1 | //===-- Double-precision atan2 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/atan2.h" |
10 | #include "atan_utils.h" |
11 | #include "src/__support/FPUtil/FEnvImpl.h" |
12 | #include "src/__support/FPUtil/FPBits.h" |
13 | #include "src/__support/FPUtil/double_double.h" |
14 | #include "src/__support/FPUtil/multiply_add.h" |
15 | #include "src/__support/FPUtil/nearest_integer.h" |
16 | #include "src/__support/macros/config.h" |
17 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
18 | |
19 | namespace LIBC_NAMESPACE_DECL { |
20 | |
21 | // There are several range reduction steps we can take for atan2(y, x) as |
22 | // follow: |
23 | |
24 | // * Range reduction 1: signness |
25 | // atan2(y, x) will return a number between -PI and PI representing the angle |
26 | // forming by the 0x axis and the vector (x, y) on the 0xy-plane. |
27 | // In particular, we have that: |
28 | // atan2(y, x) = atan( y/x ) if x >= 0 and y >= 0 (I-quadrant) |
29 | // = pi + atan( y/x ) if x < 0 and y >= 0 (II-quadrant) |
30 | // = -pi + atan( y/x ) if x < 0 and y < 0 (III-quadrant) |
31 | // = atan( y/x ) if x >= 0 and y < 0 (IV-quadrant) |
32 | // Since atan function is odd, we can use the formula: |
33 | // atan(-u) = -atan(u) |
34 | // to adjust the above conditions a bit further: |
35 | // atan2(y, x) = atan( |y|/|x| ) if x >= 0 and y >= 0 (I-quadrant) |
36 | // = pi - atan( |y|/|x| ) if x < 0 and y >= 0 (II-quadrant) |
37 | // = -pi + atan( |y|/|x| ) if x < 0 and y < 0 (III-quadrant) |
38 | // = -atan( |y|/|x| ) if x >= 0 and y < 0 (IV-quadrant) |
39 | // Which can be simplified to: |
40 | // atan2(y, x) = sign(y) * atan( |y|/|x| ) if x >= 0 |
41 | // = sign(y) * (pi - atan( |y|/|x| )) if x < 0 |
42 | |
43 | // * Range reduction 2: reciprocal |
44 | // Now that the argument inside atan is positive, we can use the formula: |
45 | // atan(1/x) = pi/2 - atan(x) |
46 | // to make the argument inside atan <= 1 as follow: |
47 | // atan2(y, x) = sign(y) * atan( |y|/|x|) if 0 <= |y| <= x |
48 | // = sign(y) * (pi/2 - atan( |x|/|y| ) if 0 <= x < |y| |
49 | // = sign(y) * (pi - atan( |y|/|x| )) if 0 <= |y| <= -x |
50 | // = sign(y) * (pi/2 + atan( |x|/|y| )) if 0 <= -x < |y| |
51 | |
52 | // * Range reduction 3: look up table. |
53 | // After the previous two range reduction steps, we reduce the problem to |
54 | // compute atan(u) with 0 <= u <= 1, or to be precise: |
55 | // atan( n / d ) where n = min(|x|, |y|) and d = max(|x|, |y|). |
56 | // An accurate polynomial approximation for the whole [0, 1] input range will |
57 | // require a very large degree. To make it more efficient, we reduce the input |
58 | // range further by finding an integer idx such that: |
59 | // | n/d - idx/64 | <= 1/128. |
60 | // In particular, |
61 | // idx := round(2^6 * n/d) |
62 | // Then for the fast pass, we find a polynomial approximation for: |
63 | // atan( n/d ) ~ atan( idx/64 ) + (n/d - idx/64) * Q(n/d - idx/64) |
