xref: /aosp_15_r20/external/llvm-libc/src/math/generic/tan.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Double-precision tan 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/tan.h"
10 #include "hdr/errno_macros.h"
11 #include "src/__support/FPUtil/FEnvImpl.h"
12 #include "src/__support/FPUtil/FPBits.h"
13 #include "src/__support/FPUtil/PolyEval.h"
14 #include "src/__support/FPUtil/double_double.h"
15 #include "src/__support/FPUtil/dyadic_float.h"
16 #include "src/__support/FPUtil/except_value_utils.h"
17 #include "src/__support/FPUtil/multiply_add.h"
18 #include "src/__support/FPUtil/rounding_mode.h"
19 #include "src/__support/common.h"
20 #include "src/__support/macros/config.h"
21 #include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
22 #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
23 #include "src/math/generic/range_reduction_double_common.h"
24 
25 #ifdef LIBC_TARGET_CPU_HAS_FMA
26 #include "range_reduction_double_fma.h"
27 #else
28 #include "range_reduction_double_nofma.h"
29 #endif // LIBC_TARGET_CPU_HAS_FMA
30 
31 namespace LIBC_NAMESPACE_DECL {
32 
33 using DoubleDouble = fputil::DoubleDouble;
34 using Float128 = typename fputil::DyadicFloat<128>;
35 
36 namespace {
37 
tan_eval(const DoubleDouble & u,DoubleDouble & result)38 LIBC_INLINE double tan_eval(const DoubleDouble &u, DoubleDouble &result) {
39   // Evaluate tan(y) = tan(x - k * (pi/128))
40   // We use the degree-9 Taylor approximation:
41   //   tan(y) ~ P(y) = y + y^3/3 + 2*y^5/15 + 17*y^7/315 + 62*y^9/2835
42   // Then the error is bounded by:
43   //   |tan(y) - P(y)| < 2^-6 * |y|^11 < 2^-6 * 2^-66 = 2^-72.
44   // For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms
45   // < ulp(u_hi^3) gives us:
46   //   P(y) = y + y^3/3 + 2*y^5/15 + 17*y^7/315 + 62*y^9/2835 = ...
47   // ~ u_hi + u_hi^3 * (1/3 + u_hi^2 * (2/15 + u_hi^2 * (17/315 +
48   //                                                     + u_hi^2 * 62/2835))) +
49   //        + u_lo (1 + u_hi^2 * (1 + u_hi^2 * 2/3))
50   double u_hi_sq = u.hi * u.hi; // Error < ulp(u_hi^2) < 2^(-6 - 52) = 2^-58.
51   // p1 ~ 17/315 + u_hi^2 62 / 2835.
52   double p1 =
53       fputil::multiply_add(u_hi_sq, 0x1.664f4882c10fap-6, 0x1.ba1ba1ba1ba1cp-5);
54   // p2 ~ 1/3 + u_hi^2 2 / 15.
55   double p2 =
56       fputil::multiply_add(u_hi_sq, 0x1.1111111111111p-3, 0x1.5555555555555p-2);
57   // q1 ~ 1 + u_hi^2 * 2/3.
58   double q1 = fputil::multiply_add(u_hi_sq, 0x1.5555555555555p-1, 1.0);
59   double u_hi_3 = u_hi_sq * u.hi;
60   double u_hi_4 = u_hi_sq * u_hi_sq;
61   // p3 ~ 1/3 + u_hi^2 * (2/15 + u_hi^2 * (17/315 + u_hi^2 * 62/2835))
62   double p3 = fputil::multiply_add(u_hi_4, p1, p2);
63   // q2 ~ 1 + u_hi^2 * (1 + u_hi^2 * 2/3)
64   double q2 = fputil::multiply_add(u_hi_sq, q1, 1.0);
65   double tan_lo = fputil::multiply_add(u_hi_3, p3, u.lo * q2);
66   // Overall, |tan(y) - (u_hi + tan_lo)| < ulp(u_hi^3) <= 2^-71.
