1 | //===-- Single-precision log(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/logf.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/FPBits.h" |
13 | #include "src/__support/FPUtil/PolyEval.h" |
14 | #include "src/__support/FPUtil/except_value_utils.h" |
15 | #include "src/__support/FPUtil/multiply_add.h" |
16 | #include "src/__support/common.h" |
17 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
18 | #include "src/__support/macros/properties/cpu_features.h" |
19 | |
20 | // This is an algorithm for log(x) in single precision which is correctly |
21 | // rounded for all rounding modes, based on the implementation of log(x) from |
22 | // the RLIBM project at: |
23 | // https://people.cs.rutgers.edu/~sn349/rlibm |
24 | |
25 | // Step 1 - Range reduction: |
26 | // For x = 2^m * 1.mant, log(x) = m * log(2) + log(1.m) |
27 | // If x is denormal, we normalize it by multiplying x by 2^23 and subtracting |
28 | // m by 23. |
29 | |
30 | // Step 2 - Another range reduction: |
31 | // To compute log(1.mant), let f be the highest 8 bits including the hidden |
32 | // bit, and d be the difference (1.mant - f), i.e. the remaining 16 bits of the |
33 | // mantissa. Then we have the following approximation formula: |
34 | // log(1.mant) = log(f) + log(1.mant / f) |
35 | // = log(f) + log(1 + d/f) |
36 | // ~ log(f) + P(d/f) |
37 | // since d/f is sufficiently small. |
38 | // log(f) and 1/f are then stored in two 2^7 = 128 entries look-up tables. |
39 | |
40 | // Step 3 - Polynomial approximation: |
41 | // To compute P(d/f), we use a single degree-5 polynomial in double precision |
42 | // which provides correct rounding for all but few exception values. |
43 | // For more detail about how this polynomial is obtained, please refer to the |
44 | // paper: |
45 | // Lim, J. and Nagarakatte, S., "One Polynomial Approximation to Produce |
46 | // Correctly Rounded Results of an Elementary Function for Multiple |
47 | // Representations and Rounding Modes", Proceedings of the 49th ACM SIGPLAN |
48 | // Symposium on Principles of Programming Languages (POPL-2022), Philadelphia, |
49 | // USA, January 16-22, 2022. |
50 | // https://people.cs.rutgers.edu/~sn349/papers/rlibmall-popl-2022.pdf |
51 | |
52 | namespace LIBC_NAMESPACE { |
53 | |
54 | LLVM_LIBC_FUNCTION(float, logf, (float x)) { |
55 | constexpr double LOG_2 = 0x1.62e42fefa39efp-1; |
56 | using FPBits = typename fputil::FPBits<float>; |
57 | |
58 | FPBits xbits(x); |
59 | uint32_t x_u = xbits.uintval(); |
60 | |
61 | int m = -FPBits::EXP_BIAS; |
62 | |
63 | using fputil::round_result_slightly_down; |
64 | using fputil::round_result_slightly_up; |
65 | |
66 | // Small inputs |
67 | if (x_u < 0x4c5d65a5U) { |
68 | // Hard-to-round cases. |
69 | switch (x_u) { |
70 | case 0x3f7f4d6fU: // x = 0x1.fe9adep-1f |
71 | return round_result_slightly_up(value_rn: -0x1.659ec8p-9f); |
72 | case 0x41178febU: // x = 0x1.2f1fd6p+3f |
73 | return round_result_slightly_up(value_rn: 0x1.1fcbcep+1f); |
74 | #ifdef LIBC_TARGET_CPU_HAS_FMA |
75 | case 0x3f800000U: // x = 1.0f |
76 | return 0.0f; |
77 | #else |
78 | case 0x1e88452dU: // x = 0x1.108a5ap-66f |
79 | return round_result_slightly_up(-0x1.6d7b18p+5f); |
80 | #endif // LIBC_TARGET_CPU_HAS_FMA |
81 | } |
82 | // Subnormal inputs. |
83 | if (LIBC_UNLIKELY(x_u < FPBits::min_normal().uintval())) { |
84 | if (x_u == 0) { |
85 | // Return -inf and raise FE_DIVBYZERO |
86 | fputil::set_errno_if_required(ERANGE); |
87 | fputil::raise_except_if_required(FE_DIVBYZERO); |
88 | return FPBits::inf(sign: Sign::NEG).get_val(); |
89 | } |
