1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
// Copyright 2018 Developers of the Rand project.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//! Low-level API for sampling indices
use alloc::vec::{self, Vec};
use core::slice;
use core::{hash::Hash, ops::AddAssign};
// BTreeMap is not as fast in tests, but better than nothing.
#[cfg(feature = "std")]
use super::WeightError;
use crate::distr::uniform::SampleUniform;
use crate::distr::{Distribution, Uniform};
use crate::Rng;
#[cfg(not(feature = "std"))]
use alloc::collections::BTreeSet;
#[cfg(feature = "serde")]
use serde::{Deserialize, Serialize};
#[cfg(feature = "std")]
use std::collections::HashSet;
/// A vector of indices.
///
/// Multiple internal representations are possible.
#[derive(Clone, Debug)]
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
pub enum IndexVec {
#[doc(hidden)]
U32(Vec<u32>),
#[doc(hidden)]
USize(Vec<usize>),
}
impl IndexVec {
/// Returns the number of indices
#[inline]
pub fn len(&self) -> usize {
match *self {
IndexVec::U32(ref v) => v.len(),
IndexVec::USize(ref v) => v.len(),
}
}
/// Returns `true` if the length is 0.
#[inline]
pub fn is_empty(&self) -> bool {
match *self {
IndexVec::U32(ref v) => v.is_empty(),
IndexVec::USize(ref v) => v.is_empty(),
}
}
/// Return the value at the given `index`.
///
/// (Note: we cannot implement [`std::ops::Index`] because of lifetime
/// restrictions.)
#[inline]
pub fn index(&self, index: usize) -> usize {
match *self {
IndexVec::U32(ref v) => v[index] as usize,
IndexVec::USize(ref v) => v[index],
}
}
/// Return result as a `Vec<usize>`. Conversion may or may not be trivial.
#[inline]
pub fn into_vec(self) -> Vec<usize> {
match self {
IndexVec::U32(v) => v.into_iter().map(|i| i as usize).collect(),
IndexVec::USize(v) => v,
}
}
/// Iterate over the indices as a sequence of `usize` values
#[inline]
pub fn iter(&self) -> IndexVecIter<'_> {
match *self {
IndexVec::U32(ref v) => IndexVecIter::U32(v.iter()),
IndexVec::USize(ref v) => IndexVecIter::USize(v.iter()),
}
}
}
impl IntoIterator for IndexVec {
type IntoIter = IndexVecIntoIter;
type Item = usize;
/// Convert into an iterator over the indices as a sequence of `usize` values
#[inline]
fn into_iter(self) -> IndexVecIntoIter {
match self {
IndexVec::U32(v) => IndexVecIntoIter::U32(v.into_iter()),
IndexVec::USize(v) => IndexVecIntoIter::USize(v.into_iter()),
}
}
}
impl PartialEq for IndexVec {
fn eq(&self, other: &IndexVec) -> bool {
use self::IndexVec::*;
match (self, other) {
(U32(v1), U32(v2)) => v1 == v2,
(USize(v1), USize(v2)) => v1 == v2,
(U32(v1), USize(v2)) => {
(v1.len() == v2.len()) && (v1.iter().zip(v2.iter()).all(|(x, y)| *x as usize == *y))
}
(USize(v1), U32(v2)) => {
(v1.len() == v2.len()) && (v1.iter().zip(v2.iter()).all(|(x, y)| *x == *y as usize))
}
}
}
}
impl From<Vec<u32>> for IndexVec {
#[inline]
fn from(v: Vec<u32>) -> Self {
IndexVec::U32(v)
}
}
impl From<Vec<usize>> for IndexVec {
#[inline]
fn from(v: Vec<usize>) -> Self {
IndexVec::USize(v)
}
}
/// Return type of `IndexVec::iter`.
