| 1 | //===-- Implementation header for asinpi ------------------------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_ASINPI_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ASINPI_H |
| 11 | |
| 12 | #include "asin_utils.h" |
| 13 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 14 | #include "src/__support/FPUtil/FPBits.h" |
| 15 | #include "src/__support/FPUtil/double_double.h" |
| 16 | #include "src/__support/FPUtil/dyadic_float.h" |
| 17 | #include "src/__support/FPUtil/multiply_add.h" |
| 18 | #include "src/__support/FPUtil/sqrt.h" |
| 19 | #include "src/__support/macros/config.h" |
| 20 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 21 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
| 22 | #include "src/__support/math/asin_utils.h" |
| 23 | |
| 24 | namespace LIBC_NAMESPACE_DECL { |
| 25 | |
| 26 | namespace math { |
| 27 | |
| 28 | LIBC_INLINE double asinpi(double x) { |
| 29 | using namespace asin_internal; |
| 30 | using FPBits = fputil::FPBits<double>; |
| 31 | |
| 32 | FPBits xbits(x); |
| 33 | int x_exp = xbits.get_biased_exponent(); |
| 34 | |
| 35 | // |x| < 0.5. |
| 36 | if (x_exp < FPBits::EXP_BIAS - 1) { |
| 37 | // |x| < 2^-26. |
| 38 | if (LIBC_UNLIKELY(x_exp < FPBits::EXP_BIAS - 26)) { |
| 39 | // asinpi(+-0) = +-0. |
| 40 | if (LIBC_UNLIKELY(xbits.abs().uintval() == 0)) |
| 41 | return x; |
| 42 | // When |x| < 2^-26, asinpi(x) ~ x/pi. |
| 43 | // The relative error of x/pi is: |
| 44 | // |asinpi(x) - x/pi| / |asinpi(x)| < x^2/6 < 2^-54. |
| 45 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 46 | return x * ASINPI_COEFFS[0]; |
| 47 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 48 | } |
| 49 | |
| 50 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 51 | return x * asinpi_eval(x * x); |
| 52 | #else |
| 53 | using DFloat128 = fputil::DyadicFloat<128>; |
| 54 | using DoubleDouble = fputil::DoubleDouble; |
| 55 | |
| 56 | // For |x| < 2^-511, x^2 would underflow to subnormal, raising a |
| 57 | // spurious underflow exception. Since asinpi(x) = x/pi with correction |
| 58 | // x^2/(6*pi) < 2^-1024 relative (negligible), compute x/pi directly |
| 59 | // in DFloat128. |
| 60 | if (LIBC_UNLIKELY(x_exp < 512)) { |
| 61 | DFloat128 x_f128(x); |
| 62 | DFloat128 r = fputil::quick_mul(a: x_f128, b: ONE_OVER_PI_F128); |
| 63 | double result = static_cast<double>(r); |
| 64 | |
| 65 | // IEEE 754 "after rounding" tininess: the 53-bit unlimited-exponent |
| 66 | // result is strictly between +-2^-1022. DyadicFloat's conversion |
| 67 | // checks the *IEEE subnormal* result (52-bit at the boundary), not |
| 68 | // the 53-bit unlimited-exponent result, so we detect it here. |
| 69 | int exp_hi = r.exponent + 127 + FPBits::EXP_BIAS; |
| 70 | if (LIBC_UNLIKELY(exp_hi <= 0) && !r.mantissa.is_zero()) { |
| 71 | bool raise_underflow = true; |
| 72 | // When exp_hi == 0, a carry in 53-bit rounding can push the |
| 73 | // result to exactly 2^-1022 (not tiny). Check for this. |
| 74 | if (exp_hi == 0) { |
