| //! A tiny, robust PRNG implementation. |
| //! |
| //! More specifically, it implements a single GOOD PRNG algorithm, |
| //! which is currently a permuted congruential generator. It has two |
| //! implementations, one that returns `u32` and one that returns |
| //! `u64`. It also has functions that return floats or integer |
| //! ranges. And that's it. What more do you need? |
| //! |
| //! For more info on PCG generators, see http://www.pcg-random.org/ |
| //! |
| //! This was designed as a minimalist utility for video games. No |
| //! promises are made about its quality, and if you use it for |
| //! cryptography you will get what you deserve. |
| //! |
| //! Works with `#![no_std]`, has no global state, no dependencies |
| //! apart from some in the unit tests, and is generally neato. |
| |
| #![forbid(unsafe_code)] |
| #![forbid(missing_docs)] |
| #![forbid(missing_debug_implementations)] |
| #![forbid(unused_results)] |
| #![no_std] |
| use core::ops::Range; |
| |
| /// A PRNG producing a 32-bit output. |
| /// |
| /// The current implementation is `PCG-XSH-RR`. |
| #[derive(Copy, Clone, Debug, PartialEq)] |
| pub struct Rand32 { |
| state: u64, |
| inc: u64, |
| } |
| |
| impl Rand32 { |
| /// The default value for `increment`. |
| /// This is basically arbitrary, it comes from the |
| /// PCG reference C implementation: |
| /// https://github.com/imneme/pcg-c/blob/master/include/pcg_variants.h#L284 |
| pub const DEFAULT_INC: u64 = 1442695040888963407; |
| |
| /// This is the number that you have to Really Get Right. |
| /// |
| /// The value used here is from the PCG C implementation: |
| /// https://github.com/imneme/pcg-c/blob/master/include/pcg_variants.h#L278 |
| pub(crate) const MULTIPLIER: u64 = 6364136223846793005; |
| |
| /// Creates a new PRNG with the given seed and a default increment. |
| pub fn new(seed: u64) -> Self { |
| Self::new_inc(seed, Self::DEFAULT_INC) |
| } |
| |
| /// Creates a new PRNG. The two inputs, `seed` and `increment`, |
| /// determine what you get; `increment` basically selects which |
| /// sequence of all those possible the PRNG will produce, and the |
| /// `seed` selects where in that sequence you start. |
| /// |
| /// Both are arbitrary; increment must be an odd number but this |
| /// handles that for you |
| pub fn new_inc(seed: u64, increment: u64) -> Self { |
| let mut rng = Self { |
| state: 0, |
| inc: increment.wrapping_shl(1) | 1, |
| }; |
| // This initialization song-and-dance is a little odd, |
| // but seems to be just how things go. |
| let _ = rng.rand_u32(); |
| rng.state = rng.state.wrapping_add(seed); |
| let _ = rng.rand_u32(); |
| rng |
| } |
| |
| /// Returns the internal state of the PRNG. This allows |
| /// you to save a PRNG and create a new one that will resume |
| /// from the same spot in the sequence. |
| pub fn state(&self) -> (u64, u64) { |
| (self.state, self.inc) |
| } |
| |
| /// Creates a new PRNG from a saved state from `Rand32::state()`. |
| /// This is NOT quite the same as `new_inc()` because `new_inc()` does |
| /// a little extra setup work to initialize the state. |
| pub fn from_state(state: (u64, u64)) -> Self { |
| let (state, inc) = state; |
| Self { state, inc } |
| } |
| |
| /// Produces a random `u32` in the range `[0, u32::MAX]`. |
| pub fn rand_u32(&mut self) -> u32 { |
| let oldstate: u64 = self.state; |
| self.state = oldstate |
| .wrapping_mul(Self::MULTIPLIER) |
| .wrapping_add(self.inc); |
| let xorshifted: u32 = (((oldstate >> 18) ^ oldstate) >> 27) as u32; |
| let rot: u32 = (oldstate >> 59) as u32; |
| xorshifted.rotate_right(rot) |
| } |
| |
| /// Produces a random `i32` in the range `[i32::MIN, i32::MAX]`. |
| pub fn rand_i32(&mut self) -> i32 { |
| self.rand_u32() as i32 |
| } |
| |
| /// Produces a random `f32` in the range `[0.0, 1.0)`. |
| pub fn rand_float(&mut self) -> f32 { |
| // This impl was taken more or less from `rand`, see |
