xref: /aosp_15_r20/external/llvm-libc/src/math/generic/cbrt.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Implementation of cbrt 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/cbrt.h"
10 #include "hdr/fenv_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/multiply_add.h"
17 #include "src/__support/common.h"
18 #include "src/__support/integer_literals.h"
19 #include "src/__support/macros/config.h"
20 #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
21 
22 #if ((LIBC_MATH & LIBC_MATH_SKIP_ACCURATE_PASS) != 0)
23 #define LIBC_MATH_CBRT_SKIP_ACCURATE_PASS
24 #endif
25 
26 namespace LIBC_NAMESPACE_DECL {
27 
28 using DoubleDouble = fputil::DoubleDouble;
29 using Float128 = fputil::DyadicFloat<128>;
30 
31 namespace {
32 
33 // Initial approximation of x^(-2/3) for 1 <= x < 2.
34 // Polynomial generated by Sollya with:
35 // > P = fpminimax(x^(-2/3), 7, [|D...|], [1, 2]);
36 // > dirtyinfnorm(P/x^(-2/3) - 1, [1, 2]);
37 // 0x1.28...p-21
intial_approximation(double x)38 double intial_approximation(double x) {
39   constexpr double COEFFS[8] = {
40       0x1.bc52aedead5c6p1,  -0x1.b52bfebf110b3p2,  0x1.1d8d71d53d126p3,
41       -0x1.de2db9e81cf87p2, 0x1.0154ca06153bdp2,   -0x1.5973c66ee6da7p0,
42       0x1.07bf6ac832552p-2, -0x1.5e53d9ce41cb8p-6,
43   };
44 
45   double x_sq = x * x;
46 
47   double c0 = fputil::multiply_add(x, COEFFS[1], COEFFS[0]);
48   double c1 = fputil::multiply_add(x, COEFFS[3], COEFFS[2]);
49   double c2 = fputil::multiply_add(x, COEFFS[5], COEFFS[4]);
50   double c3 = fputil::multiply_add(x, COEFFS[7], COEFFS[6]);
51 
52   double x_4 = x_sq * x_sq;
53   double d0 = fputil::multiply_add(x_sq, c1, c0);
54   double d1 = fputil::multiply_add(x_sq, c3, c2);
55 
56   return fputil::multiply_add(x_4, d1, d0);
57 }
58 
59 // Get the error term for Newton iteration:
60 //   h(x) = x^3 * a^2 - 1,
61 #ifdef LIBC_TARGET_CPU_HAS_FMA
get_error(const DoubleDouble & x_3,const DoubleDouble & a_sq)62 double get_error(const DoubleDouble &x_3, const DoubleDouble &a_sq) {
63   return fputil::multiply_add(x_3.hi, a_sq.hi, -1.0) +
64          fputil::multiply_add(x_3.lo, a_sq.hi, x_3.hi * a_sq.lo);
65 }
66 #else
get_error(const DoubleDouble & x_3,const DoubleDouble & a_sq)67 double get_error(const DoubleDouble &x_3, const DoubleDouble &a_sq) {
68   DoubleDouble x_3_a_sq = fputil::quick_mult(a_sq, x_3);
69   return (x_3_a_sq.hi - 1.0) + x_3_a_sq.lo;
70 }
71 #endif
72 
73 } // anonymous namespace
74 
75 // Correctly rounded cbrt algorithm:
76 //
77 // === Step 1 - Range reduction ===
78 // For x = (-1)^s * 2^e * (1.m), we get 2 reduced arguments x_r and a as:
79 //   x_r = 1.m
80 //   a   = (-1)^s * 2^(e % 3) * (1.m)
81 // Then cbrt(x) = x^(1/3) can be computed as:
82 //   x^(1/3) = 2^(e / 3) * a^(1/3).
83 //
84 // In order to avoid division, we compute a^(-2/3) using Newton method and then
85 // multiply the results by a:
86 //   a^(1/3) = a * a^(-2/3).
87 //
88 // === Step 2 - First approximation to a^(-2/3) ===
89 // First, we use a degree-7 minimax polynomial generated by Sollya to
90 // approximate x_r^(-2/3) for 1 <= x_r < 2.
91 //   p = P(x_r) ~ x_r^(-2/3),
92 // with relative errors bounded by:
93 //   | p / x_r^(-2/3) - 1 | < 1.16 * 2^-21.
