| 1 | //! Slow, fallback algorithm for cases the Eisel-Lemire algorithm cannot round. |
| 2 | |
| 3 | use crate::num::dec2flt::common::BiasedFp; |
| 4 | use crate::num::dec2flt::decimal_seq::{DecimalSeq, parse_decimal_seq}; |
| 5 | use crate::num::dec2flt::float::RawFloat; |
| 6 | |
| 7 | /// Parse the significant digits and biased, binary exponent of a float. |
| 8 | /// |
| 9 | /// This is a fallback algorithm that uses a big-integer representation |
| 10 | /// of the float, and therefore is considerably slower than faster |
| 11 | /// approximations. However, it will always determine how to round |
| 12 | /// the significant digits to the nearest machine float, allowing |
| 13 | /// use to handle near half-way cases. |
| 14 | /// |
| 15 | /// Near half-way cases are halfway between two consecutive machine floats. |
| 16 | /// For example, the float `16777217.0` has a bitwise representation of |
| 17 | /// `100000000000000000000000 1`. Rounding to a single-precision float, |
| 18 | /// the trailing `1` is truncated. Using round-nearest, tie-even, any |
| 19 | /// value above `16777217.0` must be rounded up to `16777218.0`, while |
| 20 | /// any value before or equal to `16777217.0` must be rounded down |
| 21 | /// to `16777216.0`. These near-halfway conversions therefore may require |
| 22 | /// a large number of digits to unambiguously determine how to round. |
| 23 | /// |
| 24 | /// The algorithms described here are based on "Processing Long Numbers Quickly", |
| 25 | /// available here: <https://arxiv.org/pdf/2101.11408.pdf#section.11>. |
| 26 | pub(crate) fn parse_long_mantissa<F: RawFloat>(s: &[u8]) -> BiasedFp { |
| 27 | const MAX_SHIFT: usize = 60; |
| 28 | const NUM_POWERS: usize = 19; |
| 29 | const POWERS: [u8; 19] = |
| 30 | [0, 3, 6, 9, 13, 16, 19, 23, 26, 29, 33, 36, 39, 43, 46, 49, 53, 56, 59]; |
| 31 | |
| 32 | let get_shift = |n| { |
| 33 | if n < NUM_POWERS { POWERS[n] as usize } else { MAX_SHIFT } |
| 34 | }; |
| 35 | |
| 36 | let fp_zero = BiasedFp::zero_pow2(0); |
| 37 | let fp_inf = BiasedFp::zero_pow2(F::INFINITE_POWER); |
| 38 | |
| 39 | let mut d = parse_decimal_seq(s); |
| 40 | |
| 41 | // Short-circuit if the value can only be a literal 0 or infinity. |
| 42 | if d.num_digits == 0 || d.decimal_point < -324 { |
| 43 | return fp_zero; |
| 44 | } else if d.decimal_point >= 310 { |
| 45 | return fp_inf; |
| 46 | } |
| 47 | let mut exp2 = 0_i32; |
| 48 | // Shift right toward (1/2 ... 1]. |
| 49 | while d.decimal_point > 0 { |
| 50 | let n = d.decimal_point as usize; |
| 51 | let shift = get_shift(n); |
| 52 | d.right_shift(shift); |
| 53 | if d.decimal_point < -DecimalSeq::DECIMAL_POINT_RANGE { |
| 54 | return fp_zero; |
| 55 | } |
| 56 | exp2 += shift as i32; |
| 57 | } |
| 58 | // Shift left toward (1/2 ... 1]. |
| 59 | while d.decimal_point <= 0 { |
| 60 | let shift = if d.decimal_point == 0 { |
| 61 | match d.digits[0] { |
| 62 | digit if digit >= 5 => break, |
| 63 | 0 | 1 => 2, |
| 64 | _ => 1, |
| 65 | } |
| 66 | } else { |
| 67 | get_shift((-d.decimal_point) as _) |
| 68 | }; |
| 69 | d.left_shift(shift); |
| 70 | if d.decimal_point > DecimalSeq::DECIMAL_POINT_RANGE { |
| 71 | return fp_inf; |
| 72 | } |
| 73 | exp2 -= shift as i32; |
| 74 | } |
| 75 | // We are now in the range [1/2 ... 1] but the binary format uses [1 ... 2]. |
| 76 | exp2 -= 1; |
| 77 | while F::EXP_MIN > exp2 { |
| 78 | let mut n = (F::EXP_MIN - exp2) as usize; |
| 79 | if n > MAX_SHIFT { |
| 80 | n = MAX_SHIFT; |
| 81 | } |
| 82 | d.right_shift(n); |
| 83 | exp2 += n as i32; |
| 84 | } |
| 85 | if (exp2 - F::EXP_MIN + 1) >= F::INFINITE_POWER { |
| 86 | return fp_inf; |
| 87 | } |
| 88 | // Shift the decimal to the hidden bit, and then round the value |
| 89 | // to get the high mantissa+1 bits. |
| 90 | d.left_shift(F::SIG_BITS as usize + 1); |
| 91 | let mut mantissa = d.round(); |
| 92 | if mantissa >= (1_u64 << (F::SIG_BITS + 1)) { |
| 93 | // Rounding up overflowed to the carry bit, need to |
| 94 | // shift back to the hidden bit. |
| 95 | d.right_shift(1); |
| 96 | exp2 += 1; |
| 97 | mantissa = d.round(); |
| 98 | if (exp2 - F::EXP_MIN + 1) >= F::INFINITE_POWER { |
| 99 | return fp_inf; |
| 100 | } |
| 101 | } |
| 102 | let mut power2 = exp2 - F::EXP_MIN + 1; |
| 103 | if mantissa < (1_u64 << F::SIG_BITS) { |
| 104 | power2 -= 1; |
| 105 | } |
| 106 | // Zero out all the bits above the explicit mantissa bits. |
| 107 | mantissa &= (1_u64 << F::SIG_BITS) - 1; |
| 108 | BiasedFp { m: mantissa, p_biased: power2 } |
| 109 | } |
| 110 | |