| // <https://docs.rs/rand/0.7.0/src/rand/distributions/float.rs.html#104-117> |
| // There MAY be better ways to do this, see: |
| // https://mumble.net/~campbell/2014/04/28/uniform-random-float |
| // https://mumble.net/~campbell/2014/04/28/random_real.c |
| // https://github.com/Lokathor/randomize/issues/34 |
| const TOTAL_BITS: u32 = 32; |
| const PRECISION: u32 = core::f32::MANTISSA_DIGITS + 1; |
| const MANTISSA_SCALE: f32 = 1.0 / ((1u32 << PRECISION) as f32); |
| let mut u = self.rand_u32(); |
| u >>= TOTAL_BITS - PRECISION; |
| u as f32 * MANTISSA_SCALE |
| } |
| |
| /// Produces a random within the given bounds. Like any `Range`, |
| /// it includes the lower bound and excludes the upper one. |
| /// |
| /// This should be faster than `Self::rand() % end + start`, but the |
| /// real advantage is it's more convenient. Requires that |
| /// `range.end <= range.start`. |
| pub fn rand_range(&mut self, range: Range<u32>) -> u32 { |
| // This is harder to do well than it looks, it seems. I don't |
| // trust Lokathor's implementation 'cause I don't understand |
| // it, so I went to numpy's, which points me to "Lemire's |
| // rejection algorithm": http://arxiv.org/abs/1805.10941 |
| // |
| // Algorithms 3, 4 and 5 in that paper all seem fine modulo |
| // minor performance differences, so this is algorithm 5. |
| // It uses numpy's implementation, `buffered_bounded_lemire_uint32()` |
| |
| debug_assert!(range.start < range.end); |
| let range_starting_from_zero = 0..(range.end - range.start); |
| |
| let s: u32 = range_starting_from_zero.end; |
| let mut m: u64 = u64::from(self.rand_u32()) * u64::from(s); |
| let mut leftover: u32 = (m & 0xFFFF_FFFF) as u32; |
| |
| if leftover < s { |
| // TODO: verify the wrapping_neg() here |
| let threshold: u32 = s.wrapping_neg() % s; |
| while leftover < threshold { |
| m = u64::from(self.rand_u32()).wrapping_mul(u64::from(s)); |
| leftover = (m & 0xFFFF_FFFF) as u32; |
| } |
| } |
| (m >> 32) as u32 + range.start |
| } |
| } |
| |
| /// A PRNG producing a 64-bit output. |
| /// |
| /// The current implementation is `PCG-XSH-RR`. |
| // BUGGO: The recommended algorithm is PCG-XSL-RR? |
| // See https://github.com/imneme/pcg-c/blob/master/include/pcg_variants.h#L2405 |
| // Not sure if it matters? |
| #[derive(Copy, Clone, Debug, PartialEq)] |
| pub struct Rand64 { |
| state: u128, |
| inc: u128, |
| } |
| |
| impl Rand64 { |
| /// The default value for `increment`. |
| /// |
| /// The value used here is from the PCG default C implementation: http://www.pcg-random.org/download.html |
| pub const DEFAULT_INC: u128 = 0x2FE0E169_FFBD06E3_5BC307BD_4D2F814F; |
| |
| /// This is the number that you have to Really Get Right. |
| /// |
| /// The value used here is from the PCG C implementation: |
| /// https://github.com/imneme/pcg-c/blob/master/include/pcg_variants.h#L288 |
| pub(crate) const MULTIPLIER: u128 = 47026247687942121848144207491837523525; |
| |
| /// Creates a new PRNG with the given seed and a default increment. |
| pub fn new(seed: u128) -> Self { |
| Self::new_inc(seed, Self::DEFAULT_INC) |
| } |
| |
| /// Same as `Rand32::new_inc()` |
| pub fn new_inc(seed: u128, increment: u128) -> Self { |
| let mut rng = Self { |
| state: 0, |
| inc: increment.wrapping_shl(1) | 1, |
| }; |
| let _ = rng.rand_u64(); |
| rng.state = rng.state.wrapping_add(seed); |
| let _ = rng.rand_u64(); |
| rng |
| } |
| |
| /// Returns the internal state of the PRNG. This allows |
| /// you to save a PRNG and create a new one that will resume |
| /// from the same spot in the sequence. |
| pub fn state(&self) -> (u128, u128) { |
| (self.state, self.inc) |
| } |
| |
| /// Creates a new PRNG from a saved state from `Rand32::state()`. |
| /// This is NOT quite the same as `new_inc()` because `new_inc()` does |
| /// a little extra setup work to initialize the state. |
| pub fn from_state(state: (u128, u128)) -> Self { |
| let (state, inc) = state; |
| Self { state, inc } |
