{
3455
+
3456
+ /// Compute sine and cosine simultaneously.
3457
+ /// This returns the same results as sin() and cos() but is faster than calling each function individually.
3458
+ ///
3459
+ /// This function is exact to rounding for all rounding modes.
3460
+ /// \param arg function argument
3461
+ /// \param sin variable to take sine of \a arg
3462
+ /// \param cos variable to take cosine of \a arg
3463
+ /// \exception FE_INVALID for signaling NaN or infinity
3464
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3465
+ inline void sincos(half arg, half *sin, half *cos)
3466
+ {
3467
+ #ifdef HALF_ARITHMETIC_TYPE
3468
+ detail::internal_t f = detail::half2float<detail::internal_t>(arg.data_);
3469
+ *sin = half(detail::binary, detail::float2half<half::round_style>(std::sin(f)));
3470
+ *cos = half(detail::binary, detail::float2half<half::round_style>(std::cos(f)));
3471
+ #else
3472
+ int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15, k;
3473
+ if(abs >= 0x7C00)
3474
+ *sin = *cos = half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3475
+ else if(!abs)
3476
+ {
3477
+ *sin = arg;
3478
+ *cos = half(detail::binary, 0x3C00);
3479
+ }
3480
+ else if(abs < 0x2500)
3481
+ {
3482
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
3483
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
3484
+ }
3485
+ else
3486
+ {
3487
+ if(half::round_style != std::round_to_nearest)
3488
+ {
3489
+ switch(abs)
3490
+ {
3491
+ case 0x48B7:
3492
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x1D07, 1, 1));
3493
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0xBBFF, 1, 1));
3494
+ return;
3495
+ case 0x598C:
3496
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
3497
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x80FC, 1, 1));
3498
+ return;
3499
+ case 0x6A64:
3500
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x3BFE, 1, 1));
3501
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x27FF, 1, 1));
3502
+ return;
3503
+ case 0x6D8C:
3504
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x0FE6, 1, 1));
3505
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
3506
+ return;
3507
+ }
3508
+ }
3509
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
3510
+ switch(k & 3)
3511
+ {
3512
+ case 1: sc = std::make_pair(sc.second, -sc.first); break;
3513
+ case 2: sc = std::make_pair(-sc.first, -sc.second); break;
3514
+ case 3: sc = std::make_pair(-sc.second, sc.first); break;
3515
+ }
3516
+ *sin = half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((sc.first^-static_cast<detail::uint32>(sign))+sign));
3517
+ *cos = half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>(sc.second));
3518
+ }
3519
+ #endif
3520
+ }
3521
+
3522
+ /// Sine function.
3523
+ /// This function is exact to rounding for all rounding modes.
3524
+ ///
3525
+ /// **See also:** Documentation for [std::sin](https://en.cppreference.com/w/cpp/numeric/math/sin).
3526
+ /// \param arg function argument
3527
+ /// \return sine value of \a arg
3528
+ /// \exception FE_INVALID for signaling NaN or infinity
3529
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3530
+ inline half sin(half arg)
3531
+ {
3532
+ #ifdef HALF_ARITHMETIC_TYPE
3533
+ return half(detail::binary, detail::float2half<half::round_style>(std::sin(detail::half2float<detail::internal_t>(arg.data_))));
3534
+ #else
3535
+ int abs = arg.data_ & 0x7FFF, k;
3536
+ if(!abs)
3537
+ return arg;
3538
+ if(abs >= 0x7C00)
3539
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3540
+ if(abs < 0x2900)
3541
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
3542
+ if(half::round_style != std::round_to_nearest)
3543
+ switch(abs)
3544
+ {
3545
+ case 0x48B7: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x1D07, 1, 1));
3546
+ case 0x6A64: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x3BFE, 1, 1));
3547
+ case 0x6D8C: return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x0FE6, 1, 1));
3548
+ }
3549
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
3550
+ detail::uint32 sign = -static_cast<detail::uint32>(((k>>1)&1)^(arg.data_>>15));
3551
+ return half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((((k&1) ? sc.second : sc.first)^sign) - sign));
3552
+ #endif
3553
+ }
3554
+
3555
+ /// Cosine function.
3556
+ /// This function is exact to rounding for all rounding modes.
3557
+ ///
3558
+ /// **See also:** Documentation for [std::cos](https://en.cppreference.com/w/cpp/numeric/math/cos).
3559
+ /// \param arg function argument
3560
+ /// \return cosine value of \a arg
3561
+ /// \exception FE_INVALID for signaling NaN or infinity
3562
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3563
+ inline half cos(half arg)
3564
+ {
3565
+ #ifdef HALF_ARITHMETIC_TYPE
3566
+ return half(detail::binary, detail::float2half<half::round_style>(std::cos(detail::half2float<detail::internal_t>(arg.data_))));
3567
+ #else
3568
+ int abs = arg.data_ & 0x7FFF, k;
3569
+ if(!abs)
3570
+ return half(detail::binary, 0x3C00);
3571
+ if(abs >= 0x7C00)
3572
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3573
+ if(abs < 0x2500)
3574
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
3575
+ if(half::round_style != std::round_to_nearest && abs == 0x598C)
3576
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x80FC, 1, 1));
3577
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
3578
+ detail::uint32 sign = -static_cast<detail::uint32>(((k>>1)^k)&1);
3579
+ return half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((((k&1) ? sc.first : sc.second)^sign) - sign));
3580
+ #endif
3581
+ }
3582
+
3583
+ /// Tangent function.
3584
+ /// This function is exact to rounding for all rounding modes.
3585
+ ///
3586
+ /// **See also:** Documentation for [std::tan](https://en.cppreference.com/w/cpp/numeric/math/tan).
3587
+ /// \param arg function argument
3588
+ /// \return tangent value of \a arg
3589
+ /// \exception FE_INVALID for signaling NaN or infinity
3590
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3591
+ inline half tan(half arg)
3592
+ {
3593
+ #ifdef HALF_ARITHMETIC_TYPE
3594
+ return half(detail::binary, detail::float2half<half::round_style>(std::tan(detail::half2float<detail::internal_t>(arg.data_))));
3595
+ #else
3596
+ int abs = arg.data_ & 0x7FFF, exp = 13, k;
3597
+ if(!abs)
3598
+ return arg;
3599
+ if(abs >= 0x7C00)
3600
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3601
+ if(abs < 0x2700)
3602
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
3603
+ if(half::round_style != std::round_to_nearest)
3604
+ switch(abs)
3605
+ {
3606
+ case 0x658C: return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x07E6, 1, 1));
3607
+ case 0x7330: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x4B62, 1, 1));
3608
+ }
3609
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 30);
3610
+ if(k & 1)
3611
+ sc = std::make_pair(-sc.second, sc.first);
3612
+ detail::uint32 signy = detail::sign_mask(sc.first), signx = detail::sign_mask(sc.second);
3613
+ detail::uint32 my = (sc.first^signy) - signy, mx = (sc.second^signx) - signx;
3614
+ for(; my<0x80000000; my<<=1,--exp) ;
3615
+ for(; mx<0x80000000; mx<<=1,++exp) ;
3616
+ return half(detail::binary, detail::tangent_post<half::round_style>(my, mx, exp, (signy^signx^arg.data_)&0x8000));
3617
+ #endif
3618
+ }
3619
+
3620
+ /// Arc sine.
