1 | //===-- Single-precision atan 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/atanf.h" |
10 | #include "inv_trigf_utils.h" |
11 | #include "src/__support/FPUtil/FPBits.h" |
12 | #include "src/__support/FPUtil/PolyEval.h" |
13 | #include "src/__support/FPUtil/except_value_utils.h" |
14 | #include "src/__support/FPUtil/multiply_add.h" |
15 | #include "src/__support/FPUtil/nearest_integer.h" |
16 | #include "src/__support/FPUtil/rounding_mode.h" |
17 | #include "src/__support/macros/config.h" |
18 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
19 | |
20 | namespace LIBC_NAMESPACE_DECL { |
21 | |
22 | LLVM_LIBC_FUNCTION(float, atanf, (float x)) { |
23 | using FPBits = typename fputil::FPBits<float>; |
24 | |
25 | constexpr double FINAL_SIGN[2] = {1.0, -1.0}; |
26 | constexpr double SIGNED_PI_OVER_2[2] = {0x1.921fb54442d18p0, |
27 | -0x1.921fb54442d18p0}; |
28 | |
29 | FPBits x_bits(x); |
30 | Sign sign = x_bits.sign(); |
31 | x_bits.set_sign(Sign::POS); |
32 | uint32_t x_abs = x_bits.uintval(); |
33 | |
34 | // x is inf or nan, |x| < 2^-4 or |x|= > 16. |
35 | if (LIBC_UNLIKELY(x_abs <= 0x3d80'0000U || x_abs >= 0x4180'0000U)) { |
36 | double x_d = static_cast<double>(x); |
37 | double const_term = 0.0; |
38 | if (LIBC_UNLIKELY(x_abs >= 0x4180'0000)) { |
39 | // atan(+-Inf) = +-pi/2. |
40 | if (x_bits.is_inf()) { |
41 | volatile double sign_pi_over_2 = SIGNED_PI_OVER_2[sign.is_neg()]; |
42 | return static_cast<float>(sign_pi_over_2); |
43 | } |
44 | if (x_bits.is_nan()) |
45 | return x; |
46 | // x >= 16 |
47 | x_d = -1.0 / x_d; |
48 | const_term = SIGNED_PI_OVER_2[sign.is_neg()]; |
49 | } |
50 | // 0 <= x < 1/16; |
51 | if (LIBC_UNLIKELY(x_bits.is_zero())) |
52 | return x; |
53 | // x <= 2^-12; |
54 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000)) { |
55 | #if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT) |
56 | return fputil::multiply_add(x, -0x1.0p-25f, x); |
57 | #else |
58 | double x_d = static_cast<double>(x); |
59 | return static_cast<float>(fputil::multiply_add(x_d, -0x1.0p-25, x_d)); |
60 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
61 | } |
62 | // Use Taylor polynomial: |
63 | // atan(x) ~ x * (1 - x^2 / 3 + x^4 / 5 - x^6 / 7 + x^8 / 9 - x^10 / 11). |
64 | constexpr double ATAN_TAYLOR[6] = { |
65 | 0x1.0000000000000p+0, -0x1.5555555555555p-2, 0x1.999999999999ap-3, |
66 | -0x1.2492492492492p-3, 0x1.c71c71c71c71cp-4, -0x1.745d1745d1746p-4, |
67 | }; |
68 | double x2 = x_d * x_d; |
69 | double x4 = x2 * x2; |
70 | double c0 = fputil::multiply_add(x2, ATAN_TAYLOR[1], ATAN_TAYLOR[0]); |
71 | double c1 = fputil::multiply_add(x2, ATAN_TAYLOR[3], ATAN_TAYLOR[2]); |
72 | double c2 = fputil::multiply_add(x2, ATAN_TAYLOR[5], ATAN_TAYLOR[4]); |
73 | double p = fputil::polyeval(x4, c0, c1, c2); |
74 | double r = fputil::multiply_add(x_d, p, const_term); |
75 | return static_cast<float>(r); |
76 | } |
77 | |
78 | // Range reduction steps: |
79 | // 1) atan(x) = sign(x) * atan(|x|) |
80 | // 2) If |x| > 1, atan(|x|) = pi/2 - atan(1/|x|) |
81 | // 3) For 1/16 < x <= 1, we find k such that: |x - k/16| <= 1/32. |
82 | // 4) Then we use polynomial approximation: |
83 | // atan(x) ~ atan((k/16) + (x - (k/16)) * Q(x - k/16) |
84 | // = P(x - k/16) |
85 | double x_d, const_term, final_sign; |
86 | int idx; |
87 | |
88 | if (x_abs > 0x3f80'0000U) { |
89 | // |x| > 1, we need to invert x, so we will perform range reduction in |
90 | // double precision. |
91 | x_d = 1.0 / static_cast<double>(x_bits.get_val()); |
92 | double k_d = fputil::nearest_integer(x_d * 0x1.0p4); |
93 | x_d = fputil::multiply_add(k_d, -0x1.0p-4, x_d); |
94 | idx = static_cast<int>(k_d); |
95 | final_sign = FINAL_SIGN[sign.is_pos()]; |
96 | // Adjust constant term of the polynomial by +- pi/2. |
97 | const_term = fputil::multiply_add(final_sign, ATAN_COEFFS[idx][0], |
98 | SIGNED_PI_OVER_2[sign.is_neg()]); |
99 | } else { |
100 | // Exceptional value: |
101 | if (LIBC_UNLIKELY(x_abs == 0x3d8d'6b23U)) { // |x| = 0x1.1ad646p-4 |
102 | return sign.is_pos() ? fputil::round_result_slightly_down(0x1.1a6386p-4f) |
103 | : fputil::round_result_slightly_up(-0x1.1a6386p-4f); |
104 | } |
105 | // Perform range reduction in single precision. |
106 | float x_f = x_bits.get_val(); |
107 | float k_f = fputil::nearest_integer(x_f * 0x1.0p4f); |
108 | x_f = fputil::multiply_add(k_f, -0x1.0p-4f, x_f); |
109 | x_d = static_cast<double>(x_f); |
110 | idx = static_cast<int>(k_f); |
111 | final_sign = FINAL_SIGN[sign.is_neg()]; |
112 | const_term = final_sign * ATAN_COEFFS[idx][0]; |
113 | } |
114 | |
115 | double p = atan_eval(x_d, idx); |
116 | double r = fputil::multiply_add(final_sign * x_d, p, const_term); |
117 | |
118 | return static_cast<float>(r); |
119 | } |
120 | |
121 | } // namespace LIBC_NAMESPACE_DECL |
122 | |