xref: /freebsd/contrib/arm-optimized-routines/math/log.c (revision 6c05f3a74f30934ee60919cc97e16ec69b542b06)
1 /*
2  * Double-precision log(x) function.
3  *
4  * Copyright (c) 2018-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include <float.h>
9 #include <math.h>
10 #include <stdint.h>
11 #include "math_config.h"
12 #include "test_defs.h"
13 #include "test_sig.h"
14 
15 #define T __log_data.tab
16 #define T2 __log_data.tab2
17 #define B __log_data.poly1
18 #define A __log_data.poly
19 #define Ln2hi __log_data.ln2hi
20 #define Ln2lo __log_data.ln2lo
21 #define N (1 << LOG_TABLE_BITS)
22 #define OFF 0x3fe6000000000000
23 
24 /* Top 16 bits of a double.  */
25 static inline uint32_t
26 top16 (double x)
27 {
28   return asuint64 (x) >> 48;
29 }
30 
31 double
32 log (double x)
33 {
34   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
35   double_t w, z, r, r2, r3, y, invc, logc, kd, hi, lo;
36   uint64_t ix, iz, tmp;
37   uint32_t top;
38   int k, i;
39 
40   ix = asuint64 (x);
41   top = top16 (x);
42 
43 #if LOG_POLY1_ORDER == 10 || LOG_POLY1_ORDER == 11
44 # define LO asuint64 (1.0 - 0x1p-5)
45 # define HI asuint64 (1.0 + 0x1.1p-5)
46 #elif LOG_POLY1_ORDER == 12
47 # define LO asuint64 (1.0 - 0x1p-4)
48 # define HI asuint64 (1.0 + 0x1.09p-4)
49 #endif
50   if (unlikely (ix - LO < HI - LO))
51     {
52       /* Handle close to 1.0 inputs separately.  */
53       /* Fix sign of zero with downward rounding when x==1.  */
54       if (WANT_ROUNDING && unlikely (ix == asuint64 (1.0)))
55 	return 0;
56       r = x - 1.0;
57       r2 = r * r;
58       r3 = r * r2;
59 #if LOG_POLY1_ORDER == 10
60       /* Worst-case error is around 0.516 ULP.  */
61       y = r3 * (B[1] + r * B[2] + r2 * B[3]
62 		+ r3 * (B[4] + r * B[5] + r2 * B[6] + r3 * (B[7] + r * B[8])));
63       w = B[0] * r2; /* B[0] == -0.5.  */
64       hi = r + w;
65       y += r - hi + w;
66       y += hi;
67 #elif LOG_POLY1_ORDER == 11
68       /* Worst-case error is around 0.516 ULP.  */
69       y = r3 * (B[1] + r * B[2]
70 		+ r2 * (B[3] + r * B[4] + r2 * B[5]
71 			+ r3 * (B[6] + r * B[7] + r2 * B[8] + r3 * B[9])));
72       w = B[0] * r2; /* B[0] == -0.5.  */
73       hi = r + w;
74       y += r - hi + w;
75       y += hi;
76 #elif LOG_POLY1_ORDER == 12
77       y = r3 * (B[1] + r * B[2] + r2 * B[3]
78 		+ r3 * (B[4] + r * B[5] + r2 * B[6]
79 			+ r3 * (B[7] + r * B[8] + r2 * B[9] + r3 * B[10])));
80 # if N <= 64
81       /* Worst-case error is around 0.532 ULP.  */
82       w = B[0] * r2; /* B[0] == -0.5.  */
83       hi = r + w;
84       y += r - hi + w;
85       y += hi;
86 # else
87       /* Worst-case error is around 0.507 ULP.  */
88       w = r * 0x1p27;
89       double_t rhi = r + w - w;
90       double_t rlo = r - rhi;
91       w = rhi * rhi * B[0]; /* B[0] == -0.5.  */
92       hi = r + w;
93       lo = r - hi + w;
94       lo += B[0] * rlo * (rhi + r);
95       y += lo;
96       y += hi;
97 # endif
98 #endif
99       return eval_as_double (y);
100     }
101   if (unlikely (top - 0x0010 >= 0x7ff0 - 0x0010))
102     {
103       /* x < 0x1p-1022 or inf or nan.  */
104       if (ix * 2 == 0)
105 	return __math_divzero (1);
106       if (ix == asuint64 (INFINITY)) /* log(inf) == inf.  */
107 	return x;
108       if ((top & 0x8000) || (top & 0x7ff0) == 0x7ff0)
109 	return __math_invalid (x);
110       /* x is subnormal, normalize it.  */
111       ix = asuint64 (x * 0x1p52);
112       ix -= 52ULL << 52;
113     }
114 
115   /* x = 2^k z; where z is in range [OFF,2*OFF) and exact.
116      The range is split into N subintervals.
117      The ith subinterval contains z and c is near its center.  */
118   tmp = ix - OFF;
119   i = (tmp >> (52 - LOG_TABLE_BITS)) % N;
120   k = (int64_t) tmp >> 52; /* arithmetic shift */
121   iz = ix - (tmp & 0xfffULL << 52);
122   invc = T[i].invc;
123   logc = T[i].logc;
124   z = asdouble (iz);
125 
126   /* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
127   /* r ~= z/c - 1, |r| < 1/(2*N).  */
128 #if HAVE_FAST_FMA
129   /* rounding error: 0x1p-55/N.  */
130   r = fma (z, invc, -1.0);
131 #else
132   /* rounding error: 0x1p-55/N + 0x1p-66.  */
133   r = (z - T2[i].chi - T2[i].clo) * invc;
134 #endif
135   kd = (double_t) k;
136 
137   /* hi + lo = r + log(c) + k*Ln2.  */
138   w = kd * Ln2hi + logc;
139   hi = w + r;
140   lo = w - hi + r + kd * Ln2lo;
141 
142   /* log(x) = lo + (log1p(r) - r) + hi.  */
143   r2 = r * r; /* rounding error: 0x1p-54/N^2.  */
144   /* Worst case error if |y| > 0x1p-5:
145      0.5 + 4.13/N + abs-poly-error*2^57 ULP (+ 0.002 ULP without fma)
146      Worst case error if |y| > 0x1p-4:
147      0.5 + 2.06/N + abs-poly-error*2^56 ULP (+ 0.001 ULP without fma).  */
148 #if LOG_POLY_ORDER == 6
149   y = lo + r2 * A[0] + r * r2 * (A[1] + r * A[2] + r2 * (A[3] + r * A[4])) + hi;
150 #elif LOG_POLY_ORDER == 7
151   y = lo
152       + r2 * (A[0] + r * A[1] + r2 * (A[2] + r * A[3])
153 	      + r2 * r2 * (A[4] + r * A[5]))
154       + hi;
155 #endif
156   return eval_as_double (y);
157 }
158 #if USE_GLIBC_ABI
159 strong_alias (log, __log_finite)
160 hidden_alias (log, __ieee754_log)
161 # if LDBL_MANT_DIG == 53
162 long double logl (long double x) { return log (x); }
163 # endif
164 #endif
165 
166 TEST_SIG (S, D, 1, log, 0.01, 11.1)
167 TEST_ULP (log, 0.02)
168 TEST_ULP_NONNEAREST (log, 0.5)
169 TEST_INTERVAL (log, 0, 0xffff000000000000, 10000)
170 TEST_INTERVAL (log, 0x1p-4, 0x1p4, 400000)
171 TEST_INTERVAL (log, 0, inf, 400000)
172