summaryrefslogtreecommitdiffstats
path: root/src/corelib/global/qfloat16.cpp
blob: d816d8b888d140e7b4a6ccd5b52137784196c733 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
/****************************************************************************
**
** Copyright (C) 2020 The Qt Company Ltd.
** Copyright (C) 2016 by Southwest Research Institute (R)
** Contact: http://www.qt-project.org/legal
**
** This file is part of the QtCore module of the Qt Toolkit.
**
** $QT_BEGIN_LICENSE:LGPL$
** Commercial License Usage
** Licensees holding valid commercial Qt licenses may use this file in
** accordance with the commercial license agreement provided with the
** Software or, alternatively, in accordance with the terms contained in
** a written agreement between you and The Qt Company. For licensing terms
** and conditions see https://www.qt.io/terms-conditions. For further
** information use the contact form at https://www.qt.io/contact-us.
**
** GNU Lesser General Public License Usage
** Alternatively, this file may be used under the terms of the GNU Lesser
** General Public License version 3 as published by the Free Software
** Foundation and appearing in the file LICENSE.LGPL3 included in the
** packaging of this file. Please review the following information to
** ensure the GNU Lesser General Public License version 3 requirements
** will be met: https://www.gnu.org/licenses/lgpl-3.0.html.
**
** GNU General Public License Usage
** Alternatively, this file may be used under the terms of the GNU
** General Public License version 2.0 or (at your option) the GNU General
** Public license version 3 or any later version approved by the KDE Free
** Qt Foundation. The licenses are as published by the Free Software
** Foundation and appearing in the file LICENSE.GPL2 and LICENSE.GPL3
** included in the packaging of this file. Please review the following
** information to ensure the GNU General Public License requirements will
** be met: https://www.gnu.org/licenses/gpl-2.0.html and
** https://www.gnu.org/licenses/gpl-3.0.html.
**
** $QT_END_LICENSE$
**
****************************************************************************/

#include "qfloat16.h"
#include "private/qsimd_p.h"
#include <cmath> // for fpclassify()'s return values

QT_BEGIN_NAMESPACE

QT_IMPL_METATYPE_EXTERN(qfloat16)

/*!
    \class qfloat16
    \keyword 16-bit Floating Point Support
    \ingroup funclists
    \inmodule QtCore
    \inheaderfile QFloat16
    \brief Provides 16-bit floating point support.

    The \c qfloat16 class provides support for half-precision (16-bit) floating
    point data.  It is fully compliant with IEEE 754 as a storage type.  This
    implies that any arithmetic operation on a \c qfloat16 instance results in
    the value first being converted to a \c float.  This conversion to and from
    \c float is performed by hardware when possible, but on processors that do
    not natively support half-precision, the conversion is performed through a
    sequence of lookup table operations.

    \c qfloat16 should be treated as if it were a POD (plain old data) type.
    Consequently, none of the supported operations need any elaboration beyond
    stating that it supports all arithmetic operators incident to floating point
    types.

    \note On x86 and x86-64 that to get hardware accelerated conversions you must
    compile with F16C or AVX2 enabled, or use qFloatToFloat16() and qFloatFromFloat16()
    which will detect F16C at runtime.

    \since 5.9
*/

/*!
    \fn qfloat16::qfloat16(Qt::Initialization)
    \since 6.1

    Constructs a qfloat16 without initializing the value.
*/

/*!
    \fn bool qIsInf(qfloat16 f)
    \relates qfloat16
    \overload qIsInf(float)

    Returns true if the \c qfloat16 \a {f} is equivalent to infinity.
*/

/*!
    \fn bool qIsNaN(qfloat16 f)
    \relates qfloat16
    \overload qIsNaN(float)

    Returns true if the \c qfloat16 \a {f} is not a number (NaN).
*/

/*!
    \fn bool qIsFinite(qfloat16 f)
    \relates qfloat16
    \overload qIsFinite(float)

    Returns true if the \c qfloat16 \a {f} is a finite number.
*/

/*!
    \internal
    \since 5.14
    \fn bool qfloat16::isInf() const noexcept

    Tests whether this \c qfloat16 value is an infinity.
*/

/*!
    \internal
    \since 5.14
    \fn bool qfloat16::isNaN() const noexcept

    Tests whether this \c qfloat16 value is "not a number".
*/

/*!
    \since 5.14
    \fn bool qfloat16::isNormal() const noexcept

    Returns \c true if this \c qfloat16 value is finite and in normal form.

    \sa qFpClassify()
*/

/*!
    \internal
    \since 5.14
    \fn bool qfloat16::isFinite() const noexcept

    Tests whether this \c qfloat16 value is finite.
*/

/*!
    \since 5.15
    \fn qfloat16 qfloat16::copySign(qfloat16 sign) const noexcept

    Returns a qfloat16 with the sign of \a sign but the rest of its value taken
    from this qfloat16. Serves as qfloat16's equivalent of std::copysign().
*/

/*!
    \fn int qFpClassify(qfloat16 val)
    \relates qfloat16
    \since 5.14
    \overload qFpClassify(float)

    Returns the floating-point class of \a val.
*/

/*!
    \internal
    \since 5.14
    Implements qFpClassify() for qfloat16.
*/
int qfloat16::fpClassify() const noexcept
{
    return isInf() ? FP_INFINITE : isNaN() ? FP_NAN
        : !(b16 & 0x7fff) ? FP_ZERO : isNormal() ? FP_NORMAL : FP_SUBNORMAL;
}

/*! \fn int qRound(qfloat16 value)
    \relates qfloat16
    \overload qRound(float)

    Rounds \a value to the nearest integer.
*/

/*! \fn qint64 qRound64(qfloat16 value)
    \relates qfloat16
    \overload qRound64(float)

    Rounds \a value to the nearest 64-bit integer.
*/

/*! \fn bool qFuzzyCompare(qfloat16 p1, qfloat16 p2)
    \relates qfloat16
    \overload qFuzzyCompare(float, float)

    Compares the floating point value \a p1 and \a p2 and
    returns \c true if they are considered equal, otherwise \c false.