3621
+ /// This function is exact to rounding for all rounding modes.
3622
+ ///
3623
+ /// **See also:** Documentation for [std::asin](https://en.cppreference.com/w/cpp/numeric/math/asin).
3624
+ /// \param arg function argument
3625
+ /// \return arc sine value of \a arg
3626
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
3627
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3628
+ inline half asin(half arg)
3629
+ {
3630
+ #ifdef HALF_ARITHMETIC_TYPE
3631
+ return half(detail::binary, detail::float2half<half::round_style>(std::asin(detail::half2float<detail::internal_t>(arg.data_))));
3632
+ #else
3633
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
3634
+ if(!abs)
3635
+ return arg;
3636
+ if(abs >= 0x3C00)
3637
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() :
3638
+ detail::rounded<half::round_style,true>(sign|0x3E48, 0, 1));
3639
+ if(abs < 0x2900)
3640
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
3641
+ if(half::round_style != std::round_to_nearest && (abs == 0x2B44 || abs == 0x2DC3))
3642
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_+1, 1, 1));
3643
+ std::pair<detail::uint32,detail::uint32> sc = detail::atan2_args(abs);
3644
+ detail::uint32 m = detail::atan2(sc.first, sc.second, (half::round_style==std::round_to_nearest) ? 27 : 26);
3645
+ return half(detail::binary, detail::fixed2half<half::round_style,30,false,true,true>(m, 14, sign));
3646
+ #endif
3647
+ }
3648
+
3649
+ /// Arc cosine function.
3650
+ /// This function is exact to rounding for all rounding modes.
3651
+ ///
3652
+ /// **See also:** Documentation for [std::acos](https://en.cppreference.com/w/cpp/numeric/math/acos).
3653
+ /// \param arg function argument
3654
+ /// \return arc cosine value of \a arg
3655
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
3656
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3657
+ inline half acos(half arg)
3658
+ {
3659
+ #ifdef HALF_ARITHMETIC_TYPE
3660
+ return half(detail::binary, detail::float2half<half::round_style>(std::acos(detail::half2float<detail::internal_t>(arg.data_))));
3661
+ #else
3662
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15;
3663
+ if(!abs)
3664
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x3E48, 0, 1));
3665
+ if(abs >= 0x3C00)
3666
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() :
3667
+ sign ? detail::rounded<half::round_style,true>(0x4248, 0, 1) : 0);
3668
+ std::pair<detail::uint32,detail::uint32> cs = detail::atan2_args(abs);
3669
+ detail::uint32 m = detail::atan2(cs.second, cs.first, 28);
3670
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,true,true>(sign ? (0xC90FDAA2-m) : m, 15, 0, sign));
3671
+ #endif
3672
+ }
3673
+
3674
+ /// Arc tangent function.
3675
+ /// This function is exact to rounding for all rounding modes.
3676
+ ///
3677
+ /// **See also:** Documentation for [std::atan](https://en.cppreference.com/w/cpp/numeric/math/atan).
3678
+ /// \param arg function argument
3679
+ /// \return arc tangent value of \a arg
3680
+ /// \exception FE_INVALID for signaling NaN
3681
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3682
+ inline half atan(half arg)
3683
+ {
3684
+ #ifdef HALF_ARITHMETIC_TYPE
3685
+ return half(detail::binary, detail::float2half<half::round_style>(std::atan(detail::half2float<detail::internal_t>(arg.data_))));
3686
+ #else
3687
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
3688
+ if(!abs)
3689
+ return arg;
3690
+ if(abs >= 0x7C00)
3691
+ return half(detail::binary, (abs==0x7C00) ? detail::rounded<half::round_style,true>(sign|0x3E48, 0, 1) : detail::signal(arg.data_));
3692
+ if(abs <= 0x2700)
3693
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
3694
+ int exp = (abs>>10) + (abs<=0x3FF);
3695
+ detail::uint32 my = (abs&0x3FF) | ((abs>0x3FF)<<10);
3696
+ detail::uint32 m = (exp>15) ? detail::atan2(my<<19, 0x20000000>>(exp-15), (half::round_style==std::round_to_nearest) ? 26 : 24) :
3697
+ detail::atan2(my<<(exp+4), 0x20000000, (half::round_style==std::round_to_nearest) ? 30 : 28);
3698
+ return half(detail::binary, detail::fixed2half<half::round_style,30,false,true,true>(m, 14, sign));
3699
+ #endif
3700
+ }
3701
+
3702
+ /// Arc tangent function.
3703
+ /// This function may be 1 ULP off the correctly rounded exact result in ~0.005% of inputs for `std::round_to_nearest`,
3704
+ /// in ~0.1% of inputs for `std::round_toward_zero` and in ~0.02% of inputs for any other rounding mode.
3705
+ ///
3706
+ /// **See also:** Documentation for [std::atan2](https://en.cppreference.com/w/cpp/numeric/math/atan2).
3707
+ /// \param y numerator
3708
+ /// \param x denominator
3709
+ /// \return arc tangent value
3710
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
3711
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3712
+ inline half atan2(half y, half x)
3713
+ {
3714
+ #ifdef HALF_ARITHMETIC_TYPE
3715
+ return half(detail::binary, detail::float2half<half::round_style>(std::atan2(detail::half2float<detail::internal_t>(y.data_), detail::half2float<detail::internal_t>(x.data_))));
3716
+ #else
3717
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, signx = x.data_ >> 15, signy = y.data_ & 0x8000;
3718
+ if(absx >= 0x7C00 || absy >= 0x7C00)
3719
+ {
3720
+ if(absx > 0x7C00 || absy > 0x7C00)
3721
+ return half(detail::binary, detail::signal(x.data_, y.data_));
3722
+ if(absy == 0x7C00)
3723
+ return half(detail::binary, (absx<0x7C00) ? detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1) :
3724
+ signx ? detail::rounded<half::round_style,true>(signy|0x40B6, 0, 1) :
3725
+ detail::rounded<half::round_style,true>(signy|0x3A48, 0, 1));
3726
+ return (x.data_==0x7C00) ? half(detail::binary, signy) : half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1));
3727
+ }
3728
+ if(!absy)
3729
+ return signx ? half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1)) : y;
3730
+ if(!absx)
3731
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1));
3732
+ int d = (absy>>10) + (absy<=0x3FF) - (absx>>10) - (absx<=0x3FF);
3733
+ if(d > (signx ? 18 : 12))
3734
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1));
3735
+ if(signx && d < -11)
3736
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1));
3737
+ if(!signx && d < ((half::round_style==std::round_toward_zero) ? -15 : -9))
3738
+ {
3739
+ for(; absy<0x400; absy<<=1,--d) ;
3740
+ detail::uint32 mx = ((absx<<1)&0x7FF) | 0x800, my = ((absy<<1)&0x7FF) | 0x800;
3741
+ int i = my < mx;
3742
+ d -= i;
3743
+ if(d < -25)
3744
+ return half(detail::binary, detail::underflow<half::round_style>(signy));
3745
+ my <<= 11 + i;
3746
+ return half(detail::binary, detail::fixed2half<half::round_style,11,false,false,true>(my/mx, d+14, signy, my%mx!=0));
3747
+ }
3748
+ detail::uint32 m = detail::atan2( ((absy&0x3FF)|((absy>0x3FF)<<10))<<(19+((d<0) ? d : (d>0) ? 0 : -1)),
3749
+ ((absx&0x3FF)|((absx>0x3FF)<<10))<<(19-((d>0) ? d : (d<0) ? 0 : 1)));
3750
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,true,true>(signx ? (0xC90FDAA2-m) : m, 15, signy, signx));
3751
+ #endif
3752
+ }
3753
+
3754
+ /// \}
3755
+ /// \anchor hyperbolic
3756
+ /// \name Hyperbolic functions
3757
+ /// \{
3758
+
3759
+ /// Hyperbolic sine.
