| 1 | //===-- Implementation header for atanf -------------------------*- 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___SUPPORT_MATH_ATANF_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ATANF_H |
| 11 | |
| 12 | #include "inv_trigf_utils.h" |
| 13 | #include "src/__support/FPUtil/FPBits.h" |
| 14 | #include "src/__support/FPUtil/PolyEval.h" |
| 15 | #include "src/__support/FPUtil/except_value_utils.h" |
| 16 | #include "src/__support/FPUtil/multiply_add.h" |
| 17 | #include "src/__support/FPUtil/nearest_integer.h" |
| 18 | #include "src/__support/macros/config.h" |
| 19 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 20 | |
| 21 | #if defined(LIBC_MATH_HAS_SKIP_ACCURATE_PASS) && \ |
| 22 | defined(LIBC_MATH_HAS_INTERMEDIATE_COMP_IN_FLOAT) |
| 23 | |
| 24 | // We use float-float implementation to reduce size. |
| 25 | #include "atanf_float.h" |
| 26 | |
| 27 | #else |
| 28 | |
| 29 | namespace LIBC_NAMESPACE_DECL { |
| 30 | |
| 31 | namespace math { |
| 32 | |
| 33 | LIBC_INLINE float atanf(float x) { |
| 34 | using namespace inv_trigf_utils_internal; |
| 35 | using FPBits = typename fputil::FPBits<float>; |
| 36 | |
| 37 | constexpr double FINAL_SIGN[2] = {1.0, -1.0}; |
| 38 | constexpr double SIGNED_PI_OVER_2[2] = {0x1.921fb54442d18p0, |
| 39 | -0x1.921fb54442d18p0}; |
| 40 | |
| 41 | FPBits x_bits(x); |
| 42 | Sign sign = x_bits.sign(); |
| 43 | x_bits.set_sign(Sign::POS); |
| 44 | uint32_t x_abs = x_bits.uintval(); |
| 45 | |
| 46 | // x is inf or nan, |x| < 2^-4 or |x|= > 16. |
| 47 | if (LIBC_UNLIKELY(x_abs <= 0x3d80'0000U || x_abs >= 0x4180'0000U)) { |
| 48 | double x_d = static_cast<double>(x); |
| 49 | double const_term = 0.0; |
| 50 | if (LIBC_UNLIKELY(x_abs >= 0x4180'0000)) { |
| 51 | // atan(+-Inf) = +-pi/2. |
| 52 | if (x_bits.is_inf()) { |
| 53 | volatile double sign_pi_over_2 = SIGNED_PI_OVER_2[sign.is_neg()]; |
| 54 | return static_cast<float>(sign_pi_over_2); |
| 55 | } |
| 56 | if (x_bits.is_nan()) |
| 57 | return x; |
| 58 | // x >= 16 |
| 59 | x_d = -1.0 / x_d; |
| 60 | const_term = SIGNED_PI_OVER_2[sign.is_neg()]; |
| 61 | } |
| 62 | // 0 <= x < 1/16; |
| 63 | if (LIBC_UNLIKELY(x_bits.is_zero())) |
| 64 | return x; |
| 65 | // x <= 2^-12; |
| 66 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000)) { |
| 67 | #if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT) |
| 68 | return fputil::multiply_add(x, -0x1.0p-25f, x); |
| 69 | #else |
| 70 | return static_cast<float>(fputil::multiply_add(x: x_d, y: -0x1.0p-25, z: x_d)); |
| 71 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 72 | } |
| 73 | // Use Taylor polynomial: |
| 74 | // atan(x) ~ x * (1 - x^2 / 3 + x^4 / 5 - x^6 / 7 + x^8 / 9 - x^10 / 11). |
| 75 | constexpr double ATAN_TAYLOR[6] = { |
| 76 | 0x1.0000000000000p+0, -0x1.5555555555555p-2, 0x1.999999999999ap-3, |
