1 | //===-- Single-precision atan2f 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/atan2f.h" |
10 | #include "inv_trigf_utils.h" |
11 | #include "src/__support/FPUtil/FPBits.h" |
12 | #include "src/__support/FPUtil/PolyEval.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/FPUtil/rounding_mode.h" |
17 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
18 | |
19 | namespace LIBC_NAMESPACE { |
20 | |
21 | namespace { |
22 | |
23 | // Look up tables for accurate pass: |
24 | |
25 | // atan(i/16) with i = 0..16, generated by Sollya with: |
26 | // > for i from 0 to 16 do { |
27 | // a = round(atan(i/16), D, RN); |
28 | // b = round(atan(i/16) - a, D, RN); |
29 | // print("{", b, ",", a, "},"); |
30 | // }; |
31 | constexpr fputil::DoubleDouble ATAN_I[17] = { |
32 | {.lo: 0.0, .hi: 0.0}, |
33 | {.lo: -0x1.c934d86d23f1dp-60, .hi: 0x1.ff55bb72cfdeap-5}, |
34 | {.lo: -0x1.cd37686760c17p-59, .hi: 0x1.fd5ba9aac2f6ep-4}, |
35 | {.lo: 0x1.347b0b4f881cap-58, .hi: 0x1.7b97b4bce5b02p-3}, |
36 | {.lo: 0x1.8ab6e3cf7afbdp-57, .hi: 0x1.f5b75f92c80ddp-3}, |
37 | {.lo: -0x1.963a544b672d8p-57, .hi: 0x1.362773707ebccp-2}, |
38 | {.lo: -0x1.c63aae6f6e918p-56, .hi: 0x1.6f61941e4def1p-2}, |
39 | {.lo: -0x1.24dec1b50b7ffp-56, .hi: 0x1.a64eec3cc23fdp-2}, |
40 | {.lo: 0x1.a2b7f222f65e2p-56, .hi: 0x1.dac670561bb4fp-2}, |
41 | {.lo: -0x1.d5b495f6349e6p-56, .hi: 0x1.0657e94db30dp-1}, |
42 | {.lo: -0x1.928df287a668fp-58, .hi: 0x1.1e00babdefeb4p-1}, |
43 | {.lo: 0x1.1021137c71102p-55, .hi: 0x1.345f01cce37bbp-1}, |
44 | {.lo: 0x1.2419a87f2a458p-56, .hi: 0x1.4978fa3269ee1p-1}, |
45 | {.lo: 0x1.0028e4bc5e7cap-57, .hi: 0x1.5d58987169b18p-1}, |
46 | {.lo: -0x1.8c34d25aadef6p-56, .hi: 0x1.700a7c5784634p-1}, |
47 | {.lo: -0x1.bf76229d3b917p-56, .hi: 0x1.819d0b7158a4dp-1}, |
48 | {.lo: 0x1.1a62633145c07p-55, .hi: 0x1.921fb54442d18p-1}, |
49 | }; |
50 | |
51 | // Taylor polynomial, generated by Sollya with: |
52 | // > for i from 0 to 8 do { |
53 | // j = (-1)^(i + 1)/(2*i + 1); |
54 | // a = round(j, D, RN); |
55 | // b = round(j - a, D, RN); |
56 | // print("{", b, ",", a, "},"); |
57 | // }; |
58 | constexpr fputil::DoubleDouble COEFFS[9] = { |
59 | {.lo: 0.0, .hi: 1.0}, // 1 |
60 | {.lo: -0x1.5555555555555p-56, .hi: -0x1.5555555555555p-2}, // -1/3 |
61 | {.lo: -0x1.999999999999ap-57, .hi: 0x1.999999999999ap-3}, // 1/5 |
62 | {.lo: -0x1.2492492492492p-57, .hi: -0x1.2492492492492p-3}, // -1/7 |
63 | {.lo: 0x1.c71c71c71c71cp-58, .hi: 0x1.c71c71c71c71cp-4}, // 1/9 |
64 | {.lo: 0x1.745d1745d1746p-59, .hi: -0x1.745d1745d1746p-4}, // -1/11 |
65 | {.lo: -0x1.3b13b13b13b14p-58, .hi: 0x1.3b13b13b13b14p-4}, // 1/13 |
66 | {.lo: -0x1.1111111111111p-60, .hi: -0x1.1111111111111p-4}, // -1/15 |
67 | {.lo: 0x1.e1e1e1e1e1e1ep-61, .hi: 0x1.e1e1e1e1e1e1ep-5}, // 1/17 |