64 | // For the accurate pass, we use the addition formula: |
65 | // atan( n/d ) - atan( idx/64 ) = atan( (n/d - idx/64)/(1 + (n*idx)/(64*d)) ) |
66 | // = atan( (n - d*(idx/64))/(d + n*(idx/64)) ) |
67 | // And for the fast pass, we use degree-9 Taylor polynomial to compute the RHS: |
68 | // atan(u) ~ P(u) = u - u^3/3 + u^5/5 - u^7/7 + u^9/9 |
69 | // with absolute errors bounded by: |
70 | // |atan(u) - P(u)| < |u|^11 / 11 < 2^-80 |
71 | // and relative errors bounded by: |
72 | // |(atan(u) - P(u)) / P(u)| < u^10 / 11 < 2^-73. |
73 | |
74 | LLVM_LIBC_FUNCTION(double, atan2, (double y, double x)) { |
75 | using FPBits = fputil::FPBits<double>; |
76 | |
77 | constexpr double IS_NEG[2] = {1.0, -1.0}; |
78 | constexpr DoubleDouble ZERO = {0.0, 0.0}; |
79 | constexpr DoubleDouble MZERO = {-0.0, -0.0}; |
80 | constexpr DoubleDouble PI = {0x1.1a62633145c07p-53, 0x1.921fb54442d18p+1}; |
81 | constexpr DoubleDouble MPI = {-0x1.1a62633145c07p-53, -0x1.921fb54442d18p+1}; |
82 | constexpr DoubleDouble PI_OVER_2 = {0x1.1a62633145c07p-54, |
83 | 0x1.921fb54442d18p0}; |
84 | constexpr DoubleDouble MPI_OVER_2 = {-0x1.1a62633145c07p-54, |
85 | -0x1.921fb54442d18p0}; |
86 | constexpr DoubleDouble PI_OVER_4 = {0x1.1a62633145c07p-55, |
87 | 0x1.921fb54442d18p-1}; |
88 | constexpr DoubleDouble THREE_PI_OVER_4 = {0x1.a79394c9e8a0ap-54, |
89 | 0x1.2d97c7f3321d2p+1}; |
90 | // Adjustment for constant term: |
91 | // CONST_ADJ[x_sign][y_sign][recip] |
92 | constexpr DoubleDouble CONST_ADJ[2][2][2] = { |
93 | {{ZERO, MPI_OVER_2}, {MZERO, MPI_OVER_2}}, |
94 | {{MPI, PI_OVER_2}, {MPI, PI_OVER_2}}}; |
95 | |
96 | FPBits x_bits(x), y_bits(y); |
97 | bool x_sign = x_bits.sign().is_neg(); |
98 | bool y_sign = y_bits.sign().is_neg(); |
99 | x_bits = x_bits.abs(); |
100 | y_bits = y_bits.abs(); |
101 | uint64_t x_abs = x_bits.uintval(); |
102 | uint64_t y_abs = y_bits.uintval(); |
103 | bool recip = x_abs < y_abs; |
104 | uint64_t min_abs = recip ? x_abs : y_abs; |
105 | uint64_t max_abs = !recip ? x_abs : y_abs; |
106 | unsigned min_exp = static_cast<unsigned>(min_abs >> FPBits::FRACTION_LEN); |
107 | unsigned max_exp = static_cast<unsigned>(max_abs >> FPBits::FRACTION_LEN); |
108 | |
109 | double num = FPBits(min_abs).get_val(); |
110 | double den = FPBits(max_abs).get_val(); |
111 | |
112 | // Check for exceptional cases, whether inputs are 0, inf, nan, or close to |
113 | // overflow, or close to underflow. |
114 | if (LIBC_UNLIKELY(max_exp > 0x7ffU - 128U || min_exp < 128U)) { |
115 | if (x_bits.is_nan() || y_bits.is_nan()) { |
116 | if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) |
117 | fputil::raise_except_if_required(FE_INVALID); |
118 | return FPBits::quiet_nan().get_val(); |
119 | } |
120 | unsigned x_except = x == 0.0 ? 0 : (FPBits(x_abs).is_inf() ? 2 : 1); |
121 | unsigned y_except = y == 0.0 ? 0 : (FPBits(y_abs).is_inf() ? 2 : 1); |
122 | |
123 | // Exceptional cases: |