| 75 | constexpr unsigned SHIFT_53 = 128 - FPBits::SIG_LEN - 1; |
| 76 | using MantT = typename DFloat128::MantissaType; |
| 77 | MantT m53 = r.mantissa >> SHIFT_53; |
| 78 | constexpr MantT ALL_ONES_53 = (MantT(1) << (FPBits::SIG_LEN + 1)) - 1; |
| 79 | if (m53 == ALL_ONES_53) { |
| 80 | // All 53 bits set. carry happens if rounding rounds away |
| 81 | // from zero at this precision. |
| 82 | bool round_bit = |
| 83 | static_cast<bool>((r.mantissa >> (SHIFT_53 - 1)) & 1); |
| 84 | MantT sticky_mask = (MantT(1) << (SHIFT_53 - 1)) - 1; |
| 85 | bool sticky = (r.mantissa & sticky_mask) != 0; |
| 86 | bool lsb = static_cast<bool>(m53 & 1); |
| 87 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 88 | // Carry if round_bit && (lsb || sticky) (round half to even). |
| 89 | raise_underflow = !(round_bit && (lsb || sticky)); |
| 90 | #else |
| 91 | switch (fputil::quick_get_round()) { |
| 92 | case FE_TONEAREST: |
| 93 | // Carry if round_bit && (lsb || sticky) (round half to even). |
| 94 | raise_underflow = !(round_bit && (lsb || sticky)); |
| 95 | break; |
| 96 | case FE_UPWARD: |
| 97 | raise_underflow = xbits.is_neg() || !(round_bit || sticky); |
| 98 | break; |
| 99 | case FE_DOWNWARD: |
| 100 | raise_underflow = !xbits.is_neg() || !(round_bit || sticky); |
| 101 | break; |
| 102 | case FE_TOWARDZERO: |
| 103 | default: |
| 104 | raise_underflow = true; // truncation never carries |
| 105 | break; |
| 106 | } |
| 107 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 108 | } |
| 109 | } |
| 110 | if (raise_underflow) |
| 111 | fputil::raise_except_if_required(FE_UNDERFLOW | FE_INEXACT); |
| 112 | } |
| 113 | return result; |
| 114 | } |
| 115 | |
| 116 | unsigned idx = 0; |
| 117 | DoubleDouble x_sq = fputil::exact_mult(a: x, b: x); |
| 118 | double err = xbits.abs().get_val() * 0x1.0p-51; |
| 119 | // Polynomial approximation: |
| 120 | // p ~ asin(x)/(pi*x) |
| 121 | |
| 122 | DoubleDouble p = asinpi_eval(u: x_sq, idx, err); |
| 123 | // asinpi(x) ~ x * p |
| 124 | DoubleDouble r0 = fputil::exact_mult(a: x, b: p.hi); |
| 125 | double r_lo = fputil::multiply_add(x, y: p.lo, z: r0.lo); |
| 126 | |
| 127 | // Ziv's accuracy test. |
| 128 | double r_upper = r0.hi + (r_lo + err); |
| 129 | double r_lower = r0.hi + (r_lo - err); |
| 130 | |
| 131 | if (LIBC_LIKELY(r_upper == r_lower)) |
| 132 | return r_upper; |
| 133 | |
| 134 | // Ziv's accuracy test failed, perform 128-bit calculation. |
| 135 | |
| 136 | // Recalculate mod 1/64. |
| 137 | idx = static_cast<unsigned>(fputil::nearest_integer(x: x_sq.hi * 0x1.0p6)); |
| 138 | |
| 139 | DFloat128 x_f128(x); |
| 140 | |
| 141 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 142 | DFloat128 u_hi( |
| 143 | fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, x_sq.hi)); |
| 144 | DFloat128 u = fputil::quick_add(u_hi, DFloat128(x_sq.lo)); |
| 145 | #else |
| 146 | DFloat128 x_sq_f128 = fputil::quick_mul(a: x_f128, b: x_f128); |
| 147 | DFloat128 u = fputil::quick_add( |
