1 // Copyright 2018 Developers of the Rand project.
2 //
3 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
4 // https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
5 // <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
6 // option. This file may not be copied, modified, or distributed
7 // except according to those terms.
8
9 //! Index sampling
10
11 #[cfg(feature="alloc")] use core::slice;
12
13 #[cfg(feature="std")] use std::vec;
14 #[cfg(all(feature="alloc", not(feature="std")))] use alloc::vec::{self, Vec};
15 // BTreeMap is not as fast in tests, but better than nothing.
16 #[cfg(feature="std")] use std::collections::{HashSet};
17 #[cfg(all(feature="alloc", not(feature="std")))] use alloc::collections::BTreeSet;
18
19 #[cfg(feature="alloc")] use distributions::{Distribution, Uniform};
20 use Rng;
21
22 /// A vector of indices.
23 ///
24 /// Multiple internal representations are possible.
25 #[derive(Clone, Debug)]
26 pub enum IndexVec {
27 #[doc(hidden)] U32(Vec<u32>),
28 #[doc(hidden)] USize(Vec<usize>),
29 }
30
31 impl IndexVec {
32 /// Returns the number of indices
len(&self) -> usize33 pub fn len(&self) -> usize {
34 match self {
35 &IndexVec::U32(ref v) => v.len(),
36 &IndexVec::USize(ref v) => v.len(),
37 }
38 }
39
40 /// Return the value at the given `index`.
41 ///
42 /// (Note: we cannot implement [`std::ops::Index`] because of lifetime
43 /// restrictions.)
index(&self, index: usize) -> usize44 pub fn index(&self, index: usize) -> usize {
45 match self {
46 &IndexVec::U32(ref v) => v[index] as usize,
47 &IndexVec::USize(ref v) => v[index],
48 }
49 }
50
51 /// Return result as a `Vec<usize>`. Conversion may or may not be trivial.
into_vec(self) -> Vec<usize>52 pub fn into_vec(self) -> Vec<usize> {
53 match self {
54 IndexVec::U32(v) => v.into_iter().map(|i| i as usize).collect(),
55 IndexVec::USize(v) => v,
56 }
57 }
58
59 /// Iterate over the indices as a sequence of `usize` values
iter<'a>(&'a self) -> IndexVecIter<'a>60 pub fn iter<'a>(&'a self) -> IndexVecIter<'a> {
61 match self {
62 &IndexVec::U32(ref v) => IndexVecIter::U32(v.iter()),
63 &IndexVec::USize(ref v) => IndexVecIter::USize(v.iter()),
64 }
65 }
66
67 /// Convert into an iterator over the indices as a sequence of `usize` values
into_iter(self) -> IndexVecIntoIter68 pub fn into_iter(self) -> IndexVecIntoIter {
69 match self {
70 IndexVec::U32(v) => IndexVecIntoIter::U32(v.into_iter()),
71 IndexVec::USize(v) => IndexVecIntoIter::USize(v.into_iter()),
72 }
73 }
74 }
75
76 impl PartialEq for IndexVec {
eq(&self, other: &IndexVec) -> bool77 fn eq(&self, other: &IndexVec) -> bool {
78 use self::IndexVec::*;
79 match (self, other) {
80 (&U32(ref v1), &U32(ref v2)) => v1 == v2,
81 (&USize(ref v1), &USize(ref v2)) => v1 == v2,
82 (&U32(ref v1), &USize(ref v2)) => (v1.len() == v2.len())
83 && (v1.iter().zip(v2.iter()).all(|(x, y)| *x as usize == *y)),
84 (&USize(ref v1), &U32(ref v2)) => (v1.len() == v2.len())
85 && (v1.iter().zip(v2.iter()).all(|(x, y)| *x == *y as usize)),
86 }
87 }
88 }
89
90 impl From<Vec<u32>> for IndexVec {
from(v: Vec<u32>) -> Self91 fn from(v: Vec<u32>) -> Self {
92 IndexVec::U32(v)
93 }
94 }
95
96 impl From<Vec<usize>> for IndexVec {
from(v: Vec<usize>) -> Self97 fn from(v: Vec<usize>) -> Self {
98 IndexVec::USize(v)
99 }
100 }
101
102 /// Return type of `IndexVec::iter`.