3760
+ /// This function is exact to rounding for all rounding modes.
3761
+ ///
3762
+ /// **See also:** Documentation for [std::sinh](https://en.cppreference.com/w/cpp/numeric/math/sinh).
3763
+ /// \param arg function argument
3764
+ /// \return hyperbolic sine value of \a arg
3765
+ /// \exception FE_INVALID for signaling NaN
3766
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3767
+ inline half sinh(half arg)
3768
+ {
3769
+ #ifdef HALF_ARITHMETIC_TYPE
3770
+ return half(detail::binary, detail::float2half<half::round_style>(std::sinh(detail::half2float<detail::internal_t>(arg.data_))));
3771
+ #else
3772
+ int abs = arg.data_ & 0x7FFF, exp;
3773
+ if(!abs || abs >= 0x7C00)
3774
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
3775
+ if(abs <= 0x2900)
3776
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
3777
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 29 : 27);
3778
+ detail::uint32 m = mm.first - mm.second;
3779
+ for(exp+=13; m<0x80000000 && exp; m<<=1,--exp) ;
3780
+ unsigned int sign = arg.data_ & 0x8000;
3781
+ if(exp > 29)
3782
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
3783
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,false,true>(m, exp, sign));
3784
+ #endif
3785
+ }
3786
+
3787
+ /// Hyperbolic cosine.
3788
+ /// This function is exact to rounding for all rounding modes.
3789
+ ///
3790
+ /// **See also:** Documentation for [std::cosh](https://en.cppreference.com/w/cpp/numeric/math/cosh).
3791
+ /// \param arg function argument
3792
+ /// \return hyperbolic cosine value of \a arg
3793
+ /// \exception FE_INVALID for signaling NaN
3794
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3795
+ inline half cosh(half arg)
3796
+ {
3797
+ #ifdef HALF_ARITHMETIC_TYPE
3798
+ return half(detail::binary, detail::float2half<half::round_style>(std::cosh(detail::half2float<detail::internal_t>(arg.data_))));
3799
+ #else
3800
+ int abs = arg.data_ & 0x7FFF, exp;
3801
+ if(!abs)
3802
+ return half(detail::binary, 0x3C00);
3803
+ if(abs >= 0x7C00)
3804
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : 0x7C00);
3805
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 23 : 26);
3806
+ detail::uint32 m = mm.first + mm.second, i = (~m&0xFFFFFFFF) >> 31;
3807
+ m = (m>>i) | (m&i) | 0x80000000;
3808
+ if((exp+=13+i) > 29)
3809
+ return half(detail::binary, detail::overflow<half::round_style>());
3810
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,false,true>(m, exp));
3811
+ #endif
3812
+ }
3813
+
3814
+ /// Hyperbolic tangent.
3815
+ /// This function is exact to rounding for all rounding modes.
3816
+ ///
3817
+ /// **See also:** Documentation for [std::tanh](https://en.cppreference.com/w/cpp/numeric/math/tanh).
3818
+ /// \param arg function argument
3819
+ /// \return hyperbolic tangent value of \a arg
3820
+ /// \exception FE_INVALID for signaling NaN
3821
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3822
+ inline half tanh(half arg)
3823
+ {
3824
+ #ifdef HALF_ARITHMETIC_TYPE
3825
+ return half(detail::binary, detail::float2half<half::round_style>(std::tanh(detail::half2float<detail::internal_t>(arg.data_))));
3826
+ #else
3827
+ int abs = arg.data_ & 0x7FFF, exp;
3828
+ if(!abs)
3829
+ return arg;
3830
+ if(abs >= 0x7C00)
3831
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_-0x4000));
3832
+ if(abs >= 0x4500)
3833
+ return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
3834
+ if(abs < 0x2700)
3835
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
3836
+ if(half::round_style != std::round_to_nearest && abs == 0x2D3F)
3837
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-3, 0, 1));
3838
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, 27);
3839
+ detail::uint32 my = mm.first - mm.second - (half::round_style!=std::round_to_nearest), mx = mm.first + mm.second, i = (~mx&0xFFFFFFFF) >> 31;
3840
+ for(exp=13; my<0x80000000; my<<=1,--exp) ;
3841
+ mx = (mx>>i) | 0x80000000;
3842
+ return half(detail::binary, detail::tangent_post<half::round_style>(my, mx, exp-i, arg.data_&0x8000));
3843
+ #endif
3844
+ }
3845
+
3846
+ /// Hyperbolic area sine.
3847
+ /// This function is exact to rounding for all rounding modes.
3848
+ ///
3849
+ /// **See also:** Documentation for [std::asinh](https://en.cppreference.com/w/cpp/numeric/math/asinh).
3850
+ /// \param arg function argument
3851
+ /// \return area sine value of \a arg
3852
+ /// \exception FE_INVALID for signaling NaN
3853
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3854
+ inline half asinh(half arg)
3855
+ {
3856
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3857
+ return half(detail::binary, detail::float2half<half::round_style>(std::asinh(detail::half2float<detail::internal_t>(arg.data_))));
3858
+ #else
3859
+ int abs = arg.data_ & 0x7FFF;
3860
+ if(!abs || abs >= 0x7C00)
3861
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
3862
+ if(abs <= 0x2900)
3863
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
3864
+ if(half::round_style != std::round_to_nearest)
3865
+ switch(abs)
3866
+ {
3867
+ case 0x32D4: return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-13, 1, 1));
3868
+ case 0x3B5B: return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-197, 1, 1));
3869
+ }
3870
+ return half(detail::binary, detail::area<half::round_style,true>(arg.data_));
3871
+ #endif
3872
+ }
3873
+
3874
+ /// Hyperbolic area cosine.
3875
+ /// This function is exact to rounding for all rounding modes.
3876
+ ///
3877
+ /// **See also:** Documentation for [std::acosh](https://en.cppreference.com/w/cpp/numeric/math/acosh).
3878
+ /// \param arg function argument
3879
+ /// \return area cosine value of \a arg
3880
+ /// \exception FE_INVALID for signaling NaN or arguments <1
3881
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3882
+ inline half acosh(half arg)
3883
+ {
3884
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3885
+ return half(detail::binary, detail::float2half<half::round_style>(std::acosh(detail::half2float<detail::internal_t>(arg.data_))));
3886
+ #else
3887
+ int abs = arg.data_ & 0x7FFF;
3888
+ if((arg.data_&0x8000) || abs < 0x3C00)
3889
+ return half(detail::binary, (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3890
+ if(abs == 0x3C00)
3891
+ return half(detail::binary, 0);
3892
+ if(arg.data_ >= 0x7C00)
3893
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
3894
+ return half(detail::binary, detail::area<half::round_style,false>(arg.data_));
3895
+ #endif
3896
+ }
3897
+
3898
+ /// Hyperbolic area tangent.
