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
|
/*-
* Copyright (c) 2005 David Schultz <das@FreeBSD.ORG>
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
/* __FBSDID("$FreeBSD: src/lib/msun/src/s_fma.c,v 1.4 2005/03/18 02:27:59 das Exp $"); */
#include <fenv.h>
#include <float.h>
#include <math.h>
/*
* Fused multiply-add: Compute x * y + z with a single rounding error.
*
* We use scaling to avoid overflow/underflow, along with the
* canonical precision-doubling technique adapted from:
*
* Dekker, T. A Floating-Point Technique for Extending the
* Available Precision. Numer. Math. 18, 224-242 (1971).
*
* This algorithm is sensitive to the rounding precision. FPUs such
* as the i387 must be set in double-precision mode if variables are
* to be stored in FP registers in order to avoid incorrect results.
* This is the default on FreeBSD, but not on many other systems.
*
* Hardware instructions should be used on architectures that support it,
* since this implementation will likely be several times slower.
*/
#if LDBL_MANT_DIG != 113
double
fma(double x, double y, double z)
{
static const double split = 0x1p27 + 1.0;
double xs, ys, zs;
double c, cc, hx, hy, p, q, tx, ty;
double r, rr, s;
int oround;
int ex, ey, ez;
int spread;
if (z == 0.0)
return (x * y);
if (x == 0.0 || y == 0.0)
return (x * y + z);
/* Results of frexp() are undefined for these cases. */
if (!isfinite(x) || !isfinite(y) || !isfinite(z))
return (x * y + z);
xs = frexp(x, &ex);
ys = frexp(y, &ey);
zs = frexp(z, &ez);
oround = fegetround();
spread = ex + ey - ez;
/*
* If x * y and z are many orders of magnitude apart, the scaling
* will overflow, so we handle these cases specially. Rounding
* modes other than FE_TONEAREST are painful.
*/
if (spread > DBL_MANT_DIG * 2) {
fenv_t env;
feraiseexcept(FE_INEXACT);
switch(oround) {
case FE_TONEAREST:
return (x * y);
case FE_TOWARDZERO:
if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
return (x * y);
feholdexcept(&env);
r = x * y;
if (!fetestexcept(FE_INEXACT))
r = nextafter(r, 0);
feupdateenv(&env);
return (r);
case FE_DOWNWARD:
if (z > 0.0)
return (x * y);
feholdexcept(&env);
r = x * y;
if (!fetestexcept(FE_INEXACT))
r = nextafter(r, -INFINITY);
feupdateenv(&env);
return (r);
default: /* FE_UPWARD */
if (z < 0.0)
return (x * y);
feholdexcept(&env);
r = x * y;
if (!fetestexcept(FE_INEXACT))
r = nextafter(r, INFINITY);
feupdateenv(&env);
return (r);
}
}
if (spread < -DBL_MANT_DIG) {
feraiseexcept(FE_INEXACT);
if (!isnormal(z))
feraiseexcept(FE_UNDERFLOW);
switch (oround) {
case FE_TONEAREST:
return (z);
case FE_TOWARDZERO:
if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
return (z);
else
return (nextafter(z, 0));
case FE_DOWNWARD:
if (x > 0.0 ^ y < 0.0)
return (z);
else
return (nextafter(z, -INFINITY));
default: /* FE_UPWARD */
if (x > 0.0 ^ y < 0.0)
return (nextafter(z, INFINITY));
else
return (z);
}
}
/*
* Use Dekker's algorithm to perform the multiplication and
* subsequent addition in twice the machine precision.
* Arrange so that x * y = c + cc, and x * y + z = r + rr.
*/
fesetround(FE_TONEAREST);
p = xs * split;
hx = xs - p;
hx += p;
tx = xs - hx;
p = ys * split;
hy = ys - p;
hy += p;
ty = ys - hy;
p = hx * hy;
q = hx * ty + tx * hy;
c = p + q;
cc = p - c + q + tx * ty;
zs = ldexp(zs, -spread);
r = c + zs;
s = r - c;
rr = (c - (r - s)) + (zs - s) + cc;
spread = ex + ey;
if (spread + ilogb(r) > -1023) {
fesetround(oround);
r = r + rr;
} else {
/*
* The result is subnormal, so we round before scaling to
* avoid double rounding.
*/
p = ldexp(copysign(0x1p-1022, r), -spread);
c = r + p;
s = c - r;
cc = (r - (c - s)) + (p - s) + rr;
fesetround(oround);
r = (c + cc) - p;
}
return (ldexp(r, spread));
}
#else /* LDBL_MANT_DIG == 113 */
/*
* 113 bits of precision is more than twice the precision of a double,
* so it is enough to represent the intermediate product exactly.
*/
double
fma(double x, double y, double z)
{
return ((long double)x * y + z);
}
#endif /* LDBL_MANT_DIG != 113 */
#if (LDBL_MANT_DIG == 53)
__weak_reference(fma, fmal);
#endif
|
