1 /* Single-precision 10^x function.
2    Copyright (C) 2020-2022 Free Software Foundation, Inc.
3    This file is part of the GNU C Library.
4 
5    The GNU C Library is free software; you can redistribute it and/or
6    modify it under the terms of the GNU Lesser General Public
7    License as published by the Free Software Foundation; either
8    version 2.1 of the License, or (at your option) any later version.
9 
10    The GNU C Library is distributed in the hope that it will be useful,
11    but WITHOUT ANY WARRANTY; without even the implied warranty of
12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13    Lesser General Public License for more details.
14 
15    You should have received a copy of the GNU Lesser General Public
16    License along with the GNU C Library; if not, see
17    <https://www.gnu.org/licenses/>.  */
18 
19 #include <math.h>
20 #include <math-narrow-eval.h>
21 #include <stdint.h>
22 #include <libm-alias-finite.h>
23 #include <libm-alias-float.h>
24 #include <shlib-compat.h>
25 #include <math-svid-compat.h>
26 #include "math_config.h"
27 
28 /*
29   EXP2F_TABLE_BITS 5
30   EXP2F_POLY_ORDER 3
31 
32   Max. ULP error: 0.502 (normal range, nearest rounding).
33   Max. relative error: 2^-33.240 (before rounding to float).
34   Wrong count: 169839.
35   Non-nearest ULP error: 1 (rounded ULP error).
36 
37   Detailed error analysis (for EXP2F_TABLE_BITS=3 thus N=32):
38 
39   - We first compute z = RN(InvLn10N * x) where
40 
41       InvLn10N = N*log(10)/log(2) * (1 + theta1) with |theta1| < 2^-54.150
42 
43     since z is rounded to nearest:
44 
45       z = InvLn10N * x * (1 + theta2) with |theta2| < 2^-53
46 
47     thus z =  N*log(10)/log(2) * x * (1 + theta3) with |theta3| < 2^-52.463
48 
49   - Since |x| < 38 in the main branch, we deduce:
50 
51     z = N*log(10)/log(2) * x + theta4 with |theta4| < 2^-40.483
52 
53   - We then write z = k + r where k is an integer and |r| <= 0.5 (exact)
54 
55   - We thus have
56 
57     x = z*log(2)/(N*log(10)) - theta4*log(2)/(N*log(10))
58       = z*log(2)/(N*log(10)) + theta5 with |theta5| < 2^-47.215
59 
60     10^x = 2^(k/N) * 2^(r/N) * 10^theta5
61          =  2^(k/N) * 2^(r/N) * (1 + theta6) with |theta6| < 2^-46.011
62 
63   - Then 2^(k/N) is approximated by table lookup, the maximal relative error
64     is for (k%N) = 5, with
65 
66       s = 2^(i/N) * (1 + theta7) with |theta7| < 2^-53.249
67 
68   - 2^(r/N) is approximated by a degree-3 polynomial, where the maximal
69     mathematical relative error is 2^-33.243.
70 
71   - For C[0] * r + C[1], assuming no FMA is used, since |r| <= 0.5 and
72     |C[0]| < 1.694e-6, |C[0] * r| < 8.47e-7, and the absolute error on
73     C[0] * r is bounded by 1/2*ulp(8.47e-7) = 2^-74.  Then we add C[1] with
74     |C[1]| < 0.000235, thus the absolute error on C[0] * r + C[1] is bounded
75     by 2^-65.994 (z is bounded by 0.000236).
76 
77   - For r2 = r * r, since |r| <= 0.5, we have |r2| <= 0.25 and the absolute
78     error is bounded by 1/4*ulp(0.25) = 2^-56 (the factor 1/4 is because |r2|
79     cannot exceed 1/4, and on the left of 1/4 the distance between two
80     consecutive numbers is 1/4*ulp(1/4)).
81 
82   - For y = C[2] * r + 1, assuming no FMA is used, since |r| <= 0.5 and
83     |C[2]| < 0.0217, the absolute error on C[2] * r is bounded by 2^-60,
84     and thus the absolute error on C[2] * r + 1 is bounded by 1/2*ulp(1)+2^60
85     < 2^-52.988, and |y| < 1.01085 (the error bound is better if a FMA is
86     used).
87 
88   - for z * r2 + y: the absolute error on z is bounded by 2^-65.994, with
89     |z| < 0.000236, and the absolute error on r2 is bounded by 2^-56, with
90     r2 < 0.25, thus |z*r2| < 0.000059, and the absolute error on z * r2
91     (including the rounding error) is bounded by:
92 
93       2^-65.994 * 0.25 + 0.000236 * 2^-56 + 1/2*ulp(0.000059) < 2^-66.429.
