| 1 | //===-- Implementation header for asinf -------------------------*- 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_ASINHF_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ASINHF_H |
| 11 | |
| 12 | #include "acoshf_utils.h" |
| 13 | #include "src/__support/FPUtil/FPBits.h" |
| 14 | #include "src/__support/FPUtil/PolyEval.h" |
| 15 | #include "src/__support/FPUtil/multiply_add.h" |
| 16 | #include "src/__support/FPUtil/sqrt.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 | namespace math { |
| 23 | |
| 24 | LIBC_INLINE float asinhf(float x) { |
| 25 | using namespace acoshf_internal; |
| 26 | using FPBits_t = typename fputil::FPBits<float>; |
| 27 | FPBits_t xbits(x); |
| 28 | uint32_t x_u = xbits.uintval(); |
| 29 | uint32_t x_abs = xbits.abs().uintval(); |
| 30 | |
| 31 | // |x| <= 2^-3 |
| 32 | if (LIBC_UNLIKELY(x_abs <= 0x3e00'0000U)) { |
| 33 | // |x| <= 2^-26 |
| 34 | if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) { |
| 35 | return static_cast<float>(LIBC_UNLIKELY(x_abs == 0) |
| 36 | ? x |
| 37 | : (x - 0x1.5555555555555p-3 * x * x * x)); |
| 38 | } |
| 39 | |
| 40 | // Generated by Sollya with: |
| 41 | // > P = fpminimax(asinh(x)/x, [|0, 2, 4, 6, 8, 10, 12|], [|1, D...|], |
| 42 | // [0, 2^-3]); |
| 43 | // > dirtyinfnorm((asinh(x) - x*P)/asinh(x), [0, 2^-3]); |
| 44 | // 0x1.ee29e366e2913deff32ed8fa17f94bfe277a5babbp-62 |
| 45 | constexpr double COEFFS[] = { |
| 46 | -0x1.555555555551ap-3, 0x1.333333330f782p-4, -0x1.6db6dafa7f405p-5, |
| 47 | 0x1.f1c67120a7cf1p-6, -0x1.6e4b0e52674d3p-6, 0x1.10450cf441118p-6, |
| 48 | }; |
| 49 | |
| 50 | double x_d = x; |
| 51 | double x_sq = x_d * x_d; |
| 52 | double c0 = fputil::multiply_add(x: x_sq, y: COEFFS[1], z: COEFFS[0]); |
| 53 | double c1 = fputil::multiply_add(x: x_sq, y: COEFFS[3], z: COEFFS[2]); |
| 54 | double c2 = fputil::multiply_add(x: x_sq, y: COEFFS[5], z: COEFFS[4]); |
| 55 | double x_4 = x_sq * x_sq; |
| 56 | double x_3 = x_d * x_sq; |
| 57 | double p = fputil::polyeval(x: x_4, a0: c0, a: c1, a: c2); |
| 58 | return static_cast<float>(fputil::multiply_add(x: x_3, y: p, z: x_d)); |
| 59 | } |
| 60 | |
| 61 | constexpr double SIGN[2] = {1.0, -1.0}; |
| 62 | double x_sign = SIGN[x_u >> 31]; |
| 63 | double x_a = static_cast<double>(FPBits_t(x_abs).get_val()); |
| 64 | |
| 65 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 66 | // Helper functions to set results for exceptional cases. |
| 67 | auto round_result_slightly_down = [x_sign](float r) -> float { |
| 68 | return fputil::multiply_add(x: static_cast<float>(x_sign), y: r, |
| 69 | z: static_cast<float>(x_sign) * (-0x1.0p-24f)); |
| 70 | }; |
| 71 | auto round_result_slightly_up = [x_sign](float r) -> float { |
| 72 | return fputil::multiply_add(x: static_cast<float>(x_sign), y: r, |
| 73 | z: static_cast<float>(x_sign) * 0x1.0p-24f); |
| 74 | }; |
| 75 | |
| 76 | if (LIBC_UNLIKELY(x_abs >= 0x4b80'0000U)) { |
| 77 | // |x| >= 2^24 |
| 78 | // We can approximate asinh(x) = sign(x) * log(2 * |x|). |
| 79 | if (LIBC_UNLIKELY(x_abs >= FPBits_t::inf().uintval())) { |
| 80 | if (xbits.is_signaling_nan()) { |
| 81 | fputil::raise_except_if_required(FE_INVALID); |
| 82 | return FPBits_t::quiet_nan().get_val(); |
| 83 | } |
| 84 | |
| 85 | return x; |
| 86 | } |
| 87 | |
