libc/src/math/generic/expm1f.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-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8
   9 #include "src/math/expm1f.h"
  10 #include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2.
  11 #include "src/__support/FPUtil/BasicOperations.h"
  12 #include "src/__support/FPUtil/FEnvImpl.h"
  13 #include "src/__support/FPUtil/FMA.h"
  14 #include "src/__support/FPUtil/FPBits.h"
  15 #include "src/__support/FPUtil/PolyEval.h"
  16 #include "src/__support/FPUtil/multiply_add.h"
  17 #include "src/__support/FPUtil/nearest_integer.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
  24 namespace LIBC_NAMESPACE_DECL {
  25
  26 LLVM_LIBC_FUNCTION(float, expm1f, (float x)) {
  27   using FPBits = typename fputil::FPBits<float>;
  28   FPBits xbits(x);
  29
  30   uint32_t x_u = xbits.uintval();
  31   uint32_t x_abs = x_u & 0x7fff'ffffU;
  32
  33   // Exceptional value
  34   if (LIBC_UNLIKELY(x_u == 0x3e35'bec5U)) { // x = 0x1.6b7d8ap-3f
  35     int round_mode = fputil::quick_get_round();
  36     if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
  37       return 0x1.8dbe64p-3f;
  38     return 0x1.8dbe62p-3f;
  39   }
  40
  41 #if !defined(LIBC_TARGET_CPU_HAS_FMA)
  42   if (LIBC_UNLIKELY(x_u == 0xbdc1'c6cbU)) { // x = -0x1.838d96p-4f
  43     int round_mode = fputil::quick_get_round();
  44     if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD)
  45       return -0x1.71c884p-4f;
  46     return -0x1.71c882p-4f;
  47   }
  48 #endif // LIBC_TARGET_CPU_HAS_FMA
  49
  50   // When |x| > 25*log(2), or nan
  51   if (LIBC_UNLIKELY(x_abs >= 0x418a'a123U)) {
  52     // x < log(2^-25)
  53     if (xbits.is_neg()) {
  54       // exp(-Inf) = 0
  55       if (xbits.is_inf())
  56         return -1.0f;
  57       // exp(nan) = nan
  58       if (xbits.is_nan())
  59         return x;
  60       int round_mode = fputil::quick_get_round();
  61       if (round_mode == FE_UPWARD || round_mode == FE_TOWARDZERO)
  62         return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f
  63       return -1.0f;
  64     } else {
  65       // x >= 89 or nan
  66       if (xbits.uintval() >= 0x42b2'0000) {
  67         if (xbits.uintval() < 0x7f80'0000U) {
  68           int rounding = fputil::quick_get_round();
  69           if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
  70             return FPBits::max_normal().get_val();
  71
  72           fputil::set_errno_if_required(ERANGE);
  73           fputil::raise_except_if_required(FE_OVERFLOW);
  74         }
  75         return x + FPBits::inf().get_val();
  76       }
  77     }
  78   }
  79
  80   // |x| < 2^-4
  81   if (x_abs < 0x3d80'0000U) {
  82     // |x| < 2^-25
  83     if (x_abs < 0x3300'0000U) {
  84       // x = -0.0f
  85       if (LIBC_UNLIKELY(xbits.uintval() == 0x8000'0000U))
  86         return x;
  87         // When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x
  88         // is:
  89         //   |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x|
  90         //                               = |x|
  91         //                               < 2^-25
  92         //                               < epsilon(1)/2.
  93         // So the correctly rounded values of expm1(x) are:
  94         //   = x + eps(x) if rounding mode = FE_UPWARD,
  95         //                   or (rounding mode = FE_TOWARDZERO and x is
  96         //                   negative),
  97         //   = x otherwise.
  98         // To simplify the rounding decision and make it more efficient, we use
  99         //   fma(x, x, x) ~ x + x^2 instead.
 100         // Note: to use the formula x + x^2 to decide the correct rounding, we
 101         // do need fma(x, x, x) to prevent underflow caused by x*x when |x| <
 102         // 2^-76. For targets without FMA instructions, we simply use double for
 103         // intermediate results as it is more efficient than using an emulated
 104         // version of FMA.
 105 #if defined(LIBC_TARGET_CPU_HAS_FMA)
 106       return fputil::fma<float>(x, x, x);
 107 #else
 108       double xd = x;
 109       return static_cast<float>(fputil::multiply_add(xd, xd, xd));
 110 #endif // LIBC_TARGET_CPU_HAS_FMA
 111     }
 112
 113     constexpr double COEFFS[] = {0x1p-1,
 114                                  0x1.55555555557ddp-3,
 115                                  0x1.55555555552fap-5,
 116                                  0x1.111110fcd58b7p-7,
 117                                  0x1.6c16c1717660bp-10,
 118                                  0x1.a0241f0006d62p-13,
 119                                  0x1.a01e3f8d3c06p-16};
 120
 121     // 2^-25 <= |x| < 2^-4
 122     double xd = static_cast<double>(x);
 123     double xsq = xd * xd;
 124     // Degree-8 minimax polynomial generated by Sollya with:
 125     // > display = hexadecimal;
 126     // > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]);
 127
 128     double c0 = fputil::multiply_add(xd, COEFFS[1], COEFFS[0]);
 129     double c1 = fputil::multiply_add(xd, COEFFS[3], COEFFS[2]);
 130     double c2 = fputil::multiply_add(xd, COEFFS[5], COEFFS[4]);
 131
 132     double r = fputil::polyeval(xsq, c0, c1, c2, COEFFS[6]);
 133     return static_cast<float>(fputil::multiply_add(r, xsq, xd));
 134   }
 135
 136   // For -18 < x < 89, to compute expm1(x), we perform the following range
 137   // reduction: find hi, mid, lo such that:
 138   //   x = hi + mid + lo, in which
 139   //     hi is an integer,
 140   //     mid * 2^7 is an integer
 141   //     -2^(-8) <= lo < 2^-8.
 142   // In particular,
 143   //   hi + mid = round(x * 2^7) * 2^(-7).
 144   // Then,
 145   //   expm1(x) = exp(hi + mid + lo) - 1 = exp(hi) * exp(mid) * exp(lo) - 1.
 146   // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2
 147   // respectively.  exp(lo) is computed using a degree-4 minimax polynomial
 148   // generated by Sollya.
 149
 150   // x_hi = hi + mid.
 151   float kf = fputil::nearest_integer(x * 0x1.0p7f);
 152   int x_hi = static_cast<int>(kf);
 153   // Subtract (hi + mid) from x to get lo.
 154   double xd = static_cast<double>(fputil::multiply_add(kf, -0x1.0p-7f, x));
 155   x_hi += 104 << 7;
 156   // hi = x_hi >> 7
 157   double exp_hi = EXP_M1[x_hi >> 7];
 158   // lo = x_hi & 0x0000'007fU;
 159   double exp_mid = EXP_M2[x_hi & 0x7f];
 160   double exp_hi_mid = exp_hi * exp_mid;
 161   // Degree-4 minimax polynomial generated by Sollya with the following
 162   // commands:
 163   //   > display = hexadecimal;
 164   //   > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]);
 165   //   > Q;
 166   double exp_lo =
 167       fputil::polyeval(xd, 0x1.0p0, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,
 168                        0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);
 169   return static_cast<float>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0));
 170 }
 171
 172 } // namespace LIBC_NAMESPACE_DECL