1 | //===-- Single-precision log1p(x) function --------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "src/math/log1pf.h" |
10 | #include "common_constants.h" // Lookup table for (1/f) and log(f) |
11 | #include "src/__support/FPUtil/FEnvImpl.h" |
12 | #include "src/__support/FPUtil/FMA.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/PolyEval.h" |
15 | #include "src/__support/FPUtil/except_value_utils.h" |
16 | #include "src/__support/FPUtil/multiply_add.h" |
17 | #include "src/__support/common.h" |
18 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
19 | #include "src/__support/macros/properties/cpu_features.h" |
20 | |
21 | // This is an algorithm for log10(x) in single precision which is |
22 | // correctly rounded for all rounding modes. |
23 | // - An exhaustive test show that when x >= 2^45, log1pf(x) == logf(x) |
24 | // for all rounding modes. |
25 | // - When 2^(-6) <= |x| < 2^45, the sum (double(x) + 1.0) is exact, |
26 | // so we can adapt the correctly rounded algorithm of logf to compute |
27 | // log(double(x) + 1.0) correctly. For more information about the logf |
28 | // algorithm, see `libc/src/math/generic/logf.cpp`. |
29 | // - When |x| < 2^(-6), we use a degree-8 polynomial in double precision |
30 | // generated with Sollya using the following command: |
31 | // fpminimax(log(1 + x)/x, 7, [|D...|], [-2^-6; 2^-6]); |
32 | |
33 | namespace LIBC_NAMESPACE { |
34 | |
35 | namespace internal { |
36 | |
37 | // We don't need to treat denormal and 0 |
38 | LIBC_INLINE float log(double x) { |
39 | constexpr double LOG_2 = 0x1.62e42fefa39efp-1; |
40 | |
41 | using FPBits = typename fputil::FPBits<double>; |
42 | FPBits xbits(x); |
43 | |
44 | uint64_t x_u = xbits.uintval(); |
45 | |
46 | if (LIBC_UNLIKELY(x_u > FPBits::max_normal().uintval())) { |
47 | if (xbits.is_neg() && !xbits.is_nan()) { |
48 | fputil::set_errno_if_required(EDOM); |
49 | fputil::raise_except_if_required(FE_INVALID); |
50 | return fputil::FPBits<float>::quiet_nan().get_val(); |
51 | } |
52 | return static_cast<float>(x); |
53 | } |
54 | |
55 | double m = static_cast<double>(xbits.get_exponent()); |
56 | |
57 | // Get the 8 highest bits, use 7 bits (excluding the implicit hidden bit) for |
58 | // lookup tables. |
59 | int f_index = static_cast<int>(xbits.get_mantissa() >> |
60 | (fputil::FPBits<double>::FRACTION_LEN - 7)); |
61 | |
62 | // Set bits to 1.m |
63 | xbits.set_biased_exponent(0x3FF); |
64 | FPBits f = xbits; |
65 | |
66 | // Clear the lowest 45 bits. |
67 | f.set_uintval(f.uintval() & ~0x0000'1FFF'FFFF'FFFFULL); |
68 | |
69 | double d = xbits.get_val() - f.get_val(); |
70 | d *= ONE_OVER_F[f_index]; |
71 | |
72 | double = fputil::multiply_add(x: m, y: LOG_2, z: LOG_F[f_index]); |
73 | |
74 | double r = fputil::polyeval(x: d, a0: extra_factor, a: 0x1.fffffffffffacp-1, |
75 | a: -0x1.fffffffef9cb2p-2, a: 0x1.5555513bc679ap-2, |
76 | a: -0x1.fff4805ea441p-3, a: 0x1.930180dbde91ap-3); |
77 | |
78 | return static_cast<float>(r); |
79 | } |
80 | |
81 | } // namespace internal |
82 | |
83 | LLVM_LIBC_FUNCTION(float, log1pf, (float x)) { |
84 | using FPBits = typename fputil::FPBits<float>; |
85 | FPBits xbits(x); |
86 | uint32_t x_u = xbits.uintval(); |
87 | uint32_t x_a = x_u & 0x7fff'ffffU; |