| 88 | // Exceptional cases when x > 2^24. |
| 89 | switch (x_abs) { |
| 90 | case 0x4bdd65a5: // |x| = 0x1.bacb4ap24f |
| 91 | return round_result_slightly_down(0x1.1e0696p4f); |
| 92 | case 0x4c803f2c: // |x| = 0x1.007e58p26f |
| 93 | return round_result_slightly_down(0x1.2b786cp4f); |
| 94 | case 0x4f8ffb03: // |x| = 0x1.1ff606p32f |
| 95 | return round_result_slightly_up(0x1.6fdd34p4f); |
| 96 | case 0x5c569e88: // |x| = 0x1.ad3d1p57f |
| 97 | return round_result_slightly_up(0x1.45c146p5f); |
| 98 | case 0x5e68984e: // |x| = 0x1.d1309cp61f |
| 99 | return round_result_slightly_up(0x1.5c9442p5f); |
| 100 | case 0x62f7a05a: // |x| = 0x1.ef40b4p70f |
| 101 | return round_result_slightly_up(0x1.8efc9ap5f); |
| 102 | case 0x655890d3: // |x| = 0x1.b121a6p75f |
| 103 | return round_result_slightly_down(0x1.a9a3f2p5f); |
| 104 | case 0x65de7ca6: // |x| = 0x1.bcf94cp76f |
| 105 | return round_result_slightly_up(0x1.af66cp5f); |
| 106 | case 0x6eb1a8ec: // |x| = 0x1.6351d8p94f |
| 107 | return round_result_slightly_down(0x1.08b512p6f); |
| 108 | case 0x76be09de: // |x| = 0x1.7c13bcp110f |
| 109 | return round_result_slightly_up(0x1.35569p6f); |
| 110 | case 0x7997f30a: // |x| = 0x1.2fe614p116f |
| 111 | return round_result_slightly_up(0x1.451436p6f); |
| 112 | #ifndef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 113 | case 0x7967ec37: // |x| = 0x1.cfd86ep115f |
| 114 | return round_result_slightly_up(0x1.43ff6ep6f); |
| 115 | case 0x58719e31: // |x| = 0x1.e33c62p49f |
| 116 | return round_result_slightly_down(0x1.1a576cp5f); |
| 117 | case 0x71699003: // |x| = 0x1.d32006p99f |
| 118 | return round_result_slightly_up(0x1.17aa2p6f); |
| 119 | #endif // !LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 120 | } |
| 121 | |
| 122 | return static_cast<float>(x_sign * log_eval(x: 2.0 * x_a)); |
| 123 | |
| 124 | } else { |
| 125 | // Exceptional cases when x < 2^24. |
| 126 | if (LIBC_UNLIKELY(x_abs == 0x45abaf26)) { |
| 127 | // |x| = 0x1.575e4cp12f |
| 128 | return round_result_slightly_down(0x1.29becap3f); |
| 129 | } |
| 130 | if (LIBC_UNLIKELY(x_abs == 0x49d29048)) { |
| 131 | // |x| = 0x1.a5209p20f |
| 132 | return round_result_slightly_down(0x1.e1b92p3f); |
| 133 | } |
| 134 | #ifndef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 135 | if (LIBC_UNLIKELY(x_abs == 0x45e19b90)) { |
| 136 | // |x| = 0x1.c3372p12f |
| 137 | return round_result_slightly_down(0x1.327c5cp3f); |
| 138 | } |
| 139 | #endif // !LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 140 | } |
| 141 | #else |
| 142 | if (LIBC_UNLIKELY(x_abs >= FPBits_t::inf().uintval())) { |
| 143 | if (xbits.is_signaling_nan()) { |
| 144 | fputil::raise_except_if_required(FE_INVALID); |
| 145 | return FPBits_t::quiet_nan().get_val(); |
| 146 | } |
| 147 | |
| 148 | return x; |
| 149 | } |
| 150 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 151 | |
| 152 | // asinh(x) = log(x + sqrt(x^2 + 1)) |
| 153 | return static_cast<float>( |
| 154 | x_sign * log_eval(x: x_a + fputil::sqrt<double>( |
| 155 | x: fputil::multiply_add(x: x_a, y: x_a, z: 1.0)))); |
| 156 | } |
| 157 | |
| 158 | } // namespace math |
| 159 | |
| 160 | } // namespace LIBC_NAMESPACE_DECL |
| 161 | |
| 162 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ASINHF_H |
| 163 | |