124 | // EXCEPT[y_except][x_except][x_is_neg] |
125 | // with x_except & y_except: |
126 | // 0: zero |
127 | // 1: finite, non-zero |
128 | // 2: infinity |
129 | constexpr DoubleDouble EXCEPTS[3][3][2] = { |
130 | {{ZERO, PI}, {ZERO, PI}, {ZERO, PI}}, |
131 | {{PI_OVER_2, PI_OVER_2}, {ZERO, ZERO}, {ZERO, PI}}, |
132 | {{PI_OVER_2, PI_OVER_2}, |
133 | {PI_OVER_2, PI_OVER_2}, |
134 | {PI_OVER_4, THREE_PI_OVER_4}}, |
135 | }; |
136 | |
137 | if ((x_except != 1) || (y_except != 1)) { |
138 | DoubleDouble r = EXCEPTS[y_except][x_except][x_sign]; |
139 | return fputil::multiply_add(IS_NEG[y_sign], r.hi, IS_NEG[y_sign] * r.lo); |
140 | } |
141 | bool scale_up = min_exp < 128U; |
142 | bool scale_down = max_exp > 0x7ffU - 128U; |
143 | // At least one input is denormal, multiply both numerator and denominator |
144 | // by some large enough power of 2 to normalize denormal inputs. |
145 | if (scale_up) { |
146 | num *= 0x1.0p64; |
147 | if (!scale_down) |
148 | den *= 0x1.0p64; |
149 | } else if (scale_down) { |
150 | den *= 0x1.0p-64; |
151 | if (!scale_up) |
152 | num *= 0x1.0p-64; |
153 | } |
154 | |
155 | min_abs = FPBits(num).uintval(); |
156 | max_abs = FPBits(den).uintval(); |
157 | min_exp = static_cast<unsigned>(min_abs >> FPBits::FRACTION_LEN); |
158 | max_exp = static_cast<unsigned>(max_abs >> FPBits::FRACTION_LEN); |
159 | } |
160 | |
161 | double final_sign = IS_NEG[(x_sign != y_sign) != recip]; |
162 | DoubleDouble const_term = CONST_ADJ[x_sign][y_sign][recip]; |
163 | unsigned exp_diff = max_exp - min_exp; |
164 | // We have the following bound for normalized n and d: |
165 | // 2^(-exp_diff - 1) < n/d < 2^(-exp_diff + 1). |
166 | if (LIBC_UNLIKELY(exp_diff > 54)) { |
167 | return fputil::multiply_add(final_sign, const_term.hi, |
168 | final_sign * (const_term.lo + num / den)); |
169 | } |
170 | |
171 | double k = fputil::nearest_integer(64.0 * num / den); |
172 | unsigned idx = static_cast<unsigned>(k); |
173 | // k = idx / 64 |
174 | k *= 0x1.0p-6; |
175 | |
176 | // Range reduction: |
177 | // atan(n/d) - atan(k/64) = atan((n/d - k/64) / (1 + (n/d) * (k/64))) |
178 | // = atan((n - d * k/64)) / (d + n * k/64)) |
179 | DoubleDouble num_k = fputil::exact_mult(num, k); |
180 | DoubleDouble den_k = fputil::exact_mult(den, k); |
181 | |
182 | // num_dd = n - d * k |
183 | DoubleDouble num_dd = fputil::exact_add(num - den_k.hi, -den_k.lo); |
184 | // den_dd = d + n * k |
185 | DoubleDouble den_dd = fputil::exact_add(den, num_k.hi); |
186 | den_dd.lo += num_k.lo; |
187 | |
188 | // q = (n - d * k) / (d + n * k) |
189 | DoubleDouble q = fputil::div(num_dd, den_dd); |
190 | // p ~ atan(q) |
191 | DoubleDouble p = atan_eval(q); |
192 | |
193 | DoubleDouble r = fputil::add(const_term, fputil::add(ATAN_I[idx], p)); |
194 | r.hi *= final_sign; |
195 | r.lo *= final_sign; |
196 | |
197 | return r.hi + r.lo; |
198 | } |
199 | |
200 | } // namespace LIBC_NAMESPACE_DECL |
201 | |