| 1 | // Copyright 2018-2023 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 | use crate::RngCore; |
| 10 | |
| 11 | pub(crate) struct CoinFlipper<R: RngCore> { |
| 12 | pub rng: R, |
| 13 | chunk: u32, // TODO(opt): this should depend on RNG word size |
| 14 | chunk_remaining: u32, |
| 15 | } |
| 16 | |
| 17 | impl<R: RngCore> CoinFlipper<R> { |
| 18 | pub fn new(rng: R) -> Self { |
| 19 | Self { |
| 20 | rng, |
| 21 | chunk: 0, |
| 22 | chunk_remaining: 0, |
| 23 | } |
| 24 | } |
| 25 | |
| 26 | #[inline ] |
| 27 | /// Returns true with a probability of 1 / d |
| 28 | /// Uses an expected two bits of randomness |
| 29 | /// Panics if d == 0 |
| 30 | pub fn random_ratio_one_over(&mut self, d: usize) -> bool { |
| 31 | debug_assert_ne!(d, 0); |
| 32 | // This uses the same logic as `random_ratio` but is optimized for the case that |
| 33 | // the starting numerator is one (which it always is for `Sequence::Choose()`) |
| 34 | |
| 35 | // In this case (but not `random_ratio`), this way of calculating c is always accurate |
| 36 | let c = (usize::BITS - 1 - d.leading_zeros()).min(32); |
| 37 | |
| 38 | if self.flip_c_heads(c) { |
| 39 | let numerator = 1 << c; |
| 40 | self.random_ratio(numerator, d) |
| 41 | } else { |
| 42 | false |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | #[inline ] |
| 47 | /// Returns true with a probability of n / d |
| 48 | /// Uses an expected two bits of randomness |
| 49 | fn random_ratio(&mut self, mut n: usize, d: usize) -> bool { |
| 50 | // Explanation: |
| 51 | // We are trying to return true with a probability of n / d |
| 52 | // If n >= d, we can just return true |
| 53 | // Otherwise there are two possibilities 2n < d and 2n >= d |
| 54 | // In either case we flip a coin. |
| 55 | // If 2n < d |
| 56 | // If it comes up tails, return false |
| 57 | // If it comes up heads, double n and start again |
| 58 | // This is fair because (0.5 * 0) + (0.5 * 2n / d) = n / d and 2n is less than d |
| 59 | // (if 2n was greater than d we would effectively round it down to 1 |
| 60 | // by returning true) |
| 61 | // If 2n >= d |
| 62 | // If it comes up tails, set n to 2n - d and start again |
| 63 | // If it comes up heads, return true |
| 64 | // This is fair because (0.5 * 1) + (0.5 * (2n - d) / d) = n / d |
| 65 | // Note that if 2n = d and the coin comes up tails, n will be set to 0 |
| 66 | // before restarting which is equivalent to returning false. |
| 67 | |
| 68 | // As a performance optimization we can flip multiple coins at once |
| 69 | // This is efficient because we can use the `lzcnt` intrinsic |
| 70 | // We can check up to 32 flips at once but we only receive one bit of information |
| 71 | // - all heads or at least one tail. |
| 72 | |
| 73 | // Let c be the number of coins to flip. 1 <= c <= 32 |
| 74 | // If 2n < d, n * 2^c < d |
| 75 | // If the result is all heads, then set n to n * 2^c |
| 76 | // If there was at least one tail, return false |
| 77 | // If 2n >= d, the order of results matters so we flip one coin at a time so c = 1 |
| 78 | // Ideally, c will be as high as possible within these constraints |
| 79 | |
| 80 | while n < d { |
| 81 | // Find a good value for c by counting leading zeros |
| 82 | // This will either give the highest possible c, or 1 less than that |
| 83 | let c = n |
| 84 | .leading_zeros() |
| 85 | .saturating_sub(d.leading_zeros() + 1) |
| 86 | .clamp(1, 32); |
| 87 | |
| 88 | if self.flip_c_heads(c) { |
| 89 | // All heads |
| 90 | // Set n to n * 2^c |
| 91 | // If 2n >= d, the while loop will exit and we will return `true` |
| 92 | // If n * 2^c > `usize::MAX` we always return `true` anyway |
| 93 | n = n.saturating_mul(2_usize.pow(c)); |
| 94 | } else { |
| 95 | // At least one tail |
| 96 | if c == 1 { |
| 97 | // Calculate 2n - d. |
| 98 | // We need to use wrapping as 2n might be greater than `usize::MAX` |
| 99 | let next_n = n.wrapping_add(n).wrapping_sub(d); |
| 100 | if next_n == 0 || next_n > n { |
| 101 | // This will happen if 2n < d |
| 102 | return false; |
| 103 | } |
| 104 | n = next_n; |
| 105 | } else { |
| 106 | // c > 1 so 2n < d so we can return false |
| 107 | return false; |
| 108 | } |
| 109 | } |
| 110 | } |
| 111 | true |
| 112 | } |
| 113 | |
| 114 | /// If the next `c` bits of randomness all represent heads, consume them, return true |
| 115 | /// Otherwise return false and consume the number of heads plus one. |
| 116 | /// Generates new bits of randomness when necessary (in 32 bit chunks) |
| 117 | /// Has a 1 in 2 to the `c` chance of returning true |
| 118 | /// `c` must be less than or equal to 32 |
| 119 | fn flip_c_heads(&mut self, mut c: u32) -> bool { |
| 120 | debug_assert!(c <= 32); |
| 121 | // Note that zeros on the left of the chunk represent heads. |
| 122 | // It needs to be this way round because zeros are filled in when left shifting |
| 123 | loop { |
| 124 | let zeros = self.chunk.leading_zeros(); |
| 125 | |
| 126 | if zeros < c { |
| 127 | // The happy path - we found a 1 and can return false |
| 128 | // Note that because a 1 bit was detected, |
| 129 | // We cannot have run out of random bits so we don't need to check |
| 130 | |
| 131 | // First consume all of the bits read |
| 132 | // Using shl seems to give worse performance for size-hinted iterators |
| 133 | self.chunk = self.chunk.wrapping_shl(zeros + 1); |
| 134 | |
| 135 | self.chunk_remaining = self.chunk_remaining.saturating_sub(zeros + 1); |
| 136 | return false; |
| 137 | } else { |
| 138 | // The number of zeros is larger than `c` |
| 139 | // There are two possibilities |
| 140 | if let Some(new_remaining) = self.chunk_remaining.checked_sub(c) { |
| 141 | // Those zeroes were all part of our random chunk, |
| 142 | // throw away `c` bits of randomness and return true |
| 143 | self.chunk_remaining = new_remaining; |
| 144 | self.chunk <<= c; |
| 145 | return true; |
| 146 | } else { |
| 147 | // Some of those zeroes were part of the random chunk |
| 148 | // and some were part of the space behind it |
| 149 | // We need to take into account only the zeroes that were random |
| 150 | c -= self.chunk_remaining; |
| 151 | |
| 152 | // Generate a new chunk |
| 153 | self.chunk = self.rng.next_u32(); |
| 154 | self.chunk_remaining = 32; |
| 155 | // Go back to start of loop |
| 156 | } |
| 157 | } |
| 158 | } |
| 159 | } |
| 160 | } |
| 161 | |