94 //
95 // Then we multiply with 2^(e % 3) from a small lookup table to get:
96 //   x_0 = 2^(-2*(e % 3)/3) * p
97 //       ~ 2^(-2*(e % 3)/3) * x_r^(-2/3)
98 //       = a^(-2/3)
99 // With relative errors:
100 //   | x_0 / a^(-2/3) - 1 | < 1.16 * 2^-21.
101 // This step is done in double precision.
102 //
103 // === Step 3 - First Newton iteration ===
104 // We follow the method described in:
105 //   Sibidanov, A. and Zimmermann, P., "Correctly rounded cubic root evaluation
106 //   in double precision", https://core-math.gitlabpages.inria.fr/cbrt64.pdf
107 // to derive multiplicative Newton iterations as below:
108 // Let x_n be the nth approximation to a^(-2/3).  Define the n^th error as:
109 //   h_n = x_n^3 * a^2 - 1
110 // Then:
111 //   a^(-2/3) = x_n / (1 + h_n)^(1/3)
112 //            = x_n * (1 - (1/3) * h_n + (2/9) * h_n^2 - (14/81) * h_n^3 + ...)
113 // using the Taylor series expansion of (1 + h_n)^(-1/3).
114 //
115 // Apply to x_0 above:
116 //   h_0 = x_0^3 * a^2 - 1
117 //       = a^2 * (x_0 - a^(-2/3)) * (x_0^2 + x_0 * a^(-2/3) + a^(-4/3)),
118 // it's bounded by:
119 //   |h_0| < 4 * 3 * 1.16 * 2^-21 * 4 < 2^-17.
120 // So in the first iteration step, we use:
121 //   x_1 = x_0 * (1 - (1/3) * h_n + (2/9) * h_n^2 - (14/81) * h_n^3)
122 // Its relative error is bounded by:
123 //   | x_1 / a^(-2/3) - 1 | < 35/242 * |h_0|^4 < 2^-70.
124 // Then we perform Ziv's rounding test and check if the answer is exact.
125 // This step is done in double-double precision.
126 //
127 // === Step 4 - Second Newton iteration ===
128 // If the Ziv's rounding test from the previous step fails, we define the error
129 // term:
130 //   h_1 = x_1^3 * a^2 - 1,
131 // And perform another iteration:
132 //   x_2 = x_1 * (1 - h_1 / 3)
133 // with the relative errors exceed the precision of double-double.
134 // We then check the Ziv's accuracy test with relative errors < 2^-102 to
135 // compensate for rounding errors.
136 //
137 // === Step 5 - Final iteration ===
138 // If the Ziv's accuracy test from the previous step fails, we perform another
139 // iteration in 128-bit precision and check for exact outputs.
140 //
141 // TODO: It is possible to replace this costly computation step with special
142 // exceptional handling, similar to what was done in the CORE-MATH project:
143 // https://gitlab.inria.fr/core-math/core-math/-/blob/master/src/binary64/cbrt/cbrt.c
144 
145 LLVM_LIBC_FUNCTION(double, cbrt, (double x)) {
146   using FPBits = fputil::FPBits<double>;
147 
148   uint64_t x_abs = FPBits(x).abs().uintval();
149 
150   unsigned exp_bias_correction = 682; // 1023 * 2/3
151 
152   if (LIBC_UNLIKELY(x_abs < FPBits::min_normal().uintval() ||
153                     x_abs >= FPBits::inf().uintval())) {
154     if (x == 0.0 || x_abs >= FPBits::inf().uintval())
155       // x is 0, Inf, or NaN.
156       // Make sure it works for FTZ/DAZ modes.
157       return static_cast<double>(x + x);
158 
159     // x is non-zero denormal number.
160     // Normalize x.
161     x *= 0x1.0p60;
162     exp_bias_correction -= 20;
163   }
164 
165   FPBits x_bits(x);
166 
167   // When using biased exponent of x in double precision,
168   //   x_e = real_exponent_of_x + 1023
169   // Then:
170   //   x_e / 3 = real_exponent_of_x / 3 + 1023/3
171   //           = real_exponent_of_x / 3 + 341
172   // So to make it the correct biased exponent of x^(1/3), we add
173   //   1023 - 341 = 682
174   // to the quotient x_e / 3.
