/src/moddable/xs/tools/fdlibm/s_log1p.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * ==================================================== |
3 | | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
4 | | * |
5 | | * Developed at SunPro, a Sun Microsystems, Inc. business. |
6 | | * Permission to use, copy, modify, and distribute this |
7 | | * software is freely granted, provided that this notice |
8 | | * is preserved. |
9 | | * ==================================================== |
10 | | */ |
11 | | |
12 | | /* double log1p(double x) |
13 | | * |
14 | | * Method : |
15 | | * 1. Argument Reduction: find k and f such that |
16 | | * 1+x = 2^k * (1+f), |
17 | | * where sqrt(2)/2 < 1+f < sqrt(2) . |
18 | | * |
19 | | * Note. If k=0, then f=x is exact. However, if k!=0, then f |
20 | | * may not be representable exactly. In that case, a correction |
21 | | * term is need. Let u=1+x rounded. Let c = (1+x)-u, then |
22 | | * log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u), |
23 | | * and add back the correction term c/u. |
24 | | * (Note: when x > 2**53, one can simply return log(x)) |
25 | | * |
26 | | * 2. Approximation of log1p(f). |
27 | | * Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s) |
28 | | * = 2s + 2/3 s**3 + 2/5 s**5 + ....., |
29 | | * = 2s + s*R |
30 | | * We use a special Reme algorithm on [0,0.1716] to generate |
31 | | * a polynomial of degree 14 to approximate R The maximum error |
32 | | * of this polynomial approximation is bounded by 2**-58.45. In |
33 | | * other words, |
34 | | * 2 4 6 8 10 12 14 |
35 | | * R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s +Lp6*s +Lp7*s |
36 | | * (the values of Lp1 to Lp7 are listed in the program) |
37 | | * and |
38 | | * | 2 14 | -58.45 |
39 | | * | Lp1*s +...+Lp7*s - R(z) | <= 2 |
40 | | * | | |
41 | | * Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2. |
42 | | * In order to guarantee error in log below 1ulp, we compute log |
43 | | * by |
44 | | * log1p(f) = f - (hfsq - s*(hfsq+R)). |
45 | | * |
46 | | * 3. Finally, log1p(x) = k*ln2 + log1p(f). |
47 | | * = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo))) |
48 | | * Here ln2 is split into two floating point number: |
49 | | * ln2_hi + ln2_lo, |
50 | | * where n*ln2_hi is always exact for |n| < 2000. |
51 | | * |
52 | | * Special cases: |
53 | | * log1p(x) is NaN with signal if x < -1 (including -INF) ; |
54 | | * log1p(+INF) is +INF; log1p(-1) is -INF with signal; |
55 | | * log1p(NaN) is that NaN with no signal. |
56 | | * |
57 | | * Accuracy: |
58 | | * according to an error analysis, the error is always less than |
59 | | * 1 ulp (unit in the last place). |
60 | | * |
61 | | * Constants: |
62 | | * The hexadecimal values are the intended ones for the following |
63 | | * constants. The decimal values may be used, provided that the |
64 | | * compiler will convert from decimal to binary accurately enough |
65 | | * to produce the hexadecimal values shown. |
66 | | * |
67 | | * Note: Assuming log() return accurate answer, the following |
68 | | * algorithm can be used to compute log1p(x) to within a few ULP: |
69 | | * |
70 | | * u = 1+x; |
71 | | * if(u==1.0) return x ; else |
72 | | * return log(u)*(x/(u-1.0)); |
73 | | * |
74 | | * See HP-15C Advanced Functions Handbook, p.193. |
75 | | */ |
76 | | |
77 | | #include "math_private.h" |
78 | | |
79 | | static const double |
80 | | ln2_hi = 6.93147180369123816490e-01, /* 3fe62e42 fee00000 */ |
