| 1 | //===-- Implementation header for atan2f16 ----------------------*- 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_ATAN2F16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ATAN2F16_H |
| 11 | |
| 12 | #include "include/llvm-libc-macros/float16-macros.h" |
| 13 | |
| 14 | #ifdef LIBC_TYPES_HAS_FLOAT16 |
| 15 | |
| 16 | #include "inv_trigf_utils.h" |
| 17 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 18 | #include "src/__support/FPUtil/FPBits.h" |
| 19 | #include "src/__support/FPUtil/cast.h" |
| 20 | #include "src/__support/FPUtil/multiply_add.h" |
| 21 | #include "src/__support/FPUtil/nearest_integer.h" |
| 22 | #include "src/__support/macros/optimization.h" |
| 23 | |
| 24 | namespace LIBC_NAMESPACE_DECL { |
| 25 | |
| 26 | namespace math { |
| 27 | |
| 28 | LIBC_INLINE float16 atan2f16(float16 y, float16 x) { |
| 29 | using namespace inv_trigf_utils_internal; |
| 30 | using FPBits = fputil::FPBits<float16>; |
| 31 | |
| 32 | constexpr double IS_NEG[2] = {1.0, -1.0}; |
| 33 | constexpr double PI = 0x1.921fb54442d18p+1; |
| 34 | constexpr double PI_OVER_2 = 0x1.921fb54442d18p+0; |
| 35 | constexpr double PI_OVER_4 = 0x1.921fb54442d18p-1; |
| 36 | constexpr double THREE_PI_OVER_4 = 0x1.2d97c7f3321d2p+1; |
| 37 | |
| 38 | // const_term[x_sign][recip]; recip = (|x| < |y|) |
| 39 | constexpr double CONST_TERM[2][2] = { |
| 40 | {0.0, -PI_OVER_2}, |
| 41 | {-PI, PI_OVER_2}, |
| 42 | }; |
| 43 | |
| 44 | FPBits x_bits(x), y_bits(y); |
| 45 | bool x_sign = x_bits.sign().is_neg(); |
| 46 | bool y_sign = y_bits.sign().is_neg(); |
| 47 | x_bits.set_sign(Sign::POS); |
| 48 | y_bits.set_sign(Sign::POS); |
| 49 | uint16_t x_abs = x_bits.uintval(); |
| 50 | uint16_t y_abs = y_bits.uintval(); |
| 51 | uint16_t max_abs = x_abs > y_abs ? x_abs : y_abs; |
| 52 | uint16_t min_abs = x_abs <= y_abs ? x_abs : y_abs; |
| 53 | |
| 54 | if (LIBC_UNLIKELY(max_abs >= 0x7c00U || min_abs == 0)) { |
| 55 | if (x_bits.is_nan() || y_bits.is_nan()) { |
| 56 | if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) |
| 57 | fputil::raise_except_if_required(FE_INVALID); |
| 58 | return FPBits::quiet_nan().get_val(); |
| 59 | } |
| 60 | size_t x_except = (x_abs == 0) ? 0 : (x_abs == 0x7c00U ? 2 : 1); |
| 61 | size_t y_except = (y_abs == 0) ? 0 : (y_abs == 0x7c00U ? 2 : 1); |
| 62 | constexpr double EXCEPTS[3][3][2] = { |
| 63 | {{0.0, PI}, {0.0, PI}, {0.0, PI}}, |
| 64 | {{PI_OVER_2, PI_OVER_2}, {0.0, 0.0}, {0.0, PI}}, |
| 65 | {{PI_OVER_2, PI_OVER_2}, |
| 66 | {PI_OVER_2, PI_OVER_2}, |
| 67 | {PI_OVER_4, THREE_PI_OVER_4}}, |
| 68 | }; |
| 69 | double r = IS_NEG[y_sign] * EXCEPTS[y_except][x_except][x_sign]; |
| 70 | return fputil::cast<float16>(x: r); |
| 71 | } |
| 72 | |
| 73 | bool recip = x_abs < y_abs; |
| 74 | double final_sign = IS_NEG[x_sign ^ y_sign ^ recip]; |
| 75 | double const_term = CONST_TERM[x_sign][recip]; |
| 76 | |
| 77 | // atan2(y,x) = final_sign * (const_term + atan(n/d)), |
| 78 | // where n = min(|x|,|y|), d = max(|x|,|y|), so 0 <= n/d <= 1. |
| 79 | // |
| 80 | // To compute atan(n/d), we use a lookup table with 16 equally-spaced knots: |
| 81 | // idx = round(16 * n/d), so |n/d - idx/16| <= 1/32. |
| 82 | // q_d = n/d - idx/16. |
| 83 | // Then by the atan addition formula: |
| 84 | // atan(n/d) = atan(idx/16) + q_d * Q(q_d) |
| 85 | // where Q(q_d) approximates |
| 86 | // [atan(idx/16 + q_d) - atan(idx/16)] / q_d |
| 87 | // via atan_eval(q_d, idx) from inv_trigf_utils. |
| 88 | double n = static_cast<double>(FPBits(min_abs).get_val()); |
| 89 | double d = static_cast<double>(FPBits(max_abs).get_val()); |
| 90 | double q_d = n / d; |
| 91 | |
| 92 | double k_d = fputil::nearest_integer(x: q_d * 0x1.0p4); |
| 93 | int idx = static_cast<int>(k_d); |
| 94 | q_d = fputil::multiply_add(x: k_d, y: -0x1.0p-4, z: q_d); |
| 95 | |
| 96 | double p = atan_eval(x: q_d, i: static_cast<unsigned>(idx)); |
| 97 | double r = final_sign * |
| 98 | fputil::multiply_add(x: q_d, y: p, z: const_term + ATAN_K_OVER_16[idx]); |
| 99 | return fputil::cast<float16>(x: r); |
| 100 | } |
| 101 | |
| 102 | } // namespace math |
| 103 | |
| 104 | } // namespace LIBC_NAMESPACE_DECL |
| 105 | |
| 106 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 107 | |
| 108 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ATAN2F16_H |
| 109 | |