| 1 | //===-- Implementation header for hypotf ------------------------*- 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_HYPOTF_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_HYPOTF_H |
| 11 | |
| 12 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 13 | #include "src/__support/FPUtil/FPBits.h" |
| 14 | #include "src/__support/FPUtil/double_double.h" |
| 15 | #include "src/__support/FPUtil/multiply_add.h" |
| 16 | #include "src/__support/FPUtil/sqrt.h" |
| 17 | #include "src/__support/common.h" |
| 18 | #include "src/__support/macros/config.h" |
| 19 | #include "src/__support/macros/optimization.h" |
| 20 | |
| 21 | namespace LIBC_NAMESPACE_DECL { |
| 22 | |
| 23 | namespace math { |
| 24 | |
| 25 | LIBC_INLINE float hypotf(float x, float y) { |
| 26 | using DoubleBits = fputil::FPBits<double>; |
| 27 | using FPBits = fputil::FPBits<float>; |
| 28 | using fputil::DoubleDouble; |
| 29 | |
| 30 | uint32_t x_a = FPBits(x).uintval() & 0x7fff'ffff; |
| 31 | uint32_t y_a = FPBits(y).uintval() & 0x7fff'ffff; |
| 32 | |
| 33 | // Note: replacing `x_a >= FPBits::EXP_MASK` with `x_bits.is_inf_or_nan()` |
| 34 | // generates extra exponent bit masking instructions on x86-64. |
| 35 | if (LIBC_UNLIKELY(x_a >= FPBits::EXP_MASK || y_a >= FPBits::EXP_MASK)) { |
| 36 | // x or y is inf or nan |
| 37 | FPBits x_bits(x); |
| 38 | FPBits y_bits(y); |
| 39 | if (x_bits.is_signaling_nan() || y_bits.is_signaling_nan()) { |
| 40 | fputil::raise_except_if_required(FE_INVALID); |
| 41 | return FPBits::quiet_nan().get_val(); |
| 42 | } |
| 43 | if (x_bits.is_inf() || y_bits.is_inf()) |
| 44 | return FPBits::inf().get_val(); |
| 45 | return x + y; |
| 46 | } |
| 47 | |
| 48 | double xd = static_cast<double>(x); |
| 49 | double yd = static_cast<double>(y); |
| 50 | |
| 51 | // x^2 and y^2 are exact in double precision. |
| 52 | double x_sq = xd * xd; |
| 53 | |
| 54 | double sum_sq; |
| 55 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 56 | sum_sq = fputil::multiply_add(yd, yd, x_sq); |
| 57 | #else |
| 58 | double y_sq = yd * yd; |
| 59 | sum_sq = x_sq + y_sq; |
| 60 | #endif |
| 61 | |
| 62 | // Take sqrt in double precision. |
| 63 | DoubleBits result(fputil::sqrt<double>(x: sum_sq)); |
| 64 | double r = result.get_val(); |
| 65 | float r_f = static_cast<float>(r); |
| 66 | |
| 67 | // If any of the sticky bits of the result are non-zero, except the LSB, then |
| 68 | // the rounded result is correct. |
| 69 | uint64_t r_u = result.uintval(); |
| 70 | uint32_t r_u32 = static_cast<uint32_t>(r_u); |
| 71 | |
| 72 | if (LIBC_UNLIKELY(((r_u32 + 1) & 0x0FFF'FFFE) == 0)) { |
| 73 | // Almost all the sticky bits of the results are non-zero, extra checks are |
| 74 | // needed to make sure rounding is correct. |
| 75 | |
| 76 | // Perform a quick check to see if the result rounded to float is already |
| 77 | // correct. Majority of hard-to-round cases fall in this case. If not, we |
| 78 | // will need to perform more expensive computations to get the correct error |
| 79 | // terms. |
| 80 | double r_d = static_cast<double>(r_f); |
| 81 | bool y_a_smaller = y_a < x_a; |
| 82 | |
| 83 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 84 | // Compute the missing y_sq variable for FMA code path. |
| 85 | double y_sq = yd * yd; |
| 86 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 87 | |
| 88 | double a = y_a_smaller ? x_sq : y_sq; |
| 89 | double b = y_a_smaller ? y_sq : x_sq; |
| 90 | double e = b - fputil::multiply_add(x: r_d, y: r_d, z: -a); |
| 91 | if (e == 0.0) |
| 92 | return r_f; |
| 93 | |
| 94 | // Rounding correction is needed. |
| 95 | // The errors come from two parts: |
| 96 | // - rounding errors from sqrt(sum_sq) -> D(sum_sq) |
| 97 | // - rounding errors from x_sq + y_sq -> sum_sq |
| 98 | // We use FastTwoSum algorithm to compute those errors and then combine. |
| 99 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 100 | double sum_sq_lo = (b - (sum_sq - a)); |
| 101 | double err = sum_sq_lo - fputil::multiply_add(r, r, -sum_sq); |
| 102 | #else |
| 103 | fputil::DoubleDouble r_sq = fputil::exact_mult(a: r, b: r); |
| 104 | double sum_sq_lo = (b - (sum_sq - a)); |
| 105 | double err = (sum_sq - r_sq.hi) + (sum_sq_lo - r_sq.lo); |
| 106 | #endif |
| 107 | |
| 108 | if (err > 0) { |
| 109 | r_u |= 1; |
| 110 | } else if ((err < 0) && ((r_u32 & 0x0FFF'FFFF) == 0)) { |
| 111 | r_u -= 1; |
| 112 | } |
| 113 | |
| 114 | return static_cast<float>(DoubleBits(r_u).get_val()); |
| 115 | } |
| 116 | |
| 117 | return r_f; |
| 118 | } |
| 119 | |
| 120 | } // namespace math |
| 121 | |
| 122 | } // namespace LIBC_NAMESPACE_DECL |
| 123 | |
| 124 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_HYPOTF_H |
| 125 | |