| 1 | //===-- Implementation header for hypot -------------------------*- 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_HYPOT_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_HYPOT_H |
| 11 | |
| 12 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 13 | #include "src/__support/FPUtil/FPBits.h" |
| 14 | #include "src/__support/FPUtil/Hypot.h" |
| 15 | #include "src/__support/FPUtil/double_double.h" |
| 16 | #include "src/__support/FPUtil/multiply_add.h" |
| 17 | #include "src/__support/FPUtil/sqrt.h" |
| 18 | #include "src/__support/common.h" |
| 19 | #include "src/__support/macros/config.h" |
| 20 | #include "src/__support/macros/optimization.h" |
| 21 | |
| 22 | namespace LIBC_NAMESPACE_DECL { |
| 23 | namespace math { |
| 24 | |
| 25 | LIBC_INLINE double hypot(double x, double y) { |
| 26 | using FPBits = fputil::FPBits<double>; |
| 27 | using DoubleDouble = fputil::DoubleDouble; |
| 28 | |
| 29 | uint64_t x_u = FPBits(x).uintval(); |
| 30 | uint64_t y_u = FPBits(y).uintval(); |
| 31 | |
| 32 | // Shift the exponent field to the top 11 bits of the lower 32-bit. |
| 33 | // Casting it to 32-bit effectively remove the sign bit. |
| 34 | uint32_t x_e = static_cast<uint32_t>(x_u >> 31); |
| 35 | uint32_t y_e = static_cast<uint32_t>(y_u >> 31); |
| 36 | |
| 37 | // a = maximum_mag(x, y); |
| 38 | // b = minimum_mag(x, y); |
| 39 | double a, b; |
| 40 | uint32_t a_e, b_e; |
| 41 | |
| 42 | if (x_e >= y_e) { |
| 43 | a_e = x_e; |
| 44 | b_e = y_e; |
| 45 | a = x; |
| 46 | b = y; |
| 47 | } else { |
| 48 | a_e = y_e; |
| 49 | b_e = x_e; |
| 50 | a = y; |
| 51 | b = x; |
| 52 | } |
| 53 | |
| 54 | double scale = 1.0; |
| 55 | double scale_back = 1.0; |
| 56 | |
| 57 | // For a_e, b_e, the top 11 bits are exponent fields. |
| 58 | if (LIBC_UNLIKELY(a_e >= ((500U + FPBits::EXP_BIAS) << (32 - 11)))) { |
| 59 | // The larger magnitude is above 2^500 (or Inf/NaN), need to scale down to |
| 60 | // prevent overflow when squaring. |
| 61 | if (a_e >= static_cast<uint32_t>(FPBits::EXP_MASK >> 31)) { |
| 62 | // Inf or NaN; |
| 63 | FPBits x_bits(x); |
| 64 | FPBits y_bits(y); |
| 65 | if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) { |
| 66 | fputil::raise_except_if_required(FE_INVALID); |
| 67 | return FPBits::quiet_nan().get_val(); |
| 68 | } |
| 69 | if (x_bits.is_inf() || y_bits.is_inf()) |
| 70 | return FPBits::inf().get_val(); |
| 71 | if (x_bits.is_nan()) |
| 72 | return x; |
| 73 | return y; |
| 74 | } |
| 75 | // Any scaling factor < 2^(-1024/2) = 2^-512 would work. |
| 76 | scale = 0x1.0p-600; |
| 77 | scale_back = 0x1.0p600; |
| 78 | a *= scale; |
| 79 | b *= scale; |
| 80 | } else if (LIBC_UNLIKELY(b_e <= ((FPBits::EXP_BIAS - 500) << (32 - 11)))) { |
| 81 | // The smaller magnitude is below 2^-500 (or 0), need to scale up to prevent |
| 82 | // underflow when squaring. |
| 83 | if ((x == 0.0) || (y == 0.0)) { |
| 84 | double x_abs = FPBits(x_u & FPBits::EXP_SIG_MASK).get_val(); |
| 85 | double y_abs = FPBits(y_u & FPBits::EXP_SIG_MASK).get_val(); |
| 86 | return x_abs + y_abs; |
| 87 | } |
| 88 | // Any scaling factor > 2^((1072 + 52)/2) = 2^562 would work. |
| 89 | scale = 0x1.0p600; |
| 90 | scale_back = 0x1.0p-600; |
| 91 | a *= scale; |
| 92 | b *= scale; |
| 93 | } |
| 94 | |
| 95 | // When the gap in the exponent of `a` and `b` is >= 54, |
| 96 | // |b| < ufp(a) * 2^(-53) = ulp(a)/2 |
| 97 | // So: |
| 98 | // hypot(x, y) = sqrt(a^2 + b^2) |
| 99 | // <= sqrt( (|a| + |b|)^2 ) |
| 100 | // = |a| + |b| |
| 101 | // < |a| + ulp(a) |
| 102 | // Hence, we can return: |
| 103 | // |a| + |b| = |x| + |y| |
| 104 | // to perform correct rounding to all rounding modes. |
