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/config.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 log(x) in single precision which is correctly |
22 | // rounded for all rounding modes, based on the implementation of log(x) from |
23 | // the RLIBM project at: |
24 | // https://people.cs.rutgers.edu/~sn349/rlibm |
25 | |
26 | // Step 1 - Range reduction: |
27 | // For x = 2^m * 1.mant, log(x) = m * log(2) + log(1.m) |
28 | // If x is denormal, we normalize it by multiplying x by 2^23 and subtracting |
29 | // m by 23. |
30 | |
31 | // Step 2 - Another range reduction: |
32 | // To compute log(1.mant), let f be the highest 8 bits including the hidden |
33 | // bit, and d be the difference (1.mant - f), i.e. the remaining 16 bits of the |
34 | // mantissa. Then we have the following approximation formula: |
35 | // log(1.mant) = log(f) + log(1.mant / f) |
36 | // = log(f) + log(1 + d/f) |
37 | // ~ log(f) + P(d/f) |
38 | // since d/f is sufficiently small. |
39 | // log(f) and 1/f are then stored in two 2^7 = 128 entries look-up tables. |
40 | |
41 | // Step 3 - Polynomial approximation: |
42 | // To compute P(d/f), we use a single degree-5 polynomial in double precision |
43 | // which provides correct rounding for all but few exception values. |
44 | // For more detail about how this polynomial is obtained, please refer to the |
45 | // paper: |
46 | // Lim, J. and Nagarakatte, S., "One Polynomial Approximation to Produce |
47 | // Correctly Rounded Results of an Elementary Function for Multiple |
48 | // Representations and Rounding Modes", Proceedings of the 49th ACM SIGPLAN |
49 | // Symposium on Principles of Programming Languages (POPL-2022), Philadelphia, |
50 | // USA, January 16-22, 2022. |
51 | // https://people.cs.rutgers.edu/~sn349/papers/rlibmall-popl-2022.pdf |
52 | |
53 | namespace LIBC_NAMESPACE_DECL { |
54 | |
55 | LLVM_LIBC_FUNCTION(float, logf, (float x)) { |
56 | constexpr double LOG_2 = 0x1.62e42fefa39efp-1; |
57 | using FPBits = typename fputil::FPBits<float>; |
58 | |
59 | FPBits xbits(x); |
60 | uint32_t x_u = xbits.uintval(); |
61 | |
62 | int m = -FPBits::EXP_BIAS; |
63 | |
64 | using fputil::round_result_slightly_down; |
65 | using fputil::round_result_slightly_up; |
66 | |
67 | // Small inputs |
68 | if (x_u < 0x4c5d65a5U) { |
69 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
70 | // Hard-to-round cases. |
71 | switch (x_u) { |
72 | case 0x3f7f4d6fU: // x = 0x1.fe9adep-1f |
73 | return round_result_slightly_up(-0x1.659ec8p-9f); |
74 | case 0x41178febU: // x = 0x1.2f1fd6p+3f |
75 | return round_result_slightly_up(0x1.1fcbcep+1f); |
76 | #ifdef LIBC_TARGET_CPU_HAS_FMA |
77 | case 0x3f800000U: // x = 1.0f |
78 | return 0.0f; |
79 | #else |
80 | case 0x1e88452dU: // x = 0x1.108a5ap-66f |
81 | return round_result_slightly_up(-0x1.6d7b18p+5f); |
82 | #endif // LIBC_TARGET_CPU_HAS_FMA |
83 | } |
84 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
85 | // Subnormal inputs. |
86 | if (LIBC_UNLIKELY(x_u < FPBits::min_normal().uintval())) { |
87 | if (x == 0.0f) { |
88 | // Return -inf and raise FE_DIVBYZERO |
89 | fputil::set_errno_if_required(ERANGE); |
90 | fputil::raise_except_if_required(FE_DIVBYZERO); |
91 | return FPBits::inf(Sign::NEG).get_val(); |
92 | } |
93 | // Normalize denormal inputs. |
94 | xbits = FPBits(xbits.get_val() * 0x1.0p23f); |
95 | m -= 23; |
96 | x_u = xbits.uintval(); |
97 | } |
98 | } else { |
99 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
100 | // Hard-to-round cases. |
101 | switch (x_u) { |
