| 1 |
//! An implementation of Clinger's Bellerophon algorithm. |
| 2 |
//! |
| 3 |
//! This is a moderate path algorithm that uses an extended-precision |
| 4 |
//! float, represented in 80 bits, by calculating the bits of slop |
| 5 |
//! and determining if those bits could prevent unambiguous rounding. |
| 6 |
//! |
| 7 |
//! This algorithm requires less static storage than the Lemire algorithm, |
| 8 |
//! and has decent performance, and is therefore used when non-decimal, |
| 9 |
//! non-power-of-two strings need to be parsed. Clinger's algorithm |
| 10 |
//! is described in depth in "How to Read Floating Point Numbers Accurately.", |
| 11 |
//! available online [here](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.45.4152&rep=rep1&type=pdf). |
| 12 |
//! |
| 13 |
//! This implementation is loosely based off the Golang implementation, |
| 14 |
//! found [here](https://github.com/golang/go/blob/b10849fbb97a2244c086991b4623ae9f32c212d0/src/strconv/extfloat.go). |
| 15 |
//! This code is therefore subject to a 3-clause BSD license. |
| 16 |
|
| 17 |
#![cfg(feature = "compact")] |
| 18 |
#![doc(hidden)] |
| 19 |
|
| 20 |
use crate::extended_float::ExtendedFloat; |
| 21 |
use crate::mask::{lower_n_halfway, lower_n_mask}; |
| 22 |
use crate::num::Float; |
| 23 |
use crate::number::Number; |
| 24 |
use crate::rounding::{round, round_nearest_tie_even}; |
| 25 |
use crate::table::BASE10_POWERS; |
| 26 |
|
| 27 |
// ALGORITHM |
| 28 |
// --------- |
| 29 |
|
| 30 |
/// Core implementation of the Bellerophon algorithm. |
| 31 |
/// |
| 32 |
/// Create an extended-precision float, scale it to the proper radix power, |
| 33 |
/// calculate the bits of slop, and return the representation. The value |
| 34 |
/// will always be guaranteed to be within 1 bit, rounded-down, of the real |
| 35 |
/// value. If a negative exponent is returned, this represents we were |
| 36 |
/// unable to unambiguously round the significant digits. |
| 37 |
/// |
| 38 |
/// This has been modified to return a biased, rather than unbiased exponent. |
| 39 |
pub fn bellerophon<F: Float>(num: &Number) -> ExtendedFloat { |
| 40 |
let fp_zero = ExtendedFloat { |
| 41 |
mant: 0, |
| 42 |
exp: 0, |
| 43 |
}; |
| 44 |
let fp_inf = ExtendedFloat { |
| 45 |
mant: 0, |
| 46 |
exp: F::INFINITE_POWER, |
| 47 |
}; |
| 48 |
|
| 49 |
// Early short-circuit, in case of literal 0 or infinity. |
| 50 |
// This allows us to avoid narrow casts causing numeric overflow, |
| 51 |
// and is a quick check for any radix. |
| 52 |
if num.mantissa == 0 || num.exponent <= -0x1000 { |
| 53 |
return fp_zero; |
| 54 |
} else if num.exponent >= 0x1000 { |
| 55 |
return fp_inf; |
| 56 |
} |
| 57 |
|
| 58 |
// Calculate our indexes for our extended-precision multiplication. |
| 59 |
// This narrowing cast is safe, since exponent must be in a valid range. |
| 60 |
let exponent = num.exponent as i32 + BASE10_POWERS.bias; |
| 61 |
let small_index = exponent % BASE10_POWERS.step; |
| 62 |
let large_index = exponent / BASE10_POWERS.step; |
| 63 |
|
| 64 |
if exponent < 0 { |
| 65 |
// Guaranteed underflow (assign 0). |
| 66 |
return fp_zero; |
| 67 |
} |
| 68 |
