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
|
/*
* Copyright (c) 2014,2015 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <clc/clc.h>
#include "math.h"
#include "../clcmacro.h"
_CLC_OVERLOAD _CLC_DEF float asinpi(float x) {
// Computes arcsin(x).
// The argument is first reduced by noting that arcsin(x)
// is invalid for abs(x) > 1 and arcsin(-x) = -arcsin(x).
// For denormal and small arguments arcsin(x) = x to machine
// accuracy. Remaining argument ranges are handled as follows.
// For abs(x) <= 0.5 use
// arcsin(x) = x + x^3*R(x^2)
// where R(x^2) is a rational minimax approximation to
// (arcsin(x) - x)/x^3.
// For abs(x) > 0.5 exploit the identity:
// arcsin(x) = pi/2 - 2*arcsin(sqrt(1-x)/2)
// together with the above rational approximation, and
// reconstruct the terms carefully.
const float pi = 3.1415926535897933e+00f;
const float piby2_tail = 7.5497894159e-08F; /* 0x33a22168 */
const float hpiby2_head = 7.8539812565e-01F; /* 0x3f490fda */
uint ux = as_uint(x);
uint aux = ux & EXSIGNBIT_SP32;
uint xs = ux ^ aux;
float shalf = as_float(xs | as_uint(0.5f));
int xexp = (int)(aux >> EXPSHIFTBITS_SP32) - EXPBIAS_SP32;
float y = as_float(aux);
// abs(x) >= 0.5
int transform = xexp >= -1;
float y2 = y * y;
float rt = 0.5f * (1.0f - y);
float r = transform ? rt : y2;
// Use a rational approximation for [0.0, 0.5]
float a = mad(r,
mad(r,
mad(r, -0.00396137437848476485201154797087F, -0.0133819288943925804214011424456F),
-0.0565298683201845211985026327361F),
0.184161606965100694821398249421F);
float b = mad(r, -0.836411276854206731913362287293F, 1.10496961524520294485512696706F);
float u = r * MATH_DIVIDE(a, b);
float s = MATH_SQRT(r);
float s1 = as_float(as_uint(s) & 0xffff0000);
float c = MATH_DIVIDE(mad(-s1, s1, r), s + s1);
float p = mad(2.0f*s, u, -mad(c, -2.0f, piby2_tail));
float q = mad(s1, -2.0f, hpiby2_head);
float vt = hpiby2_head - (p - q);
float v = mad(y, u, y);
v = transform ? vt : v;
v = MATH_DIVIDE(v, pi);
float xbypi = MATH_DIVIDE(x, pi);
float ret = as_float(xs | as_uint(v));
ret = aux > 0x3f800000U ? as_float(QNANBITPATT_SP32) : ret;
ret = aux == 0x3f800000U ? shalf : ret;
ret = xexp < -14 ? xbypi : ret;
return ret;
}
_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, asinpi, float)
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
_CLC_OVERLOAD _CLC_DEF double asinpi(double x) {
// Computes arcsin(x).
// The argument is first reduced by noting that arcsin(x)
// is invalid for abs(x) > 1 and arcsin(-x) = -arcsin(x).
// For denormal and small arguments arcsin(x) = x to machine
// accuracy. Remaining argument ranges are handled as follows.
// For abs(x) <= 0.5 use
// arcsin(x) = x + x^3*R(x^2)
// where R(x^2) is a rational minimax approximation to
// (arcsin(x) - x)/x^3.
// For abs(x) > 0.5 exploit the identity:
// arcsin(x) = pi/2 - 2*arcsin(sqrt(1-x)/2)
// together with the above rational approximation, and
// reconstruct the terms carefully.
const double pi = 0x1.921fb54442d18p+1;
const double piby2_tail = 6.1232339957367660e-17; /* 0x3c91a62633145c07 */
const double hpiby2_head = 7.8539816339744831e-01; /* 0x3fe921fb54442d18 */
double y = fabs(x);
int xneg = as_int2(x).hi < 0;
int xexp = (as_int2(y).hi >> 20) - EXPBIAS_DP64;
// abs(x) >= 0.5
int transform = xexp >= -1;
double rt = 0.5 * (1.0 - y);
double y2 = y * y;
double r = transform ? rt : y2;
// Use a rational approximation for [0.0, 0.5]
double un = fma(r,
fma(r,
fma(r,
fma(r,
fma(r, 0.0000482901920344786991880522822991,
0.00109242697235074662306043804220),
-0.0549989809235685841612020091328),
0.275558175256937652532686256258),
-0.445017216867635649900123110649),
0.227485835556935010735943483075);
double ud = fma(r,
fma(r,
fma(r,
fma(r, 0.105869422087204370341222318533,
-0.943639137032492685763471240072),
2.76568859157270989520376345954),
-3.28431505720958658909889444194),
1.36491501334161032038194214209);
double u = r * MATH_DIVIDE(un, ud);
// Reconstruct asin carefully in transformed region
double s = sqrt(r);
double sh = as_double(as_ulong(s) & 0xffffffff00000000UL);
double c = MATH_DIVIDE(fma(-sh, sh, r), s + sh);
double p = fma(2.0*s, u, -fma(-2.0, c, piby2_tail));
double q = fma(-2.0, sh, hpiby2_head);
double vt = hpiby2_head - (p - q);
double v = fma(y, u, y);
v = transform ? vt : v;
v = xexp < -28 ? y : v;
v = MATH_DIVIDE(v, pi);
v = xexp >= 0 ? as_double(QNANBITPATT_DP64) : v;
v = y == 1.0 ? 0.5 : v;
return xneg ? -v : v;
}
_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, asinpi, double)
#endif
|