175   unsigned x_e = static_cast<unsigned>(x_bits.get_biased_exponent());
176   unsigned out_e = (x_e / 3 + exp_bias_correction);
177   unsigned shift_e = x_e % 3;
178 
179   // Set x_r = 1.mantissa
180   double x_r =
181       FPBits(x_bits.get_mantissa() |
182              (static_cast<uint64_t>(FPBits::EXP_BIAS) << FPBits::FRACTION_LEN))
183           .get_val();
184 
185   // Set a = (-1)^x_sign * 2^(x_e % 3) * (1.mantissa)
186   uint64_t a_bits = x_bits.uintval() & 0x800F'FFFF'FFFF'FFFF;
187   a_bits |=
188       (static_cast<uint64_t>(shift_e + static_cast<unsigned>(FPBits::EXP_BIAS))
189        << FPBits::FRACTION_LEN);
190   double a = FPBits(a_bits).get_val();
191 
192   // Initial approximation of x_r^(-2/3).
193   double p = intial_approximation(x_r);
194 
195   // Look up for 2^(-2*n/3) used for first approximation step.
196   constexpr double EXP2_M2_OVER_3[3] = {1.0, 0x1.428a2f98d728bp-1,
197                                         0x1.965fea53d6e3dp-2};
198 
199   // x0 is an initial approximation of a^(-2/3) for 1 <= |a| < 8.
200   // Relative error: < 1.16 * 2^(-21).
201   double x0 = static_cast<double>(EXP2_M2_OVER_3[shift_e] * p);
202 
203   // First iteration in double precision.
204   DoubleDouble a_sq = fputil::exact_mult(a, a);
205 
206   // h0 = x0^3 * a^2 - 1
207   DoubleDouble x0_sq = fputil::exact_mult(x0, x0);
208   DoubleDouble x0_3 = fputil::quick_mult(x0, x0_sq);
209 
210   double h0 = get_error(x0_3, a_sq);
211 
212 #ifdef LIBC_MATH_CBRT_SKIP_ACCURATE_PASS
213   constexpr double REL_ERROR = 0;
214 #else
215   constexpr double REL_ERROR = 0x1.0p-51;
216 #endif // LIBC_MATH_CBRT_SKIP_ACCURATE_PASS
217 
218   // Taylor polynomial of (1 + h)^(-1/3):
219   //   (1 + h)^(-1/3) = 1 - h/3 + 2 h^2 / 9 - 14 h^3 / 81 + ...
220   constexpr double ERR_COEFFS[3] = {
221       -0x1.5555555555555p-2 - REL_ERROR, // -1/3 - relative_error
222       0x1.c71c71c71c71cp-3,              // 2/9
223       -0x1.61f9add3c0ca4p-3,             // -14/81
224   };
225   // e0 = -14 * h^2 / 81 + 2 * h / 9 - 1/3 - relative_error.
226   double e0 = fputil::polyeval(h0, ERR_COEFFS[0], ERR_COEFFS[1], ERR_COEFFS[2]);
227   double x0_h0 = x0 * h0;
228 
229   // x1 = x0 (1 - h0/3 + 2 h0^2 / 9 - 14 h0^3 / 81)
230   // x1 approximate a^(-2/3) with relative errors bounded by:
231   //   | x1 / a^(-2/3) - 1 | < (34/243) h0^4 < h0 * REL_ERROR
232   DoubleDouble x1_dd{x0_h0 * e0, x0};
233 
234   // r1 = x1 * a ~ a^(-2/3) * a = a^(1/3).
235   DoubleDouble r1 = fputil::quick_mult(a, x1_dd);
236 
237   // Lambda function to update the exponent of the result.
__anon5d527db00202(double r) 238   auto update_exponent = [=](double r) -> double {
239     uint64_t r_m = FPBits(r).uintval() - 0x3FF0'0000'0000'0000;
240     // Adjust exponent and sign.
241     uint64_t r_bits =
242         r_m + (static_cast<uint64_t>(out_e) << FPBits::FRACTION_LEN);
243     return FPBits(r_bits).get_val();
244   };
245 
246 #ifdef LIBC_MATH_CBRT_SKIP_ACCURATE_PASS
247   // TODO: We probably don't need to use double-double if accurate tests and
248   // passes are skipped.
249   return update_exponent(r1.hi + r1.lo);
250 #else
251   // Accurate checks and passes.
252   double r1_lower = r1.hi + r1.lo;
253   double r1_upper =
254       r1.hi + fputil::multiply_add(x0_h0, 2.0 * REL_ERROR * a, r1.lo);
255 
256   // Ziv's accuracy test.
257   if (LIBC_LIKELY(r1_upper == r1_lower)) {
258     // Test for exact outputs.
259     // Check if lower (52 - 17 = 35) bits are 0's.