| 77 | -0x1.2492492492492p-3, 0x1.c71c71c71c71cp-4, -0x1.745d1745d1746p-4, |
| 78 | }; |
| 79 | double x2 = x_d * x_d; |
| 80 | double x4 = x2 * x2; |
| 81 | double c0 = fputil::multiply_add(x: x2, y: ATAN_TAYLOR[1], z: ATAN_TAYLOR[0]); |
| 82 | double c1 = fputil::multiply_add(x: x2, y: ATAN_TAYLOR[3], z: ATAN_TAYLOR[2]); |
| 83 | double c2 = fputil::multiply_add(x: x2, y: ATAN_TAYLOR[5], z: ATAN_TAYLOR[4]); |
| 84 | double p = fputil::polyeval(x: x4, a0: c0, a: c1, a: c2); |
| 85 | double r = fputil::multiply_add(x: x_d, y: p, z: const_term); |
| 86 | return static_cast<float>(r); |
| 87 | } |
| 88 | |
| 89 | // Range reduction steps: |
| 90 | // 1) atan(x) = sign(x) * atan(|x|) |
| 91 | // 2) If |x| > 1, atan(|x|) = pi/2 - atan(1/|x|) |
| 92 | // 3) For 1/16 < x <= 1, we find k such that: |x - k/16| <= 1/32. |
| 93 | // 4) Then we use polynomial approximation: |
| 94 | // atan(x) ~ atan((k/16) + (x - (k/16)) * Q(x - k/16) |
| 95 | // = P(x - k/16) |
| 96 | double x_d = 0, const_term = 0, final_sign = 0; |
| 97 | int idx = 0; |
| 98 | |
| 99 | if (x_abs > 0x3f80'0000U) { |
| 100 | // |x| > 1, we need to invert x, so we will perform range reduction in |
| 101 | // double precision. |
| 102 | x_d = 1.0 / static_cast<double>(x_bits.get_val()); |
| 103 | double k_d = fputil::nearest_integer(x: x_d * 0x1.0p4); |
| 104 | x_d = fputil::multiply_add(x: k_d, y: -0x1.0p-4, z: x_d); |
| 105 | idx = static_cast<int>(k_d); |
| 106 | final_sign = FINAL_SIGN[sign.is_pos()]; |
| 107 | // Adjust constant term of the polynomial by +- pi/2. |
| 108 | const_term = fputil::multiply_add(x: final_sign, y: ATAN_COEFFS[idx][0], |
| 109 | z: SIGNED_PI_OVER_2[sign.is_neg()]); |
| 110 | } else { |
| 111 | // Exceptional value: |
| 112 | if (LIBC_UNLIKELY(x_abs == 0x3d8d'6b23U)) { // |x| = 0x1.1ad646p-4 |
| 113 | return sign.is_pos() ? fputil::round_result_slightly_down(value_rn: 0x1.1a6386p-4f) |
| 114 | : fputil::round_result_slightly_up(value_rn: -0x1.1a6386p-4f); |
| 115 | } |
| 116 | // Perform range reduction in single precision. |
| 117 | float x_f = x_bits.get_val(); |
| 118 | float k_f = fputil::nearest_integer(x: x_f * 0x1.0p4f); |
| 119 | x_f = fputil::multiply_add(x: k_f, y: -0x1.0p-4f, z: x_f); |
| 120 | x_d = static_cast<double>(x_f); |
| 121 | idx = static_cast<int>(k_f); |
| 122 | final_sign = FINAL_SIGN[sign.is_neg()]; |
| 123 | const_term = final_sign * ATAN_COEFFS[idx][0]; |
| 124 | } |
| 125 | |
| 126 | double p = atan_eval(x: x_d, i: idx); |
| 127 | double r = fputil::multiply_add(x: final_sign * x_d, y: p, z: const_term); |
| 128 | |
| 129 | return static_cast<float>(r); |
| 130 | } |
| 131 | |
| 132 | } // namespace math |
| 133 | |
| 134 | } // namespace LIBC_NAMESPACE_DECL |
| 135 | |
| 136 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 137 | |
| 138 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ATANF_H |
| 139 | |