67   // And the relative errors is:
68   // |(tan(y) - (u_hi + tan_lo)) / tan(y) | <= 2*ulp(u_hi^2) < 2^-64
69   result = fputil::exact_add(u.hi, tan_lo);
70   return fputil::multiply_add(fputil::FPBits<double>(u_hi_3).abs().get_val(),
71                               0x1.0p-51, 0x1.0p-102);
72 }
73 
74 #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
75 // Accurate evaluation of tan for small u.
tan_eval(const Float128 & u)76 [[maybe_unused]] Float128 tan_eval(const Float128 &u) {
77   Float128 u_sq = fputil::quick_mul(u, u);
78 
79   // tan(x) ~ x + x^3/3 + x^5 * 2/15 + x^7 * 17/315 + x^9 * 62/2835 +
80   //          + x^11 * 1382/155925 + x^13 * 21844/6081075 +
81   //          + x^15 * 929569/638512875 + x^17 * 6404582/10854718875
82   // Relative errors < 2^-127 for |u| < pi/256.
83   constexpr Float128 TAN_COEFFS[] = {
84       {Sign::POS, -127, 0x80000000'00000000'00000000'00000000_u128}, // 1
85       {Sign::POS, -129, 0xaaaaaaaa'aaaaaaaa'aaaaaaaa'aaaaaaab_u128}, // 1
86       {Sign::POS, -130, 0x88888888'88888888'88888888'88888889_u128}, // 2/15
87       {Sign::POS, -132, 0xdd0dd0dd'0dd0dd0d'd0dd0dd0'dd0dd0dd_u128}, // 17/315
88       {Sign::POS, -133, 0xb327a441'6087cf99'6b5dd24e'ec0b327a_u128}, // 62/2835
89       {Sign::POS, -134,
90        0x91371aaf'3611e47a'da8e1cba'7d900eca_u128}, // 1382/155925
91       {Sign::POS, -136,
92        0xeb69e870'abeefdaf'e606d2e4'd1e65fbc_u128}, // 21844/6081075
93       {Sign::POS, -137,
94        0xbed1b229'5baf15b5'0ec9af45'a2619971_u128}, // 929569/638512875
95       {Sign::POS, -138,
96        0x9aac1240'1b3a2291'1b2ac7e3'e4627d0a_u128}, // 6404582/10854718875
97   };
98 
99   return fputil::quick_mul(
100       u, fputil::polyeval(u_sq, TAN_COEFFS[0], TAN_COEFFS[1], TAN_COEFFS[2],
101                           TAN_COEFFS[3], TAN_COEFFS[4], TAN_COEFFS[5],
102                           TAN_COEFFS[6], TAN_COEFFS[7], TAN_COEFFS[8]));
103 }
104 
105 // Calculation a / b = a * (1/b) for Float128.
106 // Using the initial approximation of q ~ (1/b), then apply 2 Newton-Raphson
107 // iterations, before multiplying by a.
newton_raphson_div(const Float128 & a,Float128 b,double q)108 [[maybe_unused]] Float128 newton_raphson_div(const Float128 &a, Float128 b,
109                                              double q) {
110   Float128 q0(q);
111   constexpr Float128 TWO(2.0);
112   b.sign = (b.sign == Sign::POS) ? Sign::NEG : Sign::POS;
113   Float128 q1 =
114       fputil::quick_mul(q0, fputil::quick_add(TWO, fputil::quick_mul(b, q0)));
115   Float128 q2 =
116       fputil::quick_mul(q1, fputil::quick_add(TWO, fputil::quick_mul(b, q1)));
117   return fputil::quick_mul(a, q2);
118 }
119 #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
120 
121 } // anonymous namespace
122 
123 LLVM_LIBC_FUNCTION(double, tan, (double x)) {
124   using FPBits = typename fputil::FPBits<double>;
125   FPBits xbits(x);
126 
127   uint16_t x_e = xbits.get_biased_exponent();
128 
129   DoubleDouble y;
130   unsigned k;
131   LargeRangeReduction range_reduction_large{};
132 
133   // |x| < 2^16
134   if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) {
135     // |x| < 2^-7
136     if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 7)) {
137       // |x| < 2^-27, |tan(x) - x| < ulp(x)/2.