260     if (LIBC_UNLIKELY((FPBits(r1_lower).uintval() & 0x0000'0007'FFFF'FFFF) ==
261                       0)) {
262       double r1_err = (r1_lower - r1.hi) - r1.lo;
263       if (FPBits(r1_err).abs().get_val() < 0x1.0p69)
264         fputil::clear_except_if_required(FE_INEXACT);
265     }
266 
267     return update_exponent(r1_lower);
268   }
269 
270   // Accuracy test failed, perform another Newton iteration.
271   double x1 = x1_dd.hi + (e0 + REL_ERROR) * x0_h0;
272 
273   // Second iteration in double-double precision.
274   // h1 = x1^3 * a^2 - 1.
275   DoubleDouble x1_sq = fputil::exact_mult(x1, x1);
276   DoubleDouble x1_3 = fputil::quick_mult(x1, x1_sq);
277   double h1 = get_error(x1_3, a_sq);
278 
279   // e1 = -x1*h1/3.
280   double e1 = h1 * (x1 * -0x1.5555555555555p-2);
281   // x2 = x1*(1 - h1/3) = x1 + e1 ~ a^(-2/3) with relative errors < 2^-101.
282   DoubleDouble x2 = fputil::exact_add(x1, e1);
283   // r2 = a * x2 ~ a * a^(-2/3) = a^(1/3) with relative errors < 2^-100.
284   DoubleDouble r2 = fputil::quick_mult(a, x2);
285 
286   double r2_upper = r2.hi + fputil::multiply_add(a, 0x1.0p-102, r2.lo);
287   double r2_lower = r2.hi + fputil::multiply_add(a, -0x1.0p-102, r2.lo);
288 
289   // Ziv's accuracy test.
290   if (LIBC_LIKELY(r2_upper == r2_lower))
291     return update_exponent(r2_upper);
292 
293   // TODO: Investigate removing float128 and just list exceptional cases.
294   // Apply another Newton iteration with ~126-bit accuracy.
295   Float128 x2_f128 = fputil::quick_add(Float128(x2.hi), Float128(x2.lo));
296   // x2^3
297   Float128 x2_3 =
298       fputil::quick_mul(fputil::quick_mul(x2_f128, x2_f128), x2_f128);
299   // a^2
300   Float128 a_sq_f128 = fputil::quick_mul(Float128(a), Float128(a));
301   // x2^3 * a^2
302   Float128 x2_3_a_sq = fputil::quick_mul(x2_3, a_sq_f128);
303   // h2 = x2^3 * a^2 - 1
304   Float128 h2_f128 = fputil::quick_add(x2_3_a_sq, Float128(-1.0));
305   double h2 = static_cast<double>(h2_f128);
306   // t2 = 1 - h2 / 3
307   Float128 t2 =
308       fputil::quick_add(Float128(1.0), Float128(h2 * (-0x1.5555555555555p-2)));
309   // x3 = x2 * (1 - h2 / 3) ~ a^(-2/3)
310   Float128 x3 = fputil::quick_mul(x2_f128, t2);
311   // r3 = a * x3 ~ a * a^(-2/3) = a^(1/3)
312   Float128 r3 = fputil::quick_mul(Float128(a), x3);
313 
314   // Check for exact cases:
315   Float128::MantissaType rounding_bits =
316       r3.mantissa & 0x0000'0000'0000'03FF'FFFF'FFFF'FFFF'FFFF_u128;
317 
318   double result = static_cast<double>(r3);
319   if ((rounding_bits < 0x0000'0000'0000'0000'0000'0000'0000'000F_u128) ||
320       (rounding_bits >= 0x0000'0000'0000'03FF'FFFF'FFFF'FFFF'FFF0_u128)) {
321     // Output is exact.
322     r3.mantissa &= 0xFFFF'FFFF'FFFF'FFFF'FFFF'FFFF'FFFF'FFF0_u128;
323 
324     if (rounding_bits >= 0x0000'0000'0000'03FF'FFFF'FFFF'FFFF'FFF0_u128) {
325       Float128 tmp{r3.sign, r3.exponent - 123,
326                    0x8000'0000'0000'0000'0000'0000'0000'0000_u128};
327       Float128 r4 = fputil::quick_add(r3, tmp);
328       result = static_cast<double>(r4);
329     } else {
330       result = static_cast<double>(r3);
331     }
332 
333     fputil::clear_except_if_required(FE_INEXACT);
334   }
335 
336   return update_exponent(result);
337 #endif // LIBC_MATH_CBRT_SKIP_ACCURATE_PASS
338 }
339 
340 } // namespace LIBC_NAMESPACE_DECL
341