1/****************************************************************************
2**
3** Copyright (C) 2019 The Qt Company Ltd.
4** Copyright (C) 2016 by Southwest Research Institute (R)
5** Contact: http://www.qt-project.org/legal
6**
7** This file is part of the QtCore module of the Qt Toolkit.
8**
9** $QT_BEGIN_LICENSE:LGPL$
10** Commercial License Usage
11** Licensees holding valid commercial Qt licenses may use this file in
12** accordance with the commercial license agreement provided with the
13** Software or, alternatively, in accordance with the terms contained in
14** a written agreement between you and The Qt Company. For licensing terms
15** and conditions see https://www.qt.io/terms-conditions. For further
16** information use the contact form at https://www.qt.io/contact-us.
17**
18** GNU Lesser General Public License Usage
19** Alternatively, this file may be used under the terms of the GNU Lesser
20** General Public License version 3 as published by the Free Software
21** Foundation and appearing in the file LICENSE.LGPL3 included in the
22** packaging of this file. Please review the following information to
23** ensure the GNU Lesser General Public License version 3 requirements
24** will be met: https://www.gnu.org/licenses/lgpl-3.0.html.
25**
26** GNU General Public License Usage
27** Alternatively, this file may be used under the terms of the GNU
28** General Public License version 2.0 or (at your option) the GNU General
29** Public license version 3 or any later version approved by the KDE Free
30** Qt Foundation. The licenses are as published by the Free Software
31** Foundation and appearing in the file LICENSE.GPL2 and LICENSE.GPL3
32** included in the packaging of this file. Please review the following
33** information to ensure the GNU General Public License requirements will
34** be met: https://www.gnu.org/licenses/gpl-2.0.html and
35** https://www.gnu.org/licenses/gpl-3.0.html.
36**
37** $QT_END_LICENSE$
38**
39****************************************************************************/
40
41#ifndef QFLOAT16_H
42#define QFLOAT16_H
43
44#include <QtCore/qglobal.h>
45#include <QtCore/qmetatype.h>
46#include <limits>
47#include <string.h>
48
49#if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__AVX2__) && !defined(__F16C__)
50// All processors that support AVX2 do support F16C too. That doesn't mean
51// we're allowed to use the intrinsics directly, so we'll do it only for
52// the Intel and Microsoft's compilers.
53# if defined(Q_CC_INTEL) || defined(Q_CC_MSVC)
54# define __F16C__ 1
55# endif
56#endif
57
58#if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__F16C__)
59#include <immintrin.h>
60#endif
61
62QT_BEGIN_NAMESPACE
63
64#if 0
65#pragma qt_class(QFloat16)
66#pragma qt_no_master_include
67#endif
68
69class qfloat16
70{
71 struct Wrap
72 {
73 // To let our private constructor work, without other code seeing
74 // ambiguity when constructing from int, double &c.
75 quint16 b16;
76 constexpr inline explicit Wrap(int value) : b16(value) {}
77 };
78public:
79 constexpr inline qfloat16() noexcept : b16(0) {}
80 inline qfloat16(float f) noexcept;
81 inline operator float() const noexcept;
82
83 // Support for qIs{Inf,NaN,Finite}:
84 bool isInf() const noexcept { return (b16 & 0x7fff) == 0x7c00; }
85 bool isNaN() const noexcept { return (b16 & 0x7fff) > 0x7c00; }
86 bool isFinite() const noexcept { return (b16 & 0x7fff) < 0x7c00; }
87 Q_CORE_EXPORT int fpClassify() const noexcept;
88 // Can't specialize std::copysign() for qfloat16
89 qfloat16 copySign(qfloat16 sign) const noexcept
90 { return qfloat16(Wrap((sign.b16 & 0x8000) | (b16 & 0x7fff))); }
91 // Support for std::numeric_limits<qfloat16>
92 static constexpr qfloat16 _limit_epsilon() noexcept { return qfloat16(Wrap(0x1400)); }
