s_fma.c revision 1.2
1/* $OpenBSD: s_fma.c,v 1.2 2012/12/05 23:20:04 deraadt 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 <fenv.h> 32#include <float.h> 33#include <math.h> 34 35/* 36 * Fused multiply-add: Compute x * y + z with a single rounding error. 37 * 38 * We use scaling to avoid overflow/underflow, along with the 39 * canonical precision-doubling technique adapted from: 40 * 41 * Dekker, T. A Floating-Point Technique for Extending the 42 * Available Precision. Numer. Math. 18, 224-242 (1971). 43 * 44 * This algorithm is sensitive to the rounding precision. FPUs such 45 * as the i387 must be set in double-precision mode if variables are 46 * to be stored in FP registers in order to avoid incorrect results. 47 * This is the default on FreeBSD, but not on many other systems. 48 * 49 * Hardware instructions should be used on architectures that support it, 50 * since this implementation will likely be several times slower. 51 */ 52#if LDBL_MANT_DIG != 113 53double 54fma(double x, double y, double z) 55{ 56 static const double split = 0x1p27 + 1.0; 57 double xs, ys, zs; 58 double c, cc, hx, hy, p, q, tx, ty; 59 double r, rr, s; 60 int oround; 61 int ex, ey, ez; 62 int spread; 63 64 /* 65 * Handle special cases. The order of operations and the particular 66 * return values here are crucial in handling special cases involving 67 * infinities, NaNs, overflows, and signed zeroes correctly. 68 */ 69 if (x == 0.0 || y == 0.0) 70 return (x * y + z); 71 if (z == 0.0) 72 return (x * y); 73 if (!isfinite(x) || !isfinite(y)) 74 return (x * y + z); 75 if (!isfinite(z)) 76 return (z); 77 78 xs = frexp(x, &ex); 79 ys = frexp(y, &ey); 80 zs = frexp(z, &ez); 81 oround = fegetround(); 82 spread = ex + ey - ez; 83 84 /* 85 * If x * y and z are many orders of magnitude apart, the scaling 86 * will overflow, so we handle these cases specially. Rounding 87 * modes other than FE_TONEAREST are painful. 88 */ 89 if (spread > DBL_MANT_DIG * 2) { 90 fenv_t env; 91 feraiseexcept(FE_INEXACT); 92 switch(oround) { 93 case FE_TONEAREST: 94 return (x * y); 95 case FE_TOWARDZERO: 96 if (x > 0.0 ^ y < 0.0 ^ z < 0.0) 97 return (x * y); 98 feholdexcept(&env); 99 r = x * y; 100 if (!fetestexcept(FE_INEXACT)) 101 r = nextafter(r, 0); 102 feupdateenv(&env); 103 return (r); 104 case FE_DOWNWARD: 105 if (z > 0.0) 106 return (x * y); 107 feholdexcept(&env); 108 r = x * y; 109 if (!fetestexcept(FE_INEXACT)) 110 r = nextafter(r, -INFINITY); 111 feupdateenv(&env); 112 return (r); 113 default: /* FE_UPWARD */ 114 if (z < 0.0) 115 return (x * y); 116 feholdexcept(&env); 117 r = x * y; 118 if (!fetestexcept(FE_INEXACT)) 119 r = nextafter(r, INFINITY); 120 feupdateenv(&env); 121 return (r); 122 } 123 } 124 if (spread < -DBL_MANT_DIG) { 125 feraiseexcept(FE_INEXACT); 126 if (!isnormal(z)) 127 feraiseexcept(FE_UNDERFLOW); 128 switch (oround) { 129 case FE_TONEAREST: 130 return (z); 131 case FE_TOWARDZERO: 132 if (x > 0.0 ^ y < 0.0 ^ z < 0.0) 133 return (z); 134 else 135 return (nextafter(z, 0)); 136 case FE_DOWNWARD: 137 if (x > 0.0 ^ y < 0.0) 138 return (z); 139 else 140 return (nextafter(z, -INFINITY)); 141 default: /* FE_UPWARD */ 142 if (x > 0.0 ^ y < 0.0) 143 return (nextafter(z, INFINITY)); 144 else 145 return (z); 146 } 147 } 148 149 /* 150 * Use Dekker's algorithm to perform the multiplication and 151 * subsequent addition in twice the machine precision. 152 * Arrange so that x * y = c + cc, and x * y + z = r + rr. 153 */ 154 fesetround(FE_TONEAREST); 155 156 p = xs * split; 157 hx = xs - p; 158 hx += p; 159 tx = xs - hx; 160 161 p = ys * split; 162 hy = ys - p; 163 hy += p; 164 ty = ys - hy; 165 166 p = hx * hy; 167 q = hx * ty + tx * hy; 168 c = p + q; 169 cc = p - c + q + tx * ty; 170 171 zs = ldexp(zs, -spread); 172 r = c + zs; 173 s = r - c; 174 rr = (c - (r - s)) + (zs - s) + cc; 175 176 spread = ex + ey; 177 if (spread + ilogb(r) > -1023) { 178 fesetround(oround); 179 r = r + rr; 180 } else { 181 /* 182 * The result is subnormal, so we round before scaling to 183 * avoid double rounding. 184 */ 185 p = ldexp(copysign(0x1p-1022, r), -spread); 186 c = r + p; 187 s = c - r; 188 cc = (r - (c - s)) + (p - s) + rr; 189 fesetround(oround); 190 r = (c + cc) - p; 191 } 192 return (ldexp(r, spread)); 193} 194#else /* LDBL_MANT_DIG == 113 */ 195/* 196 * 113 bits of precision is more than twice the precision of a double, 197 * so it is enough to represent the intermediate product exactly. 198 */ 199double 200fma(double x, double y, double z) 201{ 202 return ((long double)x * y + z); 203} 204#endif /* LDBL_MANT_DIG != 113 */ 205 206#if LDBL_MANT_DIG == 53 207#ifdef lint 208/* PROTOLIB1 */ 209long double fmal(long double, long double, long double); 210#else /* lint */ 211__weak_alias(fmal, fma); 212#endif /* lint */ 213#endif /* LDBL_MANT_DIG == 53 */ 214