xref: /openbsd/lib/libm/src/s_fma.c (revision cca36db2)
1 /*	$OpenBSD: s_fma.c,v 1.1 2011/07/06 00:02:42 martynas Exp $	*/
2 
3 /*-
4  * Copyright (c) 2005 David Schultz <das@FreeBSD.ORG>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /* LINTLIBRARY */
30 
31 #include <sys/cdefs.h>
32 #if 0
33 __FBSDID("$FreeBSD: src/lib/msun/src/s_fma.c,v 1.5 2008/04/03 06:14:51 das Exp $");
34 #endif
35 
36 #include <fenv.h>
37 #include <float.h>
38 #include <math.h>
39 
40 /*
41  * Fused multiply-add: Compute x * y + z with a single rounding error.
42  *
43  * We use scaling to avoid overflow/underflow, along with the
44  * canonical precision-doubling technique adapted from:
45  *
46  *	Dekker, T.  A Floating-Point Technique for Extending the
47  *	Available Precision.  Numer. Math. 18, 224-242 (1971).
48  *
49  * This algorithm is sensitive to the rounding precision.  FPUs such
50  * as the i387 must be set in double-precision mode if variables are
51  * to be stored in FP registers in order to avoid incorrect results.
52  * This is the default on FreeBSD, but not on many other systems.
53  *
54  * Hardware instructions should be used on architectures that support it,
55  * since this implementation will likely be several times slower.
56  */
57 #if LDBL_MANT_DIG != 113
58 double
59 fma(double x, double y, double z)
60 {
61 	static const double split = 0x1p27 + 1.0;
62 	double xs, ys, zs;
63 	double c, cc, hx, hy, p, q, tx, ty;
64 	double r, rr, s;
65 	int oround;
66 	int ex, ey, ez;
67 	int spread;
68 
69 	/*
70 	 * Handle special cases. The order of operations and the particular
71 	 * return values here are crucial in handling special cases involving
72 	 * infinities, NaNs, overflows, and signed zeroes correctly.
73 	 */
74 	if (x == 0.0 || y == 0.0)
75 		return (x * y + z);
76 	if (z == 0.0)
77 		return (x * y);
78 	if (!isfinite(x) || !isfinite(y))
79 		return (x * y + z);
80 	if (!isfinite(z))
81 		return (z);
82 
83 	xs = frexp(x, &ex);
84 	ys = frexp(y, &ey);
85 	zs = frexp(z, &ez);
86 	oround = fegetround();
87 	spread = ex + ey - ez;
88 
89 	/*
90 	 * If x * y and z are many orders of magnitude apart, the scaling
91 	 * will overflow, so we handle these cases specially.  Rounding
92 	 * modes other than FE_TONEAREST are painful.
93 	 */
94 	if (spread > DBL_MANT_DIG * 2) {
95 		fenv_t env;
96 		feraiseexcept(FE_INEXACT);
97 		switch(oround) {
98 		case FE_TONEAREST:
99 			return (x * y);
100 		case FE_TOWARDZERO:
101 			if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
102 				return (x * y);
103 			feholdexcept(&env);
104 			r = x * y;
105 			if (!fetestexcept(FE_INEXACT))
106 				r = nextafter(r, 0);
107 			feupdateenv(&env);
108 			return (r);
109 		case FE_DOWNWARD:
110 			if (z > 0.0)
111 				return (x * y);
112 			feholdexcept(&env);
113 			r = x * y;
114 			if (!fetestexcept(FE_INEXACT))
115 				r = nextafter(r, -INFINITY);
116 			feupdateenv(&env);
117 			return (r);
118 		default:	/* FE_UPWARD */
119 			if (z < 0.0)
120 				return (x * y);
121 			feholdexcept(&env);
122 			r = x * y;
123 			if (!fetestexcept(FE_INEXACT))
124 				r = nextafter(r, INFINITY);
125 			feupdateenv(&env);
126 			return (r);
127 		}
128 	}
129 	if (spread < -DBL_MANT_DIG) {
130 		feraiseexcept(FE_INEXACT);
131 		if (!isnormal(z))
132 			feraiseexcept(FE_UNDERFLOW);
133 		switch (oround) {
134 		case FE_TONEAREST:
135 			return (z);
136 		case FE_TOWARDZERO:
137 			if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
138 				return (z);
139 			else
140 				return (nextafter(z, 0));
141 		case FE_DOWNWARD:
142 			if (x > 0.0 ^ y < 0.0)
143 				return (z);
144 			else
145 				return (nextafter(z, -INFINITY));
146 		default:	/* FE_UPWARD */
147 			if (x > 0.0 ^ y < 0.0)
148 				return (nextafter(z, INFINITY));
149 			else
150 				return (z);
151 		}
152 	}
153 
154 	/*
155 	 * Use Dekker's algorithm to perform the multiplication and
156 	 * subsequent addition in twice the machine precision.
157 	 * Arrange so that x * y = c + cc, and x * y + z = r + rr.
158 	 */
159 	fesetround(FE_TONEAREST);
160 
161 	p = xs * split;
162 	hx = xs - p;
163 	hx += p;
164 	tx = xs - hx;
165 
166 	p = ys * split;
167 	hy = ys - p;
168 	hy += p;
169 	ty = ys - hy;
170 
171 	p = hx * hy;
172 	q = hx * ty + tx * hy;
173 	c = p + q;
174 	cc = p - c + q + tx * ty;
175 
176 	zs = ldexp(zs, -spread);
177 	r = c + zs;
178 	s = r - c;
179 	rr = (c - (r - s)) + (zs - s) + cc;
180 
181 	spread = ex + ey;
182 	if (spread + ilogb(r) > -1023) {
183 		fesetround(oround);
184 		r = r + rr;
185 	} else {
186 		/*
187 		 * The result is subnormal, so we round before scaling to
188 		 * avoid double rounding.
189 		 */
190 		p = ldexp(copysign(0x1p-1022, r), -spread);
191 		c = r + p;
192 		s = c - r;
193 		cc = (r - (c - s)) + (p - s) + rr;
194 		fesetround(oround);
195 		r = (c + cc) - p;
196 	}
197 	return (ldexp(r, spread));
198 }
199 #else	/* LDBL_MANT_DIG == 113 */
200 /*
201  * 113 bits of precision is more than twice the precision of a double,
202  * so it is enough to represent the intermediate product exactly.
203  */
204 double
205 fma(double x, double y, double z)
206 {
207 	return ((long double)x * y + z);
208 }
209 #endif	/* LDBL_MANT_DIG != 113 */
210 
211 #if	LDBL_MANT_DIG == 53
212 #ifdef	lint
213 /* PROTOLIB1 */
214 long double fmal(long double, long double, long double);
215 #else	/* lint */
216 __weak_alias(fmal, fma);
217 #endif	/* lint */
218 #endif	/* LDBL_MANT_DIG == 53 */
219