138       if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {
139         // Signed zeros.
140         if (LIBC_UNLIKELY(x == 0.0))
141           return x + x; // Make sure it works with FTZ/DAZ.
142 
143 #ifdef LIBC_TARGET_CPU_HAS_FMA
144         return fputil::multiply_add(x, 0x1.0p-54, x);
145 #else
146         if (LIBC_UNLIKELY(x_e < 4)) {
147           int rounding_mode = fputil::quick_get_round();
148           if ((xbits.sign() == Sign::POS && rounding_mode == FE_UPWARD) ||
149               (xbits.sign() == Sign::NEG && rounding_mode == FE_DOWNWARD))
150             return FPBits(xbits.uintval() + 1).get_val();
151         }
152         return fputil::multiply_add(x, 0x1.0p-54, x);
153 #endif // LIBC_TARGET_CPU_HAS_FMA
154       }
155       // No range reduction needed.
156       k = 0;
157       y.lo = 0.0;
158       y.hi = x;
159     } else {
160       // Small range reduction.
161       k = range_reduction_small(x, y);
162     }
163   } else {
164     // Inf or NaN
165     if (LIBC_UNLIKELY(x_e > 2 * FPBits::EXP_BIAS)) {
166       // tan(+-Inf) = NaN
167       if (xbits.get_mantissa() == 0) {
168         fputil::set_errno_if_required(EDOM);
169         fputil::raise_except_if_required(FE_INVALID);
170       }
171       return x + FPBits::quiet_nan().get_val();
172     }
173 
174     // Large range reduction.
175     k = range_reduction_large.fast(x, y);
176   }
177 
178   DoubleDouble tan_y;
179   [[maybe_unused]] double err = tan_eval(y, tan_y);
180 
181   // Look up sin(k * pi/128) and cos(k * pi/128)
182 #ifdef LIBC_MATH_HAS_SMALL_TABLES
183   // Memory saving versions. Use 65-entry table:
__anon2c32d4080202(unsigned kk) 184   auto get_idx_dd = [](unsigned kk) -> DoubleDouble {
185     unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
186     DoubleDouble ans = SIN_K_PI_OVER_128[idx];
187     if (kk & 128) {
188       ans.hi = -ans.hi;
189       ans.lo = -ans.lo;
190     }
191     return ans;
192   };
193   DoubleDouble msin_k = get_idx_dd(k + 128);
194   DoubleDouble cos_k = get_idx_dd(k + 64);
195 #else
196   // Fast look up version, but needs 256-entry table.
197   // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
198   DoubleDouble msin_k = SIN_K_PI_OVER_128[(k + 128) & 255];
199   DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255];
200 #endif // LIBC_MATH_HAS_SMALL_TABLES
201 
202   // After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128).
203   // So k is an integer and -pi / 256 <= y <= pi / 256.
204   // Then tan(x) = sin(x) / cos(x)
205   //             = sin((k * pi/128 + y) / cos((k * pi/128 + y)
206   //             = (cos(y) * sin(k*pi/128) + sin(y) * cos(k*pi/128)) /
207   //               / (cos(y) * cos(k*pi/128) - sin(y) * sin(k*pi/128))
208   //             = (sin(k*pi/128) + tan(y) * cos(k*pi/128)) /
209   //               / (cos(k*pi/128) - tan(y) * sin(k*pi/128))
210   DoubleDouble cos_k_tan_y = fputil::quick_mult(tan_y, cos_k);
211   DoubleDouble msin_k_tan_y = fputil::quick_mult(tan_y, msin_k);
212 
213   // num_dd = sin(k*pi/128) + tan(y) * cos(k*pi/128)
214   DoubleDouble num_dd = fputil::exact_add<false>(cos_k_tan_y.hi, -msin_k.hi);
215   // den_dd = cos(k*pi/128) - tan(y) * sin(k*pi/128)
216   DoubleDouble den_dd = fputil::exact_add<false>(msin_k_tan_y.hi, cos_k.hi);
217   num_dd.lo += cos_k_tan_y.lo - msin_k.lo;
218   den_dd.lo += msin_k_tan_y.lo + cos_k.lo;
219 
220 #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
221   double tan_x = (num_dd.hi + num_dd.lo) / (den_dd.hi + den_dd.lo);
222   return tan_x;
223 #else
224   // Accurate test and pass for correctly rounded implementation.
