brintos

brintos / llvm-project-archived public Read only

0
0
Text · 9.2 KiB · b5ba9ea Raw
277 lines · cpp
1//===-- Double-precision asin 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-exception6//7//===----------------------------------------------------------------------===//8 9#include "src/math/asin.h"10#include "src/__support/math/asin.h"11 12namespace LIBC_NAMESPACE_DECL {13 14LLVM_LIBC_FUNCTION(double, asin, (double x)) {15  using namespace asin_internal;16  using FPBits = fputil::FPBits<double>;17 18  FPBits xbits(x);19  int x_exp = xbits.get_biased_exponent();20 21  // |x| < 0.5.22  if (x_exp < FPBits::EXP_BIAS - 1) {23    // |x| < 2^-26.24    if (LIBC_UNLIKELY(x_exp < FPBits::EXP_BIAS - 26)) {25      // When |x| < 2^-26, the relative error of the approximation asin(x) ~ x26      // is:27      //   |asin(x) - x| / |asin(x)| < |x^3| / (6|x|)28      //                             = x^2 / 629      //                             < 2^-5430      //                             < epsilon(1)/2.31      // So the correctly rounded values of asin(x) are:32      //   = x + sign(x)*eps(x) if rounding mode = FE_TOWARDZERO,33      //                        or (rounding mode = FE_UPWARD and x is34      //                        negative),35      //   = x otherwise.36      // To simplify the rounding decision and make it more efficient, we use37      //   fma(x, 2^-54, x) instead.38      // Note: to use the formula x + 2^-54*x to decide the correct rounding, we39      // do need fma(x, 2^-54, x) to prevent underflow caused by 2^-54*x when40      // |x| < 2^-1022. For targets without FMA instructions, when x is close to41      // denormal range, we normalize x,42#if defined(LIBC_MATH_HAS_SKIP_ACCURATE_PASS)43      return x;44#elif defined(LIBC_TARGET_CPU_HAS_FMA_DOUBLE)45      return fputil::multiply_add(x, 0x1.0p-54, x);46#else47      if (xbits.abs().uintval() == 0)48        return x;49      // Get sign(x) * min_normal.50      FPBits eps_bits = FPBits::min_normal();51      eps_bits.set_sign(xbits.sign());52      double eps = eps_bits.get_val();53      double normalize_const = (x_exp == 0) ? eps : 0.0;54      double scaled_normal =55          fputil::multiply_add(x + normalize_const, 0x1.0p54, eps);56      return fputil::multiply_add(scaled_normal, 0x1.0p-54, -normalize_const);57#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS58    }59 60#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS61    return x * asin_eval(x * x);62#else63    unsigned idx;64    DoubleDouble x_sq = fputil::exact_mult(x, x);65    double err = xbits.abs().get_val() * 0x1.0p-51;66    // Polynomial approximation:67    //   p ~ asin(x)/x68 69    DoubleDouble p = asin_eval(x_sq, idx, err);70    // asin(x) ~ x * (ASIN_COEFFS[idx][0] + p)71    DoubleDouble r0 = fputil::exact_mult(x, p.hi);72    double r_lo = fputil::multiply_add(x, p.lo, r0.lo);73 74    // Ziv's accuracy test.75 76    double r_upper = r0.hi + (r_lo + err);77    double r_lower = r0.hi + (r_lo - err);78 79    if (LIBC_LIKELY(r_upper == r_lower))80      return r_upper;81 82    // Ziv's accuracy test failed, perform 128-bit calculation.83 84    // Recalculate mod 1/64.85    idx = static_cast<unsigned>(fputil::nearest_integer(x_sq.hi * 0x1.0p6));86 87    // Get x^2 - idx/64 exactly.  When FMA is available, double-double88    // multiplication will be correct for all rounding modes.  