#[derive(Debug)]
pub enum IndexVecIter<'a> {
#[doc(hidden)]
U32(slice::Iter<'a, u32>),
#[doc(hidden)]
USize(slice::Iter<'a, usize>),
}
impl<'a> Iterator for IndexVecIter<'a> {
type Item = usize;
#[inline]
fn next(&mut self) -> Option<usize> {
use self::IndexVecIter::*;
match *self {
U32(ref mut iter) => iter.next().map(|i| *i as usize),
USize(ref mut iter) => iter.next().cloned(),
}
}
#[inline]
fn size_hint(&self) -> (usize, Option<usize>) {
match *self {
IndexVecIter::U32(ref v) => v.size_hint(),
IndexVecIter::USize(ref v) => v.size_hint(),
}
}
}
impl<'a> ExactSizeIterator for IndexVecIter<'a> {}
/// Return type of `IndexVec::into_iter`.
#[derive(Clone, Debug)]
pub enum IndexVecIntoIter {
#[doc(hidden)]
U32(vec::IntoIter<u32>),
#[doc(hidden)]
USize(vec::IntoIter<usize>),
}
impl Iterator for IndexVecIntoIter {
type Item = usize;
#[inline]
fn next(&mut self) -> Option<Self::Item> {
use self::IndexVecIntoIter::*;
match *self {
U32(ref mut v) => v.next().map(|i| i as usize),
USize(ref mut v) => v.next(),
}
}
#[inline]
fn size_hint(&self) -> (usize, Option<usize>) {
use self::IndexVecIntoIter::*;
match *self {
U32(ref v) => v.size_hint(),
USize(ref v) => v.size_hint(),
}
}
}
impl ExactSizeIterator for IndexVecIntoIter {}
/// Randomly sample exactly `amount` distinct indices from `0..length`, and
/// return them in random order (fully shuffled).
///
/// This method is used internally by the slice sampling methods, but it can
/// sometimes be useful to have the indices themselves so this is provided as
/// an alternative.
///
/// The implementation used is not specified; we automatically select the
/// fastest available algorithm for the `length` and `amount` parameters
/// (based on detailed profiling on an Intel Haswell CPU). Roughly speaking,
/// complexity is `O(amount)`, except that when `amount` is small, performance
/// is closer to `O(amount^2)`, and when `length` is close to `amount` then
/// `O(length)`.
///
/// Note that performance is significantly better over `u32` indices than over
/// `u64` indices. Because of this we hide the underlying type behind an
/// abstraction, `IndexVec`.
///
/// If an allocation-free `no_std` function is required, it is suggested
/// to adapt the internal `sample_floyd` implementation.
///
/// Panics if `amount > length`.
#[track_caller]
pub fn sample<R>(rng: &mut R, length: usize, amount: usize) -> IndexVec
where
R: Rng + ?Sized,
{
if amount > length {
panic!("`amount` of samples must be less than or equal to `length`");
}
if length > (u32::MAX as usize) {
// We never want to use inplace here, but could use floyd's alg
// Lazy version: always use the cache alg.
return sample_rejection(rng, length, amount);
}
let amount = amount as u32;
let length = length as u32;
// Choice of algorithm here depends on both length and amount. See:
// https://github.com/rust-random/rand/pull/479
// We do some calculations with f32. Accuracy is not very important.
if amount < 163 {
const C: [[f32; 2]; 2] = [[1.6, 8.0 / 45.0], [10.0, 70.0 / 9.0]];
let j = usize::from(length >= 500_000);
let amount_fp = amount as f32;
let m4 = C[0][j] * amount_fp;
// Short-cut: when amount < 12, floyd's is always faster
if amount > 11 && (length as f32) < (C[1][j] + m4) * amount_fp {
sample_inplace(rng, length, amount)
} else {
sample_floyd(rng, length, amount)
}
} else {
const C: [f32; 2] = [270.0, 330.0 / 9.0];
let j = usize::from(length >= 500_000);
if (length as f32) < C[j] * (amount as f32) {
sample_inplace(rng, length, amount)
} else {
sample_rejection(rng, length, amount)
}
}
}
/// Randomly sample exactly `amount` distinct indices from `0..length`, and
/// return them in an arbitrary order (there is no guarantee of shuffling or
/// ordering). The weights are to be provided by the input function `weights`,
/// which will be called once for each index.
///
/// This method is used internally by the slice sampling methods, but it can
/// sometimes be useful to have the indices themselves so this is provided as
/// an alternative.
///
/// Error cases:
/// - [`WeightError::InvalidWeight`] when a weight is not-a-number or negative.
