| 1 | //===-- Implementation header for atan2f128 ---------------------*- 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_ATAN2F128_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ATAN2F128_H |
| 11 | |
| 12 | #include "include/llvm-libc-types/float128.h" |
| 13 | |
| 14 | #ifdef LIBC_TYPES_HAS_FLOAT128 |
| 15 | |
| 16 | #include "atan_utils.h" |
| 17 | #include "src/__support/FPUtil/FPBits.h" |
| 18 | #include "src/__support/FPUtil/dyadic_float.h" |
| 19 | #include "src/__support/FPUtil/nearest_integer.h" |
| 20 | #include "src/__support/integer_literals.h" |
| 21 | #include "src/__support/macros/config.h" |
| 22 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 23 | #include "src/__support/uint128.h" |
| 24 | |
| 25 | namespace LIBC_NAMESPACE_DECL { |
| 26 | |
| 27 | namespace math { |
| 28 | |
| 29 | // There are several range reduction steps we can take for atan2(y, x) as |
| 30 | // follow: |
| 31 | |
| 32 | // * Range reduction 1: signness |
| 33 | // atan2(y, x) will return a number between -PI and PI representing the angle |
| 34 | // forming by the 0x axis and the vector (x, y) on the 0xy-plane. |
| 35 | // In particular, we have that: |
| 36 | // atan2(y, x) = atan( y/x ) if x >= 0 and y >= 0 (I-quadrant) |
| 37 | // = pi + atan( y/x ) if x < 0 and y >= 0 (II-quadrant) |
| 38 | // = -pi + atan( y/x ) if x < 0 and y < 0 (III-quadrant) |
| 39 | // = atan( y/x ) if x >= 0 and y < 0 (IV-quadrant) |
| 40 | // Since atan function is odd, we can use the formula: |
| 41 | // atan(-u) = -atan(u) |
| 42 | // to adjust the above conditions a bit further: |
| 43 | // atan2(y, x) = atan( |y|/|x| ) if x >= 0 and y >= 0 (I-quadrant) |
| 44 | // = pi - atan( |y|/|x| ) if x < 0 and y >= 0 (II-quadrant) |
| 45 | // = -pi + atan( |y|/|x| ) if x < 0 and y < 0 (III-quadrant) |
| 46 | // = -atan( |y|/|x| ) if x >= 0 and y < 0 (IV-quadrant) |
| 47 | // Which can be simplified to: |
| 48 | // atan2(y, x) = sign(y) * atan( |y|/|x| ) if x >= 0 |
| 49 | // = sign(y) * (pi - atan( |y|/|x| )) if x < 0 |
| 50 | |
| 51 | // * Range reduction 2: reciprocal |
| 52 | // Now that the argument inside atan is positive, we can use the formula: |
| 53 | // atan(1/x) = pi/2 - atan(x) |
| 54 | // to make the argument inside atan <= 1 as follow: |
| 55 | // atan2(y, x) = sign(y) * atan( |y|/|x|) if 0 <= |y| <= x |
| 56 | // = sign(y) * (pi/2 - atan( |x|/|y| ) if 0 <= x < |y| |
| 57 | // = sign(y) * (pi - atan( |y|/|x| )) if 0 <= |y| <= -x |
| 58 | // = sign(y) * (pi/2 + atan( |x|/|y| )) if 0 <= -x < |y| |
| 59 | |
| 60 | // * Range reduction 3: look up table. |
| 61 | // After the previous two range reduction steps, we reduce the problem to |
| 62 | // compute atan(u) with 0 <= u <= 1, or to be precise: |
| 63 | // atan( n / d ) where n = min(|x|, |y|) and d = max(|x|, |y|). |
| 64 | // An accurate polynomial approximation for the whole [0, 1] input range will |
| 65 | // require a very large degree. To make it more efficient, we reduce the input |
| 66 | // range further by finding an integer idx such that: |
| 67 | // | n/d - idx/64 | <= 1/128. |
| 68 | // In particular, |
| 69 | // idx := round(2^6 * n/d) |
| 70 | // Then for the fast pass, we find a polynomial approximation for: |