if large_index as usize >= BASE10_POWERS.large.len() { |
| 69 |
// Overflow (assign infinity) |
| 70 |
return fp_inf; |
| 71 |
} |
| 72 |
|
| 73 |
// Within the valid exponent range, multiply by the large and small |
| 74 |
// exponents and return the resulting value. |
| 75 |
|
| 76 |
// Track errors to as a factor of unit in last-precision. |
| 77 |
let mut errors: u32 = 0; |
| 78 |
if num.many_digits { |
| 79 |
errors += error_halfscale(); |
| 80 |
} |
| 81 |
|
| 82 |
// Multiply by the small power. |
| 83 |
// Check if we can directly multiply by an integer, if not, |
| 84 |
// use extended-precision multiplication. |
| 85 |
let mut fp = ExtendedFloat { |
| 86 |
mant: num.mantissa, |
| 87 |
exp: 0, |
| 88 |
}; |
| 89 |
match fp.mant.overflowing_mul(BASE10_POWERS.get_small_int(small_index as usize)) { |
| 90 |
// Overflow, multiplication unsuccessful, go slow path. |
| 91 |
(_, true) => { |
| 92 |
normalize(&mut fp); |
| 93 |
fp = mul(&fp, &BASE10_POWERS.get_small(small_index as usize)); |
| 94 |
errors += error_halfscale(); |
| 95 |
}, |
| 96 |
// No overflow, multiplication successful. |
| 97 |
(mant, false) => { |
| 98 |
fp.mant = mant; |
| 99 |
normalize(&mut fp); |
| 100 |
}, |
| 101 |
} |
| 102 |
|
| 103 |
// Multiply by the large power. |
| 104 |
fp = mul(&fp, &BASE10_POWERS.get_large(large_index as usize)); |
| 105 |
if errors > 0 { |
| 106 |
errors += 1; |
| 107 |
} |
| 108 |
errors += error_halfscale(); |
| 109 |
|
| 110 |
// Normalize the floating point (and the errors). |
| 111 |
let shift = normalize(&mut fp); |
| 112 |
errors <<= shift; |
| 113 |
fp.exp += F::EXPONENT_BIAS; |
| 114 |
|
| 115 |
// Check for literal overflow, even with halfway cases. |
| 116 |
if -fp.exp + 1 > 65 { |
| 117 |
return fp_zero; |
| 118 |
} |
| 119 |
|
| 120 |
// Too many errors accumulated, return an error. |
| 121 |
if !error_is_accurate::<F>(errors, &fp) { |
| 122 |
// Bias the exponent so we know it's invalid. |
| 123 |
fp.exp += F::INVALID_FP; |
| 124 |
return fp; |
| 125 |
} |
| 126 |
|
| 127 |
// Check if we have a literal 0 or overflow here. |
| 128 |
// If we have an exponent of -63, we can still have a valid shift, |
| 129 |
// giving a case where we have too many errors and need to round-up. |
| 130 |
if -fp.exp + 1 == 65 { |
| 131 |
// Have more than 64 bits below the minimum exponent, must be 0. |
| 132 |
return fp_zero; |
| 133 |
} |
| 134 |
|
| 135 |
round::<F, _>(&mut fp, |f, s| { |
| 136 |
round_nearest_tie_even(f, s, |is_odd, is_halfway, is_above| { |
| 137 |
is_above || (is_odd && is_halfway) |
| 138 |
}); |
| 139 |
}); |
| 140 |
fp |
| 141 |
} |
| 142 |
|
| 143 |
// ERRORS |
| 144 |
// ------ |
| 145 |
|
| 146 |
// Calculate if the errors in calculating the extended-precision float. |
| 147 |
// |
| 148 |
// Specifically, we want to know if we are close to a halfway representation, |
| 149 |
// or halfway between `b` and `b+1`, or `b+h`. The halfway representation |
| 150 |
// has the form: |
| 151 |
// SEEEEEEEHMMMMMMMMMMMMMMMMMMMMMMM100... |
| 152 |
// where: |
| 153 |
// S = Sign Bit |
| 154 |
// E = Exponent Bits |
| 155 |
// H = Hidden Bit |
| 156 |
// M = Mantissa Bits |
