1*181254a7Smrg /* expm1q.c
2*181254a7Smrg *
3*181254a7Smrg * Exponential function, minus 1
4*181254a7Smrg * 128-bit long double precision
5*181254a7Smrg *
6*181254a7Smrg *
7*181254a7Smrg *
8*181254a7Smrg * SYNOPSIS:
9*181254a7Smrg *
10*181254a7Smrg * long double x, y, expm1q();
11*181254a7Smrg *
12*181254a7Smrg * y = expm1q( x );
13*181254a7Smrg *
14*181254a7Smrg *
15*181254a7Smrg *
16*181254a7Smrg * DESCRIPTION:
17*181254a7Smrg *
18*181254a7Smrg * Returns e (2.71828...) raised to the x power, minus one.
19*181254a7Smrg *
20*181254a7Smrg * Range reduction is accomplished by separating the argument
21*181254a7Smrg * into an integer k and fraction f such that
22*181254a7Smrg *
23*181254a7Smrg * x k f
24*181254a7Smrg * e = 2 e.
25*181254a7Smrg *
26*181254a7Smrg * An expansion x + .5 x^2 + x^3 R(x) approximates exp(f) - 1
27*181254a7Smrg * in the basic range [-0.5 ln 2, 0.5 ln 2].
28*181254a7Smrg *
29*181254a7Smrg *
30*181254a7Smrg * ACCURACY:
31*181254a7Smrg *
32*181254a7Smrg * Relative error:
33*181254a7Smrg * arithmetic domain # trials peak rms
34*181254a7Smrg * IEEE -79,+MAXLOG 100,000 1.7e-34 4.5e-35
35*181254a7Smrg *
36*181254a7Smrg */
37*181254a7Smrg
38*181254a7Smrg /* Copyright 2001 by Stephen L. Moshier
39*181254a7Smrg
40*181254a7Smrg This library is free software; you can redistribute it and/or
41*181254a7Smrg modify it under the terms of the GNU Lesser General Public
42*181254a7Smrg License as published by the Free Software Foundation; either
43*181254a7Smrg version 2.1 of the License, or (at your option) any later version.
44*181254a7Smrg
45*181254a7Smrg This library is distributed in the hope that it will be useful,
46*181254a7Smrg but WITHOUT ANY WARRANTY; without even the implied warranty of
47*181254a7Smrg MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
48*181254a7Smrg Lesser General Public License for more details.
49*181254a7Smrg
50*181254a7Smrg You should have received a copy of the GNU Lesser General Public
51*181254a7Smrg License along with this library; if not, see
52*181254a7Smrg <http://www.gnu.org/licenses/>. */
53*181254a7Smrg
54*181254a7Smrg #include "quadmath-imp.h"
55*181254a7Smrg
56*181254a7Smrg /* exp(x) - 1 = x + 0.5 x^2 + x^3 P(x)/Q(x)
57*181254a7Smrg -.5 ln 2 < x < .5 ln 2
58*181254a7Smrg Theoretical peak relative error = 8.1e-36 */
59*181254a7Smrg
60*181254a7Smrg static const __float128
61*181254a7Smrg P0 = 2.943520915569954073888921213330863757240E8Q,
62*181254a7Smrg P1 = -5.722847283900608941516165725053359168840E7Q,
63*181254a7Smrg P2 = 8.944630806357575461578107295909719817253E6Q,
64*181254a7Smrg P3 = -7.212432713558031519943281748462837065308E5Q,
65*181254a7Smrg P4 = 4.578962475841642634225390068461943438441E4Q,
66*181254a7Smrg P5 = -1.716772506388927649032068540558788106762E3Q,
67*181254a7Smrg P6 = 4.401308817383362136048032038528753151144E1Q,
68*181254a7Smrg P7 = -4.888737542888633647784737721812546636240E-1Q,
69*181254a7Smrg Q0 = 1.766112549341972444333352727998584753865E9Q,
70*181254a7Smrg Q1 = -7.848989743695296475743081255027098295771E8Q,
71*181254a7Smrg Q2 = 1.615869009634292424463780387327037251069E8Q,
72*181254a7Smrg Q3 = -2.019684072836541751428967854947019415698E7Q,
73*181254a7Smrg Q4 = 1.682912729190313538934190635536631941751E6Q,
74*181254a7Smrg Q5 = -9.615511549171441430850103489315371768998E4Q,
75*181254a7Smrg Q6 = 3.697714952261803935521187272204485251835E3Q,