| 148 | a: x_sq_f128, b: DFloat128(static_cast<double>(idx) * (-0x1.0p-6))); |
| 149 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 150 | |
| 151 | DFloat128 p_f128 = asinpi_eval(u, idx); |
| 152 | DFloat128 r = fputil::quick_mul(a: x_f128, b: p_f128); |
| 153 | |
| 154 | return static_cast<double>(r); |
| 155 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 156 | } |
| 157 | // |x| >= 0.5 |
| 158 | |
| 159 | double x_abs = xbits.abs().get_val(); |
| 160 | |
| 161 | // Maintaining the sign: |
| 162 | constexpr double SIGN[2] = {1.0, -1.0}; |
| 163 | double x_sign = SIGN[xbits.is_neg()]; |
| 164 | |
| 165 | // |x| >= 1 |
| 166 | if (LIBC_UNLIKELY(x_exp >= FPBits::EXP_BIAS)) { |
| 167 | // x = +-1, asinpi(x) = +- 0.5 |
| 168 | if (x_abs == 1.0) { |
| 169 | return x_sign * 0.5; |
| 170 | } |
| 171 | // |x| > 1, return NaN. |
| 172 | if (xbits.is_quiet_nan()) |
| 173 | return x; |
| 174 | |
| 175 | // Set domain error for non-NaN input. |
| 176 | if (!xbits.is_nan()) |
| 177 | fputil::set_errno_if_required(EDOM); |
| 178 | |
| 179 | fputil::raise_except_if_required(FE_INVALID); |
| 180 | return FPBits::quiet_nan().get_val(); |
| 181 | } |
| 182 | |
| 183 | // When |x| >= 0.5, we perform range reduction as follow: |
| 184 | // |
| 185 | // Assume further that 0.5 <= x < 1, and let: |
| 186 | // y = asin(x) |
| 187 | // Using the identity: |
| 188 | // asin(x) = pi/2 - 2 * asin( sqrt( (1 - x)/2 ) ) |
| 189 | // We get: |
| 190 | // asinpi(x) = asin(x)/pi = 0.5 - 2 * asin(sqrt(u)) / pi |
| 191 | // = 0.5 - 2 * sqrt(u) * [asin(sqrt(u)) / (pi * sqrt(u))] |
| 192 | // = 0.5 - 2 * sqrt(u) * asinpi_eval(u) |
| 193 | // where u = (1 - |x|) / 2. |
| 194 | |
| 195 | // u = (1 - |x|)/2 |
| 196 | double u = fputil::multiply_add(x: x_abs, y: -0.5, z: 0.5); |
| 197 | // v_hi ~ sqrt(u). |
| 198 | double v_hi = fputil::sqrt<double>(x: u); |
| 199 | |
| 200 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 201 | double p = asinpi_eval(u); |
| 202 | double r = x_sign * fputil::multiply_add(-2.0 * v_hi, p, 0.5); |
| 203 | return r; |
| 204 | #else |
| 205 | using DFloat128 = fputil::DyadicFloat<128>; |
| 206 | using DoubleDouble = fputil::DoubleDouble; |
| 207 | |
| 208 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 209 | double h = fputil::multiply_add(v_hi, -v_hi, u); |
| 210 | #else |
| 211 | DoubleDouble v_hi_sq = fputil::exact_mult(a: v_hi, b: v_hi); |
| 212 | double h = (u - v_hi_sq.hi) - v_hi_sq.lo; |
| 213 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 214 | |
| 215 | // Scale v_lo and v_hi by 2 from the formula: |
| 216 | // vh = v_hi * 2 |
| 217 | // vl = 2*v_lo = h / v_hi. |
| 218 | double vh = v_hi * 2.0; |
| 219 | double vl = h / v_hi; |
| 220 | |
| 221 | // Polynomial approximation: |
| 222 | // p ~ asin(sqrt(u))/(pi*sqrt(u)) |
| 223 | unsigned idx = 0; |
| 224 | double err = vh * 0x1.0p-51; |
| 225 | |
| 226 | DoubleDouble p = asinpi_eval(u: DoubleDouble{.lo: 0.0, .hi: u}, idx, err); |
| 227 | |
| 228 | // Perform computations in double-double arithmetic: |