88 | double xd = static_cast<double>(x); |
89 | |
90 | // Use log1p(x) = log(1 + x) for |x| > 2^-6; |
91 | if (x_a > 0x3c80'0000U) { |
92 | // Hard-to-round cases. |
93 | switch (x_u) { |
94 | case 0x41078febU: // x = 0x1.0f1fd6p3 |
95 | return fputil::round_result_slightly_up(value_rn: 0x1.1fcbcep1f); |
96 | case 0x5cd69e88U: // x = 0x1.ad3d1p+58f |
97 | return fputil::round_result_slightly_up(value_rn: 0x1.45c146p+5f); |
98 | case 0x65d890d3U: // x = 0x1.b121a6p+76f |
99 | return fputil::round_result_slightly_down(value_rn: 0x1.a9a3f2p+5f); |
100 | case 0x6f31a8ecU: // x = 0x1.6351d8p+95f |
101 | return fputil::round_result_slightly_down(value_rn: 0x1.08b512p+6f); |
102 | case 0x7a17f30aU: // x = 0x1.2fe614p+117f |
103 | return fputil::round_result_slightly_up(value_rn: 0x1.451436p+6f); |
104 | case 0xbd1d20afU: // x = -0x1.3a415ep-5f |
105 | return fputil::round_result_slightly_up(value_rn: -0x1.407112p-5f); |
106 | case 0xbf800000U: // x = -1.0 |
107 | fputil::set_errno_if_required(ERANGE); |
108 | fputil::raise_except_if_required(FE_DIVBYZERO); |
109 | return FPBits::inf(sign: Sign::NEG).get_val(); |
110 | #ifndef LIBC_TARGET_CPU_HAS_FMA |
111 | case 0x4cc1c80bU: // x = 0x1.839016p+26f |
112 | return fputil::round_result_slightly_down(0x1.26fc04p+4f); |
113 | case 0x5ee8984eU: // x = 0x1.d1309cp+62f |
114 | return fputil::round_result_slightly_up(0x1.5c9442p+5f); |
115 | case 0x665e7ca6U: // x = 0x1.bcf94cp+77f |
116 | return fputil::round_result_slightly_up(0x1.af66cp+5f); |
117 | case 0x79e7ec37U: // x = 0x1.cfd86ep+116f |
118 | return fputil::round_result_slightly_up(0x1.43ff6ep+6f); |
119 | #endif // LIBC_TARGET_CPU_HAS_FMA |
120 | } |
121 | |
122 | return internal::log(x: xd + 1.0); |
123 | } |
124 | |
125 | // |x| <= 2^-6. |
126 | // Hard-to round cases. |
127 | switch (x_u) { |
128 | case 0x35400003U: // x = 0x1.800006p-21f |
129 | return fputil::round_result_slightly_down(value_rn: 0x1.7ffffep-21f); |
130 | case 0x3710001bU: // x = 0x1.200036p-17f |
131 | return fputil::round_result_slightly_down(value_rn: 0x1.1fffe6p-17f); |
132 | case 0xb53ffffdU: // x = -0x1.7ffffap-21 |
133 | return fputil::round_result_slightly_down(value_rn: -0x1.800002p-21f); |
134 | case 0xb70fffe5U: // x = -0x1.1fffcap-17 |
135 | return fputil::round_result_slightly_down(value_rn: -0x1.20001ap-17f); |
136 | case 0xbb0ec8c4U: // x = -0x1.1d9188p-9 |
137 | return fputil::round_result_slightly_up(value_rn: -0x1.1de14ap-9f); |
138 | } |
139 | |
140 | // Polymial generated by Sollya with: |
141 | // > fpminimax(log(1 + x)/x, 7, [|D...|], [-2^-6; 2^-6]); |
142 | const double COEFFS[7] = {-0x1.0000000000000p-1, 0x1.5555555556aadp-2, |
143 | -0x1.000000000181ap-2, 0x1.999998998124ep-3, |
144 | -0x1.55555452e2a2bp-3, 0x1.24adb8cde4aa7p-3, |
145 | -0x1.0019db915ef6fp-3}; |
146 | |
147 | double xsq = xd * xd; |
148 | double c0 = fputil::multiply_add(x: xd, y: COEFFS[1], z: COEFFS[0]); |
149 | double c1 = fputil::multiply_add(x: xd, y: COEFFS[3], z: COEFFS[2]); |
150 | double c2 = fputil::multiply_add(x: xd, y: COEFFS[5], z: COEFFS[4]); |
151 | double r = fputil::polyeval(x: xsq, a0: xd, a: c0, a: c1, a: c2, a: COEFFS[6]); |
152 | |
153 | return static_cast<float>(r); |
154 | } |
155 | |
156 | } // namespace LIBC_NAMESPACE |
157 | |