3899
+ /// This function is exact to rounding for all rounding modes.
3900
+ ///
3901
+ /// **See also:** Documentation for [std::atanh](https://en.cppreference.com/w/cpp/numeric/math/atanh).
3902
+ /// \param arg function argument
3903
+ /// \return area tangent value of \a arg
3904
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
3905
+ /// \exception FE_DIVBYZERO for +/-1
3906
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3907
+ inline half atanh(half arg)
3908
+ {
3909
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3910
+ return half(detail::binary, detail::float2half<half::round_style>(std::atanh(detail::half2float<detail::internal_t>(arg.data_))));
3911
+ #else
3912
+ int abs = arg.data_ & 0x7FFF, exp = 0;
3913
+ if(!abs)
3914
+ return arg;
3915
+ if(abs >= 0x3C00)
3916
+ return half(detail::binary, (abs==0x3C00) ? detail::pole(arg.data_&0x8000) : (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_));
3917
+ if(abs < 0x2700)
3918
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
3919
+ detail::uint32 m = static_cast<detail::uint32>((abs&0x3FF)|((abs>0x3FF)<<10)) << ((abs>>10)+(abs<=0x3FF)+6), my = 0x80000000 + m, mx = 0x80000000 - m;
3920
+ for(; mx<0x80000000; mx<<=1,++exp) ;
3921
+ int i = my >= mx, s;
3922
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(detail::log2(
3923
+ (detail::divide64(my>>i, mx, s)+1)>>1, 27)+0x10, exp+i-1, 16, arg.data_&0x8000));
3924
+ #endif
3925
+ }
3926
+
3927
+ /// \}
3928
+ /// \anchor special
3929
+ /// \name Error and gamma functions
3930
+ /// \{
3931
+
3932
+ /// Error function.
3933
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs.
3934
+ ///
3935
+ /// **See also:** Documentation for [std::erf](https://en.cppreference.com/w/cpp/numeric/math/erf).
3936
+ /// \param arg function argument
3937
+ /// \return error function value of \a arg
3938
+ /// \exception FE_INVALID for signaling NaN
3939
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3940
+ inline half erf(half arg)
3941
+ {
3942
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3943
+ return half(detail::binary, detail::float2half<half::round_style>(std::erf(detail::half2float<detail::internal_t>(arg.data_))));
3944
+ #else
3945
+ unsigned int abs = arg.data_ & 0x7FFF;
3946
+ if(!abs || abs >= 0x7C00)
3947
+ return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (arg.data_-0x4000) : detail::signal(arg.data_)) : arg;
3948
+ if(abs >= 0x4200)
3949
+ return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
3950
+ return half(detail::binary, detail::erf<half::round_style,false>(arg.data_));
3951
+ #endif
3952
+ }
3953
+
3954
+ /// Complementary error function.
3955
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs.
3956
+ ///
3957
+ /// **See also:** Documentation for [std::erfc](https://en.cppreference.com/w/cpp/numeric/math/erfc).
3958
+ /// \param arg function argument
3959
+ /// \return 1 minus error function value of \a arg
3960
+ /// \exception FE_INVALID for signaling NaN
3961
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3962
+ inline half erfc(half arg)
3963
+ {
3964
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3965
+ return half(detail::binary, detail::float2half<half::round_style>(std::erfc(detail::half2float<detail::internal_t>(arg.data_))));
3966
+ #else
3967
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
3968
+ if(abs >= 0x7C00)
3969
+ return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (sign>>1) : detail::signal(arg.data_)) : arg;
3970
+ if(!abs)
3971
+ return half(detail::binary, 0x3C00);
3972
+ if(abs >= 0x4400)
3973
+ return half(detail::binary, detail::rounded<half::round_style,true>((sign>>1)-(sign>>15), sign>>15, 1));
3974
+ return half(detail::binary, detail::erf<half::round_style,true>(arg.data_));
3975
+ #endif
3976
+ }
3977
+
3978
+ /// Natural logarithm of gamma function.
3979
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.025% of inputs.
3980
+ ///
3981
+ /// **See also:** Documentation for [std::lgamma](https://en.cppreference.com/w/cpp/numeric/math/lgamma).
3982
+ /// \param arg function argument
3983
+ /// \return natural logarith of gamma function for \a arg
3984
+ /// \exception FE_INVALID for signaling NaN
3985
+ /// \exception FE_DIVBYZERO for 0 or negative integer arguments
3986
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
3987
+ inline half lgamma(half arg)
3988
+ {
3989
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
3990
+ return half(detail::binary, detail::float2half<half::round_style>(std::lgamma(detail::half2float<detail::internal_t>(arg.data_))));
3991
+ #else
3992
+ int abs = arg.data_ & 0x7FFF;
3993
+ if(abs >= 0x7C00)
3994
+ return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_));
3995
+ if(!abs || arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1))))
3996
+ return half(detail::binary, detail::pole());
3997
+ if(arg.data_ == 0x3C00 || arg.data_ == 0x4000)
3998
+ return half(detail::binary, 0);
3999
+ return half(detail::binary, detail::gamma<half::round_style,true>(arg.data_));
4000
+ #endif
4001
+ }
4002
+
4003
+ /// Gamma function.
4004
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.25% of inputs.
4005
+ ///
4006
+ /// **See also:** Documentation for [std::tgamma](https://en.cppreference.com/w/cpp/numeric/math/tgamma).
4007
+ /// \param arg function argument
4008
+ /// \return gamma function value of \a arg
4009
+ /// \exception FE_INVALID for signaling NaN, negative infinity or negative integer arguments
4010
+ /// \exception FE_DIVBYZERO for 0
4011
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4012
+ inline half tgamma(half arg)
4013
+ {
4014
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
4015
+ return half(detail::binary, detail::float2half<half::round_style>(std::tgamma(detail::half2float<detail::internal_t>(arg.data_))));
4016
+ #else
4017
+ unsigned int abs = arg.data_ & 0x7FFF;
4018
+ if(!abs)
4019
+ return half(detail::binary, detail::pole(arg.data_));
4020
+ if(abs >= 0x7C00)
4021
+ return (arg.data_==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
4022
+ if(arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1))))
4023
+ return half(detail::binary, detail::invalid());
4024
+ if(arg.data_ >= 0xCA80)
4025
+ return half(detail::binary, detail::underflow<half::round_style>((1-((abs>>(25-(abs>>10)))&1))<<15));
4026
+ if(arg.data_ <= 0x100 || (arg.data_ >= 0x4900 && arg.data_ < 0x8000))
4027
+ return half(detail::binary, detail::overflow<half::round_style>());
4028
+ if(arg.data_ == 0x3C00)
4029
+ return arg;
4030
+ return half(detail::binary, detail::gamma<half::round_style,false>(arg.data_));
4031
+ #endif
4032
+ }
4033
+
4034
+ /// \}
4035
+ /// \anchor rounding
4036
+ /// \name Rounding
4037
+ /// \{
4038
+
4039
+ /// Nearest integer not less than half value.
4040
+ /// **See also:** Documentation for [std::ceil](https://en.cppreference.com/w/cpp/numeric/math/ceil).
