blob: c09e58042294edeb4b6458a3a152b2d5961f6f51 [file] [log] [blame]
// 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
#[cfg(feature = "alloc")] use core::slice;
#[cfg(feature = "alloc")] use alloc::vec::{self, Vec};
// BTreeMap is not as fast in tests, but better than nothing.
#[cfg(all(feature = "alloc", not(feature = "std")))]
use alloc::collections::BTreeSet;
#[cfg(feature = "std")] use std::collections::HashSet;
#[cfg(feature = "alloc")]
use crate::distributions::{uniform::SampleUniform, Distribution, Uniform, WeightedError};
use crate::Rng;
#[cfg(feature = "serde1")]
use serde::{Serialize, Deserialize};
/// A vector of indices.
///
/// Multiple internal representations are possible.
#[derive(Clone, Debug)]
#[cfg_attr(feature = "serde1", 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 Item = usize;
type IntoIter = IndexVecIntoIter;
/// 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(ref v1), &U32(ref v2)) => v1 == v2,
(&USize(ref v1), &USize(ref v2)) => v1 == v2,
(&U32(ref v1), &USize(ref v2)) => {
(v1.len() == v2.len()) && (v1.iter().zip(v2.iter()).all(|(x, y)| *x as usize == *y))
}
(&USize(ref v1), &U32(ref 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`.
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 > (::core::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 = if length < 500_000 { 0 } else { 1 };
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 = if length < 500_000 { 0 } else { 1 };
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.
///
/// This implementation uses `O(length + amount)` space and `O(length)` time
/// if the "nightly" feature is enabled, or `O(length)` space and
/// `O(length + amount * log length)` time otherwise.
///
/// Panics if `amount > length`.
pub fn sample_weighted<R, F, X>(
rng: &mut R, length: usize, weight: F, amount: usize,
) -> Result<IndexVec, WeightedError>
where
R: Rng + ?Sized,
F: Fn(usize) -> X,
X: Into<f64>,
{
if length > (core::u32::MAX as usize) {
sample_efraimidis_spirakis(rng, length, weight, amount)
} else {
assert!(amount <= core::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 if the
/// "nightly" feature is enabled, or `O(length)` space and `O(length
/// + amount * log length)` time otherwise.
///
/// Panics if `amount > length`.
fn sample_efraimidis_spirakis<R, F, X, N>(
rng: &mut R, length: N, weight: F, amount: N,
) -> Result<IndexVec, WeightedError>
where
R: Rng + ?Sized,
F: Fn(usize) -> X,
X: Into<f64>,
N: UInt,
IndexVec: From<Vec<N>>,
{
if amount == N::zero() {
return Ok(IndexVec::U32(Vec::new()));
}
if amount > length {
panic!("`amount` of samples must be less than or equal to `length`");
}
struct Element<N> {
index: N,
key: f64,
}
impl<N> PartialOrd for Element<N> {
fn partial_cmp(&self, other: &Self) -> Option<core::cmp::Ordering> {
self.key.partial_cmp(&other.key)
}
}
impl<N> Ord for Element<N> {
fn cmp(&self, other: &Self) -> core::cmp::Ordering {
// partial_cmp will always produce a value,
// because we check that the weights are not nan
self.partial_cmp(other).unwrap()
}
}
impl<N> PartialEq for Element<N> {
fn eq(&self, other: &Self) -> bool {
self.key == other.key
}
}
impl<N> Eq for Element<N> {}
#[cfg(feature = "nightly")]
{
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.) {
return Err(WeightedError::InvalidWeight);
}
let key = rng.gen::<f64>().powf(1.0 / weight);
candidates.push(Element { index, key });
index += N::one();
}
// 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(length.as_usize() - 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))
}
#[cfg(not(feature = "nightly"))]
{
#[cfg(all(feature = "alloc", not(feature = "std")))]
use crate::alloc::collections::BinaryHeap;
#[cfg(feature = "std")]
use std::collections::BinaryHeap;
// Partially sort the array such that the `amount` elements with the largest
// keys are first using a binary max heap.
let mut candidates = BinaryHeap::with_capacity(length.as_usize());
let mut index = N::zero();
while index < length {
let weight = weight(index.as_usize()).into();
if !(weight >= 0.) {
return Err(WeightedError::InvalidWeight);
}
let key = rng.gen::<f64>().powf(1.0 / weight);
candidates.push(Element { index, key });
index += N::one();
}
let mut result: Vec<N> = Vec::with_capacity(amount.as_usize());
while result.len() < amount.as_usize() {
result.push(candidates.pop().unwrap().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 {
// For small amount we use Floyd's fully-shuffled variant. For larger
// amounts this is slow due to Vec::insert performance, so we shuffle
// afterwards. Benchmarks show little overhead from extra logic.
let floyd_shuffle = amount < 50;
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 floyd_shuffle {
if let Some(pos) = indices.iter().position(|&x| x == t) {
indices.insert(pos, j);
continue;
}
} else if indices.contains(&t) {
indices.push(j);
continue;
}
indices.push(t);
}
if !floyd_shuffle {
// Reimplement SliceRandom::shuffle with smaller indices
for i in (1..amount).rev() {
// invariant: elements with index > i have been locked in place.
indices.swap(i as usize, rng.gen_range(0..=i) as usize);
}
}
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
+ core::hash::Hash + core::ops::AddAssign {
fn zero() -> Self;
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);
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::*;
#[test]
#[cfg(feature = "serde1")]
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`")}
}
}
#[cfg(feature = "alloc")] use alloc::vec;
#[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);
}
#[test]
fn test_sample_weighted() {
let seed_rng = crate::test::rng;
for &(amount, len) in &[(0, 10), (5, 10), (10, 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();
indices.dedup();
assert_eq!(indices.len(), amount);
for &i in &indices {
assert!((i as usize) < len);
}
},
IndexVec::USize(_) => panic!("expected `IndexVec::U32`"),
}
}
}
#[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, &[8, 0, 3, 5, 9, 6]); // floyd
do_test(25, 10, &[18, 15, 14, 9, 0, 13, 5, 24]); // floyd
do_test(300, 8, &[30, 283, 150, 1, 73, 13, 285, 35]); // floyd
do_test(300, 80, &[31, 289, 248, 154, 5, 78, 19, 286]); // inplace
do_test(300, 180, &[31, 289, 248, 154, 5, 78, 19, 286]); // inplace
do_test(1000_000, 8, &[
103717, 963485, 826422, 509101, 736394, 807035, 5327, 632573,
]); // floyd
do_test(1000_000, 180, &[
103718, 963490, 826426, 509103, 736396, 807036, 5327, 632573,
]); // rejection
}
}