| 157 |
// |
| 158 |
// The halfway representation has a bit set 1-after the mantissa digits, |
| 159 |
// and no bits set immediately afterward, making it impossible to |
| 160 |
// round between `b` and `b+1` with this representation. |
| 161 |
|
| 162 |
/// Get the full error scale. |
| 163 |
#[inline(always)] |
| 164 |
const fn error_scale() -> u32 { |
| 165 |
8 |
| 166 |
} |
| 167 |
|
| 168 |
/// Get the half error scale. |
| 169 |
#[inline(always)] |
| 170 |
const fn error_halfscale() -> u32 { |
| 171 |
error_scale() / 2 |
| 172 |
} |
| 173 |
|
| 174 |
/// Determine if the number of errors is tolerable for float precision. |
| 175 |
fn error_is_accurate<F: Float>(errors: u32, fp: &ExtendedFloat) -> bool { |
| 176 |
// Check we can't have a literal 0 denormal float. |
| 177 |
debug_assert!(fp.exp >= -64); |
| 178 |
|
| 179 |
// Determine if extended-precision float is a good approximation. |
| 180 |
// If the error has affected too many units, the float will be |
| 181 |
// inaccurate, or if the representation is too close to halfway |
| 182 |
// that any operations could affect this halfway representation. |
| 183 |
// See the documentation for dtoa for more information. |
| 184 |
|
| 185 |
// This is always a valid u32, since `fp.exp >= -64` |
| 186 |
// will always be positive and the significand size is {23, 52}. |
| 187 |
let mantissa_shift = 64 - F::MANTISSA_SIZE - 1; |
| 188 |
|
| 189 |
// The unbiased exponent checks is `unbiased_exp <= F::MANTISSA_SIZE |
| 190 |
// - F::EXPONENT_BIAS -64 + 1`, or `biased_exp <= F::MANTISSA_SIZE - 63`, |
| 191 |
// or `biased_exp <= mantissa_shift`. |
| 192 |
let extrabits = match fp.exp <= -mantissa_shift { |
| 193 |
// Denormal, since shifting to the hidden bit still has a negative exponent. |
| 194 |
// The unbiased check calculation for bits is `1 - F::EXPONENT_BIAS - unbiased_exp`, |
| 195 |
// or `1 - biased_exp`. |
| 196 |
true => 1 - fp.exp, |
| 197 |
false => 64 - F::MANTISSA_SIZE - 1, |
| 198 |
}; |
| 199 |
|
| 200 |
// Our logic is as follows: we want to determine if the actual |
| 201 |
// mantissa and the errors during calculation differ significantly |
| 202 |
// from the rounding point. The rounding point for round-nearest |
| 203 |
// is the halfway point, IE, this when the truncated bits start |
| 204 |
// with b1000..., while the rounding point for the round-toward |
| 205 |
// is when the truncated bits are equal to 0. |
| 206 |
// To do so, we can check whether the rounding point +/- the error |
| 207 |
// are >/< the actual lower n bits. |
| 208 |
// |
| 209 |
// For whether we need to use signed or unsigned types for this |
| 210 |
// analysis, see this example, using u8 rather than u64 to simplify |
| 211 |
// things. |
| 212 |
// |
| 213 |
// # Comparisons |
| 214 |
// cmp1 = (halfway - errors) < extra |
| 215 |
// cmp1 = extra < (halfway + errors) |
| 216 |
// |
| 217 |
// # Large Extrabits, Low Errors |
| 218 |
// |
| 219 |
// extrabits = 8 |
| 220 |
// halfway = 0b10000000 |
| 221 |
// extra = 0b10000010 |
| 222 |
// errors = 0b00000100 |
| 223 |
// halfway - errors = 0b01111100 |
| 224 |
// halfway + errors = 0b10000100 |