81 | | ln2_lo = 1.90821492927058770002e-10, /* 3dea39ef 35793c76 */ |
82 | | two54 = 1.80143985094819840000e+16, /* 43500000 00000000 */ |
83 | | Lp1 = 6.666666666666735130e-01, /* 3FE55555 55555593 */ |
84 | | Lp2 = 3.999999999940941908e-01, /* 3FD99999 9997FA04 */ |
85 | | Lp3 = 2.857142874366239149e-01, /* 3FD24924 94229359 */ |
86 | | Lp4 = 2.222219843214978396e-01, /* 3FCC71C5 1D8E78AF */ |
87 | | Lp5 = 1.818357216161805012e-01, /* 3FC74664 96CB03DE */ |
88 | | Lp6 = 1.531383769920937332e-01, /* 3FC39A09 D078C69F */ |
89 | | Lp7 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */ |
90 | | |
91 | | static const double zero = 0.0; |
92 | | static volatile double vzero = 0.0; |
93 | | |
94 | | double |
95 | | s_log1p(double x) |
96 | 434k | { |
97 | 434k | double hfsq,f,c,s,z,R,u; |
98 | 434k | int32_t k,hx,hu,ax; |
99 | | |
100 | 434k | GET_HIGH_WORD(hx,x); |
101 | 434k | ax = hx&0x7fffffff; |
102 | | |
103 | 434k | k = 1; |
104 | 434k | if (hx < 0x3FDA827A) { /* 1+x < sqrt(2)+ */ |
105 | 3.02k | if(ax>=0x3ff00000) { /* x <= -1.0 */ |
106 | 234 | if(x==-1.0) return -two54/vzero; /* log1p(-1)=+inf */ |
107 | 85 | else return (x-x)/(x-x); /* log1p(x<-1)=NaN */ |
108 | 234 | } |
109 | 2.78k | if(ax<0x3e200000) { /* |x| < 2**-29 */ |
110 | 461 | if(two54+x>zero /* raise inexact */ |
111 | 461 | &&ax<0x3c900000) /* |x| < 2**-54 */ |
112 | 461 | return x; |
113 | 0 | else |
114 | 0 | return x - x*x*0.5; |
115 | 461 | } |
116 | 2.32k | if(hx>0||hx<=((int32_t)0xbfd2bec4)) { |
117 | 2.32k | k=0;f=x;hu=1;} /* sqrt(2)/2- <= 1+x < sqrt(2)+ */ |
118 | 2.32k | } |
119 | 434k | if (hx >= 0x7ff00000) return x+x; |
120 | 433k | if(k!=0) { |
121 | 430k | if(hx<0x43400000) { |
122 | 155k | STRICT_ASSIGN(double,u,1.0+x); |
123 | 155k | GET_HIGH_WORD(hu,u); |
124 | 155k | k = (hu>>20)-1023; |
125 | 155k | c = (k>0)? 1.0-(u-x):x-(u-1.0);/* correction term */ |
126 | 155k | c /= u; |
127 | 275k | } else { |
128 | 275k | u = x; |
129 | 275k | GET_HIGH_WORD(hu,u); |
130 | 275k | k = (hu>>20)-1023; |
131 | 275k | c = 0; |
132 | 275k | } |
133 | 430k | hu &= 0x000fffff; |
134 | | /* |
135 | | * The approximation to sqrt(2) used in thresholds is not |
136 | | * critical. However, the ones used above must give less |
137 | | * strict bounds than the one here so that the k==0 case is |
138 | | * never reached from here, since here we have committed to |
139 | | * using the correction term but don't use it if k==0. |
140 | | */ |
141 | 430k | if(hu<0x6a09e) { /* u ~< sqrt(2) */ |
142 | 405k | SET_HIGH_WORD(u,hu|0x3ff00000); /* normalize u */ |
143 | 405k | } else { |
144 | 25.2k | k += 1; |
145 | 25.2k | SET_HIGH_WORD(u,hu|0x3fe00000); /* normalize u/2 */ |
146 | 25.2k | hu = (0x00100000-hu)>>2; |
147 | 25.2k | } |
148 | 430k | f = u-1.0; |
149 | 430k | } |
150 | 433k | hfsq=0.5*f*f; |
151 | 433k | if(hu==0) { /* |f| < 2**-20 */ |
152 | 128k | if(f==zero) { |
153 | 127k | if(k==0) { |
154 | 0 | return zero; |
155 | 127k | } else { |
156 | 127k | c += k*ln2_lo; |
157 | 127k | return k*ln2_hi+c; |
158 | 127k | } |
159 | 127k | } |
160 | 1.73k | R = hfsq*(1.0-0.66666666666666666*f); |
161 | 1.73k | if(k==0) return f-R; else |
162 | 1.73k | return k*ln2_hi-((R-(k*ln2_lo+c))-f); |
163 | 1.73k | } |
164 | 304k | s = f/(2.0+f); |
165 | 304k | z = s*s; |
166 | 304k | R = z*(Lp1+z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7)))))); |
167 | 304k | if(k==0) return f-(hfsq-s*(hfsq+R)); else |
168 | 301k | return k*ln2_hi-((hfsq-(s*(hfsq+R)+(k*ln2_lo+c)))-f); |
169 | 304k | } |