brintos

brintos / llvm-project-archived public Read only

0
0
Text · 6.7 KiB · 72c8aa3 Raw
175 lines · cpp
1//===-- Single-precision e^x - 1 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/expm1f.h"10#include "src/__support/FPUtil/BasicOperations.h"11#include "src/__support/FPUtil/FEnvImpl.h"12#include "src/__support/FPUtil/FMA.h"13#include "src/__support/FPUtil/FPBits.h"14#include "src/__support/FPUtil/PolyEval.h"15#include "src/__support/FPUtil/multiply_add.h"16#include "src/__support/FPUtil/nearest_integer.h"17#include "src/__support/FPUtil/rounding_mode.h"18#include "src/__support/common.h"19#include "src/__support/macros/config.h"20#include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY21#include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA22#include "src/__support/math/common_constants.h" // Lookup tables EXP_M1 and EXP_M2.23 24namespace LIBC_NAMESPACE_DECL {25 26LLVM_LIBC_FUNCTION(float, expm1f, (float x)) {27  using namespace common_constants_internal;28  using FPBits = typename fputil::FPBits<float>;29  FPBits xbits(x);30 31  uint32_t x_u = xbits.uintval();32  uint32_t x_abs = x_u & 0x7fff'ffffU;33 34#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS35  // Exceptional value36  if (LIBC_UNLIKELY(x_u == 0x3e35'bec5U)) { // x = 0x1.6b7d8ap-3f37    int round_mode = fputil::quick_get_round();38    if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)39      return 0x1.8dbe64p-3f;40    return 0x1.8dbe62p-3f;41  }42#if !defined(LIBC_TARGET_CPU_HAS_FMA_DOUBLE)43  if (LIBC_UNLIKELY(x_u == 0xbdc1'c6cbU)) { // x = -0x1.838d96p-4f44    int round_mode = fputil::quick_get_round();45    if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD)46      return -0x1.71c884p-4f;47    return -0x1.71c882p-4f;48  }49#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE50#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS51 52  // When |x| > 25*log(2), or nan53  if (LIBC_UNLIKELY(x_abs >= 0x418a'a123U)) {54    // x < log(2^-25)55    if (xbits.is_neg()) {56      // exp(-Inf) = 057      if (xbits.is_inf())58        return -1.0f;59      // exp(nan) = nan60      if (xbits.is_nan())61        return x;62      int round_mode = fputil::quick_get_round();63      if (round_mode == FE_UPWARD || round_mode == FE_TOWARDZERO)64        return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f65      return -1.0f;66    } else {67      // x >= 89 or nan68      if (xbits.uintval() >= 0x42b2'0000) {69        if (xbits.uintval() < 0x7f80'0000U) {70          int rounding = fputil::quick_get_round();71          if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)72            return FPBits::max_normal().get_val();73 74          fputil::set_errno_if_required(ERANGE);75          fputil::raise_except_if_required(FE_OVERFLOW);76        }77        return x + FPBits::inf().get_val();78      }79    }80  }81 82  // |x| < 2^-483  if (x_abs < 0x3d80'0000U) {84    // |x| < 2^-2585    if (x_abs < 0x3300'0000U) {86      // x = -0.0f87      if (LIBC_UNLIKELY(xbits.uintval() == 0x8000'0000U))88        return x;89        // When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x90        // is:91        //   |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x|92        //                               = |x|93        //                               < 2^-2594        //                               < epsilon(1)/2.95        // So the correctly rounded values of expm1(x) are:96        //   = x + eps(x) if rounding mode = FE_UPWARD,97        //                   or (rounding mode = FE_TOWARDZERO and x is98        //                   negative),99        //   = x otherwise.100        // To simplify the rounding decision and make it more efficient, we use101        //   fma(x, x, x) ~ x + x^2 instead.102        // Note: to use the formula x + x^2 to decide the correct rounding, we103        // do need fma(x, x, x) to prevent underflow caused by x*x when |x| <104        // 2^-76. For targets without FMA instructions, we simply use double for105        // intermediate results as it is more efficient than using an emulated106        // version of FMA.107#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)108      return fputil::multiply_add(x, x, x);109#else110      double xd = x;111      return static_cast<float>(fputil::multiply_add(xd, xd, xd));112#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT113    }114 115    constexpr double COEFFS[] = {0x1p-1,116                                 0x1.55555555557ddp-3,117                                 0x1.55555555552fap-5,118                                 0x1.111110fcd58b7p-7,119                                 0x1.6c16c1717660bp-10,120                                 0x1.a0241f0006d62p-13,121                                 0x1.a01e3f8d3c06p-16};122 123    // 2^-25 <= |x| < 2^-4124    double xd = static_cast<double>(x);125    double xsq = xd * xd;126    // Degree-8 minimax polynomial generated by Sollya with:127    // > display = hexadecimal;128    // > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]);129 130    double c0 = fputil::multiply_add(xd, COEFFS[1], COEFFS[0]);131    double c1 = fputil::multiply_add(xd, COEFFS[3], COEFFS[2]);132    double c2 = fputil::multiply_add(xd, COEFFS[5], COEFFS[4]);133 134    double r = fputil::polyeval(xsq, c0, c1, c2, COEFFS[6]);135    return static_cast<float>(fputil::multiply_add(r, xsq, xd));136  }137 138  // For -18 < x < 89, to compute expm1(x), we perform the following range139  // reduction: find hi, mid, lo such that:140  //   x = hi + mid + lo, in which141  //     hi is an integer,142  //     mid * 2^7 is an integer143  //     -2^(-8) <= lo < 2^-8.144  // In particular,145  //   hi + mid = round(x * 2^7) * 2^(-7).146  // Then,147  //   expm1(x) = exp(hi + mid + lo) - 1 = exp(hi) * exp(mid) * exp(lo) - 1.148  // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2149  // respectively.  exp(lo) is computed using a degree-4 minimax polynomial150  // generated by Sollya.151 152  // x_hi = hi + mid.153  float kf = fputil::nearest_integer(x * 0x1.0p7f);154  int x_hi = static_cast<int>(kf);155  // Subtract (hi + mid) from x to get lo.156  double xd = static_cast<double>(fputil::multiply_add(kf, -0x1.0p-7f, x));157  x_hi += 104 << 7;158  // hi = x_hi >> 7159  double exp_hi = EXP_M1[x_hi >> 7];160  // lo = x_hi & 0x0000'007fU;161  double exp_mid = EXP_M2[x_hi & 0x7f];162  double exp_hi_mid = exp_hi * exp_mid;163  // Degree-4 minimax polynomial generated by Sollya with the following164  // commands:165  //   > display = hexadecimal;166  //   > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]);167  //   > Q;168  double exp_lo =169      fputil::polyeval(xd, 0x1.0p0, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,170                       0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);171  return static_cast<float>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0));172}173 174} // namespace LIBC_NAMESPACE_DECL175