4041
+ /// \param arg half to round
4042
+ /// \return nearest integer not less than \a arg
4043
+ /// \exception FE_INVALID for signaling NaN
4044
+ /// \exception FE_INEXACT if value had to be rounded
4045
+ inline half ceil(half arg) { return half(detail::binary, detail::integral<std::round_toward_infinity,true,true>(arg.data_)); }
4046
+
4047
+ /// Nearest integer not greater than half value.
4048
+ /// **See also:** Documentation for [std::floor](https://en.cppreference.com/w/cpp/numeric/math/floor).
4049
+ /// \param arg half to round
4050
+ /// \return nearest integer not greater than \a arg
4051
+ /// \exception FE_INVALID for signaling NaN
4052
+ /// \exception FE_INEXACT if value had to be rounded
4053
+ inline half floor(half arg) { return half(detail::binary, detail::integral<std::round_toward_neg_infinity,true,true>(arg.data_)); }
4054
+
4055
+ /// Nearest integer not greater in magnitude than half value.
4056
+ /// **See also:** Documentation for [std::trunc](https://en.cppreference.com/w/cpp/numeric/math/trunc).
4057
+ /// \param arg half to round
4058
+ /// \return nearest integer not greater in magnitude than \a arg
4059
+ /// \exception FE_INVALID for signaling NaN
4060
+ /// \exception FE_INEXACT if value had to be rounded
4061
+ inline half trunc(half arg) { return half(detail::binary, detail::integral<std::round_toward_zero,true,true>(arg.data_)); }
4062
+
4063
+ /// Nearest integer.
4064
+ /// **See also:** Documentation for [std::round](https://en.cppreference.com/w/cpp/numeric/math/round).
4065
+ /// \param arg half to round
4066
+ /// \return nearest integer, rounded away from zero in half-way cases
4067
+ /// \exception FE_INVALID for signaling NaN
4068
+ /// \exception FE_INEXACT if value had to be rounded
4069
+ inline half round(half arg) { return half(detail::binary, detail::integral<std::round_to_nearest,false,true>(arg.data_)); }
4070
+
4071
+ /// Nearest integer.
4072
+ /// **See also:** Documentation for [std::lround](https://en.cppreference.com/w/cpp/numeric/math/round).
4073
+ /// \param arg half to round
4074
+ /// \return nearest integer, rounded away from zero in half-way cases
4075
+ /// \exception FE_INVALID if value is not representable as `long`
4076
+ inline long lround(half arg) { return detail::half2int<std::round_to_nearest,false,false,long>(arg.data_); }
4077
+
4078
+ /// Nearest integer using half's internal rounding mode.
4079
+ /// **See also:** Documentation for [std::rint](https://en.cppreference.com/w/cpp/numeric/math/rint).
4080
+ /// \param arg half expression to round
4081
+ /// \return nearest integer using default rounding mode
4082
+ /// \exception FE_INVALID for signaling NaN
4083
+ /// \exception FE_INEXACT if value had to be rounded
4084
+ inline half rint(half arg) { return half(detail::binary, detail::integral<half::round_style,true,true>(arg.data_)); }
4085
+
4086
+ /// Nearest integer using half's internal rounding mode.
4087
+ /// **See also:** Documentation for [std::lrint](https://en.cppreference.com/w/cpp/numeric/math/rint).
4088
+ /// \param arg half expression to round
4089
+ /// \return nearest integer using default rounding mode
4090
+ /// \exception FE_INVALID if value is not representable as `long`
4091
+ /// \exception FE_INEXACT if value had to be rounded
4092
+ inline long lrint(half arg) { return detail::half2int<half::round_style,true,true,long>(arg.data_); }
4093
+
4094
+ /// Nearest integer using half's internal rounding mode.
4095
+ /// **See also:** Documentation for [std::nearbyint](https://en.cppreference.com/w/cpp/numeric/math/nearbyint).
4096
+ /// \param arg half expression to round
4097
+ /// \return nearest integer using default rounding mode
4098
+ /// \exception FE_INVALID for signaling NaN
4099
+ inline half nearbyint(half arg) { return half(detail::binary, detail::integral<half::round_style,true,false>(arg.data_)); }
4100
+ #if HALF_ENABLE_CPP11_LONG_LONG
4101
+ /// Nearest integer.
4102
+ /// **See also:** Documentation for [std::llround](https://en.cppreference.com/w/cpp/numeric/math/round).
4103
+ /// \param arg half to round
4104
+ /// \return nearest integer, rounded away from zero in half-way cases
4105
+ /// \exception FE_INVALID if value is not representable as `long long`
4106
+ inline long long llround(half arg) { return detail::half2int<std::round_to_nearest,false,false,long long>(arg.data_); }
4107
+
4108
+ /// Nearest integer using half's internal rounding mode.
4109
+ /// **See also:** Documentation for [std::llrint](https://en.cppreference.com/w/cpp/numeric/math/rint).
4110
+ /// \param arg half expression to round
4111
+ /// \return nearest integer using default rounding mode
4112
+ /// \exception FE_INVALID if value is not representable as `long long`
4113
+ /// \exception FE_INEXACT if value had to be rounded
4114
+ inline long long llrint(half arg) { return detail::half2int<half::round_style,true,true,long long>(arg.data_); }
4115
+ #endif
4116
+
4117
+ /// \}
4118
+ /// \anchor float
4119
+ /// \name Floating point manipulation
4120
+ /// \{
4121
+
4122
+ /// Decompress floating-point number.
4123
+ /// **See also:** Documentation for [std::frexp](https://en.cppreference.com/w/cpp/numeric/math/frexp).
4124
+ /// \param arg number to decompress
4125
+ /// \param exp address to store exponent at
4126
+ /// \return significant in range [0.5, 1)
4127
+ /// \exception FE_INVALID for signaling NaN
4128
+ inline half frexp(half arg, int *exp)
4129
+ {
4130
+ *exp = 0;
4131
+ unsigned int abs = arg.data_ & 0x7FFF;
4132
+ if(abs >= 0x7C00 || !abs)
4133
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
4134
+ for(; abs<0x400; abs<<=1,--*exp) ;
4135
+ *exp += (abs>>10) - 14;
4136
+ return half(detail::binary, (arg.data_&0x8000)|0x3800|(abs&0x3FF));
4137
+ }
4138
+
4139
+ /// Multiply by power of two.
4140
+ /// This function is exact to rounding for all rounding modes.
4141
+ ///
4142
+ /// **See also:** Documentation for [std::scalbln](https://en.cppreference.com/w/cpp/numeric/math/scalbn).
4143
+ /// \param arg number to modify
4144
+ /// \param exp power of two to multiply with
4145
+ /// \return \a arg multplied by 2 raised to \a exp
4146
+ /// \exception FE_INVALID for signaling NaN
4147
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4148
+ inline half scalbln(half arg, long exp)
4149
+ {
4150
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
4151
+ if(abs >= 0x7C00 || !abs)
4152
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
4153
+ for(; abs<0x400; abs<<=1,--exp) ;
4154
+ exp += abs >> 10;
4155
+ if(exp > 30)
4156
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
4157
+ else if(exp < -10)
4158
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
4159
+ else if(exp > 0)
4160
+ return half(detail::binary, sign|(exp<<10)|(abs&0x3FF));
4161
+ unsigned int m = (abs&0x3FF) | 0x400;
4162
+ return half(detail::binary, detail::rounded<half::round_style,false>(sign|(m>>(1-exp)), (m>>-exp)&1, (m&((1<<-exp)-1))!=0));
4163
+ }
4164
+
4165
+ /// Multiply by power of two.
