xref: /freebsd/contrib/arm-optimized-routines/math/log10f.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner  * Single-precision log10 function.
3*f3087befSAndrew Turner  *
4*f3087befSAndrew Turner  * Copyright (c) 2022-2024, Arm Limited.
5*f3087befSAndrew Turner  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner  */
7*f3087befSAndrew Turner 
8*f3087befSAndrew Turner #include <math.h>
9*f3087befSAndrew Turner #include <stdint.h>
10*f3087befSAndrew Turner 
11*f3087befSAndrew Turner #include "math_config.h"
12*f3087befSAndrew Turner #include "test_sig.h"
13*f3087befSAndrew Turner #include "test_defs.h"
14*f3087befSAndrew Turner 
15*f3087befSAndrew Turner /* Data associated to logf:
16*f3087befSAndrew Turner 
17*f3087befSAndrew Turner    LOGF_TABLE_BITS = 4
18*f3087befSAndrew Turner    LOGF_POLY_ORDER = 4
19*f3087befSAndrew Turner 
20*f3087befSAndrew Turner    ULP error: 0.818 (nearest rounding.)
21*f3087befSAndrew Turner    Relative error: 1.957 * 2^-26 (before rounding.).  */
22*f3087befSAndrew Turner 
23*f3087befSAndrew Turner #define T __logf_data.tab
24*f3087befSAndrew Turner #define A __logf_data.poly
25*f3087befSAndrew Turner #define Ln2 __logf_data.ln2
26*f3087befSAndrew Turner #define InvLn10 __logf_data.invln10
27*f3087befSAndrew Turner #define N (1 << LOGF_TABLE_BITS)
28*f3087befSAndrew Turner #define OFF 0x3f330000
29*f3087befSAndrew Turner 
30*f3087befSAndrew Turner /* This naive implementation of log10f mimics that of log
31*f3087befSAndrew Turner    then simply scales the result by 1/log(10) to switch from base e to
32*f3087befSAndrew Turner    base 10. Hence, most computations are carried out in double precision.
33*f3087befSAndrew Turner    Scaling before rounding to single precision is both faster and more
34*f3087befSAndrew Turner    accurate.
35*f3087befSAndrew Turner 
36*f3087befSAndrew Turner    ULP error: 0.797 ulp (nearest rounding.).  */
37*f3087befSAndrew Turner float
log10f(float x)38*f3087befSAndrew Turner log10f (float x)
39*f3087befSAndrew Turner {
40*f3087befSAndrew Turner   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
41*f3087befSAndrew Turner   double_t z, r, r2, y, y0, invc, logc;
42*f3087befSAndrew Turner   uint32_t ix, iz, tmp;
43*f3087befSAndrew Turner   int k, i;
44*f3087befSAndrew Turner 
45*f3087befSAndrew Turner   ix = asuint (x);
46*f3087befSAndrew Turner #if WANT_ROUNDING
47*f3087befSAndrew Turner   /* Fix sign of zero with downward rounding when x==1.  */
48*f3087befSAndrew Turner   if (unlikely (ix == 0x3f800000))
49*f3087befSAndrew Turner     return 0;
50*f3087befSAndrew Turner #endif
51*f3087befSAndrew Turner   if (unlikely (ix - 0x00800000 >= 0x7f800000 - 0x00800000))
52*f3087befSAndrew Turner     {
53*f3087befSAndrew Turner       /* x < 0x1p-126 or inf or nan.  */
54*f3087befSAndrew Turner       if (ix * 2 == 0)
55*f3087befSAndrew Turner 	return __math_divzerof (1);
56*f3087befSAndrew Turner       if (ix == 0x7f800000) /* log(inf) == inf.  */
57*f3087befSAndrew Turner 	return x;
58*f3087befSAndrew Turner       if ((ix & 0x80000000) || ix * 2 >= 0xff000000)
59*f3087befSAndrew Turner 	return __math_invalidf (x);
60*f3087befSAndrew Turner       /* x is subnormal, normalize it.  */
61*f3087befSAndrew Turner       ix = asuint (x * 0x1p23f);
62*f3087befSAndrew Turner       ix -= 23 << 23;
63*f3087befSAndrew Turner     }
64*f3087befSAndrew Turner 
65*f3087befSAndrew Turner   /* x = 2^k z; where z is in range [OFF,2*OFF] and exact.
66*f3087befSAndrew Turner      The range is split into N subintervals.
67*f3087befSAndrew Turner      The ith subinterval contains z and c is near its center.  */
68*f3087befSAndrew Turner   tmp = ix - OFF;
69*f3087befSAndrew Turner   i = (tmp >> (23 - LOGF_TABLE_BITS)) % N;
70*f3087befSAndrew Turner   k = (int32_t) tmp >> 23; /* arithmetic shift.  */
71*f3087befSAndrew Turner   iz = ix - (tmp & 0xff800000);
72*f3087befSAndrew Turner   invc = T[i].invc;
73*f3087befSAndrew Turner   logc = T[i].logc;
74*f3087befSAndrew Turner   z = (double_t) asfloat (iz);
75*f3087befSAndrew Turner 
76*f3087befSAndrew Turner   /* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
77*f3087befSAndrew Turner   r = z * invc - 1;
78*f3087befSAndrew Turner   y0 = logc + (double_t) k * Ln2;
79*f3087befSAndrew Turner 
80*f3087befSAndrew Turner   /* Pipelined polynomial evaluation to approximate log1p(r).  */
81*f3087befSAndrew Turner   r2 = r * r;
82*f3087befSAndrew Turner   y = A[1] * r + A[2];
83*f3087befSAndrew Turner   y = A[0] * r2 + y;
84*f3087befSAndrew Turner   y = y * r2 + (y0 + r);
85*f3087befSAndrew Turner 
86*f3087befSAndrew Turner   /* Multiply by 1/log(10).  */
87*f3087befSAndrew Turner   y = y * InvLn10;
88*f3087befSAndrew Turner 
89*f3087befSAndrew Turner   return eval_as_float (y);
90*f3087befSAndrew Turner }
91*f3087befSAndrew Turner 
92*f3087befSAndrew Turner TEST_SIG (S, F, 1, log10, 0.01, 11.1)
93*f3087befSAndrew Turner TEST_ULP (log10f, 0.30)
94*f3087befSAndrew Turner TEST_ULP_NONNEAREST (log10f, 0.5)
95*f3087befSAndrew Turner TEST_INTERVAL (log10f, 0, 0xffff0000, 10000)
96*f3087befSAndrew Turner TEST_INTERVAL (log10f, 0x1p-127, 0x1p-26, 50000)
97*f3087befSAndrew Turner TEST_INTERVAL (log10f, 0x1p-26, 0x1p3, 50000)
98*f3087befSAndrew Turner TEST_INTERVAL (log10f, 0x1p-4, 0x1p4, 50000)
99*f3087befSAndrew Turner TEST_INTERVAL (log10f, 0, inf, 50000)
100