90 | // Normalize denormal inputs. |
91 | xbits = FPBits(xbits.get_val() * 0x1.0p23f); |
92 | m -= 23; |
93 | x_u = xbits.uintval(); |
94 | } |
95 | } else { |
96 | // Hard-to-round cases. |
97 | switch (x_u) { |
98 | case 0x4c5d65a5U: // x = 0x1.bacb4ap+25f |
99 | return round_result_slightly_down(value_rn: 0x1.1e0696p+4f); |
100 | case 0x65d890d3U: // x = 0x1.b121a6p+76f |
101 | return round_result_slightly_down(value_rn: 0x1.a9a3f2p+5f); |
102 | case 0x6f31a8ecU: // x = 0x1.6351d8p+95f |
103 | return round_result_slightly_down(value_rn: 0x1.08b512p+6f); |
104 | case 0x7a17f30aU: // x = 0x1.2fe614p+117f |
105 | return round_result_slightly_up(value_rn: 0x1.451436p+6f); |
106 | #ifndef LIBC_TARGET_CPU_HAS_FMA |
107 | case 0x500ffb03U: // x = 0x1.1ff606p+33f |
108 | return round_result_slightly_up(0x1.6fdd34p+4f); |
109 | case 0x5cd69e88U: // x = 0x1.ad3d1p+58f |
110 | return round_result_slightly_up(0x1.45c146p+5f); |
111 | case 0x5ee8984eU: // x = 0x1.d1309cp+62f; |
112 | return round_result_slightly_up(0x1.5c9442p+5f); |
113 | #endif // LIBC_TARGET_CPU_HAS_FMA |
114 | } |
115 | // Exceptional inputs. |
116 | if (LIBC_UNLIKELY(x_u > FPBits::max_normal().uintval())) { |
117 | if (x_u == 0x8000'0000U) { |
118 | // Return -inf and raise FE_DIVBYZERO |
119 | fputil::set_errno_if_required(ERANGE); |
120 | fputil::raise_except_if_required(FE_DIVBYZERO); |
121 | return FPBits::inf(sign: Sign::NEG).get_val(); |
122 | } |
123 | if (xbits.is_neg() && !xbits.is_nan()) { |
124 | // Return NaN and raise FE_INVALID |
125 | fputil::set_errno_if_required(EDOM); |
126 | fputil::raise_except_if_required(FE_INVALID); |
127 | return FPBits::quiet_nan().get_val(); |
128 | } |
129 | // x is +inf or nan |
130 | return x; |
131 | } |
132 | } |
133 | |
134 | #ifndef LIBC_TARGET_CPU_HAS_FMA |
135 | // Returning the correct +0 when x = 1.0 for non-FMA targets with FE_DOWNWARD |
136 | // rounding mode. |
137 | if (LIBC_UNLIKELY((x_u & 0x007f'ffffU) == 0)) |
138 | return static_cast<float>( |
139 | static_cast<double>(m + xbits.get_biased_exponent()) * LOG_2); |
140 | #endif // LIBC_TARGET_CPU_HAS_FMA |
141 | |
142 | uint32_t mant = xbits.get_mantissa(); |
143 | // Extract 7 leading fractional bits of the mantissa |
144 | int index = mant >> 16; |
145 | // Add unbiased exponent. Add an extra 1 if the 7 leading fractional bits are |
146 | // all 1's. |
147 | m += static_cast<int>((x_u + (1 << 16)) >> 23); |
148 | |
149 | // Set bits to 1.m |
150 | xbits.set_biased_exponent(0x7F); |
151 | |
152 | float u = xbits.get_val(); |
153 | double v; |
154 | #ifdef LIBC_TARGET_CPU_HAS_FMA |
155 | v = static_cast<double>(fputil::multiply_add(x: u, y: R[index], z: -1.0f)); // Exact. |
156 | #else |
157 | v = fputil::multiply_add(static_cast<double>(u), RD[index], -1.0); // Exact |
158 | #endif // LIBC_TARGET_CPU_HAS_FMA |
159 | |
160 | // Degree-5 polynomial approximation of log generated by Sollya with: |
161 | // > P = fpminimax(log(1 + x)/x, 4, [|1, D...|], [-2^-8, 2^-7]); |
162 | constexpr double COEFFS[4] = {-0x1.000000000fe63p-1, 0x1.555556e963c16p-2, |
163 | -0x1.000028dedf986p-2, 0x1.966681bfda7f7p-3}; |
164 | double v2 = v * v; // Exact |
165 | double p2 = fputil::multiply_add(x: v, y: COEFFS[3], z: COEFFS[2]); |
166 | double p1 = fputil::multiply_add(x: v, y: COEFFS[1], z: COEFFS[0]); |
167 | double p0 = LOG_R[index] + v; |
168 | double r = fputil::multiply_add(x: static_cast<double>(m), y: LOG_2, |
169 | z: fputil::polyeval(x: v2, a0: p0, a: p1, a: p2)); |
170 | return static_cast<float>(r); |
171 | } |
172 | |
173 | } // namespace LIBC_NAMESPACE |
174 | |