68 | }; |
69 | |
70 | // Veltkamp's splitting of a double precision into hi + lo, where the hi part is |
71 | // slightly smaller than an even split, so that the product of |
72 | // hi * (s1 * k + s2) is exact, |
73 | // where: |
74 | // s1, s2 are single precsion, |
75 | // 1/16 <= s1/s2 <= 1 |
76 | // 1/16 <= k <= 1 is an integer. |
77 | // So the maximal precision of (s1 * k + s2) is: |
78 | // prec(s1 * k + s2) = 2 + log2(msb(s2)) - log2(lsb(k_d * s1)) |
79 | // = 2 + log2(msb(s1)) + 4 - log2(lsb(k_d)) - log2(lsb(s1)) |
80 | // = 2 + log2(lsb(s1)) + 23 + 4 - (-4) - log2(lsb(s1)) |
81 | // = 33. |
82 | // Thus, the Veltkamp splitting constant is C = 2^33 + 1. |
83 | // This is used when FMA instruction is not available. |
84 | [[maybe_unused]] constexpr fputil::DoubleDouble split_d(double a) { |
85 | fputil::DoubleDouble r{.lo: 0.0, .hi: 0.0}; |
86 | constexpr double C = 0x1.0p33 + 1.0; |
87 | double t1 = C * a; |
88 | double t2 = a - t1; |
89 | r.hi = t1 + t2; |
90 | r.lo = a - r.hi; |
91 | return r; |
92 | } |
93 | |
94 | // Compute atan( num_d / den_d ) in double-double precision. |
95 | // num_d = min(|x|, |y|) |
96 | // den_d = max(|x|, |y|) |
97 | // q_d = num_d / den_d |
98 | // idx, k_d = round( 2^4 * num_d / den_d ) |
99 | // final_sign = sign of the final result |
100 | // const_term = the constant term in the final expression. |
101 | float atan2f_double_double(double num_d, double den_d, double q_d, int idx, |
102 | double k_d, double final_sign, |
103 | const fputil::DoubleDouble &const_term) { |
104 | fputil::DoubleDouble q; |
105 | double num_r, den_r; |
106 | |
107 | if (idx != 0) { |
108 | // The following range reduction is accurate even without fma for |
109 | // 1/16 <= n/d <= 1. |
110 | // atan(n/d) - atan(idx/16) = atan((n/d - idx/16) / (1 + (n/d) * (idx/16))) |
111 | // = atan((n - d*(idx/16)) / (d + n*idx/16)) |
112 | k_d *= 0x1.0p-4; |
113 | num_r = fputil::multiply_add(x: k_d, y: -den_d, z: num_d); // Exact |
114 | den_r = fputil::multiply_add(x: k_d, y: num_d, z: den_d); // Exact |
115 | q.hi = num_r / den_r; |
116 | } else { |
117 | // For 0 < n/d < 1/16, we just need to calculate the lower part of their |
118 | // quotient. |
119 | q.hi = q_d; |
120 | num_r = num_d; |
121 | den_r = den_d; |
122 | } |
123 | #ifdef LIBC_TARGET_CPU_HAS_FMA |
124 | q.lo = fputil::multiply_add(x: q.hi, y: -den_r, z: num_r) / den_r; |
125 | #else |
126 | // Compute `(num_r - q.hi * den_r) / den_r` accurately without FMA |
127 | // instructions. |
128 | fputil::DoubleDouble q_hi_dd = split_d(q.hi); |
129 | double t1 = fputil::multiply_add(q_hi_dd.hi, -den_r, num_r); // Exact |
130 | double t2 = fputil::multiply_add(q_hi_dd.lo, -den_r, t1); |
131 | q.lo = t2 / den_r; |
132 | #endif // LIBC_TARGET_CPU_HAS_FMA |
133 | |