103 #[derive(Debug)]
104 pub enum IndexVecIter<'a> {
105 #[doc(hidden)] U32(slice::Iter<'a, u32>),
106 #[doc(hidden)] USize(slice::Iter<'a, usize>),
107 }
108
109 impl<'a> Iterator for IndexVecIter<'a> {
110 type Item = usize;
next(&mut self) -> Option<usize>111 fn next(&mut self) -> Option<usize> {
112 use self::IndexVecIter::*;
113 match self {
114 &mut U32(ref mut iter) => iter.next().map(|i| *i as usize),
115 &mut USize(ref mut iter) => iter.next().cloned(),
116 }
117 }
118
size_hint(&self) -> (usize, Option<usize>)119 fn size_hint(&self) -> (usize, Option<usize>) {
120 match self {
121 &IndexVecIter::U32(ref v) => v.size_hint(),
122 &IndexVecIter::USize(ref v) => v.size_hint(),
123 }
124 }
125 }
126
127 impl<'a> ExactSizeIterator for IndexVecIter<'a> {}
128
129 /// Return type of `IndexVec::into_iter`.
130 #[derive(Clone, Debug)]
131 pub enum IndexVecIntoIter {
132 #[doc(hidden)] U32(vec::IntoIter<u32>),
133 #[doc(hidden)] USize(vec::IntoIter<usize>),
134 }
135
136 impl Iterator for IndexVecIntoIter {
137 type Item = usize;
138
next(&mut self) -> Option<Self::Item>139 fn next(&mut self) -> Option<Self::Item> {
140 use self::IndexVecIntoIter::*;
141 match self {
142 &mut U32(ref mut v) => v.next().map(|i| i as usize),
143 &mut USize(ref mut v) => v.next(),
144 }
145 }
146
size_hint(&self) -> (usize, Option<usize>)147 fn size_hint(&self) -> (usize, Option<usize>) {
148 use self::IndexVecIntoIter::*;
149 match self {
150 &U32(ref v) => v.size_hint(),
151 &USize(ref v) => v.size_hint(),
152 }
153 }
154 }
155
156 impl ExactSizeIterator for IndexVecIntoIter {}
157
158
159 /// Randomly sample exactly `amount` distinct indices from `0..length`, and
160 /// return them in random order (fully shuffled).
161 ///
162 /// This method is used internally by the slice sampling methods, but it can
163 /// sometimes be useful to have the indices themselves so this is provided as
164 /// an alternative.
165 ///
166 /// The implementation used is not specified; we automatically select the
167 /// fastest available algorithm for the `length` and `amount` parameters
168 /// (based on detailed profiling on an Intel Haswell CPU). Roughly speaking,
169 /// complexity is `O(amount)`, except that when `amount` is small, performance
170 /// is closer to `O(amount^2)`, and when `length` is close to `amount` then
171 /// `O(length)`.
172 ///
173 /// Note that performance is significantly better over `u32` indices than over
174 /// `u64` indices. Because of this we hide the underlying type behind an
175 /// abstraction, `IndexVec`.
176 ///
177 /// If an allocation-free `no_std` function is required, it is suggested
178 /// to adapt the internal `sample_floyd` implementation.
179 ///
180 /// Panics if `amount > length`.
sample<R>(rng: &mut R, length: usize, amount: usize) -> IndexVec where R: Rng + ?Sized,181 pub fn sample<R>(rng: &mut R, length: usize, amount: usize) -> IndexVec
182 where R: Rng + ?Sized,
183 {
184 if amount > length {
185 panic!("`amount` of samples must be less than or equal to `length`");
186 }
187 if length > (::core::u32::MAX as usize) {
188 // We never want to use inplace here, but could use floyd's alg
189 // Lazy version: always use the cache alg.