| 71 | // atan( n/d ) ~ atan( idx/64 ) + (n/d - idx/64) * Q(n/d - idx/64) |
| 72 | // For the accurate pass, we use the addition formula: |
| 73 | // atan( n/d ) - atan( idx/64 ) = atan( (n/d - idx/64)/(1 + (n*idx)/(64*d)) ) |
| 74 | // = atan( (n - d*(idx/64))/(d + n*(idx/64)) ) |
| 75 | // And for the fast pass, we use degree-13 minimax polynomial to compute the |
| 76 | // RHS: |
| 77 | // atan(u) ~ P(u) = u - c_3 * u^3 + c_5 * u^5 - c_7 * u^7 + c_9 *u^9 - |
| 78 | // - c_11 * u^11 + c_13 * u^13 |
| 79 | // with absolute errors bounded by: |
| 80 | // |atan(u) - P(u)| < 2^-121 |
| 81 | // and relative errors bounded by: |
| 82 | // |(atan(u) - P(u)) / P(u)| < 2^-114. |
| 83 | |
| 84 | LIBC_INLINE float128 atan2f128(float128 y, float128 x) { |
| 85 | using DFloat128 = fputil::DyadicFloat<128>; |
| 86 | |
| 87 | constexpr DFloat128 ZERO = {Sign::POS, 0, 0_u128}; |
| 88 | constexpr DFloat128 MZERO = {Sign::NEG, 0, 0_u128}; |
| 89 | constexpr DFloat128 PI = {Sign::POS, -126, |
| 90 | 0xc90fdaa2'2168c234'c4c6628b'80dc1cd1_u128}; |
| 91 | constexpr DFloat128 MPI = {Sign::NEG, -126, |
| 92 | 0xc90fdaa2'2168c234'c4c6628b'80dc1cd1_u128}; |
| 93 | constexpr DFloat128 PI_OVER_2 = {Sign::POS, -127, |
| 94 | 0xc90fdaa2'2168c234'c4c6628b'80dc1cd1_u128}; |
| 95 | constexpr DFloat128 MPI_OVER_2 = {Sign::NEG, -127, |
| 96 | 0xc90fdaa2'2168c234'c4c6628b'80dc1cd1_u128}; |
| 97 | constexpr DFloat128 PI_OVER_4 = {Sign::POS, -128, |
| 98 | 0xc90fdaa2'2168c234'c4c6628b'80dc1cd1_u128}; |
| 99 | constexpr DFloat128 THREE_PI_OVER_4 = { |
| 100 | Sign::POS, -128, 0x96cbe3f9'990e91a7'9394c9e8'a0a5159d_u128}; |
| 101 | |
| 102 | // Adjustment for constant term: |
| 103 | // CONST_ADJ[x_sign][y_sign][recip] |
| 104 | constexpr DFloat128 CONST_ADJ[2][2][2] = { |
| 105 | {{ZERO, MPI_OVER_2}, {MZERO, MPI_OVER_2}}, |
| 106 | {{MPI, PI_OVER_2}, {MPI, PI_OVER_2}}}; |
| 107 | |
| 108 | using namespace atan_internal; |
| 109 | using FPBits = fputil::FPBits<float128>; |
| 110 | using DFloat128 = fputil::DyadicFloat<128>; |
| 111 | |
| 112 | FPBits x_bits(x), y_bits(y); |
| 113 | bool x_sign = x_bits.sign().is_neg(); |
| 114 | bool y_sign = y_bits.sign().is_neg(); |
| 115 | x_bits = x_bits.abs(); |
| 116 | y_bits = y_bits.abs(); |
| 117 | UInt128 x_abs = x_bits.uintval(); |
| 118 | UInt128 y_abs = y_bits.uintval(); |
| 119 | bool recip = x_abs < y_abs; |
| 120 | UInt128 min_abs = recip ? x_abs : y_abs; |
| 121 | UInt128 max_abs = !recip ? x_abs : y_abs; |
| 122 | unsigned min_exp = static_cast<unsigned>(min_abs >> FPBits::FRACTION_LEN); |
| 123 | unsigned max_exp = static_cast<unsigned>(max_abs >> FPBits::FRACTION_LEN); |
| 124 | |
| 125 | DFloat128 num(FPBits(min_abs).get_val()); |
| 126 | DFloat128 den(FPBits(max_abs).get_val()); |
| 127 | |
| 128 | // Check for exceptional cases, whether inputs are 0, inf, nan, or close to |
| 129 | // overflow, or close to underflow. |
| 130 | if (LIBC_UNLIKELY(max_exp >= 0x7fffU || min_exp == 0U)) { |
| 131 | if (x_bits.is_nan() || y_bits.is_nan()) |
| 132 | return FPBits::quiet_nan().get_val(); |
| 133 | unsigned x_except = x == 0 ? 0 : (FPBits(x_abs).is_inf() ? 2 : 1); |
| 134 | unsigned y_except = y == 0 ? 0 : (FPBits(y_abs).is_inf() ? 2 : 1); |
| 135 | |
| 136 | // Exceptional cases: |
| 137 | // EXCEPT[y_except][x_except][x_is_neg] |