| } |
| |
| /// Produces a random `u64` in the range`[0, u64::MAX]`. |
| pub fn rand_u64(&mut self) -> u64 { |
| let oldstate: u128 = self.state; |
| self.state = oldstate |
| .wrapping_mul(Self::MULTIPLIER) |
| .wrapping_add(self.inc); |
| let xorshifted: u64 = (((oldstate >> 29) ^ oldstate) >> 58) as u64; |
| let rot: u32 = (oldstate >> 122) as u32; |
| xorshifted.rotate_right(rot) |
| } |
| |
| /// Produces a random `i64` in the range `[i64::MIN, i64::MAX]`. |
| pub fn rand_i64(&mut self) -> i64 { |
| self.rand_u64() as i64 |
| } |
| |
| /// Produces a random `f64` in the range `[0.0, 1.0)`. |
| pub fn rand_float(&mut self) -> f64 { |
| const TOTAL_BITS: u32 = 64; |
| const PRECISION: u32 = core::f64::MANTISSA_DIGITS + 1; |
| const MANTISSA_SCALE: f64 = 1.0 / ((1u64 << PRECISION) as f64); |
| let mut u = self.rand_u64(); |
| u >>= TOTAL_BITS - PRECISION; |
| u as f64 * MANTISSA_SCALE |
| } |
| |
| /// Produces a random within the given bounds. Like any `Range`, |
| /// it includes the lower bound and excludes the upper one. |
| /// |
| /// This should be faster than `Self::rand() % end + start`, but the |
| /// real advantage is it's more convenient. Requires that |
| /// `range.end <= range.start`. |
| pub fn rand_range(&mut self, range: Range<u64>) -> u64 { |
| // Same as `Rand32::rand_range()` |
| debug_assert!(range.start < range.end); |
| let range_starting_from_zero = 0..(range.end - range.start); |
| |
| let s: u64 = range_starting_from_zero.end; |
| let mut m: u128 = u128::from(self.rand_u64()) * u128::from(s); |
| let mut leftover: u64 = (m & 0xFFFFFFFF_FFFFFFFF) as u64; |
| |
| if leftover < s { |
| // TODO: Verify the wrapping_negate() here |
| let threshold: u64 = s.wrapping_neg() % s; |
| while leftover < threshold { |
| m = u128::from(self.rand_u64()) * u128::from(s); |
| leftover = (m & 0xFFFFFFFF_FFFFFFFF) as u64; |
| } |
| } |
| (m.wrapping_shr(64)) as u64 + range.start |
| } |
| } |
| |
| #[cfg(test)] |
| mod tests { |
| use super::*; |
| use randomize::{self, PCG32, PCG64}; |
| |
| #[test] |
| fn test_rand32_vs_randomize() { |
| // Generate some random numbers and validate them against |
| // a known-good generator. |
| { |
| let seed = 54321; |
| let mut r1 = Rand32::new(seed); |
| let mut r2 = PCG32::seed(seed, Rand32::DEFAULT_INC); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.next_u32()); |
| assert_eq!(r1.rand_i32(), r2.next_u32() as i32); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand32::new_inc(seed, inc); |
| let mut r2 = PCG32::seed(seed, inc); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.next_u32()); |
| assert_eq!(r1.rand_i32(), r2.next_u32() as i32); |
| } |
| } |
| } |
| |
| #[test] |
| fn test_rand64_vs_randomize() { |
| // Generate some random numbers and validate them against |
| // a known-good generator. |
| { |
| let seed = 54321; |
| let mut r1 = Rand64::new(seed); |
| let mut r2 = PCG64::seed(seed, Rand64::DEFAULT_INC); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u64(), r2.next_u64()); |
| assert_eq!(r1.rand_i64(), r2.next_u64() as i64); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand64::new_inc(seed, inc); |
| let mut r2 = PCG64::seed(seed, inc); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u64(), r2.next_u64()); |
| assert_eq!(r1.rand_i64(), r2.next_u64() as i64); |
| } |
| } |
| } |
| |
| #[test] |
| fn test_float32() { |
| { |
| let seed = 2718281828; |
| let mut r1 = Rand32::new(seed); |
| let mut r2 = PCG32::seed(seed, Rand32::DEFAULT_INC); |
| for _ in 0..1000 { |
| // First just make sure they both work with randomize's |
| // f32 conversion function -- sanity checks. |
| let i1 = r1.rand_u32(); |
| let i2 = r2.next_u32(); |
| assert_eq!(i1, i2); |
| let f1 = randomize::f32_half_open_right(i1); |
| let f2 = randomize::f32_half_open_right(i2); |
| // We can directly compare floats 'cause we do no math, it's |
| // literally the same bitwise algorithm with the same inputs. |