190 return sample_rejection(rng, length, amount);
191 }
192 let amount = amount as u32;
193 let length = length as u32;
194
195 // Choice of algorithm here depends on both length and amount. See:
196 // https://github.com/rust-random/rand/pull/479
197 // We do some calculations with f32. Accuracy is not very important.
198
199 if amount < 163 {
200 const C: [[f32; 2]; 2] = [[1.6, 8.0/45.0], [10.0, 70.0/9.0]];
201 let j = if length < 500_000 { 0 } else { 1 };
202 let amount_fp = amount as f32;
203 let m4 = C[0][j] * amount_fp;
204 // Short-cut: when amount < 12, floyd's is always faster
205 if amount > 11 && (length as f32) < (C[1][j] + m4) * amount_fp {
206 sample_inplace(rng, length, amount)
207 } else {
208 sample_floyd(rng, length, amount)
209 }
210 } else {
211 const C: [f32; 2] = [270.0, 330.0/9.0];
212 let j = if length < 500_000 { 0 } else { 1 };
213 if (length as f32) < C[j] * (amount as f32) {
214 sample_inplace(rng, length, amount)
215 } else {
216 // note: could have a specific u32 impl, but I'm lazy and
217 // generics don't have usable conversions
218 sample_rejection(rng, length as usize, amount as usize)
219 }
220 }
221 }
222
223 /// Randomly sample exactly `amount` indices from `0..length`, using Floyd's
224 /// combination algorithm.
225 ///
226 /// The output values are fully shuffled. (Overhead is under 50%.)
227 ///
228 /// This implementation uses `O(amount)` memory and `O(amount^2)` time.
sample_floyd<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec where R: Rng + ?Sized,229 fn sample_floyd<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec
230 where R: Rng + ?Sized,
231 {
232 // For small amount we use Floyd's fully-shuffled variant. For larger
233 // amounts this is slow due to Vec::insert performance, so we shuffle
234 // afterwards. Benchmarks show little overhead from extra logic.
235 let floyd_shuffle = amount < 50;
236
237 debug_assert!(amount <= length);
238 let mut indices = Vec::with_capacity(amount as usize);
239 for j in length - amount .. length {
240 let t = rng.gen_range(0, j + 1);
241 if floyd_shuffle {
242 if let Some(pos) = indices.iter().position(|&x| x == t) {
243 indices.insert(pos, j);
244 continue;
245 }
246 } else {
247 if indices.contains(&t) {
248 indices.push(j);
249 continue;
250 }
251 }
252 indices.push(t);
253 }
254 if !floyd_shuffle {
255 // Reimplement SliceRandom::shuffle with smaller indices
256 for i in (1..amount).rev() {
257 // invariant: elements with index > i have been locked in place.
258 indices.swap(i as usize, rng.gen_range(0, i + 1) as usize);
259 }
260 }
261 IndexVec::from(indices)
262 }
263
264 /// Randomly sample exactly `amount` indices from `0..length`, using an inplace
265 /// partial Fisher-Yates method.
266 /// Sample an amount of indices using an inplace partial fisher yates method.
267 ///
268 /// This allocates the entire `length` of indices and randomizes only the first `amount`.
269 /// It then truncates to `amount` and returns.
270 ///
271 /// This method is not appropriate for large `length` and potentially uses a lot
272 /// of memory; because of this we only implement for `u32` index (which improves
273 /// performance in all cases).
274 ///
275 /// Set-up is `O(length)` time and memory and shuffling is `O(amount)` time.
sample_inplace<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec where R: Rng + ?Sized,276 fn sample_inplace<R>(rng: &mut R, length: u32, amount: u32) -> IndexVec
277 where R: Rng + ?Sized,
278 {
279 debug_assert!(amount <= length);
280 let mut indices: Vec<u32> = Vec::with_capacity(length as usize);
281 indices.extend(0..length);
282 for i in 0..amount {
283 let j: u32 = rng.gen_range(i, length);
284 indices.swap(i as usize, j as usize);
285 }
286 indices.truncate(amount as usize);
287 debug_assert_eq!(indices.len(), amount as usize);
288 IndexVec::from(indices)
289 }
290
291 /// Randomly sample exactly `amount` indices from `0..length`, using rejection
292 /// sampling.