| 138 | // with x_except & y_except: |
| 139 | // 0: zero |
| 140 | // 1: finite, non-zero |
| 141 | // 2: infinity |
| 142 | constexpr DFloat128 EXCEPTS[3][3][2] = { |
| 143 | {{ZERO, PI}, {ZERO, PI}, {ZERO, PI}}, |
| 144 | {{PI_OVER_2, PI_OVER_2}, {ZERO, ZERO}, {ZERO, PI}}, |
| 145 | {{PI_OVER_2, PI_OVER_2}, |
| 146 | {PI_OVER_2, PI_OVER_2}, |
| 147 | {PI_OVER_4, THREE_PI_OVER_4}}, |
| 148 | }; |
| 149 | |
| 150 | if ((x_except != 1) || (y_except != 1)) { |
| 151 | DFloat128 r = EXCEPTS[y_except][x_except][x_sign]; |
| 152 | if (y_sign) |
| 153 | r.sign = r.sign.negate(); |
| 154 | return static_cast<float128>(r); |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | bool final_sign = ((x_sign != y_sign) != recip); |
| 159 | DFloat128 const_term = CONST_ADJ[x_sign][y_sign][recip]; |
| 160 | int exp_diff = den.exponent - num.exponent; |
| 161 | // We have the following bound for normalized n and d: |
| 162 | // 2^(-exp_diff - 1) < n/d < 2^(-exp_diff + 1). |
| 163 | if (LIBC_UNLIKELY(exp_diff > FPBits::FRACTION_LEN + 2)) { |
| 164 | DFloat128 quotient = rounded_div(af: num, bf: den); |
| 165 | DFloat128 result = quick_add(a: const_term, b: quotient); |
| 166 | if (final_sign) |
| 167 | result.sign = result.sign.negate(); |
| 168 | return static_cast<float128>(result); |
| 169 | } |
| 170 | |
| 171 | // Take 24 leading bits of num and den to convert to float for fast division. |
| 172 | // We also multiply the numerator by 64 using integer addition directly to the |
| 173 | // exponent field. |
| 174 | float num_f = |
| 175 | cpp::bit_cast<float>(from: static_cast<uint32_t>(num.mantissa >> 104) + |
| 176 | (6U << fputil::FPBits<float>::FRACTION_LEN)); |
| 177 | float den_f = cpp::bit_cast<float>( |
| 178 | from: static_cast<uint32_t>(den.mantissa >> 104) + |
| 179 | (static_cast<uint32_t>(exp_diff) << fputil::FPBits<float>::FRACTION_LEN)); |
| 180 | |
| 181 | float k = fputil::nearest_integer(x: num_f / den_f); |
| 182 | unsigned idx = static_cast<unsigned>(k); |
| 183 | |
| 184 | // k_f128 = idx / 64 |
| 185 | DFloat128 k_f128(Sign::POS, -6, DFloat128::MantissaType(idx)); |
| 186 | |
| 187 | // Range reduction: |
| 188 | // atan(n/d) - atan(k) = atan((n/d - k/64) / (1 + (n/d) * (k/64))) |
| 189 | // = atan((n - d * k/64)) / (d + n * k/64)) |
| 190 | // num_f128 = n - d * k/64 |
| 191 | DFloat128 num_f128 = fputil::multiply_add(a: den, b: -k_f128, c: num); |
| 192 | // den_f128 = d + n * k/64 |
| 193 | DFloat128 den_f128 = fputil::multiply_add(a: num, b: k_f128, c: den); |
| 194 | |
| 195 | // q = (n - d * k) / (d + n * k) |
| 196 | DFloat128 q = |
| 197 | fputil::quick_mul(a: num_f128, b: fputil::approx_reciprocal(a: den_f128)); |
| 198 | // p ~ atan(q) |
| 199 | DFloat128 p = atan_eval(x: q); |
| 200 | |
| 201 | DFloat128 r = |
| 202 | fputil::quick_add(a: const_term, b: fputil::quick_add(a: ATAN_I_F128[idx], b: p)); |
| 203 | if (final_sign) |
| 204 | r.sign = r.sign.negate(); |
| 205 | |
| 206 | return static_cast<float128>(r); |
| 207 | } |
| 208 | |
| 209 | } // namespace math |
| 210 | |
| 211 | } // namespace LIBC_NAMESPACE_DECL |
| 212 | |
| 213 | #endif // LIBC_TYPES_HAS_FLOAT128 |
| 214 | |
| 215 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ATAN2F128_H |
| 216 | |