xref: /freebsd/contrib/arm-optimized-routines/pl/math/log.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*
2  * Double-precision log(x) function.
3  *
4  * Copyright (c) 2018-2023, 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 
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 optr_aor_log_f64 (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
60 	  * (B[1] + r * B[2] + r2 * B[3]
61 	     + r3 * (B[4] + r * B[5] + r2 * B[6] + r3 * (B[7] + r * B[8])));
62       w = B[0] * r2; /* B[0] == -0.5.  */
63       hi = r + w;
64       y += r - hi + w;
65       y += hi;
66 #elif LOG_POLY1_ORDER == 11
67       /* Worst-case error is around 0.516 ULP.  */
68       y = r3
69 	  * (B[1] + r * B[2]
70 	     + r2
71 		 * (B[3] + r * B[4] + r2 * B[5]
72 		    + r3 * (B[6] + r * B[7] + r2 * B[8] + r3 * B[9])));
73       w = B[0] * r2; /* B[0] == -0.5.  */
74       hi = r + w;
75       y += r - hi + w;
76       y += hi;
77 #elif LOG_POLY1_ORDER == 12
78       y = r3
79 	  * (B[1] + r * B[2] + r2 * B[3]
80 	     + r3
81 		 * (B[4] + r * B[5] + r2 * B[6]
82 		    + r3 * (B[7] + r * B[8] + r2 * B[9] + r3 * B[10])));
83 #if N <= 64
84       /* Worst-case error is around 0.532 ULP.  */
85       w = B[0] * r2; /* B[0] == -0.5.  */
86       hi = r + w;
87       y += r - hi + w;
88       y += hi;
89 #else
90       /* Worst-case error is around 0.507 ULP.  */
91       w = r * 0x1p27;
92       double_t rhi = r + w - w;
93       double_t rlo = r - rhi;
94       w = rhi * rhi * B[0]; /* B[0] == -0.5.  */
95       hi = r + w;
96       lo = r - hi + w;
97       lo += B[0] * rlo * (rhi + r);
98       y += lo;
99       y += hi;
100 #endif
101 #endif
102       return eval_as_double (y);
103     }
104   if (unlikely (top - 0x0010 >= 0x7ff0 - 0x0010))
105     {
106       /* x < 0x1p-1022 or inf or nan.  */
107       if (ix * 2 == 0)
108 	return __math_divzero (1);
109       if (ix == asuint64 (INFINITY)) /* log(inf) == inf.  */
110 	return x;
111       if ((top & 0x8000) || (top & 0x7ff0) == 0x7ff0)
112 	return __math_invalid (x);
113       /* x is subnormal, normalize it.  */
114       ix = asuint64 (x * 0x1p52);
115       ix -= 52ULL << 52;
116     }
117 
118   /* x = 2^k z; where z is in range [OFF,2*OFF) and exact.
119      The range is split into N subintervals.
120      The ith subinterval contains z and c is near its center.  */
121   tmp = ix - OFF;
122   i = (tmp >> (52 - LOG_TABLE_BITS)) % N;
123   k = (int64_t) tmp >> 52; /* arithmetic shift */
124   iz = ix - (tmp & 0xfffULL << 52);
125   invc = T[i].invc;
126   logc = T[i].logc;
127   z = asdouble (iz);
128 
129   /* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
130   /* r ~= z/c - 1, |r| < 1/(2*N).  */
131 #if HAVE_FAST_FMA
132   /* rounding error: 0x1p-55/N.  */
133   r = fma (z, invc, -1.0);
134 #else
135   /* rounding error: 0x1p-55/N + 0x1p-66.  */
136   r = (z - T2[i].chi - T2[i].clo) * invc;
137 #endif
138   kd = (double_t) k;
139 
140   /* hi + lo = r + log(c) + k*Ln2.  */
141   w = kd * Ln2hi + logc;
142   hi = w + r;
143   lo = w - hi + r + kd * Ln2lo;
144 
145   /* log(x) = lo + (log1p(r) - r) + hi.  */
146   r2 = r * r; /* rounding error: 0x1p-54/N^2.  */
147   /* Worst case error if |y| > 0x1p-5:
148      0.5 + 4.13/N + abs-poly-error*2^57 ULP (+ 0.002 ULP without fma)
149      Worst case error if |y| > 0x1p-4:
150      0.5 + 2.06/N + abs-poly-error*2^56 ULP (+ 0.001 ULP without fma).  */
151 #if LOG_POLY_ORDER == 6
152   y = lo + r2 * A[0] + r * r2 * (A[1] + r * A[2] + r2 * (A[3] + r * A[4])) + hi;
153 #elif LOG_POLY_ORDER == 7
154   y = lo
155       + r2
156 	  * (A[0] + r * A[1] + r2 * (A[2] + r * A[3])
157 	     + r2 * r2 * (A[4] + r * A[5]))
158       + hi;
159 #endif
160   return eval_as_double (y);
161 }
162