134 | // Taylor polynomial, evaluating using Horner's scheme: |
135 | // P = x - x^3/3 + x^5/5 -x^7/7 + x^9/9 - x^11/11 + x^13/13 - x^15/15 |
136 | // + x^17/17 |
137 | // = x*(1 + x^2*(-1/3 + x^2*(1/5 + x^2*(-1/7 + x^2*(1/9 + x^2* |
138 | // *(-1/11 + x^2*(1/13 + x^2*(-1/15 + x^2 * 1/17)))))))) |
139 | fputil::DoubleDouble q2 = fputil::quick_mult(a: q, b: q); |
140 | fputil::DoubleDouble p_dd = |
141 | fputil::polyeval(x: q2, a0: COEFFS[0], a: COEFFS[1], a: COEFFS[2], a: COEFFS[3], |
142 | a: COEFFS[4], a: COEFFS[5], a: COEFFS[6], a: COEFFS[7], a: COEFFS[8]); |
143 | fputil::DoubleDouble r_dd = |
144 | fputil::add(a: const_term, b: fputil::multiply_add(a: q, b: p_dd, c: ATAN_I[idx])); |
145 | r_dd.hi *= final_sign; |
146 | r_dd.lo *= final_sign; |
147 | |
148 | // Make sure the sum is normalized: |
149 | fputil::DoubleDouble rr = fputil::exact_add(a: r_dd.hi, b: r_dd.lo); |
150 | // Round to odd. |
151 | uint64_t rr_bits = cpp::bit_cast<uint64_t>(from: rr.hi); |
152 | if (LIBC_UNLIKELY(((rr_bits & 0xfff'ffff) == 0) && (rr.lo != 0.0))) { |
153 | Sign hi_sign = fputil::FPBits<double>(rr.hi).sign(); |
154 | Sign lo_sign = fputil::FPBits<double>(rr.lo).sign(); |
155 | if (hi_sign == lo_sign) { |
156 | ++rr_bits; |
157 | } else if ((rr_bits & fputil::FPBits<double>::FRACTION_MASK) > 0) { |
158 | --rr_bits; |
159 | } |
160 | } |
161 | |
162 | return static_cast<float>(cpp::bit_cast<double>(from: rr_bits)); |
163 | } |
164 | |
165 | } // anonymous namespace |
166 | |
167 | // There are several range reduction steps we can take for atan2(y, x) as |
168 | // follow: |
169 | |
170 | // * Range reduction 1: signness |
171 | // atan2(y, x) will return a number between -PI and PI representing the angle |
172 | // forming by the 0x axis and the vector (x, y) on the 0xy-plane. |
173 | // In particular, we have that: |
174 | // atan2(y, x) = atan( y/x ) if x >= 0 and y >= 0 (I-quadrant) |
175 | // = pi + atan( y/x ) if x < 0 and y >= 0 (II-quadrant) |
176 | // = -pi + atan( y/x ) if x < 0 and y < 0 (III-quadrant) |
177 | // = atan( y/x ) if x >= 0 and y < 0 (IV-quadrant) |
178 | // Since atan function is odd, we can use the formula: |
179 | // atan(-u) = -atan(u) |
180 | // to adjust the above conditions a bit further: |
181 | // atan2(y, x) = atan( |y|/|x| ) if x >= 0 and y >= 0 (I-quadrant) |
182 | // = pi - atan( |y|/|x| ) if x < 0 and y >= 0 (II-quadrant) |
183 | // = -pi + atan( |y|/|x| ) if x < 0 and y < 0 (III-quadrant) |
184 | // = -atan( |y|/|x| ) if x >= 0 and y < 0 (IV-quadrant) |
185 | // Which can be simplified to: |
186 | // atan2(y, x) = sign(y) * atan( |y|/|x| ) if x >= 0 |
187 | // = sign(y) * (pi - atan( |y|/|x| )) if x < 0 |
188 | |
189 | // * Range reduction 2: reciprocal |
190 | // Now that the argument inside atan is positive, we can use the formula: |