| 225 |
// |
| 226 |
// Unsigned: |
| 227 |
// halfway - errors = 124 |
| 228 |
// halfway + errors = 132 |
| 229 |
// extra = 130 |
| 230 |
// cmp1 = true |
| 231 |
// cmp2 = true |
| 232 |
// Signed: |
| 233 |
// halfway - errors = 124 |
| 234 |
// halfway + errors = -124 |
| 235 |
// extra = -126 |
| 236 |
// cmp1 = false |
| 237 |
// cmp2 = true |
| 238 |
// |
| 239 |
// # Conclusion |
| 240 |
// |
| 241 |
// Since errors will always be small, and since we want to detect |
| 242 |
// if the representation is accurate, we need to use an **unsigned** |
| 243 |
// type for comparisons. |
| 244 |
let maskbits = extrabits as u64; |
| 245 |
let errors = errors as u64; |
| 246 |
|
| 247 |
// Round-to-nearest, need to use the halfway point. |
| 248 |
if extrabits > 64 { |
| 249 |
// Underflow, we have a shift larger than the mantissa. |
| 250 |
// Representation is valid **only** if the value is close enough |
| 251 |
// overflow to the next bit within errors. If it overflows, |
| 252 |
// the representation is **not** valid. |
| 253 |
!fp.mant.overflowing_add(errors).1 |
| 254 |
} else { |
| 255 |
let mask = lower_n_mask(maskbits); |
| 256 |
let extra = fp.mant & mask; |
| 257 |
|
| 258 |
// Round-to-nearest, need to check if we're close to halfway. |
| 259 |
// IE, b10100 | 100000, where `|` signifies the truncation point. |
| 260 |
let halfway = lower_n_halfway(maskbits); |
| 261 |
let cmp1 = halfway.wrapping_sub(errors) < extra; |
| 262 |
let cmp2 = extra < halfway.wrapping_add(errors); |
| 263 |
|
| 264 |
// If both comparisons are true, we have significant rounding error, |
| 265 |
// and the value cannot be exactly represented. Otherwise, the |
| 266 |
// representation is valid. |
| 267 |
!(cmp1 && cmp2) |
| 268 |
} |
| 269 |
} |
| 270 |
|
| 271 |
// MATH |
| 272 |
// ---- |
| 273 |
|
| 274 |
/// Normalize float-point number. |
| 275 |
/// |
| 276 |
/// Shift the mantissa so the number of leading zeros is 0, or the value |
| 277 |
/// itself is 0. |
| 278 |
/// |
| 279 |
/// Get the number of bytes shifted. |
| 280 |
pub fn normalize(fp: &mut ExtendedFloat) -> i32 { |
| 281 |
// Note: |
| 282 |
// Using the ctlz intrinsic via leading_zeros is way faster (~10x) |
| 283 |
// than shifting 1-bit at a time, via while loop, and also way |
| 284 |
// faster (~2x) than an unrolled loop that checks at 32, 16, 4, |
| 285 |
// 2, and 1 bit. |
| 286 |
// |
| 287 |
// Using a modulus of pow2 (which will get optimized to a bitwise |
| 288 |
// and with 0x3F or faster) is slightly slower than an if/then, |
| 289 |
// however, removing the if/then will likely optimize more branched |
| 290 |
// code as it removes conditional logic. |
| 291 |
|
| 292 |
// Calculate the number of leading zeros, and then zero-out |
| 293 |
// any overflowing bits, to avoid shl overflow when self.mant == 0. |
| 294 |
if fp.mant != 0 { |
| 295 |
let shift = fp.mant.leading_zeros() as i32; |
| 296 |
fp.mant <<= shift; |
| 297 |
fp.exp -= shift; |
| 298 |
shift |
| 299 |
} else { |
| 300 |
0 |
| 301 |
} |
| 302 |
} |
| 303 |
|
| 304 |
/// Multiply two normalized extended-precision floats, as if by `a*b`. |
| 305 |
/// |
| 306 |