225 
226   // Accurate double-double division
227   DoubleDouble tan_x = fputil::div(num_dd, den_dd);
228 
229   // Simple error bound: |1 / den_dd| < 2^(1 + floor(-log2(den_dd)))).
230   uint64_t den_inv = (static_cast<uint64_t>(FPBits::EXP_BIAS + 1)
231                       << (FPBits::FRACTION_LEN + 1)) -
232                      (FPBits(den_dd.hi).uintval() & FPBits::EXP_MASK);
233 
234   // For tan_x = (num_dd + err) / (den_dd + err), the error is bounded by:
235   //   | tan_x - num_dd / den_dd |  <= err * ( 1 + | tan_x * den_dd | ).
236   double tan_err =
237       err * fputil::multiply_add(FPBits(den_inv).get_val(),
238                                  FPBits(tan_x.hi).abs().get_val(), 1.0);
239 
240   double err_higher = tan_x.lo + tan_err;
241   double err_lower = tan_x.lo - tan_err;
242 
243   double tan_upper = tan_x.hi + err_higher;
244   double tan_lower = tan_x.hi + err_lower;
245 
246   // Ziv's rounding test.
247   if (LIBC_LIKELY(tan_upper == tan_lower))
248     return tan_upper;
249 
250   Float128 u_f128;
251   if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT))
252     u_f128 = range_reduction_small_f128(x);
253   else
254     u_f128 = range_reduction_large.accurate();
255 
256   Float128 tan_u = tan_eval(u_f128);
257 
__anon2c32d4080302(unsigned kk) 258   auto get_sin_k = [](unsigned kk) -> Float128 {
259     unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
260     Float128 ans = SIN_K_PI_OVER_128_F128[idx];
261     if (kk & 128)
262       ans.sign = Sign::NEG;
263     return ans;
264   };
265 
266   // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
267   Float128 sin_k_f128 = get_sin_k(k);
268   Float128 cos_k_f128 = get_sin_k(k + 64);
269   Float128 msin_k_f128 = get_sin_k(k + 128);
270 
271   // num_f128 = sin(k*pi/128) + tan(y) * cos(k*pi/128)
272   Float128 num_f128 =
273       fputil::quick_add(sin_k_f128, fputil::quick_mul(cos_k_f128, tan_u));
274   // den_f128 = cos(k*pi/128) - tan(y) * sin(k*pi/128)
275   Float128 den_f128 =
276       fputil::quick_add(cos_k_f128, fputil::quick_mul(msin_k_f128, tan_u));
277 
278   // tan(x) = (sin(k*pi/128) + tan(y) * cos(k*pi/128)) /
279   //          / (cos(k*pi/128) - tan(y) * sin(k*pi/128))
280   // TODO: The initial seed 1.0/den_dd.hi for Newton-Raphson reciprocal can be
281   // reused from DoubleDouble fputil::div in the fast pass.
282   Float128 result = newton_raphson_div(num_f128, den_f128, 1.0 / den_dd.hi);
283 
284   // TODO: Add assertion if Ziv's accuracy tests fail in debug mode.
285   // https://github.com/llvm/llvm-project/issues/96452.
286   return static_cast<double>(result);
287 
288 #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
289 }
290 
291 } // namespace LIBC_NAMESPACE_DECL
292