Otherwise we use89    // Float128 directly.90    Float128 x_f128(x);91 92#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE93    // u = x^2 - idx/6494    Float128 u_hi(95        fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, x_sq.hi));96    Float128 u = fputil::quick_add(u_hi, Float128(x_sq.lo));97#else98    Float128 x_sq_f128 = fputil::quick_mul(x_f128, x_f128);99    Float128 u = fputil::quick_add(100        x_sq_f128, Float128(static_cast<double>(idx) * (-0x1.0p-6)));101#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE102 103    Float128 p_f128 = asin_eval(u, idx);104    Float128 r = fputil::quick_mul(x_f128, p_f128);105 106    return static_cast<double>(r);107#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS108  }109  // |x| >= 0.5110 111  double x_abs = xbits.abs().get_val();112 113  // Maintaining the sign:114  constexpr double SIGN[2] = {1.0, -1.0};115  double x_sign = SIGN[xbits.is_neg()];116 117  // |x| >= 1118  if (LIBC_UNLIKELY(x_exp >= FPBits::EXP_BIAS)) {119    // x = +-1, asin(x) = +- pi/2120    if (x_abs == 1.0) {121      // return +- pi/2122      return fputil::multiply_add(x_sign, PI_OVER_TWO.hi,123                                  x_sign * PI_OVER_TWO.lo);124    }125    // |x| > 1, return NaN.126    if (xbits.is_quiet_nan())127      return x;128 129    // Set domain error for non-NaN input.130    if (!xbits.is_nan())131      fputil::set_errno_if_required(EDOM);132 133    fputil::raise_except_if_required(FE_INVALID);134    return FPBits::quiet_nan().get_val();135  }136 137  // When |x| >= 0.5, we perform range reduction as follow:138  //139  // Assume further that 0.5 <= x < 1, and let:140  //   y = asin(x)141  // We will use the double angle formula:142  //   cos(2y) = 1 - 2 sin^2(y)143  // and the complement angle identity:144  //   x = sin(y) = cos(pi/2 - y)145  //              = 1 - 2 sin^2 (pi/4 - y/2)146  // So:147  //   sin(pi/4 - y/2) = sqrt( (1 - x)/2 )148  // And hence:149  //   pi/4 - y/2 = asin( sqrt( (1 - x)/2 ) )150  // Equivalently:151  //   asin(x) = y = pi/2 - 2 * asin( sqrt( (1 - x)/2 ) )152  // Let u = (1 - x)/2, then:153  //   asin(x) = pi/2 - 2 * asin( sqrt(u) )154  // Moreover, since 0.5 <= x < 1:155  //   0 < u <= 1/4, and 0 < sqrt(u) <= 0.5,156  // And hence we can reuse the same polynomial approximation of asin(x) when157  // |x| <= 0.5:158  //   asin(x) ~ pi/2 - 2 * sqrt(u) * P(u),159 160  // u = (1 - |x|)/2161  double u = fputil::multiply_add(x_abs, -0.5, 0.5);162  // v_hi + v_lo ~ sqrt(u).163  // Let:164  //   h = u - v_hi^2 = (sqrt(u) - v_hi) * (sqrt(u) + v_hi)165  // Then:166  //   sqrt(u) = v_hi + h / (sqrt(u) + v_hi)167  //           ~ v_hi + h / (2 * v_hi)168  // So we can use:169  //   v_lo = h / (2 * v_hi).170  // Then,171  //   asin(x) ~ pi/2 - 2*(v_hi + v_lo) * P(u)172  double v_hi = fputil::sqrt<double>(u);173 174#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS175  double p = asin_eval(u);176  double r = x_sign * fputil::multiply_add(-2.0 * v_hi, p, PI_OVER_TWO.hi);177  return r;178#else179 180#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE181  double h = fputil::multiply_add(v_hi, -v_hi, u);182#else183  DoubleDouble v_hi_sq = fputil::exact_mult(v_hi, v_hi);184  double h = (u - v_hi_sq.hi) - v_hi_sq.lo;185#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE186 187  // Scale v_lo