/// The precision is maximal when the numbers are normalized, however, |
| 307 |
/// decent precision will occur as long as both values have high bits |
| 308 |
/// set. The result is not normalized. |
| 309 |
/// |
| 310 |
/// Algorithm: |
| 311 |
/// 1. Non-signed multiplication of mantissas (requires 2x as many bits as input). |
| 312 |
/// 2. Normalization of the result (not done here). |
| 313 |
/// 3. Addition of exponents. |
| 314 |
pub fn mul(x: &ExtendedFloat, y: &ExtendedFloat) -> ExtendedFloat { |
| 315 |
// Logic check, values must be decently normalized prior to multiplication. |
| 316 |
debug_assert!(x.mant >> 32 != 0); |
| 317 |
debug_assert!(y.mant >> 32 != 0); |
| 318 |
|
| 319 |
// Extract high-and-low masks. |
| 320 |
// Mask is u32::MAX for older Rustc versions. |
| 321 |
const LOMASK: u64 = 0xffff_ffff; |
| 322 |
let x1 = x.mant >> 32; |
| 323 |
let x0 = x.mant & LOMASK; |
| 324 |
let y1 = y.mant >> 32; |
| 325 |
let y0 = y.mant & LOMASK; |
| 326 |
|
| 327 |
// Get our products |
| 328 |
let x1_y0 = x1 * y0; |
| 329 |
let x0_y1 = x0 * y1; |
| 330 |
let x0_y0 = x0 * y0; |
| 331 |
let x1_y1 = x1 * y1; |
| 332 |
|
| 333 |
let mut tmp = (x1_y0 & LOMASK) + (x0_y1 & LOMASK) + (x0_y0 >> 32); |
| 334 |
// round up |
| 335 |
tmp += 1 << (32 - 1); |
| 336 |
|
| 337 |
ExtendedFloat { |
| 338 |
mant: x1_y1 + (x1_y0 >> 32) + (x0_y1 >> 32) + (tmp >> 32), |
| 339 |
exp: x.exp + y.exp + 64, |
| 340 |
} |
| 341 |
} |
| 342 |
|
| 343 |
// POWERS |
| 344 |
// ------ |
| 345 |
|
| 346 |
/// Precalculated powers of base N for the Bellerophon algorithm. |
| 347 |
pub struct BellerophonPowers { |
| 348 |
// Pre-calculated small powers. |
| 349 |
pub small: &'static [u64], |
| 350 |
// Pre-calculated large powers. |
| 351 |
pub large: &'static [u64], |
| 352 |
/// Pre-calculated small powers as 64-bit integers |
| 353 |
pub small_int: &'static [u64], |
| 354 |
// Step between large powers and number of small powers. |
| 355 |
pub step: i32, |
| 356 |
// Exponent bias for the large powers. |
| 357 |
pub bias: i32, |
| 358 |
/// ceil(log2(radix)) scaled as a multiplier. |
| 359 |
pub log2: i64, |
| 360 |
/// Bitshift for the log2 multiplier. |
| 361 |
pub log2_shift: i32, |
| 362 |
} |
| 363 |
|
| 364 |
/// Allow indexing of values without bounds checking |
| 365 |
impl BellerophonPowers { |
| 366 |
#[inline] |
| 367 |
pub fn get_small(&self, index: usize) -> ExtendedFloat { |
| 368 |
let mant = self.small[index]; |
| 369 |
let exp = (1 - 64) + ((self.log2 * index as i64) >> self.log2_shift); |
| 370 |
ExtendedFloat { |
| 371 |
mant, |
| 372 |
exp: exp as i32, |
| 373 |
} |
| 374 |
} |
| 375 |
|
| 376 |
#[inline] |
| 377 |
pub fn get_large(&self, index: usize) -> ExtendedFloat { |
| 378 |
let mant = self.large[index]; |
| 379 |
let biased_e = index as i64 * self.step as i64 - self.bias as i64; |
| 380 |
let exp = (1 - 64) + ((self.log2 * biased_e) >> self.log2_shift); |
| 381 |
ExtendedFloat { |
| 382 |
mant, |
| 383 |
exp: exp as i32, |
| 384 |
} |
| 385 |
} |
| 386 |
|
| 387 |
#[inline] |
| 388 |
pub fn get_small_int(&self, index: usize) -> u64 { |
| 389 |
self.small_int[index] |
| 390 |
} |
| 391 |
} |