| 1 |
// Adapted from https://github.com/Alexhuszagh/rust-lexical. |
| 2 |
|
| 3 |
// FLOAT TYPE |
| 4 |
|
| 5 |
use super::num::*; |
| 6 |
use super::rounding::*; |
| 7 |
use super::shift::*; |
| 8 |
|
| 9 |
/// Extended precision floating-point type. |
| 10 |
/// |
| 11 |
/// Private implementation, exposed only for testing purposes. |
| 12 |
#[doc(hidden)] |
| 13 |
#[derive(Clone, Copy, Debug, PartialEq, Eq)] |
| 14 |
pub(crate) struct ExtendedFloat { |
| 15 |
/// Mantissa for the extended-precision float. |
| 16 |
pub mant: u64, |
| 17 |
/// Binary exponent for the extended-precision float. |
| 18 |
pub exp: i32, |
| 19 |
} |
| 20 |
|
| 21 |
impl ExtendedFloat { |
| 22 |
// PROPERTIES |
| 23 |
|
| 24 |
// OPERATIONS |
| 25 |
|
| 26 |
/// Multiply two normalized extended-precision floats, as if by `a*b`. |
| 27 |
/// |
| 28 |
/// The precision is maximal when the numbers are normalized, however, |
| 29 |
/// decent precision will occur as long as both values have high bits |
| 30 |
/// set. The result is not normalized. |
| 31 |
/// |
| 32 |
/// Algorithm: |
| 33 |
/// 1. Non-signed multiplication of mantissas (requires 2x as many bits as input). |
| 34 |
/// 2. Normalization of the result (not done here). |
| 35 |
/// 3. Addition of exponents. |
| 36 |
pub(crate) fn mul(&self, b: &ExtendedFloat) -> ExtendedFloat { |
| 37 |
// Logic check, values must be decently normalized prior to multiplication. |
| 38 |
debug_assert!((self.mant & u64::HIMASK != 0) && (b.mant & u64::HIMASK != 0)); |
| 39 |
|
| 40 |
// Extract high-and-low masks. |
| 41 |
let ah = self.mant >> u64::HALF; |
| 42 |
let al = self.mant & u64::LOMASK; |
| 43 |
let bh = b.mant >> u64::HALF; |
| 44 |
let bl = b.mant & u64::LOMASK; |
| 45 |
|
| 46 |
// Get our products |
| 47 |
let ah_bl = ah * bl; |
| 48 |
let al_bh = al * bh; |
| 49 |
let al_bl = al * bl; |
| 50 |
let ah_bh = ah * bh; |
| 51 |
|
| 52 |
let mut tmp = (ah_bl & u64::LOMASK) + (al_bh & u64::LOMASK) + (al_bl >> u64::HALF); |
| 53 |
// round up |
| 54 |
tmp += 1 << (u64::HALF - 1); |
| 55 |
|
| 56 |
ExtendedFloat { |
| 57 |
mant: ah_bh + (ah_bl >> u64::HALF) + (al_bh >> u64::HALF) + (tmp >> u64::HALF), |
| 58 |
exp: self.exp + b.exp + u64::FULL, |
| 59 |
} |
| 60 |
} |
| 61 |
|
| 62 |
/// Multiply in-place, as if by `a*b`. |
| 63 |
/// |
| 64 |
/// The result is not normalized. |
| 65 |
#[inline] |
| 66 |
pub(crate) fn imul(&mut self, b: &ExtendedFloat) { |
| 67 |
*self = self.mul(b); |
| 68 |
} |
| 69 |
|
| 70 |
// NORMALIZE |
| 71 |
|
| 72 |
/// Normalize float-point number. |
| 73 |
/// |
| 74 |
/// Shift the mantissa so the number of leading zeros is 0, or the value |
| 75 |
/// itself is 0. |
| 76 |
/// |
| 77 |
/// Get the number of bytes shifted. |
| 78 |
#[inline] |
| 79 |
pub(crate) fn normalize(&mut self) -> u32 { |
| 80 |
// Note: |
| 81 |
// Using the cltz intrinsic via leading_zeros is way faster (~10x) |
| 82 |
// than shifting 1-bit at a time, via while loop, and also way |
| 83 |
// faster (~2x) than an unrolled loop that checks at 32, 16, 4, |
| 84 |
// 2, and 1 bit. |
| 85 |
// |
| 86 |
// Using a modulus of pow2 (which will get optimized to a bitwise |
| 87 |