94 
95     Now we add y, with |y| < 1.01085, and error on y bounded by 2^-52.988,
96     thus the absolute error is bounded by:
97 
98       2^-66.429 + 2^-52.988 + 1/2*ulp(1.01085) < 2^-51.993
99 
100   - Now we convert the error on y into relative error.  Recall that y
101     approximates 2^(r/N), for |r| <= 0.5 and N=32. Thus 2^(-0.5/32) <= y,
102     and the relative error on y is bounded by
103 
104       2^-51.993/2^(-0.5/32) < 2^-51.977
105 
106   - Taking into account the mathematical relative error of 2^-33.243 we have:
107 
108       y = 2^(r/N) * (1 + theta8) with |theta8| < 2^-33.242
109 
110   - Since we had s = 2^(k/N) * (1 + theta7) with |theta7| < 2^-53.249,
111     after y = y * s we get y = 2^(k/N) * 2^(r/N) * (1 + theta9)
112     with |theta9| < 2^-33.241
113 
114   - Finally, taking into account the error theta6 due to the rounding error on
115     InvLn10N, and when multiplying InvLn10N * x, we get:
116 
117       y = 10^x * (1 + theta10) with |theta10| < 2^-33.240
118 
119   - Converting into binary64 ulps, since y < 2^53*ulp(y), the error is at most
120     2^19.76 ulp(y)
121 
122   - If the result is a binary32 value in the normal range (i.e., >= 2^-126),
123     then the error is at most 2^-9.24 ulps, i.e., less than 0.00166 (in the
124     subnormal range, the error in ulps might be larger).
125 
126   Note that this bound is tight, since for x=-0x25.54ac0p0 the final value of
127   y (before conversion to float) differs from 879853 ulps from the correctly
128   rounded value, and 879853 ~ 2^19.7469.  */
129 
130 #define N (1 << EXP2F_TABLE_BITS)
131 
132 #define InvLn10N (0x3.5269e12f346e2p0 * N) /* log(10)/log(2) to nearest */
133 #define T __exp2f_data.tab
134 #define C __exp2f_data.poly_scaled
135 #define SHIFT __exp2f_data.shift
136 
137 static inline uint32_t
top13(float x)138 top13 (float x)
139 {
140   return asuint (x) >> 19;
141 }
142 
143 float
__exp10f(float x)144 __exp10f (float x)
145 {
146   uint32_t abstop;
147   uint64_t ki, t;
148   double kd, xd, z, r, r2, y, s;
149 
150   xd = (double) x;
151   abstop = top13 (x) & 0xfff; /* Ignore sign.  */
152   if (__glibc_unlikely (abstop >= top13 (38.0f)))
153     {
154       /* |x| >= 38 or x is nan.  */
155       if (asuint (x) == asuint (-INFINITY))
156         return 0.0f;
157       if (abstop >= top13 (INFINITY))
158         return x + x;
159       /* 0x26.8826ap0 is the largest value such that 10^x < 2^128.  */
160       if (x > 0x26.8826ap0f)
161         return __math_oflowf (0);
162       /* -0x2d.278d4p0 is the smallest value such that 10^x > 2^-150.  */
163       if (x < -0x2d.278d4p0f)
164         return __math_uflowf (0);
165 #if WANT_ERRNO_UFLOW
166       if (x < -0x2c.da7cfp0)
167         return __math_may_uflowf (0);
168 #endif
169       /* the smallest value such that 10^x >= 2^-126 (normal range)
170          is x = -0x25.ee060p0 */
171       /* we go through here for 2014929 values out of 2060451840
172          (not counting NaN and infinities, i.e., about 0.1% */
173     }
174 
175   /* x*N*Ln10/Ln2 = k + r with r in [-1/2, 1/2] and int k.  */
176   z = InvLn10N * xd;
177   /* |xd| < 38 thus |z| < 1216 */
178 #if TOINT_INTRINSICS
179   kd = roundtoint (z);
180   ki = converttoint (z);
181 #else
182 # define SHIFT __exp2f_data.shift
183   kd = math_narrow_eval ((double) (z + SHIFT)); /* Needs to be double.  */
184   ki = asuint64 (kd);
185   kd -= SHIFT;
186 #endif
187   r = z - kd;
188 
189   /* 10^x = 10^(k/N) * 10^(r/N) ~= s * (C0*r^3 + C1*r^2 + C2*r + 1)  */
190   t = T[ki % N];
191   t += ki << (52 - EXP2F_TABLE_BITS);
192   s = asdouble (t);
193   z = C[0] * r + C[1];
194   r2 = r * r;
195   y = C[2] * r + 1;
196   y = z * r2 + y;
197   y = y * s;
198   return (float) y;
199 }
200 #ifndef __exp10f
201 strong_alias (__exp10f, __ieee754_exp10f)
202 libm_alias_finite (__ieee754_exp10f, __exp10f)
203 /* For architectures that already provided exp10f without SVID support, there
204    is no need to add a new version.  */
205 #if !LIBM_SVID_COMPAT
206 # define EXP10F_VERSION GLIBC_2_26
207 #else
208 # define EXP10F_VERSION GLIBC_2_32
209 #endif
210 versioned_symbol (libm, __exp10f, exp10f, EXP10F_VERSION);
211 libm_alias_float_other (__exp10, exp10)
212 #endif
213