4166
+ /// This function is exact to rounding for all rounding modes.
4167
+ ///
4168
+ /// **See also:** Documentation for [std::scalbn](https://en.cppreference.com/w/cpp/numeric/math/scalbn).
4169
+ /// \param arg number to modify
4170
+ /// \param exp power of two to multiply with
4171
+ /// \return \a arg multplied by 2 raised to \a exp
4172
+ /// \exception FE_INVALID for signaling NaN
4173
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4174
+ inline half scalbn(half arg, int exp) { return scalbln(arg, exp); }
4175
+
4176
+ /// Multiply by power of two.
4177
+ /// This function is exact to rounding for all rounding modes.
4178
+ ///
4179
+ /// **See also:** Documentation for [std::ldexp](https://en.cppreference.com/w/cpp/numeric/math/ldexp).
4180
+ /// \param arg number to modify
4181
+ /// \param exp power of two to multiply with
4182
+ /// \return \a arg multplied by 2 raised to \a exp
4183
+ /// \exception FE_INVALID for signaling NaN
4184
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4185
+ inline half ldexp(half arg, int exp) { return scalbln(arg, exp); }
4186
+
4187
+ /// Extract integer and fractional parts.
4188
+ /// **See also:** Documentation for [std::modf](https://en.cppreference.com/w/cpp/numeric/math/modf).
4189
+ /// \param arg number to decompress
4190
+ /// \param iptr address to store integer part at
4191
+ /// \return fractional part
4192
+ /// \exception FE_INVALID for signaling NaN
4193
+ inline half modf(half arg, half *iptr)
4194
+ {
4195
+ unsigned int abs = arg.data_ & 0x7FFF;
4196
+ if(abs > 0x7C00)
4197
+ {
4198
+ arg = half(detail::binary, detail::signal(arg.data_));
4199
+ return *iptr = arg, arg;
4200
+ }
4201
+ if(abs >= 0x6400)
4202
+ return *iptr = arg, half(detail::binary, arg.data_&0x8000);
4203
+ if(abs < 0x3C00)
4204
+ return iptr->data_ = arg.data_ & 0x8000, arg;
4205
+ unsigned int exp = abs >> 10, mask = (1<<(25-exp)) - 1, m = arg.data_ & mask;
4206
+ iptr->data_ = arg.data_ & ~mask;
4207
+ if(!m)
4208
+ return half(detail::binary, arg.data_&0x8000);
4209
+ for(; m<0x400; m<<=1,--exp) ;
4210
+ return half(detail::binary, (arg.data_&0x8000)|(exp<<10)|(m&0x3FF));
4211
+ }
4212
+
4213
+ /// Extract exponent.
4214
+ /// **See also:** Documentation for [std::ilogb](https://en.cppreference.com/w/cpp/numeric/math/ilogb).
4215
+ /// \param arg number to query
4216
+ /// \return floating-point exponent
4217
+ /// \retval FP_ILOGB0 for zero
4218
+ /// \retval FP_ILOGBNAN for NaN
4219
+ /// \retval INT_MAX for infinity
4220
+ /// \exception FE_INVALID for 0 or infinite values
4221
+ inline int ilogb(half arg)
4222
+ {
4223
+ int abs = arg.data_ & 0x7FFF, exp;
4224
+ if(!abs || abs >= 0x7C00)
4225
+ {
4226
+ detail::raise(FE_INVALID);
4227
+ return !abs ? FP_ILOGB0 : (abs==0x7C00) ? INT_MAX : FP_ILOGBNAN;
4228
+ }
4229
+ for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ;
4230
+ return exp;
4231
+ }
4232
+
4233
+ /// Extract exponent.
4234
+ /// **See also:** Documentation for [std::logb](https://en.cppreference.com/w/cpp/numeric/math/logb).
4235
+ /// \param arg number to query
4236
+ /// \return floating-point exponent
4237
+ /// \exception FE_INVALID for signaling NaN
4238
+ /// \exception FE_DIVBYZERO for 0
4239
+ inline half logb(half arg)
4240
+ {
4241
+ int abs = arg.data_ & 0x7FFF, exp;
4242
+ if(!abs)
4243
+ return half(detail::binary, detail::pole(0x8000));
4244
+ if(abs >= 0x7C00)
4245
+ return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_));
4246
+ for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ;
4247
+ unsigned int value = static_cast<unsigned>(exp<0) << 15;
4248
+ if(exp)
4249
+ {
4250
+ unsigned int m = std::abs(exp) << 6;
4251
+ for(exp=18; m<0x400; m<<=1,--exp) ;
4252
+ value |= (exp<<10) + m;
4253
+ }
4254
+ return half(detail::binary, value);
4255
+ }
4256
+
4257
+ /// Next representable value.
4258
+ /// **See also:** Documentation for [std::nextafter](https://en.cppreference.com/w/cpp/numeric/math/nextafter).
4259
+ /// \param from value to compute next representable value for
4260
+ /// \param to direction towards which to compute next value
4261
+ /// \return next representable value after \a from in direction towards \a to
4262
+ /// \exception FE_INVALID for signaling NaN
4263
+ /// \exception FE_OVERFLOW for infinite result from finite argument
4264
+ /// \exception FE_UNDERFLOW for subnormal result
4265
+ inline half nextafter(half from, half to)
4266
+ {
4267
+ int fabs = from.data_ & 0x7FFF, tabs = to.data_ & 0x7FFF;
4268
+ if(fabs > 0x7C00 || tabs > 0x7C00)
4269
+ return half(detail::binary, detail::signal(from.data_, to.data_));
4270
+ if(from.data_ == to.data_ || !(fabs|tabs))
4271
+ return to;
4272
+ if(!fabs)
4273
+ {
4274
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT);
4275
+ return half(detail::binary, (to.data_&0x8000)+1);
4276
+ }
4277
+ unsigned int out = from.data_ + (((from.data_>>15)^static_cast<unsigned>(
4278
+ (from.data_^(0x8000|(0x8000-(from.data_>>15))))<(to.data_^(0x8000|(0x8000-(to.data_>>15))))))<<1) - 1;
4279
+ detail::raise(FE_OVERFLOW, fabs<0x7C00 && (out&0x7C00)==0x7C00);
4280
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT && (out&0x7C00)<0x400);
4281
+ return half(detail::binary, out);
4282
+ }
4283
+
4284
+ /// Next representable value.
4285
+ /// **See also:** Documentation for [std::nexttoward](https://en.cppreference.com/w/cpp/numeric/math/nexttoward).