/// - [`WeightError::InsufficientNonZero`] when fewer than `amount` weights are positive.
///
/// This implementation uses `O(length + amount)` space and `O(length)` time.
#[cfg(feature = "std")]
pub fn sample_weighted<R, F, X>(
rng: &mut R,
length: usize,
weight: F,
amount: usize,
) -> Result<IndexVec, WeightError>
where
R: Rng + ?Sized,
F: Fn(usize) -> X,
X: Into<f64>,
{
if length > (u32::MAX as usize) {
sample_efraimidis_spirakis(rng, length, weight, amount)
} else {
assert!(amount <= u32::MAX as usize);
let amount = amount as u32;
let length = length as u32;
sample_efraimidis_spirakis(rng, length, weight, amount)
}
}
/// Randomly sample exactly `amount` distinct indices from `0..length`, and
/// return them in an arbitrary order (there is no guarantee of shuffling or
/// ordering). The weights are to be provided by the input function `weights`,
/// which will be called once for each index.
///
/// This implementation uses the algorithm described by Efraimidis and Spirakis
/// in this paper: <https://doi.org/10.1016/j.ipl.2005.11.003>
/// It uses `O(length + amount)` space and `O(length)` time.
///
/// Error cases:
/// - [`WeightError::InvalidWeight`] when a weight is not-a-number or negative.
/// - [`WeightError::InsufficientNonZero`] when fewer than `amount` weights are positive.
#[cfg(feature = "std")]
fn sample_efraimidis_spirakis<R, F, X, N>(
rng: &mut R,
length: N,
weight: F,
amount: N,
) -> Result<IndexVec, WeightError>
where
R: Rng + ?Sized,
F: Fn(usize) -> X,
X: Into<f64>,
N: UInt,
IndexVec: From<Vec<N>>,
{
use std::cmp::Ordering;
if amount == N::zero() {
return Ok(IndexVec::U32(Vec::new()));
}
struct Element<N> {
index: N,
key: f64,
}
impl<N> PartialOrd for Element<N> {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl<N> Ord for Element<N> {
fn cmp(&self, other: &Self) -> Ordering {
// partial_cmp will always produce a value,
// because we check that the weights are not nan
self.key.partial_cmp(&other.key).unwrap()
}
}
impl<N> PartialEq for Element<N> {
fn eq(&self, other: &Self) -> bool {
self.key == other.key
}
}
impl<N> Eq for Element<N> {}
let mut candidates = Vec::with_capacity(length.as_usize());
let mut index = N::zero();
while index < length {
let weight = weight(index.as_usize()).into();
if weight > 0.0 {
let key = rng.random::<f64>().powf(1.0 / weight);
candidates.push(Element { index, key });
} else if !(weight >= 0.0) {
return Err(WeightError::InvalidWeight);
}
index += N::one();
}
let avail = candidates.len();
if avail < amount.as_usize() {
return Err(WeightError::InsufficientNonZero);
}
// Partially sort the array to find the `amount` elements with the greatest
// keys. Do this by using `select_nth_unstable` to put the elements with
// the *smallest* keys at the beginning of the list in `O(n)` time, which
// provides equivalent information about the elements with the *greatest* keys.
let (_, mid, greater) = candidates.select_nth_unstable(avail - amount.as_usize());
let mut result: Vec<N> = Vec::with_capacity(amount.as_usize());
result.push(mid.index);
for element in greater {
result.push(element.index);
}
Ok(IndexVec::from(result))
}
/// Randomly sample exactly `amount` indices from `0..length`, using Floyd's
/// combination algorithm.
///
/// The output values are fully shuffled. (Overhead is under 50%.)
///
/// This implementation uses `O(amount)` memory and `O(amount^2)` time.
fn sample_floyd<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec
where
R: Rng + ?Sized,
{
// Note that the values returned by `rng.gen_range()` can be
// inferred from the returned vector by working backwards from
// the last entry. This bijection proves the algorithm fair.
debug_assert!(amount <= length);
let mut indices = Vec::with_capacity(amount as usize);
for j in length - amount..length {
let t = rng.gen_range(0..=j);
if let Some(pos) = indices.iter().position(|&x| x == t) {
indices[pos] = j;
}
indices.push(t);
}
IndexVec::from(indices)
}
/// Randomly sample exactly `amount` indices from `0..length`, using an inplace
/// partial Fisher-Yates method.