and v_hi by 2 from the formula:188  //   vh = v_hi * 2189  //   vl = 2*v_lo = h / v_hi.190  double vh = v_hi * 2.0;191  double vl = h / v_hi;192 193  // Polynomial approximation:194  //   p ~ asin(sqrt(u))/sqrt(u)195  unsigned idx;196  double err = vh * 0x1.0p-51;197 198  DoubleDouble p = asin_eval(DoubleDouble{0.0, u}, idx, err);199 200  // Perform computations in double-double arithmetic:201  //   asin(x) = pi/2 - (v_hi + v_lo) * (ASIN_COEFFS[idx][0] + p)202  DoubleDouble r0 = fputil::quick_mult(DoubleDouble{vl, vh}, p);203  DoubleDouble r = fputil::exact_add(PI_OVER_TWO.hi, -r0.hi);204 205  double r_lo = PI_OVER_TWO.lo - r0.lo + r.lo;206 207  // Ziv's accuracy test.208 209#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE210  double r_upper = fputil::multiply_add(211      r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, err));212  double r_lower = fputil::multiply_add(213      r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, -err));214#else215  r_lo *= x_sign;216  r.hi *= x_sign;217  double r_upper = r.hi + (r_lo + err);218  double r_lower = r.hi + (r_lo - err);219#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE220 221  if (LIBC_LIKELY(r_upper == r_lower))222    return r_upper;223 224  // Ziv's accuracy test failed, we redo the computations in Float128.225  // Recalculate mod 1/64.226  idx = static_cast<unsigned>(fputil::nearest_integer(u * 0x1.0p6));227 228  // After the first step of Newton-Raphson approximating v = sqrt(u), we have229  // that:230  //   sqrt(u) = v_hi + h / (sqrt(u) + v_hi)231  //      v_lo = h / (2 * v_hi)232  // With error:233  //   sqrt(u) - (v_hi + v_lo) = h * ( 1/(sqrt(u) + v_hi) - 1/(2*v_hi) )234  //                           = -h^2 / (2*v * (sqrt(u) + v)^2).235  // Since:236  //   (sqrt(u) + v_hi)^2 ~ (2sqrt(u))^2 = 4u,237  // we can add another correction term to (v_hi + v_lo) that is:238  //   v_ll = -h^2 / (2*v_hi * 4u)239  //        = -v_lo * (h / 4u)240  //        = -vl * (h / 8u),241  // making the errors:242  //   sqrt(u) - (v_hi + v_lo + v_ll) = O(h^3)243  // well beyond 128-bit precision needed.244 245  // Get the rounding error of vl = 2 * v_lo ~ h / vh246  // Get full product of vh * vl247#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE248  double vl_lo = fputil::multiply_add(-v_hi, vl, h) / v_hi;249#else250  DoubleDouble vh_vl = fputil::exact_mult(v_hi, vl);251  double vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi;252#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE253  // vll = 2*v_ll = -vl * (h / (4u)).254  double t = h * (-0.25) / u;255  double vll = fputil::multiply_add(vl, t, vl_lo);256  // m_v = -(v_hi + v_lo + v_ll).257  Float128 m_v = fputil::quick_add(258      Float128(vh), fputil::quick_add(Float128(vl), Float128(vll)));259  m_v.sign = Sign::NEG;260 261  // Perform computations in Float128:262  //   asin(x) = pi/2 - (v_hi + v_lo + vll) * P(u).263  Float128 y_f128(fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, u));264 265  Float128 p_f128 = asin_eval(y_f128, idx);266  Float128 r0_f128 = fputil::quick_mul(m_v, p_f128);267  Float128 r_f128 = fputil::quick_add(PI_OVER_TWO_F128, r0_f128);268 269  if (xbits.is_neg())270    r_f128.sign = Sign::NEG;271 272  return static_cast<double>(r_f128);273#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS274}275 276} // namespace LIBC_NAMESPACE_DECL277