191 | // atan(1/x) = pi/2 - atan(x) |
192 | // to make the argument inside atan <= 1 as follow: |
193 | // atan2(y, x) = sign(y) * atan( |y|/|x|) if 0 <= |y| <= x |
194 | // = sign(y) * (pi/2 - atan( |x|/|y| ) if 0 <= x < |y| |
195 | // = sign(y) * (pi - atan( |y|/|x| )) if 0 <= |y| <= -x |
196 | // = sign(y) * (pi/2 + atan( |x|/|y| )) if 0 <= -x < |y| |
197 | |
198 | // * Range reduction 3: look up table. |
199 | // After the previous two range reduction steps, we reduce the problem to |
200 | // compute atan(u) with 0 <= u <= 1, or to be precise: |
201 | // atan( n / d ) where n = min(|x|, |y|) and d = max(|x|, |y|). |
202 | // An accurate polynomial approximation for the whole [0, 1] input range will |
203 | // require a very large degree. To make it more efficient, we reduce the input |
204 | // range further by finding an integer idx such that: |
205 | // | n/d - idx/16 | <= 1/32. |
206 | // In particular, |
207 | // idx := 2^-4 * round(2^4 * n/d) |
208 | // Then for the fast pass, we find a polynomial approximation for: |
209 | // atan( n/d ) ~ atan( idx/16 ) + (n/d - idx/16) * Q(n/d - idx/16) |
210 | // For the accurate pass, we use the addition formula: |
211 | // atan( n/d ) - atan( idx/16 ) = atan( (n/d - idx/16)/(1 + (n*idx)/(16*d)) ) |
212 | // = atan( (n - d * idx/16)/(d + n * idx/16) ) |
213 | // And finally we use Taylor polynomial to compute the RHS in the accurate pass: |
214 | // atan(u) ~ P(u) = u - u^3/3 + u^5/5 - u^7/7 + u^9/9 - u^11/11 + u^13/13 - |
215 | // - u^15/15 + u^17/17 |
216 | // It's error in double-double precision is estimated in Sollya to be: |
217 | // > P = x - x^3/3 + x^5/5 -x^7/7 + x^9/9 - x^11/11 + x^13/13 - x^15/15 |
218 | // + x^17/17; |
219 | // > dirtyinfnorm(atan(x) - P, [-2^-5, 2^-5]); |
220 | // 0x1.aec6f...p-100 |
221 | // which is about rounding errors of double-double (2^-104). |
222 | |
223 | LLVM_LIBC_FUNCTION(float, atan2f, (float y, float x)) { |
224 | using FPBits = typename fputil::FPBits<float>; |
225 | constexpr double IS_NEG[2] = {1.0, -1.0}; |
226 | constexpr double PI = 0x1.921fb54442d18p1; |
227 | constexpr double PI_LO = 0x1.1a62633145c07p-53; |
228 | constexpr double PI_OVER_4 = 0x1.921fb54442d18p-1; |
229 | constexpr double PI_OVER_2 = 0x1.921fb54442d18p0; |
230 | constexpr double THREE_PI_OVER_4 = 0x1.2d97c7f3321d2p+1; |
231 | // Adjustment for constant term: |
232 | // CONST_ADJ[x_sign][y_sign][recip] |
233 | constexpr fputil::DoubleDouble CONST_ADJ[2][2][2] = { |
234 | {{{.lo: 0.0, .hi: 0.0}, {.lo: -PI_LO / 2, .hi: -PI_OVER_2}}, |
235 | {{.lo: -0.0, .hi: -0.0}, {.lo: -PI_LO / 2, .hi: -PI_OVER_2}}}, |
236 | {{{.lo: -PI_LO, .hi: -PI}, {.lo: PI_LO / 2, .hi: PI_OVER_2}}, |
237 | {{.lo: -PI_LO, .hi: -PI}, {.lo: PI_LO / 2, .hi: PI_OVER_2}}}}; |
