xref: /freebsd/contrib/arm-optimized-routines/math/log.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 /*
2  * Double-precision log(x) function.
3  *
4  * Copyright (c) 2018-2019, Arm Limited.
5  * SPDX-License-Identifier: MIT
6  */
7 
8 #include <float.h>
9 #include <math.h>
10 #include <stdint.h>
11 #include "math_config.h"
12 
13 #define T __log_data.tab
14 #define T2 __log_data.tab2
15 #define B __log_data.poly1
16 #define A __log_data.poly
17 #define Ln2hi __log_data.ln2hi
18 #define Ln2lo __log_data.ln2lo
19 #define N (1 << LOG_TABLE_BITS)
20 #define OFF 0x3fe6000000000000
21 
22 /* Top 16 bits of a double.  */
23 static inline uint32_t
24 top16 (double x)
25 {
26   return asuint64 (x) >> 48;
27 }
28 
29 double
30 log (double x)
31 {
32   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
33   double_t w, z, r, r2, r3, y, invc, logc, kd, hi, lo;
34   uint64_t ix, iz, tmp;
35   uint32_t top;
36   int k, i;
37 
38   ix = asuint64 (x);
39   top = top16 (x);
40 
41 #if LOG_POLY1_ORDER == 10 || LOG_POLY1_ORDER == 11
42 # define LO asuint64 (1.0 - 0x1p-5)
43 # define HI asuint64 (1.0 + 0x1.1p-5)
44 #elif LOG_POLY1_ORDER == 12
45 # define LO asuint64 (1.0 - 0x1p-4)
46 # define HI asuint64 (1.0 + 0x1.09p-4)
47 #endif
48   if (unlikely (ix - LO < HI - LO))
49     {
50       /* Handle close to 1.0 inputs separately.  */
51       /* Fix sign of zero with downward rounding when x==1.  */
52       if (WANT_ROUNDING && unlikely (ix == asuint64 (1.0)))
53 	return 0;
54       r = x - 1.0;
55       r2 = r * r;
56       r3 = r * r2;
57 #if LOG_POLY1_ORDER == 10
58       /* Worst-case error is around 0.516 ULP.  */
59       y = r3 * (B[1] + r * B[2] + r2 * B[3]
60 		+ r3 * (B[4] + r * B[5] + r2 * B[6] + r3 * (B[7] + r * B[8])));
61       w = B[0] * r2; /* B[0] == -0.5.  */
62       hi = r + w;
63       y += r - hi + w;
64       y += hi;
65 #elif LOG_POLY1_ORDER == 11
66       /* Worst-case error is around 0.516 ULP.  */
67       y = r3 * (B[1] + r * B[2]
68 		+ r2 * (B[3] + r * B[4] + r2 * B[5]
69 			+ r3 * (B[6] + r * B[7] + r2 * B[8] + r3 * B[9])));
70       w = B[0] * r2; /* B[0] == -0.5.  */
71       hi = r + w;
72       y += r - hi + w;
73       y += hi;
74 #elif LOG_POLY1_ORDER == 12
75       y = r3 * (B[1] + r * B[2] + r2 * B[3]
76 		+ r3 * (B[4] + r * B[5] + r2 * B[6]
77 			+ r3 * (B[7] + r * B[8] + r2 * B[9] + r3 * B[10])));
78 # if N <= 64
79       /* Worst-case error is around 0.532 ULP.  */
80       w = B[0] * r2; /* B[0] == -0.5.  */
81       hi = r + w;
82       y += r - hi + w;
83       y += hi;
84 # else
85       /* Worst-case error is around 0.507 ULP.  */
86       w = r * 0x1p27;
87       double_t rhi = r + w - w;
88       double_t rlo = r - rhi;
89       w = rhi * rhi * B[0]; /* B[0] == -0.5.  */
90       hi = r + w;
91       lo = r - hi + w;
92       lo += B[0] * rlo * (rhi + r);
93       y += lo;
94       y += hi;
95 # endif
96 #endif
97       return eval_as_double (y);
98     }
99   if (unlikely (top - 0x0010 >= 0x7ff0 - 0x0010))
100     {
101       /* x < 0x1p-1022 or inf or nan.  */
102       if (ix * 2 == 0)
103 	return __math_divzero (1);
104       if (ix == asuint64 (INFINITY)) /* log(inf) == inf.  */
105 	return x;
106       if ((top & 0x8000) || (top & 0x7ff0) == 0x7ff0)
107 	return __math_invalid (x);
108       /* x is subnormal, normalize it.  */
109       ix = asuint64 (x * 0x1p52);
110       ix -= 52ULL << 52;
111     }
112 
113   /* x = 2^k z; where z is in range [OFF,2*OFF) and exact.
114      The range is split into N subintervals.
115      The ith subinterval contains z and c is near its center.  */
116   tmp = ix - OFF;
117   i = (tmp >> (52 - LOG_TABLE_BITS)) % N;
118   k = (int64_t) tmp >> 52; /* arithmetic shift */
119   iz = ix - (tmp & 0xfffULL << 52);
120   invc = T[i].invc;
121   logc = T[i].logc;
122   z = asdouble (iz);
123 
124   /* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
125   /* r ~= z/c - 1, |r| < 1/(2*N).  */
126 #if HAVE_FAST_FMA
127   /* rounding error: 0x1p-55/N.  */
128   r = fma (z, invc, -1.0);
129 #else
130   /* rounding error: 0x1p-55/N + 0x1p-66.  */
131   r = (z - T2[i].chi - T2[i].clo) * invc;
132 #endif
133   kd = (double_t) k;
134 
135   /* hi + lo = r + log(c) + k*Ln2.  */
136   w = kd * Ln2hi + logc;
137   hi = w + r;
138   lo = w - hi + r + kd * Ln2lo;
139 
140   /* log(x) = lo + (log1p(r) - r) + hi.  */
141   r2 = r * r; /* rounding error: 0x1p-54/N^2.  */
142   /* Worst case error if |y| > 0x1p-5:
143      0.5 + 4.13/N + abs-poly-error*2^57 ULP (+ 0.002 ULP without fma)
144      Worst case error if |y| > 0x1p-4:
145      0.5 + 2.06/N + abs-poly-error*2^56 ULP (+ 0.001 ULP without fma).  */
146 #if LOG_POLY_ORDER == 6
147   y = lo + r2 * A[0] + r * r2 * (A[1] + r * A[2] + r2 * (A[3] + r * A[4])) + hi;
148 #elif LOG_POLY_ORDER == 7
149   y = lo
150       + r2 * (A[0] + r * A[1] + r2 * (A[2] + r * A[3])
151 	      + r2 * r2 * (A[4] + r * A[5]))
152       + hi;
153 #endif
154   return eval_as_double (y);
155 }
156 #if USE_GLIBC_ABI
157 strong_alias (log, __log_finite)
158 hidden_alias (log, __ieee754_log)
159 # if LDBL_MANT_DIG == 53
160 long double logl (long double x) { return log (x); }
161 # endif
162 #endif
163