    The two numbers are compared in a relative way, where the
    exactness is stronger the smaller the numbers are.
 */

#if QT_COMPILER_SUPPORTS(F16C)
static inline bool hasFastF16()
{
    // qsimd.cpp:detectProcessorFeatures() turns off this feature if AVX
    // state-saving is not enabled by the OS
    return qCpuHasFeature(F16C);
}

QT_FUNCTION_TARGET(F16C)
void qFloatToFloat16_fast(quint16 *out, const float *in, qsizetype len) Q_DECL_NOEXCEPT
{
    qsizetype i = 0;
    int epilog_i;
    for (; i < len - 7; i += 8)
        _mm_storeu_si128((__m128i *)(out + i), _mm256_cvtps_ph(_mm256_loadu_ps(in + i), 0));
    if (i < len - 3) {
        _mm_storel_epi64((__m128i *)(out + i), _mm_cvtps_ph(_mm_loadu_ps(in + i), 0));
        i += 4;
    }
    // Inlining "qfloat16::qfloat16(float f)":
    for (epilog_i = 0; i < len && epilog_i < 3; ++i, ++epilog_i)
        out[i] = _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(in[i]), 0), 0);
}

QT_FUNCTION_TARGET(F16C)
void qFloatFromFloat16_fast(float *out, const quint16 *in, qsizetype len) Q_DECL_NOEXCEPT
{
    qsizetype i = 0;
    int epilog_i;
    for (; i < len - 7; i += 8)
        _mm256_storeu_ps(out + i, _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(in + i))));
    if (i < len - 3) {
        _mm_storeu_ps(out + i, _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)(in + i))));
        i += 4;
    }
    // Inlining "qfloat16::operator float()":
    for (epilog_i = 0; i < len && epilog_i < 3; ++i, ++epilog_i)
        out[i] = _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(in[i])));
}

#elif defined(__ARM_FP16_FORMAT_IEEE) && defined(__ARM_NEON__) && (__ARM_FP & 2)
static inline bool hasFastF16()
{
    return true;
}

static void qFloatToFloat16_fast(quint16 *out, const float *in, qsizetype len) noexcept
{
    __fp16 *out_f16 = reinterpret_cast<__fp16 *>(out);
    qsizetype i = 0;
    for (; i < len - 3; i += 4)
        vst1_f16(out_f16 + i, vcvt_f16_f32(vld1q_f32(in + i)));
    SIMD_EPILOGUE(i, len, 3)
        out_f16[i] = __fp16(in[i]);
}

static void qFloatFromFloat16_fast(float *out, const quint16 *in, qsizetype len) noexcept
{
    const __fp16 *in_f16 = reinterpret_cast<const __fp16 *>(in);
    qsizetype i = 0;
    for (; i < len - 3; i += 4)
        vst1q_f32(out + i, vcvt_f32_f16(vld1_f16(in_f16 + i)));
    SIMD_EPILOGUE(i, len, 3)
        out[i] = float(in_f16[i]);
}
#else
static inline bool hasFastF16()
{
    return false;
}

static void qFloatToFloat16_fast(quint16 *, const float *, qsizetype) noexcept
{
    Q_UNREACHABLE();
}

static void qFloatFromFloat16_fast(float *, const quint16 *, qsizetype) noexcept
{
    Q_UNREACHABLE();
}
#endif
/*!
    \since 5.11
    \relates qfloat16

    Converts \a len floats from \a in to qfloat16 and stores them in \a out.
    Both \a in and \a out must have \a len allocated entries.

    This function is faster than converting values one by one, and will do runtime
    F16C detection on x86 and x86-64 hardware.
*/
Q_CORE_EXPORT void qFloatToFloat16(qfloat16 *out, const float *in, qsizetype len) noexcept
{
    if (hasFastF16())
        return qFloatToFloat16_fast(reinterpret_cast<quint16 *>(out), in, len);

    for (qsizetype i = 0; i < len; ++i)
        out[i] = qfloat16(in[i]);
}

/*!
    \since 5.11
    \relates qfloat16

    Converts \a len qfloat16 from \a in to floats and stores them in \a out.
    Both \a in and \a out must have \a len allocated entries.

    This function is faster than converting values one by one, and will do runtime
    F16C detection on x86 and x86-64 hardware.
*/
Q_CORE_EXPORT void qFloatFromFloat16(float *out, const qfloat16 *in, qsizetype len) noexcept
{
    if (hasFastF16())
        return qFloatFromFloat16_fast(out, reinterpret_cast<const quint16 *>(in), len);

    for (qsizetype i = 0; i < len; ++i)
        out[i] = float(in[i]);
}

QT_END_NAMESPACE

#include "qfloat16tables.cpp"