93 static constexpr qfloat16 _limit_min() noexcept { return qfloat16(Wrap(0x400)); }
94 static constexpr qfloat16 _limit_denorm_min() noexcept { return qfloat16(Wrap(1)); }
95 static constexpr qfloat16 _limit_max() noexcept { return qfloat16(Wrap(0x7bff)); }
96 static constexpr qfloat16 _limit_lowest() noexcept { return qfloat16(Wrap(0xfbff)); }
97 static constexpr qfloat16 _limit_infinity() noexcept { return qfloat16(Wrap(0x7c00)); }
98 static constexpr qfloat16 _limit_quiet_NaN() noexcept { return qfloat16(Wrap(0x7e00)); }
99#if QT_CONFIG(signaling_nan)
100 static constexpr qfloat16 _limit_signaling_NaN() noexcept { return qfloat16(Wrap(0x7d00)); }
101#endif
102 inline constexpr bool isNormal() const noexcept
103 { return (b16 & 0x7c00) && (b16 & 0x7c00) != 0x7c00; }
104private:
105 quint16 b16;
106 constexpr inline explicit qfloat16(Wrap nibble) noexcept : b16(nibble.b16) {}
107
108 Q_CORE_EXPORT static const quint32 mantissatable[];
109 Q_CORE_EXPORT static const quint32 exponenttable[];
110 Q_CORE_EXPORT static const quint32 offsettable[];
111 Q_CORE_EXPORT static const quint32 basetable[];
112 Q_CORE_EXPORT static const quint32 shifttable[];
113
114 friend bool qIsNull(qfloat16 f) noexcept;
115#if !defined(QT_NO_FLOAT16_OPERATORS)
116 friend qfloat16 operator-(qfloat16 a) noexcept;
117#endif
118};
119
120Q_DECLARE_TYPEINFO(qfloat16, Q_PRIMITIVE_TYPE);
121
122Q_CORE_EXPORT void qFloatToFloat16(qfloat16 *, const float *, qsizetype length) noexcept;
123Q_CORE_EXPORT void qFloatFromFloat16(float *, const qfloat16 *, qsizetype length) noexcept;
124
125// Complement qnumeric.h:
126Q_REQUIRED_RESULT inline bool qIsInf(qfloat16 f) noexcept { return f.isInf(); }
127Q_REQUIRED_RESULT inline bool qIsNaN(qfloat16 f) noexcept { return f.isNaN(); }
128Q_REQUIRED_RESULT inline bool qIsFinite(qfloat16 f) noexcept { return f.isFinite(); }
129Q_REQUIRED_RESULT inline int qFpClassify(qfloat16 f) noexcept { return f.fpClassify(); }
130// Q_REQUIRED_RESULT quint32 qFloatDistance(qfloat16 a, qfloat16 b);
131
132// The remainder of these utility functions complement qglobal.h
133Q_REQUIRED_RESULT inline int qRound(qfloat16 d) noexcept
134{ return qRound(static_cast<float>(d)); }
135
136Q_REQUIRED_RESULT inline qint64 qRound64(qfloat16 d) noexcept
137{ return qRound64(static_cast<float>(d)); }
138
139Q_REQUIRED_RESULT inline bool qFuzzyCompare(qfloat16 p1, qfloat16 p2) noexcept
140{
141 float f1 = static_cast<float>(p1);
142 float f2 = static_cast<float>(p2);
143 // The significand precision for IEEE754 half precision is
144 // 11 bits (10 explicitly stored), or approximately 3 decimal
145 // digits. In selecting the fuzzy comparison factor of 102.5f
146 // (that is, (2^10+1)/10) below, we effectively select a
147 // window of about 1 (least significant) decimal digit about
148 // which the two operands can vary and still return true.
149 return (qAbs(f1 - f2) * 102.5f <= qMin(qAbs(f1), qAbs(f2)));
150}
151
152Q_REQUIRED_RESULT inline bool qIsNull(qfloat16 f) noexcept
153{
154 return (f.b16 & static_cast<quint16>(0x7fff)) == 0;
155}
156
157inline int qIntCast(qfloat16 f) noexcept
158{ return int(static_cast<float>(f)); }
159
160#ifndef Q_QDOC
161QT_WARNING_PUSH
162QT_WARNING_DISABLE_CLANG("-Wc99-extensions")
163QT_WARNING_DISABLE_GCC("-Wold-style-cast")
164inline qfloat16::qfloat16(float f) noexcept
165{
166#if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__F16C__)
167 __m128 packsingle = _mm_set_ss(f);
168 __m128i packhalf = _mm_cvtps_ph(packsingle, 0);
169 b16 = _mm_extract_epi16(packhalf, 0);
170#elif defined (__ARM_FP16_FORMAT_IEEE)
171 __fp16 f16 = __fp16(f);
172 memcpy(&b16, &f16, sizeof(quint16));
173#else
174 quint32 u;
175 memcpy(&u, &f, sizeof(quint32));
