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};
getNumDims() const15 // BTreeMap is not as fast in tests, but better than nothing.
16 #[cfg(feature="std")] use std::collections::{HashSet};
getNumInputs() const17 #[cfg(all(feature="alloc", not(feature="std")))] use alloc::collections::BTreeSet;
18 
19 #[cfg(feature="alloc")] use distributions::{Distribution, Uniform};
20 use Rng;
getNumConstraints() const21 
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
getNumInequalities() const33     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.)
getConstraints() const44     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.
52     pub fn into_vec(self) -> Vec<usize> {
53         match self {
getEqFlags() const54             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
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     }
walkExprs(function_ref<void (AffineExpr)> callback) const66 
67     /// Convert into an iterator over the indices as a sequence of `usize` values
68     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 {
77     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 {
91     fn from(v: Vec<u32>) -> Self {
92         IndexVec::U32(v)
93     }
94 }
95 
96 impl From<Vec<usize>> for IndexVec {
97     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;
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 
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 
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 
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`.
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.
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.
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.
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]
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]
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