1 | //===-- Utilities for trigonometric functions with FMA ----------*- C++ -*-===// |
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 | #ifndef LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H |
10 | #define LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H |
11 | |
12 | #include "src/__support/FPUtil/FMA.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/nearest_integer.h" |
15 | #include "src/__support/common.h" |
16 | |
17 | namespace LIBC_NAMESPACE { |
18 | |
19 | namespace fma { |
20 | |
21 | static constexpr uint32_t FAST_PASS_BOUND = 0x5600'0000U; // 2^45 |
22 | |
23 | // Digits of 32/pi, generated by Sollya with: |
24 | // > a0 = D(32/pi); |
25 | // > a1 = D(32/pi - a0); |
26 | // > a2 = D(32/pi - a0 - a1); |
27 | // > a3 = D(32/pi - a0 - a1 - a2); |
28 | static constexpr double THIRTYTWO_OVER_PI[5] = { |
29 | 0x1.45f306dc9c883p+3, -0x1.6b01ec5417056p-51, -0x1.6447e493ad4cep-105, |
30 | 0x1.e21c820ff28b2p-159, -0x1.508510ea79237p-214}; |
31 | |
32 | // Return k and y, where |
33 | // k = round(x * 32 / pi) and y = (x * 32 / pi) - k. |
34 | LIBC_INLINE int64_t small_range_reduction(double x, double &y) { |
35 | double kd = fputil::nearest_integer(x: x * THIRTYTWO_OVER_PI[0]); |
36 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[0], z: -kd); |
37 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[1], z: y); |
38 | return static_cast<int64_t>(kd); |
39 | } |
40 | |
41 | // Return k and y, where |
42 | // k = round(x * 32 / pi) and y = (x * 32 / pi) - k. |
43 | // This is used for sinf, cosf, sincosf. |
44 | LIBC_INLINE int64_t large_range_reduction(double x, int x_exp, double &y) { |
45 | // 2^45 <= |x| < 2^99 |
46 | if (x_exp < 99) { |
47 | // - When x < 2^99, the full exact product of x * THIRTYTWO_OVER_PI[0] |
48 | // contains at least one integral bit <= 2^5. |
49 | // - When 2^45 <= |x| < 2^55, the lowest 6 unit bits are contained |
50 | // in the last 12 bits of double(x * THIRTYTWO_OVER_PI[0]). |
51 | // - When |x| >= 2^55, the LSB of double(x * THIRTYTWO_OVER_PI[0]) is at |
52 | // least 2^6. |
53 | fputil::FPBits<double> prod_hi(x * THIRTYTWO_OVER_PI[0]); |
54 | prod_hi.set_uintval(prod_hi.uintval() & |
55 | ((x_exp < 55) ? (~0xfffULL) : (~0ULL))); // |x| < 2^55 |
56 | double k_hi = fputil::nearest_integer(x: prod_hi.get_val()); |
57 | double truncated_prod = fputil::fma(x, y: THIRTYTWO_OVER_PI[0], z: -k_hi); |
58 | double prod_lo = fputil::fma(x, y: THIRTYTWO_OVER_PI[1], z: truncated_prod); |
59 | double k_lo = fputil::nearest_integer(x: prod_lo); |
60 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[1], z: truncated_prod - k_lo); |
61 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[2], z: y); |
62 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[3], z: y); |
63 | |
64 | return static_cast<int64_t>(k_lo); |
65 | } |
66 | |
67 | // - When x >= 2^110, the full exact product of x * THIRTYTWO_OVER_PI[0] does |
68 | // not contain any of the lowest 6 unit bits, so we can ignore it completely. |
69 | // - When 2^99 <= |x| < 2^110, the lowest 6 unit bits are contained |
70 | // in the last 12 bits of double(x * THIRTYTWO_OVER_PI[1]). |
71 | // - When |x| >= 2^110, the LSB of double(x * THIRTYTWO_OVER_PI[1]) is at |
72 | // least 64. |
73 | fputil::FPBits<double> prod_hi(x * THIRTYTWO_OVER_PI[1]); |
74 | prod_hi.set_uintval(prod_hi.uintval() & |
75 | ((x_exp < 110) ? (~0xfffULL) : (~0ULL))); // |x| < 2^110 |
76 | double k_hi = fputil::nearest_integer(x: prod_hi.get_val()); |
77 | double truncated_prod = fputil::fma(x, y: THIRTYTWO_OVER_PI[1], z: -k_hi); |
78 | double prod_lo = fputil::fma(x, y: THIRTYTWO_OVER_PI[2], z: truncated_prod); |
79 | double k_lo = fputil::nearest_integer(x: prod_lo); |
80 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[2], z: truncated_prod - k_lo); |
81 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[3], z: y); |
82 | y = fputil::fma(x, y: THIRTYTWO_OVER_PI[4], z: y); |
83 | |
84 | return static_cast<int64_t>(k_lo); |
85 | } |
86 | |
87 | } // namespace fma |
88 | |
89 | } // namespace LIBC_NAMESPACE |
90 | |
91 | #endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H |
92 | |