238 | |
239 | FPBits x_bits(x), y_bits(y); |
240 | bool x_sign = x_bits.sign().is_neg(); |
241 | bool y_sign = y_bits.sign().is_neg(); |
242 | x_bits.set_sign(Sign::POS); |
243 | y_bits.set_sign(Sign::POS); |
244 | uint32_t x_abs = x_bits.uintval(); |
245 | uint32_t y_abs = y_bits.uintval(); |
246 | uint32_t max_abs = x_abs > y_abs ? x_abs : y_abs; |
247 | uint32_t min_abs = x_abs <= y_abs ? x_abs : y_abs; |
248 | |
249 | if (LIBC_UNLIKELY(max_abs >= 0x7f80'0000U || min_abs == 0U)) { |
250 | if (x_bits.is_nan() || y_bits.is_nan()) |
251 | return FPBits::quiet_nan().get_val(); |
252 | size_t x_except = x_abs == 0 ? 0 : (x_abs == 0x7f80'0000 ? 2 : 1); |
253 | size_t y_except = y_abs == 0 ? 0 : (y_abs == 0x7f80'0000 ? 2 : 1); |
254 | |
255 | // Exceptional cases: |
256 | // EXCEPT[y_except][x_except][x_is_neg] |
257 | // with x_except & y_except: |
258 | // 0: zero |
259 | // 1: finite, non-zero |
260 | // 2: infinity |
261 | constexpr double EXCEPTS[3][3][2] = { |
262 | {{0.0, PI}, {0.0, PI}, {0.0, PI}}, |
263 | {{PI_OVER_2, PI_OVER_2}, {0.0, 0.0}, {0.0, PI}}, |
264 | {{PI_OVER_2, PI_OVER_2}, |
265 | {PI_OVER_2, PI_OVER_2}, |
266 | {PI_OVER_4, THREE_PI_OVER_4}}, |
267 | }; |
268 | |
269 | double r = IS_NEG[y_sign] * EXCEPTS[y_except][x_except][x_sign]; |
270 | |
271 | return static_cast<float>(r); |
272 | } |
273 | |
274 | bool recip = x_abs < y_abs; |
275 | double final_sign = IS_NEG[(x_sign != y_sign) != recip]; |
276 | fputil::DoubleDouble const_term = CONST_ADJ[x_sign][y_sign][recip]; |
277 | double num_d = static_cast<double>(FPBits(min_abs).get_val()); |
278 | double den_d = static_cast<double>(FPBits(max_abs).get_val()); |
279 | double q_d = num_d / den_d; |
280 | |
281 | double k_d = fputil::nearest_integer(x: q_d * 0x1.0p4f); |
282 | int idx = static_cast<int>(k_d); |
283 | q_d = fputil::multiply_add(x: k_d, y: -0x1.0p-4, z: q_d); |
284 | |
285 | double p = atan_eval(x: q_d, i: idx); |
286 | double r = final_sign * |
287 | fputil::multiply_add(x: q_d, y: p, z: const_term.hi + ATAN_COEFFS[idx][0]); |
288 | |
289 | constexpr uint32_t LOWER_ERR = 4; |
290 | // Mask sticky bits in double precision before rounding to single precision. |
291 | constexpr uint32_t MASK = |
292 | mask_trailing_ones<uint32_t, fputil::FPBits<double>::SIG_LEN - |
293 | FPBits::SIG_LEN - 1>(); |
294 | constexpr uint32_t UPPER_ERR = MASK - LOWER_ERR; |
295 | |
296 | uint32_t r_bits = static_cast<uint32_t>(cpp::bit_cast<uint64_t>(from: r)) & MASK; |
297 | |
298 | // Ziv's rounding test. |
299 | if (LIBC_LIKELY(r_bits > LOWER_ERR && r_bits < UPPER_ERR)) |
300 | return static_cast<float>(r); |
301 | |
302 | return atan2f_double_double(num_d, den_d, q_d, idx, k_d, final_sign, |
303 | const_term); |
304 | } |
305 | |
306 | } // namespace LIBC_NAMESPACE |
307 | |