4286
+ /// \param from value to compute next representable value for
4287
+ /// \param to direction towards which to compute next value
4288
+ /// \return next representable value after \a from in direction towards \a to
4289
+ /// \exception FE_INVALID for signaling NaN
4290
+ /// \exception FE_OVERFLOW for infinite result from finite argument
4291
+ /// \exception FE_UNDERFLOW for subnormal result
4292
+ inline half nexttoward(half from, long double to)
4293
+ {
4294
+ int fabs = from.data_ & 0x7FFF;
4295
+ if(fabs > 0x7C00)
4296
+ return half(detail::binary, detail::signal(from.data_));
4297
+ long double lfrom = static_cast<long double>(from);
4298
+ if(detail::builtin_isnan(to) || lfrom == to)
4299
+ return half(static_cast<float>(to));
4300
+ if(!fabs)
4301
+ {
4302
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT);
4303
+ return half(detail::binary, (static_cast<unsigned>(detail::builtin_signbit(to))<<15)+1);
4304
+ }
4305
+ unsigned int out = from.data_ + (((from.data_>>15)^static_cast<unsigned>(lfrom<to))<<1) - 1;
4306
+ detail::raise(FE_OVERFLOW, (out&0x7FFF)==0x7C00);
4307
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT && (out&0x7FFF)<0x400);
4308
+ return half(detail::binary, out);
4309
+ }
4310
+
4311
+ /// Take sign.
4312
+ /// **See also:** Documentation for [std::copysign](https://en.cppreference.com/w/cpp/numeric/math/copysign).
4313
+ /// \param x value to change sign for
4314
+ /// \param y value to take sign from
4315
+ /// \return value equal to \a x in magnitude and to \a y in sign
4316
+ inline HALF_CONSTEXPR half copysign(half x, half y) { return half(detail::binary, x.data_^((x.data_^y.data_)&0x8000)); }
4317
+
4318
+ /// \}
4319
+ /// \anchor classification
4320
+ /// \name Floating point classification
4321
+ /// \{
4322
+
4323
+ /// Classify floating-point value.
4324
+ /// **See also:** Documentation for [std::fpclassify](https://en.cppreference.com/w/cpp/numeric/math/fpclassify).
4325
+ /// \param arg number to classify
4326
+ /// \retval FP_ZERO for positive and negative zero
4327
+ /// \retval FP_SUBNORMAL for subnormal numbers
4328
+ /// \retval FP_INFINITY for positive and negative infinity
4329
+ /// \retval FP_NAN for NaNs
4330
+ /// \retval FP_NORMAL for all other (normal) values
4331
+ inline HALF_CONSTEXPR int fpclassify(half arg)
4332
+ {
4333
+ return !(arg.data_&0x7FFF) ? FP_ZERO :
4334
+ ((arg.data_&0x7FFF)<0x400) ? FP_SUBNORMAL :
4335
+ ((arg.data_&0x7FFF)<0x7C00) ? FP_NORMAL :
4336
+ ((arg.data_&0x7FFF)==0x7C00) ? FP_INFINITE :
4337
+ FP_NAN;
4338
+ }
4339
+
4340
+ /// Check if finite number.
4341
+ /// **See also:** Documentation for [std::isfinite](https://en.cppreference.com/w/cpp/numeric/math/isfinite).
4342
+ /// \param arg number to check
4343
+ /// \retval true if neither infinity nor NaN
4344
+ /// \retval false else
4345
+ inline HALF_CONSTEXPR bool isfinite(half arg) { return (arg.data_&0x7C00) != 0x7C00; }
4346
+
4347
+ /// Check for infinity.
4348
+ /// **See also:** Documentation for [std::isinf](https://en.cppreference.com/w/cpp/numeric/math/isinf).
4349
+ /// \param arg number to check
4350
+ /// \retval true for positive or negative infinity
4351
+ /// \retval false else
4352
+ inline HALF_CONSTEXPR bool isinf(half arg) { return (arg.data_&0x7FFF) == 0x7C00; }
4353
+
4354
+ /// Check for NaN.
4355
+ /// **See also:** Documentation for [std::isnan](https://en.cppreference.com/w/cpp/numeric/math/isnan).
4356
+ /// \param arg number to check
4357
+ /// \retval true for NaNs
4358
+ /// \retval false else
4359
+ inline HALF_CONSTEXPR bool isnan(half arg) { return (arg.data_&0x7FFF) > 0x7C00; }
4360
+
4361
+ /// Check if normal number.
4362
+ /// **See also:** Documentation for [std::isnormal](https://en.cppreference.com/w/cpp/numeric/math/isnormal).
4363
+ /// \param arg number to check
4364
+ /// \retval true if normal number
4365
+ /// \retval false if either subnormal, zero, infinity or NaN
4366
+ inline HALF_CONSTEXPR bool isnormal(half arg) { return ((arg.data_&0x7C00)!=0) & ((arg.data_&0x7C00)!=0x7C00); }
4367
+
4368
+ /// Check sign.
4369
+ /// **See also:** Documentation for [std::signbit](https://en.cppreference.com/w/cpp/numeric/math/signbit).
4370
+ /// \param arg number to check
4371
+ /// \retval true for negative number
4372
+ /// \retval false for positive number
4373
+ inline HALF_CONSTEXPR bool signbit(half arg) { return (arg.data_&0x8000) != 0; }
4374
+
4375
+ /// \}
4376
+ /// \anchor compfunc
4377
+ /// \name Comparison
4378
+ /// \{
4379
+
4380
+ /// Quiet comparison for greater than.
4381
+ /// **See also:** Documentation for [std::isgreater](https://en.cppreference.com/w/cpp/numeric/math/isgreater).
4382
+ /// \param x first operand
4383
+ /// \param y second operand
4384
+ /// \retval true if \a x greater than \a y
4385
+ /// \retval false else
4386
+ inline HALF_CONSTEXPR bool isgreater(half x, half y)
4387
+ {
4388
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
4389
+ }
4390
+
4391
+ /// Quiet comparison for greater equal.
4392
+ /// **See also:** Documentation for [std::isgreaterequal](https://en.cppreference.com/w/cpp/numeric/math/isgreaterequal).
4393
+ /// \param x first operand
4394
+ /// \param y second operand
4395
+ /// \retval true if \a x greater equal \a y
4396
+ /// \retval false else
4397
+ inline HALF_CONSTEXPR bool isgreaterequal(half x, half y)
4398
+ {
4399
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
4400
+ }
4401
+
4402
+ /// Quiet comparison for less than.
4403
+ /// **See also:** Documentation for [std::isless](https://en.cppreference.com/w/cpp/numeric/math/isless).
4404
+ /// \param x first operand
4405
+ /// \param y second operand
4406
+ /// \retval true if \a x less than \a y
4407
+ /// \retval false else
4408
+ inline HALF_CONSTEXPR bool isless(half x, half y)
4409
+ {
4410
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
4411
+ }
4412
+
4413
+ /// Quiet comparison for less equal.
4414
+ /// **See also:** Documentation for [std::islessequal](https://en.cppreference.com/w/cpp/numeric/math/islessequal).
4415
+ /// \param x first operand
4416
+ /// \param y second operand
4417
+ /// \retval true if \a x less equal \a y
4418
+ /// \retval false else
4419
+ inline HALF_CONSTEXPR bool islessequal(half x, half y)
4420
+ {
4421
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
4422
+ }
4423
+
4424
+ /// Quiet comarison for less or greater.
4425
+ /// **See also:** Documentation for [std::islessgreater](https://en.cppreference.com/w/cpp/numeric/math/islessgreater).