176 b16 = basetable[(u >> 23) & 0x1ff]
177 + ((u & 0x007fffff) >> shifttable[(u >> 23) & 0x1ff]);
178#endif
179}
180QT_WARNING_POP
181
182inline qfloat16::operator float() const noexcept
183{
184#if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__F16C__)
185 __m128i packhalf = _mm_cvtsi32_si128(b16);
186 __m128 packsingle = _mm_cvtph_ps(packhalf);
187 return _mm_cvtss_f32(packsingle);
188#elif defined (__ARM_FP16_FORMAT_IEEE)
189 __fp16 f16;
190 memcpy(&f16, &b16, sizeof(quint16));
191 return float(f16);
192#else
193 quint32 u = mantissatable[offsettable[b16 >> 10] + (b16 & 0x3ff)]
194 + exponenttable[b16 >> 10];
195 float f;
196 memcpy(&f, &u, sizeof(quint32));
197 return f;
198#endif
199}
200#endif
201
202#if !defined(QT_NO_FLOAT16_OPERATORS)
203inline qfloat16 operator-(qfloat16 a) noexcept
204{
205 qfloat16 f;
206 f.b16 = a.b16 ^ quint16(0x8000);
207 return f;
208}
209
210inline qfloat16 operator+(qfloat16 a, qfloat16 b) noexcept { return qfloat16(static_cast<float>(a) + static_cast<float>(b)); }
211inline qfloat16 operator-(qfloat16 a, qfloat16 b) noexcept { return qfloat16(static_cast<float>(a) - static_cast<float>(b)); }
212inline qfloat16 operator*(qfloat16 a, qfloat16 b) noexcept { return qfloat16(static_cast<float>(a) * static_cast<float>(b)); }
213inline qfloat16 operator/(qfloat16 a, qfloat16 b) noexcept { return qfloat16(static_cast<float>(a) / static_cast<float>(b)); }
214
215#define QF16_MAKE_ARITH_OP_FP(FP, OP) \
216 inline FP operator OP(qfloat16 lhs, FP rhs) noexcept { return static_cast<FP>(lhs) OP rhs; } \
217 inline FP operator OP(FP lhs, qfloat16 rhs) noexcept { return lhs OP static_cast<FP>(rhs); }
218#define QF16_MAKE_ARITH_OP_EQ_FP(FP, OP_EQ, OP) \
219 inline qfloat16& operator OP_EQ(qfloat16& lhs, FP rhs) noexcept \
220 { lhs = qfloat16(float(static_cast<FP>(lhs) OP rhs)); return lhs; }
221#define QF16_MAKE_ARITH_OP(FP) \
222 QF16_MAKE_ARITH_OP_FP(FP, +) \
223 QF16_MAKE_ARITH_OP_FP(FP, -) \
224 QF16_MAKE_ARITH_OP_FP(FP, *) \
225 QF16_MAKE_ARITH_OP_FP(FP, /) \
226 QF16_MAKE_ARITH_OP_EQ_FP(FP, +=, +) \
227 QF16_MAKE_ARITH_OP_EQ_FP(FP, -=, -) \
228 QF16_MAKE_ARITH_OP_EQ_FP(FP, *=, *) \
229 QF16_MAKE_ARITH_OP_EQ_FP(FP, /=, /)
230QF16_MAKE_ARITH_OP(long double)
231QF16_MAKE_ARITH_OP(double)
232QF16_MAKE_ARITH_OP(float)
233#undef QF16_MAKE_ARITH_OP
234#undef QF16_MAKE_ARITH_OP_FP
235
236#define QF16_MAKE_ARITH_OP_INT(OP) \
237 inline double operator OP(qfloat16 lhs, int rhs) noexcept { return static_cast<double>(lhs) OP rhs; } \
238 inline double operator OP(int lhs, qfloat16 rhs) noexcept { return lhs OP static_cast<double>(rhs); }
239QF16_MAKE_ARITH_OP_INT(+)
240QF16_MAKE_ARITH_OP_INT(-)
241QF16_MAKE_ARITH_OP_INT(*)
242QF16_MAKE_ARITH_OP_INT(/)
243#undef QF16_MAKE_ARITH_OP_INT
244
245QT_WARNING_PUSH
246QT_WARNING_DISABLE_CLANG("-Wfloat-equal")
247QT_WARNING_DISABLE_GCC("-Wfloat-equal")
248QT_WARNING_DISABLE_INTEL(1572)
249
250inline bool operator>(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) > static_cast<float>(b); }
251inline bool operator<(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) < static_cast<float>(b); }
252inline bool operator>=(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) >= static_cast<float>(b); }
253inline bool operator<=(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) <= static_cast<float>(b); }
254inline bool operator==(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) == static_cast<float>(b); }
255inline bool operator!=(qfloat16 a, qfloat16 b) noexcept { return static_cast<float>(a) != static_cast<float>(b); }
256
257#define QF16_MAKE_BOOL_OP_FP(FP, OP) \
258 inline bool operator OP(qfloat16 lhs, FP rhs) noexcept { return static_cast<FP>(lhs) OP rhs; } \
259 inline bool operator OP(FP lhs, qfloat16 rhs) noexcept { return lhs OP static_cast<FP>(rhs); }
