brintos

brintos / llvm-project-archived public Read only

0
0
Text · 7.5 KiB · 38661de Raw
224 lines · cpp
1//===-- Double-precision sincos 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/sincos.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/double_double.h"14#include "src/__support/FPUtil/dyadic_float.h"15#include "src/__support/FPUtil/except_value_utils.h"16#include "src/__support/FPUtil/multiply_add.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/range_reduction_double_common.h"23#include "src/__support/math/sincos_eval.h"24 25#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE26#include "src/__support/math/range_reduction_double_fma.h"27#else28#include "src/__support/math/range_reduction_double_nofma.h"29#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE30 31namespace LIBC_NAMESPACE_DECL {32 33using DoubleDouble = fputil::DoubleDouble;34using Float128 = typename fputil::DyadicFloat<128>;35 36LLVM_LIBC_FUNCTION(void, sincos, (double x, double *sin_x, double *cos_x)) {37  using namespace math::range_reduction_double_internal;38  using FPBits = typename fputil::FPBits<double>;39  FPBits xbits(x);40 41  uint16_t x_e = xbits.get_biased_exponent();42 43  DoubleDouble y;44  unsigned k;45  LargeRangeReduction range_reduction_large{};46 47  // |x| < 2^1648  if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) {49    // |x| < 2^-750    if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 7)) {51      // |x| < 2^-2752      if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {53        // Signed zeros.54        if (LIBC_UNLIKELY(x == 0.0)) {55          *sin_x = x;56          *cos_x = 1.0;57          return;58        }59 60        // For |x| < 2^-27, max(|sin(x) - x|, |cos(x) - 1|) < ulp(x)/2.61#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE62        *sin_x = fputil::multiply_add(x, -0x1.0p-54, x);63        *cos_x = fputil::multiply_add(x, -x, 1.0);64#else65        *cos_x = fputil::round_result_slightly_down(1.0);66 67        if (LIBC_UNLIKELY(x_e < 4)) {68          int rounding_mode = fputil::quick_get_round();69          if (rounding_mode == FE_TOWARDZERO ||70              (xbits.sign() == Sign::POS && rounding_mode == FE_DOWNWARD) ||71              (xbits.sign() == Sign::NEG && rounding_mode == FE_UPWARD))72            *sin_x = FPBits(xbits.uintval() - 1).get_val();73        }74        *sin_x = fputil::multiply_add(x, -0x1.0p-54, x);75#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE76        return;77      }78      // No range reduction needed.79      k = 0;80      y.lo = 0.0;81      y.hi = x;82    } else {83      // Small range reduction.84      k = range_reduction_small(x, y);85    }86  } else {87    // Inf or NaN88    if (LIBC_UNLIKELY(x_e > 2 * FPBits::EXP_BIAS)) {89      if (xbits.is_signaling_nan()) {90        fputil::raise_except_if_required(FE_INVALID);91        *sin_x = *cos_x = FPBits::quiet_nan().get_val();92        return;93      }94 95      // sin(+-Inf) = NaN96      if (xbits.get_mantissa() == 0) {97        fputil::set_errno_if_required(EDOM);98        fputil::raise_except_if_required(FE_INVALID);99      }100      *sin_x = *cos_x = x + FPBits::quiet_nan().get_val();101      return;102    }103 104    // Large range reduction.105    k = range_reduction_large.fast(x, y);106  }107 108  DoubleDouble sin_y, cos_y;109 110  [[maybe_unused]] double err =111      math::sincos_eval_internal::sincos_eval(y, sin_y, cos_y);112 113  // Look up sin(k * pi/128) and cos(k * pi/128)114#ifdef LIBC_MATH_HAS_SMALL_TABLES115  // Memory saving versions.  