/// Sample an amount of indices using an inplace partial fisher yates method.
///
/// This allocates the entire `length` of indices and randomizes only the first `amount`.
/// It then truncates to `amount` and returns.
///
/// This method is not appropriate for large `length` and potentially uses a lot
/// of memory; because of this we only implement for `u32` index (which improves
/// performance in all cases).
///
/// Set-up is `O(length)` time and memory and shuffling is `O(amount)` time.
fn sample_inplace<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec
where
R: Rng + ?Sized,
{
debug_assert!(amount <= length);
let mut indices: Vec<u32> = Vec::with_capacity(length as usize);
indices.extend(0..length);
for i in 0..amount {
let j: u32 = rng.gen_range(i..length);
indices.swap(i as usize, j as usize);
}
indices.truncate(amount as usize);
debug_assert_eq!(indices.len(), amount as usize);
IndexVec::from(indices)
}
trait UInt: Copy + PartialOrd + Ord + PartialEq + Eq + SampleUniform + Hash + AddAssign {
fn zero() -> Self;
#[cfg_attr(feature = "alloc", allow(dead_code))]
fn one() -> Self;
fn as_usize(self) -> usize;
}
impl UInt for u32 {
#[inline]
fn zero() -> Self {
0
}
#[inline]
fn one() -> Self {
1
}
#[inline]
fn as_usize(self) -> usize {
self as usize
}
}
impl UInt for usize {
#[inline]
fn zero() -> Self {
0
}
#[inline]
fn one() -> Self {
1
}
#[inline]
fn as_usize(self) -> usize {
self
}
}
/// Randomly sample exactly `amount` indices from `0..length`, using rejection
/// sampling.
///
/// Since `amount <<< length` there is a low chance of a random sample in
/// `0..length` being a duplicate. We test for duplicates and resample where
/// necessary. The algorithm is `O(amount)` time and memory.
///
/// This function is generic over X primarily so that results are value-stable
/// over 32-bit and 64-bit platforms.
fn sample_rejection<X: UInt, R>(rng: &mut R, length: X, amount: X) -> IndexVec
where
R: Rng + ?Sized,
IndexVec: From<Vec<X>>,
{
debug_assert!(amount < length);
#[cfg(feature = "std")]
let mut cache = HashSet::with_capacity(amount.as_usize());
#[cfg(not(feature = "std"))]
let mut cache = BTreeSet::new();
let distr = Uniform::new(X::zero(), length).unwrap();
let mut indices = Vec::with_capacity(amount.as_usize());
for _ in 0..amount.as_usize() {
let mut pos = distr.sample(rng);
while !cache.insert(pos) {
pos = distr.sample(rng);
}
indices.push(pos);
}
debug_assert_eq!(indices.len(), amount.as_usize());
IndexVec::from(indices)
}
#[cfg(test)]
mod test {
use super::*;
use alloc::vec;
#[test]
#[cfg(feature = "serde")]
fn test_serialization_index_vec() {
let some_index_vec = IndexVec::from(vec![254_usize, 234, 2, 1]);
let de_some_index_vec: IndexVec =
bincode::deserialize(&bincode::serialize(&some_index_vec).unwrap()).unwrap();
match (some_index_vec, de_some_index_vec) {
(IndexVec::U32(a), IndexVec::U32(b)) => {
assert_eq!(a, b);
}
(IndexVec::USize(a), IndexVec::USize(b)) => {
assert_eq!(a, b);
}
_ => {
panic!("failed to seralize/deserialize `IndexVec`")
}
}
}
#[test]
fn test_sample_boundaries() {
let mut r = crate::test::rng(404);
assert_eq!(sample_inplace(&mut r, 0, 0).len(), 0);
assert_eq!(sample_inplace(&mut r, 1, 0).len(), 0);
assert_eq!(sample_inplace(&mut r, 1, 1).into_vec(), vec![0]);
assert_eq!(sample_rejection(&mut r, 1u32, 0).len(), 0);
assert_eq!(sample_floyd(&mut r, 0, 0).len(), 0);
assert_eq!(sample_floyd(&mut r, 1, 0).len(), 0);
assert_eq!(sample_floyd(&mut r, 1, 1).into_vec(), vec![0]);
// These algorithms should be fast with big numbers. Test average.