| 229 | // asinpi(x) = 0.5 - (vh + vl) * p |
| 230 | DoubleDouble r0 = fputil::quick_mult(a: DoubleDouble{.lo: vl, .hi: vh}, b: p); |
| 231 | DoubleDouble r = fputil::exact_add(a: 0.5, b: -r0.hi); |
| 232 | |
| 233 | double r_lo = -r0.lo + r.lo; |
| 234 | |
| 235 | // Ziv's accuracy test. |
| 236 | |
| 237 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 238 | double r_upper = fputil::multiply_add( |
| 239 | r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, err)); |
| 240 | double r_lower = fputil::multiply_add( |
| 241 | r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, -err)); |
| 242 | #else |
| 243 | r_lo *= x_sign; |
| 244 | r.hi *= x_sign; |
| 245 | double r_upper = r.hi + (r_lo + err); |
| 246 | double r_lower = r.hi + (r_lo - err); |
| 247 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 248 | |
| 249 | if (LIBC_LIKELY(r_upper == r_lower)) |
| 250 | return r_upper; |
| 251 | |
| 252 | // Ziv's accuracy test failed, we redo the computations in DFloat128. |
| 253 | // Recalculate mod 1/64. |
| 254 | idx = static_cast<unsigned>(fputil::nearest_integer(x: u * 0x1.0p6)); |
| 255 | |
| 256 | // After the first step of Newton-Raphson approximating v = sqrt(u): |
| 257 | // sqrt(u) = v_hi + h / (sqrt(u) + v_hi) |
| 258 | // v_lo = h / (2 * v_hi) |
| 259 | // Add second-order correction: |
| 260 | // v_ll = -v_lo * (h / (4u)) |
| 261 | |
| 262 | // Get the rounding error of vl = 2 * v_lo ~ h / vh |
| 263 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 264 | double vl_lo = fputil::multiply_add(-v_hi, vl, h) / v_hi; |
| 265 | #else |
| 266 | DoubleDouble vh_vl = fputil::exact_mult(a: v_hi, b: vl); |
| 267 | double vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi; |
| 268 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 269 | // vll = 2*v_ll = -vl * (h / (4u)). |
| 270 | double t = h * (-0.25) / u; |
| 271 | double vll = fputil::multiply_add(x: vl, y: t, z: vl_lo); |
| 272 | // m_v = -(v_hi + v_lo + v_ll). |
| 273 | DFloat128 m_v = fputil::quick_add( |
| 274 | a: DFloat128(vh), b: fputil::quick_add(a: DFloat128(vl), b: DFloat128(vll))); |
| 275 | m_v.sign = Sign::NEG; |
| 276 | |
| 277 | // Perform computations in DFloat128: |
| 278 | // asinpi(x) = 0.5 - (v_hi + v_lo + vll) * P_pi(u). |
| 279 | DFloat128 y_f128( |
| 280 | fputil::multiply_add(x: static_cast<double>(idx), y: -0x1.0p-6, z: u)); |
| 281 | |
| 282 | DFloat128 p_f128 = asinpi_eval(u: y_f128, idx); |
| 283 | DFloat128 r0_f128 = fputil::quick_mul(a: m_v, b: p_f128); |
| 284 | DFloat128 r_f128 = fputil::quick_add(a: HALF_F128, b: r0_f128); |
| 285 | |
| 286 | if (xbits.is_neg()) |
| 287 | r_f128.sign = Sign::NEG; |
| 288 | |
| 289 | return static_cast<double>(r_f128); |
| 290 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 291 | } |
| 292 | |
| 293 | } // namespace math |
| 294 | |
| 295 | } // namespace LIBC_NAMESPACE_DECL |
| 296 | |
| 297 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ASINPI_H |
| 298 | |