Use 65-entry table.116  auto get_idx_dd = [](unsigned kk) -> DoubleDouble {117    unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);118    DoubleDouble ans = SIN_K_PI_OVER_128[idx];119    if (kk & 128) {120      ans.hi = -ans.hi;121      ans.lo = -ans.lo;122    }123    return ans;124  };125  DoubleDouble sin_k = get_idx_dd(k);126  DoubleDouble cos_k = get_idx_dd(k + 64);127#else128  // Fast look up version, but needs 256-entry table.129  // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).130  DoubleDouble sin_k = SIN_K_PI_OVER_128[k & 255];131  DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255];132#endif // LIBC_MATH_HAS_SMALL_TABLES133 134  DoubleDouble msin_k{-sin_k.lo, -sin_k.hi};135 136  // After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128).137  // So k is an integer and -pi / 256 <= y <= pi / 256.138  // Then sin(x) = sin((k * pi/128 + y)139  //             = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128)140  DoubleDouble sin_k_cos_y = fputil::quick_mult(cos_y, sin_k);141  DoubleDouble cos_k_sin_y = fputil::quick_mult(sin_y, cos_k);142  //      cos(x) = cos((k * pi/128 + y)143  //             = cos(y) * cos(k*pi/128) - sin(y) * sin(k*pi/128)144  DoubleDouble cos_k_cos_y = fputil::quick_mult(cos_y, cos_k);145  DoubleDouble msin_k_sin_y = fputil::quick_mult(sin_y, msin_k);146 147  DoubleDouble sin_dd =148      fputil::exact_add<false>(sin_k_cos_y.hi, cos_k_sin_y.hi);149  DoubleDouble cos_dd =150      fputil::exact_add<false>(cos_k_cos_y.hi, msin_k_sin_y.hi);151  sin_dd.lo += sin_k_cos_y.lo + cos_k_sin_y.lo;152  cos_dd.lo += msin_k_sin_y.lo + cos_k_cos_y.lo;153 154#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS155  *sin_x = sin_dd.hi + sin_dd.lo;156  *cos_x = cos_dd.hi + cos_dd.lo;157  return;158#else159  // Accurate test and pass for correctly rounded implementation.160 161  double sin_lp = sin_dd.lo + err;162  double sin_lm = sin_dd.lo - err;163  double cos_lp = cos_dd.lo + err;164  double cos_lm = cos_dd.lo - err;165 166  double sin_upper = sin_dd.hi + sin_lp;167  double sin_lower = sin_dd.hi + sin_lm;168  double cos_upper = cos_dd.hi + cos_lp;169  double cos_lower = cos_dd.hi + cos_lm;170 171  // Ziv's rounding test.172  if (LIBC_LIKELY(sin_upper == sin_lower && cos_upper == cos_lower)) {173    *sin_x = sin_upper;174    *cos_x = cos_upper;175    return;176  }177 178  Float128 u_f128, sin_u, cos_u;179  if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT))180    u_f128 = range_reduction_small_f128(x);181  else182    u_f128 = range_reduction_large.accurate();183 184  math::sincos_eval_internal::sincos_eval(u_f128, sin_u, cos_u);185 186  auto get_sin_k = [](unsigned kk) -> Float128 {187    unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);188    Float128 ans = SIN_K_PI_OVER_128_F128[idx];189    if (kk & 128)190      ans.sign = Sign::NEG;191    return ans;192  };193 194  // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).195  Float128 sin_k_f128 = get_sin_k(k);196  Float128 cos_k_f128 = get_sin_k(k + 64);197  Float128 msin_k_f128 = get_sin_k(k + 128);198 199  // TODO: Add assertion if Ziv's accuracy tests fail in debug mode.200  // https://github.com/llvm/llvm-project/issues/96452.201 202  if (sin_upper == sin_lower)203    *sin_x = sin_upper;204  else205    // sin(x) = sin((k * pi/128 + u)206    //        = sin(u) * cos(k*pi/128) + cos(u) * sin(k*pi/128)207    *sin_x = static_cast<double>(208        fputil::quick_add(fputil::quick_mul(sin_k_f128, cos_u),209                          fputil::quick_mul(cos_k_f128, sin_u)));210 211  if (cos_upper == cos_lower)212    *cos_x = cos_upper;213  else214    // cos(x) = cos((k * pi/128 + u)215    //        = cos(u) * cos(k*pi/128) - sin(u) * sin(k*pi/128)216    *cos_x = static_cast<double>(217        fputil::quick_add(fputil::quick_mul(cos_k_f128, cos_u),218                          fputil::quick_mul(msin_k_f128, sin_u)));219 220#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS221}222 223} // namespace LIBC_NAMESPACE_DECL224