102 | case 0x4c5d65a5U: // x = 0x1.bacb4ap+25f |
103 | return round_result_slightly_down(0x1.1e0696p+4f); |
104 | case 0x65d890d3U: // x = 0x1.b121a6p+76f |
105 | return round_result_slightly_down(0x1.a9a3f2p+5f); |
106 | case 0x6f31a8ecU: // x = 0x1.6351d8p+95f |
107 | return round_result_slightly_down(0x1.08b512p+6f); |
108 | case 0x7a17f30aU: // x = 0x1.2fe614p+117f |
109 | return round_result_slightly_up(0x1.451436p+6f); |
110 | #ifndef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
111 | case 0x500ffb03U: // x = 0x1.1ff606p+33f |
112 | return round_result_slightly_up(0x1.6fdd34p+4f); |
113 | case 0x5cd69e88U: // x = 0x1.ad3d1p+58f |
114 | return round_result_slightly_up(0x1.45c146p+5f); |
115 | case 0x5ee8984eU: // x = 0x1.d1309cp+62f; |
116 | return round_result_slightly_up(0x1.5c9442p+5f); |
117 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
118 | } |
119 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
120 | // Exceptional inputs. |
121 | if (LIBC_UNLIKELY(x_u > FPBits::max_normal().uintval())) { |
122 | if (x_u == 0x8000'0000U) { |
123 | // Return -inf and raise FE_DIVBYZERO |
124 | fputil::set_errno_if_required(ERANGE); |
125 | fputil::raise_except_if_required(FE_DIVBYZERO); |
126 | return FPBits::inf(Sign::NEG).get_val(); |
127 | } |
128 | if (xbits.is_neg() && !xbits.is_nan()) { |
129 | // Return NaN and raise FE_INVALID |
130 | fputil::set_errno_if_required(EDOM); |
131 | fputil::raise_except_if_required(FE_INVALID); |
132 | return FPBits::quiet_nan().get_val(); |
133 | } |
134 | // x is +inf or nan |
135 | if (xbits.is_signaling_nan()) { |
136 | fputil::raise_except_if_required(FE_INVALID); |
137 | return FPBits::quiet_nan().get_val(); |
138 | } |
139 | |
140 | return x; |
141 | } |
142 | } |
143 | |
144 | #ifndef LIBC_TARGET_CPU_HAS_FMA |
145 | // Returning the correct +0 when x = 1.0 for non-FMA targets with FE_DOWNWARD |
146 | // rounding mode. |
147 | if (LIBC_UNLIKELY((x_u & 0x007f'ffffU) == 0)) |
148 | return static_cast<float>( |
149 | static_cast<double>(m + xbits.get_biased_exponent()) * LOG_2); |
150 | #endif // LIBC_TARGET_CPU_HAS_FMA |
151 | |
152 | uint32_t mant = xbits.get_mantissa(); |
153 | // Extract 7 leading fractional bits of the mantissa |
154 | int index = mant >> 16; |
155 | // Add unbiased exponent. Add an extra 1 if the 7 leading fractional bits are |
156 | // all 1's. |
157 | m += static_cast<int>((x_u + (1 << 16)) >> 23); |
158 | |
159 | // Set bits to 1.m |
160 | xbits.set_biased_exponent(0x7F); |
161 | |
162 | float u = xbits.get_val(); |
163 | double v; |
164 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
165 | v = static_cast<double>(fputil::multiply_add(u, R[index], -1.0f)); // Exact. |
166 | #else |
167 | v = fputil::multiply_add(static_cast<double>(u), RD[index], -1.0); // Exact |
168 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
169 | |
170 | // Degree-5 polynomial approximation of log generated by Sollya with: |
171 | // > P = fpminimax(log(1 + x)/x, 4, [|1, D...|], [-2^-8, 2^-7]); |
172 | constexpr double COEFFS[4] = {-0x1.000000000fe63p-1, 0x1.555556e963c16p-2, |
173 | -0x1.000028dedf986p-2, 0x1.966681bfda7f7p-3}; |
174 | double v2 = v * v; // Exact |
175 | double p2 = fputil::multiply_add(v, COEFFS[3], COEFFS[2]); |
176 | double p1 = fputil::multiply_add(v, COEFFS[1], COEFFS[0]); |
177 | double p0 = LOG_R[index] + v; |
178 | double r = fputil::multiply_add(static_cast<double>(m), LOG_2, |
179 | fputil::polyeval(v2, p0, p1, p2)); |
180 | return static_cast<float>(r); |
181 | } |
182 | |
183 | } // namespace LIBC_NAMESPACE_DECL |
184 | |