4426
+ /// \param x first operand
4427
+ /// \param y second operand
4428
+ /// \retval true if either less or greater
4429
+ /// \retval false else
4430
+ inline HALF_CONSTEXPR bool islessgreater(half x, half y)
4431
+ {
4432
+ return x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF) && !isnan(x) && !isnan(y);
4433
+ }
4434
+
4435
+ /// Quiet check if unordered.
4436
+ /// **See also:** Documentation for [std::isunordered](https://en.cppreference.com/w/cpp/numeric/math/isunordered).
4437
+ /// \param x first operand
4438
+ /// \param y second operand
4439
+ /// \retval true if unordered (one or two NaN operands)
4440
+ /// \retval false else
4441
+ inline HALF_CONSTEXPR bool isunordered(half x, half y) { return isnan(x) || isnan(y); }
4442
+
4443
+ /// \}
4444
+ /// \anchor casting
4445
+ /// \name Casting
4446
+ /// \{
4447
+
4448
+ /// Cast to or from half-precision floating-point number.
4449
+ /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted
4450
+ /// directly using the default rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do.
4451
+ ///
4452
+ /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types
4453
+ /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler
4454
+ /// error and casting between [half](\ref half_float::half)s returns the argument unmodified.
4455
+ /// \tparam T destination type (half or built-in arithmetic type)
4456
+ /// \tparam U source type (half or built-in arithmetic type)
4457
+ /// \param arg value to cast
4458
+ /// \return \a arg converted to destination type
4459
+ /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T
4460
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4461
+ template<typename T,typename U> T half_cast(U arg) { return detail::half_caster<T,U>::cast(arg); }
4462
+
4463
+ /// Cast to or from half-precision floating-point number.
4464
+ /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted
4465
+ /// directly using the specified rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do.
4466
+ ///
4467
+ /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types
4468
+ /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler
4469
+ /// error and casting between [half](\ref half_float::half)s returns the argument unmodified.
4470
+ /// \tparam T destination type (half or built-in arithmetic type)
4471
+ /// \tparam R rounding mode to use.
4472
+ /// \tparam U source type (half or built-in arithmetic type)
4473
+ /// \param arg value to cast
4474
+ /// \return \a arg converted to destination type
4475
+ /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T
4476
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
4477
+ template<typename T,std::float_round_style R,typename U> T half_cast(U arg) { return detail::half_caster<T,U,R>::cast(arg); }
4478
+ /// \}
4479
+
4480
+ /// \}
4481
+ /// \anchor errors
4482
+ /// \name Error handling
4483
+ /// \{
4484
+
4485
+ /// Clear exception flags.
4486
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4487
+ /// but in that case manual flag management is the only way to raise flags.
4488
+ ///
4489
+ /// **See also:** Documentation for [std::feclearexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feclearexcept).
4490
+ /// \param excepts OR of exceptions to clear
4491
+ /// \retval 0 all selected flags cleared successfully
4492
+ inline int feclearexcept(int excepts) { detail::errflags() &= ~excepts; return 0; }
4493
+
4494
+ /// Test exception flags.
4495
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4496
+ /// but in that case manual flag management is the only way to raise flags.
4497
+ ///
4498
+ /// **See also:** Documentation for [std::fetestexcept](https://en.cppreference.com/w/cpp/numeric/fenv/fetestexcept).
4499
+ /// \param excepts OR of exceptions to test
4500
+ /// \return OR of selected exceptions if raised
4501
+ inline int fetestexcept(int excepts) { return detail::errflags() & excepts; }
4502
+
4503
+ /// Raise exception flags.
4504
+ /// This raises the specified floating point exceptions and also invokes any additional automatic exception handling as
4505
+ /// configured with the [HALF_ERRHANDLIG_...](\ref HALF_ERRHANDLING_ERRNO) preprocessor symbols.
4506
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4507
+ /// but in that case manual flag management is the only way to raise flags.
4508
+ ///
4509
+ /// **See also:** Documentation for [std::feraiseexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feraiseexcept).
4510
+ /// \param excepts OR of exceptions to raise
4511
+ /// \retval 0 all selected exceptions raised successfully
4512
+ inline int feraiseexcept(int excepts) { detail::errflags() |= excepts; detail::raise(excepts); return 0; }
4513
+
4514
+ /// Save exception flags.
4515
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4516
+ /// but in that case manual flag management is the only way to raise flags.
4517
+ ///
4518
+ /// **See also:** Documentation for [std::fegetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag).
4519
+ /// \param flagp adress to store flag state at
4520
+ /// \param excepts OR of flags to save
4521
+ /// \retval 0 for success
4522
+ inline int fegetexceptflag(int *flagp, int excepts) { *flagp = detail::errflags() & excepts; return 0; }
4523
+
4524
+ /// Restore exception flags.
4525
+ /// This only copies the specified exception state (including unset flags) without incurring any additional exception handling.
4526
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4527
+ /// but in that case manual flag management is the only way to raise flags.
4528
+ ///
4529
+ /// **See also:** Documentation for [std::fesetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag).
4530
+ /// \param flagp adress to take flag state from
4531
+ /// \param excepts OR of flags to restore
4532
+ /// \retval 0 for success
4533
+ inline int fesetexceptflag(const int *flagp, int excepts) { detail::errflags() = (detail::errflags()|(*flagp&excepts)) & (*flagp|~excepts); return 0; }
4534
+
4535
+ /// Throw C++ exceptions based on set exception flags.
4536
+ /// This function manually throws a corresponding C++ exception if one of the specified flags is set,
4537
+ /// no matter if automatic throwing (via [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID)) is enabled or not.
4538
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
4539
+ /// but in that case manual flag management is the only way to raise flags.
4540
+ /// \param excepts OR of exceptions to test
4541
+ /// \param msg error message to use for exception description
4542
+ /// \throw std::domain_error if `FE_INVALID` or `FE_DIVBYZERO` is selected and set
4543
+ /// \throw std::overflow_error if `FE_OVERFLOW` is selected and set
4544
+ /// \throw std::underflow_error if `FE_UNDERFLOW` is selected and set
4545
+ /// \throw std::range_error if `FE_INEXACT` is selected and set
4546
+ inline void fethrowexcept(int excepts, const char *msg = "")
4547
+ {
4548
+ excepts &= detail::errflags();
4549
+ if(excepts & (FE_INVALID|FE_DIVBYZERO))
4550
+ throw std::domain_error(msg);
4551
+ if(excepts & FE_OVERFLOW)
4552
+ throw std::overflow_error(msg);
4553
+ if(excepts & FE_UNDERFLOW)
4554
+ throw std::underflow_error(msg);
4555
+ if(excepts & FE_INEXACT)
4556
+ throw std::range_error(msg);
4557
+ }
4558
+ /// \}
4559
+ }
4560
+
4561
+
4562
+ #undef HALF_UNUSED_NOERR
4563
+ #undef HALF_CONSTEXPR
4564
+ #undef HALF_CONSTEXPR_CONST
4565
+ #undef HALF_CONSTEXPR_NOERR
4566
+ #undef HALF_NOEXCEPT
4567
+ #undef HALF_NOTHROW
4568
+ #undef HALF_THREAD_LOCAL
4569
+ #undef HALF_TWOS_COMPLEMENT_INT
4570
+ #ifdef HALF_POP_WARNINGS
4571
+ #pragma warning(pop)
4572
+ #undef HALF_POP_WARNINGS
4573
+ #endif
4574
+
4575
+ #endif