// and with 0x3F or faster) is slightly slower than an if/then, |
| 88 |
// however, removing the if/then will likely optimize more branched |
| 89 |
// code as it removes conditional logic. |
| 90 |
|
| 91 |
// Calculate the number of leading zeros, and then zero-out |
| 92 |
// any overflowing bits, to avoid shl overflow when self.mant == 0. |
| 93 |
let shift = if self.mant == 0 { |
| 94 |
0 |
| 95 |
} else { |
| 96 |
self.mant.leading_zeros() |
| 97 |
}; |
| 98 |
shl(self, shift as i32); |
| 99 |
shift |
| 100 |
} |
| 101 |
|
| 102 |
// ROUND |
| 103 |
|
| 104 |
/// Lossy round float-point number to native mantissa boundaries. |
| 105 |
#[inline] |
| 106 |
pub(crate) fn round_to_native<F, Algorithm>(&mut self, algorithm: Algorithm) |
| 107 |
where |
| 108 |
F: Float, |
| 109 |
Algorithm: FnOnce(&mut ExtendedFloat, i32), |
| 110 |
{ |
| 111 |
round_to_native::<F, _>(self, algorithm); |
| 112 |
} |
| 113 |
|
| 114 |
// FROM |
| 115 |
|
| 116 |
/// Create extended float from native float. |
| 117 |
#[inline] |
| 118 |
pub fn from_float<F: Float>(f: F) -> ExtendedFloat { |
| 119 |
from_float(f) |
| 120 |
} |
| 121 |
|
| 122 |
// INTO |
| 123 |
|
| 124 |
/// Convert into default-rounded, lower-precision native float. |
| 125 |
#[inline] |
| 126 |
pub(crate) fn into_float<F: Float>(mut self) -> F { |
| 127 |
self.round_to_native::<F, _>(round_nearest_tie_even); |
| 128 |
into_float(self) |
| 129 |
} |
| 130 |
|
| 131 |
/// Convert into downward-rounded, lower-precision native float. |
| 132 |
#[inline] |
| 133 |
pub(crate) fn into_downward_float<F: Float>(mut self) -> F { |
| 134 |
self.round_to_native::<F, _>(round_downward); |
| 135 |
into_float(self) |
| 136 |
} |
| 137 |
} |
| 138 |
|
| 139 |
// FROM FLOAT |
| 140 |
|
| 141 |
// Import ExtendedFloat from native float. |
| 142 |
#[inline] |
| 143 |
pub(crate) fn from_float<F>(f: F) -> ExtendedFloat |
| 144 |
where |
| 145 |
F: Float, |
| 146 |
{ |
| 147 |
ExtendedFloat { |
| 148 |
mant: u64::as_cast(f.mantissa()), |
| 149 |
exp: f.exponent(), |
| 150 |
} |
| 151 |
} |
| 152 |
|
| 153 |
// INTO FLOAT |
| 154 |
|
| 155 |
// Export extended-precision float to native float. |
| 156 |
// |
| 157 |
// The extended-precision float must be in native float representation, |
| 158 |
// with overflow/underflow appropriately handled. |
| 159 |
#[inline] |
| 160 |
pub(crate) fn into_float<F>(fp: ExtendedFloat) -> F |
| 161 |
where |
| 162 |
F: Float, |
| 163 |
{ |
| 164 |
// Export floating-point number. |
| 165 |
if fp.mant == 0 || fp.exp < F::DENORMAL_EXPONENT { |
| 166 |
// sub-denormal, underflow |
| 167 |
F::ZERO |
| 168 |
} else if fp.exp >= F::MAX_EXPONENT { |
| 169 |
// overflow |
| 170 |
F::from_bits(F::INFINITY_BITS) |
| 171 |
} else { |
| 172 |
// calculate the exp and fraction bits, and return a float from bits. |
| 173 |
let exp: u64; |
| 174 |
if (fp.exp == F::DENORMAL_EXPONENT) && (fp.mant & F::HIDDEN_BIT_MASK.as_u64()) == 0 { |
| 175 |
exp = 0; |
| 176 |
} else { |
| 177 |
exp = (fp.exp + F::EXPONENT_BIAS) as u64; |
| 178 |
} |
| 179 |
let exp = exp << F::MANTISSA_SIZE; |
| 180 |
let mant = fp.mant & F::MANTISSA_MASK.as_u64(); |
| 181 |
F::from_bits(F::Unsigned::as_cast(mant | exp)) |
| 182 |
} |
| 183 |
} |