let sum: usize = sample_rejection(&mut r, 1 << 25, 10u32).into_iter().sum();
assert!(1 << 25 < sum && sum < (1 << 25) * 25);
let sum: usize = sample_floyd(&mut r, 1 << 25, 10).into_iter().sum();
assert!(1 << 25 < sum && sum < (1 << 25) * 25);
}
#[test]
#[cfg_attr(miri, ignore)] // Miri is too slow
fn test_sample_alg() {
let seed_rng = crate::test::rng;
// We can't test which algorithm is used directly, but Floyd's alg
// should produce different results from the others. (Also, `inplace`
// and `cached` currently use different sizes thus produce different results.)
// A small length and relatively large amount should use inplace
let (length, amount): (usize, usize) = (100, 50);
let v1 = sample(&mut seed_rng(420), length, amount);
let v2 = sample_inplace(&mut seed_rng(420), length as u32, amount as u32);
assert!(v1.iter().all(|e| e < length));
assert_eq!(v1, v2);
// Test Floyd's alg does produce different results
let v3 = sample_floyd(&mut seed_rng(420), length as u32, amount as u32);
assert!(v1 != v3);
// A large length and small amount should use Floyd
let (length, amount): (usize, usize) = (1 << 20, 50);
let v1 = sample(&mut seed_rng(421), length, amount);
let v2 = sample_floyd(&mut seed_rng(421), length as u32, amount as u32);
assert!(v1.iter().all(|e| e < length));
assert_eq!(v1, v2);
// A large length and larger amount should use cache
let (length, amount): (usize, usize) = (1 << 20, 600);
let v1 = sample(&mut seed_rng(422), length, amount);
let v2 = sample_rejection(&mut seed_rng(422), length as u32, amount as u32);
assert!(v1.iter().all(|e| e < length));
assert_eq!(v1, v2);
}
#[cfg(feature = "std")]
#[test]
fn test_sample_weighted() {
let seed_rng = crate::test::rng;
for &(amount, len) in &[(0, 10), (5, 10), (9, 10)] {
let v = sample_weighted(&mut seed_rng(423), len, |i| i as f64, amount).unwrap();
match v {
IndexVec::U32(mut indices) => {
assert_eq!(indices.len(), amount);
indices.sort_unstable();
indices.dedup();
assert_eq!(indices.len(), amount);
for &i in &indices {
assert!((i as usize) < len);
}
}
IndexVec::USize(_) => panic!("expected `IndexVec::U32`"),
}
}
let r = sample_weighted(&mut seed_rng(423), 10, |i| i as f64, 10);
assert_eq!(r.unwrap_err(), WeightError::InsufficientNonZero);
}
#[test]
fn value_stability_sample() {
let do_test = |length, amount, values: &[u32]| {
let mut buf = [0u32; 8];
let mut rng = crate::test::rng(410);
let res = sample(&mut rng, length, amount);
let len = res.len().min(buf.len());
for (x, y) in res.into_iter().zip(buf.iter_mut()) {
*y = x as u32;
}
assert_eq!(
&buf[0..len],
values,
"failed sampling {}, {}",
length,
amount
);
};
do_test(10, 6, &[0, 9, 5, 4, 6, 8]); // floyd
do_test(25, 10, &[24, 20, 19, 9, 22, 16, 0, 14]); // floyd
do_test(300, 8, &[30, 283, 243, 150, 218, 240, 1, 189]); // floyd
do_test(300, 80, &[31, 289, 248, 154, 221, 243, 7, 192]); // inplace
do_test(300, 180, &[31, 289, 248, 154, 221, 243, 7, 192]); // inplace
do_test(
1_000_000,
8,
&[103717, 963485, 826422, 509101, 736394, 807035, 5327, 632573],
); // floyd
do_test(
1_000_000,
180,
&[103718, 963490, 826426, 509103, 736396, 807036, 5327, 632573],
); // rejection
}
}