293 ///
294 /// Since `amount <<< length` there is a low chance of a random sample in
295 /// `0..length` being a duplicate. We test for duplicates and resample where
296 /// necessary. The algorithm is `O(amount)` time and memory.
sample_rejection<R>(rng: &mut R, length: usize, amount: usize) -> IndexVec where R: Rng + ?Sized,297 fn sample_rejection<R>(rng: &mut R, length: usize, amount: usize) -> IndexVec
298 where R: Rng + ?Sized,
299 {
300 debug_assert!(amount < length);
301 #[cfg(feature="std")] let mut cache = HashSet::with_capacity(amount);
302 #[cfg(not(feature="std"))] let mut cache = BTreeSet::new();
303 let distr = Uniform::new(0, length);
304 let mut indices = Vec::with_capacity(amount);
305 for _ in 0..amount {
306 let mut pos = distr.sample(rng);
307 while !cache.insert(pos) {
308 pos = distr.sample(rng);
309 }
310 indices.push(pos);
311 }
312
313 debug_assert_eq!(indices.len(), amount);
314 IndexVec::from(indices)
315 }
316
317 #[cfg(test)]
318 mod test {
319 use super::*;
320
321 #[test]
test_sample_boundaries()322 fn test_sample_boundaries() {
323 let mut r = ::test::rng(404);
324
325 assert_eq!(sample_inplace(&mut r, 0, 0).len(), 0);
326 assert_eq!(sample_inplace(&mut r, 1, 0).len(), 0);
327 assert_eq!(sample_inplace(&mut r, 1, 1).into_vec(), vec![0]);
328
329 assert_eq!(sample_rejection(&mut r, 1, 0).len(), 0);
330
331 assert_eq!(sample_floyd(&mut r, 0, 0).len(), 0);
332 assert_eq!(sample_floyd(&mut r, 1, 0).len(), 0);
333 assert_eq!(sample_floyd(&mut r, 1, 1).into_vec(), vec![0]);
334
335 // These algorithms should be fast with big numbers. Test average.
336 let sum: usize = sample_rejection(&mut r, 1 << 25, 10)
337 .into_iter().sum();
338 assert!(1 << 25 < sum && sum < (1 << 25) * 25);
339
340 let sum: usize = sample_floyd(&mut r, 1 << 25, 10)
341 .into_iter().sum();
342 assert!(1 << 25 < sum && sum < (1 << 25) * 25);
343 }
344
345 #[test]
test_sample_alg()346 fn test_sample_alg() {
347 let seed_rng = ::test::rng;
348
349 // We can't test which algorithm is used directly, but Floyd's alg
350 // should produce different results from the others. (Also, `inplace`
351 // and `cached` currently use different sizes thus produce different results.)
352
353 // A small length and relatively large amount should use inplace
354 let (length, amount): (usize, usize) = (100, 50);
355 let v1 = sample(&mut seed_rng(420), length, amount);
356 let v2 = sample_inplace(&mut seed_rng(420), length as u32, amount as u32);
357 assert!(v1.iter().all(|e| e < length));
358 assert_eq!(v1, v2);
359
360 // Test Floyd's alg does produce different results
361 let v3 = sample_floyd(&mut seed_rng(420), length as u32, amount as u32);
362 assert!(v1 != v3);
363
364 // A large length and small amount should use Floyd
365 let (length, amount): (usize, usize) = (1<<20, 50);
366 let v1 = sample(&mut seed_rng(421), length, amount);
367 let v2 = sample_floyd(&mut seed_rng(421), length as u32, amount as u32);
368 assert!(v1.iter().all(|e| e < length));
369 assert_eq!(v1, v2);
370
371 // A large length and larger amount should use cache
372 let (length, amount): (usize, usize) = (1<<20, 600);
373 let v1 = sample(&mut seed_rng(422), length, amount);
374 let v2 = sample_rejection(&mut seed_rng(422), length, amount);
375 assert!(v1.iter().all(|e| e < length));
376 assert_eq!(v1, v2);
377 }
378 }
379