1 | //===-- Single-precision log2(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/log2f.h" |
10 | #include "common_constants.h" // Lookup table for (1/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 | |
19 | // This is a correctly-rounded algorithm for log2(x) in single precision with |
20 | // round-to-nearest, tie-to-even mode from the RLIBM project at: |
21 | // https://people.cs.rutgers.edu/~sn349/rlibm |
22 | |
23 | // Step 1 - Range reduction: |
24 | // For x = 2^m * 1.mant, log2(x) = m + log2(1.m) |
25 | // If x is denormal, we normalize it by multiplying x by 2^23 and subtracting |
26 | // m by 23. |
27 | |
28 | // Step 2 - Another range reduction: |
29 | // To compute log(1.mant), let f be the highest 8 bits including the hidden |
30 | // bit, and d be the difference (1.mant - f), i.e. the remaining 16 bits of the |
31 | // mantissa. Then we have the following approximation formula: |
32 | // log2(1.mant) = log2(f) + log2(1.mant / f) |
33 | // = log2(f) + log2(1 + d/f) |
34 | // ~ log2(f) + P(d/f) |
35 | // since d/f is sufficiently small. |
36 | // log2(f) and 1/f are then stored in two 2^7 = 128 entries look-up tables. |
37 | |
38 | // Step 3 - Polynomial approximation: |
39 | // To compute P(d/f), we use a single degree-5 polynomial in double precision |
40 | // which provides correct rounding for all but few exception values. |
41 | // For more detail about how this polynomial is obtained, please refer to the |
42 | // papers: |
43 | // Lim, J. and Nagarakatte, S., "One Polynomial Approximation to Produce |
44 | // Correctly Rounded Results of an Elementary Function for Multiple |
45 | // Representations and Rounding Modes", Proceedings of the 49th ACM SIGPLAN |
46 | // Symposium on Principles of Programming Languages (POPL-2022), Philadelphia, |
47 | // USA, Jan. 16-22, 2022. |
48 | // https://people.cs.rutgers.edu/~sn349/papers/rlibmall-popl-2022.pdf |
49 | // Aanjaneya, M., Lim, J., and Nagarakatte, S., "RLibm-Prog: Progressive |
50 | // Polynomial Approximations for Fast Correctly Rounded Math Libraries", |
51 | // Dept. of Comp. Sci., Rutgets U., Technical Report DCS-TR-758, Nov. 2021. |
52 | // https://arxiv.org/pdf/2111.12852.pdf. |
53 | |
54 | namespace LIBC_NAMESPACE { |
55 | |
56 | LLVM_LIBC_FUNCTION(float, log2f, (float x)) { |
57 | using FPBits = typename fputil::FPBits<float>; |
58 | |
59 | FPBits xbits(x); |
60 | uint32_t x_u = xbits.uintval(); |
61 | |
62 | // Hard to round value(s). |
63 | using fputil::round_result_slightly_up; |
64 | |
65 | int m = -FPBits::EXP_BIAS; |
66 | |
67 | // log2(1.0f) = 0.0f. |
68 | if (LIBC_UNLIKELY(x_u == 0x3f80'0000U)) |
69 | return 0.0f; |
70 | |
71 | // Exceptional inputs. |
72 | if (LIBC_UNLIKELY(x_u < FPBits::min_normal().uintval() || |
73 | x_u > FPBits::max_normal().uintval())) { |
74 | if (xbits.is_zero()) { |
75 | fputil::set_errno_if_required(ERANGE); |
76 | fputil::raise_except_if_required(FE_DIVBYZERO); |
77 | return FPBits::inf(sign: Sign::NEG).get_val(); |
78 | } |
79 | if (xbits.is_neg() && !xbits.is_nan()) { |
80 | fputil::set_errno_if_required(EDOM); |
81 | fputil::raise_except(FE_INVALID); |
82 | return FPBits::quiet_nan().get_val(); |
83 | } |
84 | if (xbits.is_inf_or_nan()) { |
85 | return x; |
86 | } |
87 | // Normalize denormal inputs. |
88 | xbits = FPBits(xbits.get_val() * 0x1.0p23f); |
89 | m -= 23; |
90 | } |
91 | |
92 | m += xbits.get_biased_exponent(); |
93 | int index = xbits.get_mantissa() >> 16; |
94 | // Set bits to 1.m |
95 | xbits.set_biased_exponent(0x7F); |
96 | |
97 | float u = xbits.get_val(); |
98 | double v; |
99 | #ifdef LIBC_TARGET_CPU_HAS_FMA |
100 | v = static_cast<double>(fputil::multiply_add(x: u, y: R[index], z: -1.0f)); // Exact. |
101 | #else |
102 | v = fputil::multiply_add(static_cast<double>(u), RD[index], -1.0); // Exact |
103 | #endif // LIBC_TARGET_CPU_HAS_FMA |
104 | |
105 | double = static_cast<double>(m) + LOG2_R[index]; |
106 | |
107 | // Degree-5 polynomial approximation of log2 generated by Sollya with: |
108 | // > P = fpminimax(log2(1 + x)/x, 4, [|1, D...|], [-2^-8, 2^-7]); |
109 | constexpr double COEFFS[5] = {0x1.71547652b8133p0, -0x1.71547652d1e33p-1, |
110 | 0x1.ec70a098473dep-2, -0x1.7154c5ccdf121p-2, |
111 | 0x1.2514fd90a130ap-2}; |
112 | |
113 | double vsq = v * v; // Exact |
114 | double c0 = fputil::multiply_add(x: v, y: COEFFS[0], z: extra_factor); |
115 | double c1 = fputil::multiply_add(x: v, y: COEFFS[2], z: COEFFS[1]); |
116 | double c2 = fputil::multiply_add(x: v, y: COEFFS[4], z: COEFFS[3]); |
117 | |
118 | double r = fputil::polyeval(x: vsq, a0: c0, a: c1, a: c2); |
119 | |
120 | return static_cast<float>(r); |
121 | } |
122 | |
123 | } // namespace LIBC_NAMESPACE |
124 | |