260#define QF16_MAKE_BOOL_OP(FP) \
261 QF16_MAKE_BOOL_OP_FP(FP, <) \
262 QF16_MAKE_BOOL_OP_FP(FP, >) \
263 QF16_MAKE_BOOL_OP_FP(FP, >=) \
264 QF16_MAKE_BOOL_OP_FP(FP, <=) \
265 QF16_MAKE_BOOL_OP_FP(FP, ==) \
266 QF16_MAKE_BOOL_OP_FP(FP, !=)
267QF16_MAKE_BOOL_OP(long double)
268QF16_MAKE_BOOL_OP(double)
269QF16_MAKE_BOOL_OP(float)
270#undef QF16_MAKE_BOOL_OP
271#undef QF16_MAKE_BOOL_OP_FP
272
273#define QF16_MAKE_BOOL_OP_INT(OP) \
274 inline bool operator OP(qfloat16 a, int b) noexcept { return static_cast<float>(a) OP b; } \
275 inline bool operator OP(int a, qfloat16 b) noexcept { return a OP static_cast<float>(b); }
276QF16_MAKE_BOOL_OP_INT(>)
277QF16_MAKE_BOOL_OP_INT(<)
278QF16_MAKE_BOOL_OP_INT(>=)
279QF16_MAKE_BOOL_OP_INT(<=)
280QF16_MAKE_BOOL_OP_INT(==)
281QF16_MAKE_BOOL_OP_INT(!=)
282#undef QF16_MAKE_BOOL_OP_INT
283
284QT_WARNING_POP
285#endif // QT_NO_FLOAT16_OPERATORS
286
287/*!
288 \internal
289*/
290Q_REQUIRED_RESULT inline bool qFuzzyIsNull(qfloat16 f) noexcept
291{
292 return qAbs(static_cast<float>(f)) <= 0.001f;
293}
294
295QT_END_NAMESPACE
296
297Q_DECLARE_METATYPE(qfloat16)
298
299namespace std {
300template<>
301class numeric_limits<QT_PREPEND_NAMESPACE(qfloat16)> : public numeric_limits<float>
302{
303public:
304 /*
305 Treat quint16 b16 as if it were:
306 uint S: 1; // b16 >> 15 (sign); can be set for zero
307 uint E: 5; // (b16 >> 10) & 0x1f (offset exponent)
308 uint M: 10; // b16 & 0x3ff (adjusted mantissa)
309
310 for E == 0: magnitude is M / 2.^{24}
311 for 0 < E < 31: magnitude is (1. + M / 2.^{10}) * 2.^{E - 15)
312 for E == 31: not finite
313 */
314 static constexpr int digits = 11;
315 static constexpr int min_exponent = -13;
316 static constexpr int max_exponent = 16;
317
318 static constexpr int digits10 = 3;
319 static constexpr int max_digits10 = 5;
320 static constexpr int min_exponent10 = -4;
321 static constexpr int max_exponent10 = 4;
322
323 static constexpr QT_PREPEND_NAMESPACE(qfloat16) epsilon()
324 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_epsilon(); }
325 static constexpr QT_PREPEND_NAMESPACE(qfloat16) (min)()
326 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_min(); }
327 static constexpr QT_PREPEND_NAMESPACE(qfloat16) denorm_min()
328 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_denorm_min(); }
329 static constexpr QT_PREPEND_NAMESPACE(qfloat16) (max)()
330 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_max(); }
331 static constexpr QT_PREPEND_NAMESPACE(qfloat16) lowest()
332 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_lowest(); }
333 static constexpr QT_PREPEND_NAMESPACE(qfloat16) infinity()
334 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_infinity(); }
335 static constexpr QT_PREPEND_NAMESPACE(qfloat16) quiet_NaN()
336 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_quiet_NaN(); }
337#if QT_CONFIG(signaling_nan)
338 static constexpr QT_PREPEND_NAMESPACE(qfloat16) signaling_NaN()
339 { return QT_PREPEND_NAMESPACE(qfloat16)::_limit_signaling_NaN(); }
340#else
341 static constexpr bool has_signaling_NaN = false;
342#endif
343};
344
345template<> class numeric_limits<const QT_PREPEND_NAMESPACE(qfloat16)>
346 : public numeric_limits<QT_PREPEND_NAMESPACE(qfloat16)> {};
347template<> class numeric_limits<volatile QT_PREPEND_NAMESPACE(qfloat16)>
348 : public numeric_limits<QT_PREPEND_NAMESPACE(qfloat16)> {};
349template<> class numeric_limits<const volatile QT_PREPEND_NAMESPACE(qfloat16)>
350 : public numeric_limits<QT_PREPEND_NAMESPACE(qfloat16)> {};
351
352// Adding overloads to std isn't allowed, so we can't extend this to support
353// for fpclassify(), isnormal() &c. (which, furthermore, are macros on MinGW).
354} // namespace std
355
356#endif // QFLOAT16_H
357