| 105 | if (LIBC_UNLIKELY(a_e - b_e >= (54U << (32 - 11)))) { |
| 106 | double x_abs = FPBits(x_u & FPBits::EXP_SIG_MASK).get_val(); |
| 107 | double y_abs = FPBits(y_u & FPBits::EXP_SIG_MASK).get_val(); |
| 108 | return x_abs + y_abs; |
| 109 | } |
| 110 | |
| 111 | // sum.hi + sum.lo ~ a^2 + b^2. |
| 112 | DoubleDouble a_sq = fputil::exact_mult(a, b: a); |
| 113 | DoubleDouble b_sq = fputil::exact_mult(a: b, b); |
| 114 | DoubleDouble sum = fputil::exact_add(a: a_sq.hi, b: b_sq.hi); |
| 115 | sum.lo += a_sq.lo + b_sq.lo; |
| 116 | |
| 117 | // Let hi = sum.hi and lo = sum.lo. |
| 118 | // To compute r_hi + r_lo ~ sqrt(hi + lo): |
| 119 | // - First we use fast sqrt instruction to get: |
| 120 | // r_hi ~ sqrt(hi) |
| 121 | // - Then use Taylor expansion: |
| 122 | // f(hi + lo) = f(hi) + f'(hi) * lo + f''(hi) * lo^2 / 2 + ... |
| 123 | // with f(x) = sqrt(x): |
| 124 | // sqrt(hi + lo) ~ sqrt(hi) + lo / (2 * sqrt(hi)). |
| 125 | // - Subtract by r_hi to find the correction term: |
| 126 | // sqrt(hi + lo) - r_hi ~ (sqrt(hi) - r_hi) + lo / (2 * sqrt(hi)) |
| 127 | // - Instead of finding the rounding errors sqrt(hi) - r_hi, we use the |
| 128 | // squared residual d = hi - r_hi^2, which can be calculated accurately in |
| 129 | // double-double. Then, using the same Taylor approximation of sqrt(x) as |
| 130 | // above: |
| 131 | // sqrt(hi) - r_hi = sqrt(r_hi^2 + d) - r_hi |
| 132 | // ~ sqrt(r_hi^2) + d / (2 * sqrt(r_hi^2)) - r_hi |
| 133 | // = d / (2 * r_hi). |
| 134 | // - Similarly, |
| 135 | // 1 / sqrt(hi) = 1 / sqrt(r_hi^2 + d) |
| 136 | // ~ 1 / sqrt(r_hi^2) - d / (2 * (r_hi^2)^(3/2)) |
| 137 | // = 1 / r_hi - d / (2 * r_hi^3) |
| 138 | // - Putting them together, we have the correction term: |
| 139 | // sqrt(hi + lo) - r_hi + lo / (2 * sqrt(hi)) ~ |
| 140 | // ~ (lo + d) / (2 * r_hi) + lo * d / (4 * r_hi^3) |
| 141 | // ~ (hi + lo - r_hi^2) / (2 * r_hi). |
| 142 | // - When computing hi + lo - r_hi^2, we will pair (hi - r_sq.hi) and |
| 143 | // (lo - r_sq.lo), since `r_sq.hi` is very close to `hi`, and the |
| 144 | // subtraction is exact. |
| 145 | // - Taking intermediate roundings with directed rounding modes into |
| 146 | // consideration, the overall errors should be bounded by |
| 147 | // (2^-51)^2 = 2^-102. |
| 148 | |
| 149 | // |sqrt(sum.hi) - r_hi| < 2^-52. |
| 150 | double r_hi = fputil::sqrt<double>(x: sum.hi); |
| 151 | // r_inv ~ 1 / (2 * r_hi) |
| 152 | double r_inv = 0.5 / r_hi; |
| 153 | // r_hi^2 |
| 154 | DoubleDouble r_sq = fputil::exact_mult(a: r_hi, b: r_hi); |
| 155 | // (hi + lo - r_hi^2) |
| 156 | double num_lo = (sum.lo - r_sq.lo) - (r_sq.hi - sum.hi); |
| 157 | // (hi + lo - r_hi^2) / (2 * r_hi) |
| 158 | double r_lo = num_lo * r_inv; |
| 159 | |
| 160 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 161 | // TODO: What's the worst error if we just do: |
| 162 | // return sqrt(a*a + b*b) * scale_back; |
| 163 | // without all the double-double computations? |
| 164 | return (r_hi + r_lo) * scale_back; |
| 165 | #else |
| 166 | constexpr double ERR = 0x1.0p-102; |
| 167 | |
| 168 | // Ziv's rounding test. |
| 169 | double upper = r_hi + fputil::multiply_add(x: r_hi, y: ERR, z: r_lo); |
| 170 | double lower = r_hi + fputil::multiply_add(x: r_hi, y: -ERR, z: r_lo); |
| 171 | |
| 172 | if (LIBC_LIKELY(upper == lower)) { |
| 173 | return upper * scale_back; |
| 174 | } |
| 175 | |
| 176 | return fputil::hypot(x, y); |
| 177 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 178 | } |
| 179 | |
| 180 | } // namespace math |
| 181 | } // namespace LIBC_NAMESPACE_DECL |
| 182 | |
| 183 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_HYPOT_H |
| 184 | |