76*181254a7Smrg Q7 = -8.802340681794263968892934703309274564037E1Q,
77*181254a7Smrg /* Q8 = 1.000000000000000000000000000000000000000E0 */
78*181254a7Smrg /* C1 + C2 = ln 2 */
79*181254a7Smrg
80*181254a7Smrg C1 = 6.93145751953125E-1Q,
81*181254a7Smrg C2 = 1.428606820309417232121458176568075500134E-6Q,
82*181254a7Smrg /* ln 2^-114 */
83*181254a7Smrg minarg = -7.9018778583833765273564461846232128760607E1Q, big = 1e4932Q;
84*181254a7Smrg
85*181254a7Smrg
86*181254a7Smrg __float128
expm1q(__float128 x)87*181254a7Smrg expm1q (__float128 x)
88*181254a7Smrg {
89*181254a7Smrg __float128 px, qx, xx;
90*181254a7Smrg int32_t ix, sign;
91*181254a7Smrg ieee854_float128 u;
92*181254a7Smrg int k;
93*181254a7Smrg
94*181254a7Smrg /* Detect infinity and NaN. */
95*181254a7Smrg u.value = x;
96*181254a7Smrg ix = u.words32.w0;
97*181254a7Smrg sign = ix & 0x80000000;
98*181254a7Smrg ix &= 0x7fffffff;
99*181254a7Smrg if (!sign && ix >= 0x40060000)
100*181254a7Smrg {
101*181254a7Smrg /* If num is positive and exp >= 6 use plain exp. */
102*181254a7Smrg return expq (x);
103*181254a7Smrg }
104*181254a7Smrg if (ix >= 0x7fff0000)
105*181254a7Smrg {
106*181254a7Smrg /* Infinity (which must be negative infinity). */
107*181254a7Smrg if (((ix & 0xffff) | u.words32.w1 | u.words32.w2 | u.words32.w3) == 0)
108*181254a7Smrg return -1;
109*181254a7Smrg /* NaN. Invalid exception if signaling. */
110*181254a7Smrg return x + x;
111*181254a7Smrg }
112*181254a7Smrg
113*181254a7Smrg /* expm1(+- 0) = +- 0. */
114*181254a7Smrg if ((ix == 0) && (u.words32.w1 | u.words32.w2 | u.words32.w3) == 0)
115*181254a7Smrg return x;
116*181254a7Smrg
117*181254a7Smrg /* Minimum value. */
118*181254a7Smrg if (x < minarg)
119*181254a7Smrg return (4.0/big - 1);
120*181254a7Smrg
121*181254a7Smrg /* Avoid internal underflow when result does not underflow, while
122*181254a7Smrg ensuring underflow (without returning a zero of the wrong sign)
123*181254a7Smrg when the result does underflow. */
124*181254a7Smrg if (fabsq (x) < 0x1p-113Q)
125*181254a7Smrg {
126*181254a7Smrg math_check_force_underflow (x);
127*181254a7Smrg return x;
128*181254a7Smrg }
129*181254a7Smrg
130*181254a7Smrg /* Express x = ln 2 (k + remainder), remainder not exceeding 1/2. */
131*181254a7Smrg xx = C1 + C2; /* ln 2. */
132*181254a7Smrg px = floorq (0.5 + x / xx);
133*181254a7Smrg k = px;
134*181254a7Smrg /* remainder times ln 2 */
135*181254a7Smrg x -= px * C1;
136*181254a7Smrg x -= px * C2;
137*181254a7Smrg
138*181254a7Smrg /* Approximate exp(remainder ln 2). */
139*181254a7Smrg px = (((((((P7 * x
140*181254a7Smrg + P6) * x
141*181254a7Smrg + P5) * x + P4) * x + P3) * x + P2) * x + P1) * x + P0) * x;
142*181254a7Smrg
143*181254a7Smrg qx = (((((((x
144*181254a7Smrg + Q7) * x
145*181254a7Smrg + Q6) * x + Q5) * x + Q4) * x + Q3) * x + Q2) * x + Q1) * x + Q0;
146*181254a7Smrg
147*181254a7Smrg xx = x * x;
148*181254a7Smrg qx = x + (0.5 * xx + xx * px / qx);
149*181254a7Smrg
150*181254a7Smrg /* exp(x) = exp(k ln 2) exp(remainder ln 2) = 2^k exp(remainder ln 2).
151*181254a7Smrg
152*181254a7Smrg We have qx = exp(remainder ln 2) - 1, so
153*181254a7Smrg exp(x) - 1 = 2^k (qx + 1) - 1
154*181254a7Smrg = 2^k qx + 2^k - 1. */
155*181254a7Smrg
156*181254a7Smrg px = ldexpq (1, k);
157*181254a7Smrg x = px * qx + (px - 1.0);
158*181254a7Smrg return x;
159*181254a7Smrg }
160