| assert_eq!(f1, f2); |
| |
| // Make sure result is in [0.0, 1.0) |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| |
| // At least make sure our float's from rand_float() are in the valid range. |
| for _ in 0..1000 { |
| let f1 = r1.rand_float(); |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| |
| /* |
| TODO: Randomize changed its int-to-float conversion functions and now they don't |
| match ours. |
| for _ in 0..1000 { |
| // Now make sure our own float conversion function works. |
| let f1 = r1.rand_float(); |
| //let f2 = randomize::f32_half_open_right(r2.next_u32()); |
| let f2 = randomize::f32_open(r2.next_u32()); |
| assert_eq!(f1, f2); |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| */ |
| } |
| } |
| |
| #[test] |
| fn test_float64() { |
| { |
| let seed = 2718281828; |
| let mut r1 = Rand64::new(seed); |
| let mut r2 = PCG64::seed(seed, Rand64::DEFAULT_INC); |
| for _ in 0..1000 { |
| // First just make sure they both work with randomize's |
| // f64 conversion function -- sanity checks. |
| let i1 = r1.rand_u64(); |
| let i2 = r2.next_u64(); |
| assert_eq!(i1, i2); |
| let f1 = randomize::f64_half_open_right(i1); |
| let f2 = randomize::f64_half_open_right(i2); |
| // We can directly compare floats 'cause we do no math, it's |
| // literally the same bitwise algorithm with the same inputs. |
| assert_eq!(f1, f2); |
| |
| // Make sure result is in [0.0, 1.0) |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| |
| // At least make sure our float's from rand_float() are in the valid range. |
| for _ in 0..1000 { |
| let f1 = r1.rand_float(); |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| |
| /* |
| TODO: Randomize changed its int-to-float conversion functions and now they don't |
| match ours. |
| for _ in 0..1000 { |
| // Now make sure our own float conversion function works. |
| let f1 = r1.rand_float(); |
| //let f2 = randomize::f32_half_open_right(r2.next_u32()); |
| let f2 = randomize::f32_open(r2.next_u32()); |
| assert_eq!(f1, f2); |
| assert!(f1 >= 0.0); |
| assert!(f1 < 1.0); |
| } |
| */ |
| } |
| } |
| |
| #[test] |
| fn test_randrange32() { |
| // Make sure ranges are valid and in the given range |
| let seed = 2342_3141; |
| let mut r1 = Rand32::new(seed); |
| for _ in 0..1000 { |
| // Generate our bounds at random |
| let a = r1.rand_u32(); |
| let b = r1.rand_u32(); |
| if a == b { |
| continue; |
| } |
| let (low, high) = if a < b { (a, b) } else { (b, a) }; |
| |
| // Get a number in that range |
| let in_range = r1.rand_range(low..high); |
| assert!(in_range >= low); |
| assert!(in_range < high); |
| } |
| } |
| |
| #[test] |
| fn test_randrange64() { |
| // Make sure ranges are valid and in the given range |
| let seed = 2342_2718; |
| let mut r1 = Rand64::new(seed); |
| for _ in 0..1000 { |
| // Generate our bounds at random |
| let a = r1.rand_u64(); |
| let b = r1.rand_u64(); |
| if a == b { |
| continue; |
| } |
| let (low, high) = if a < b { (a, b) } else { (b, a) }; |
| |
| // Get a number in that range |
| let in_range = r1.rand_range(low..high); |
| assert!(in_range >= low); |
| assert!(in_range < high); |
| } |
| } |
| |
| #[test] |
| fn test_rand32_vs_rand() { |
| use rand_core::RngCore; |
| use rand_pcg; |
| { |
| let seed = 54321; |
| let mut r1 = Rand32::new(seed); |
| let mut r2 = rand_pcg::Pcg32::new(seed, Rand32::DEFAULT_INC); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.next_u32()); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand32::new_inc(seed, inc); |
| let mut r2 = rand_pcg::Pcg32::new(seed, inc); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.next_u32()); |
| } |
| } |
| } |
| |
| // This doesn't work 'cause for 64-bit output `rand` uses |
| // PCG-XSL-RR |
| // and we use |
| // PCG-XSH-RR |
| /* |
| #[test] |
| fn test_rand64_vs_rand() { |
| use rand_pcg; |
| use rand_core::RngCore; |
| { |
| let seed = 54321; |
| let mut r1 = Rand64::new(seed); |
| let mut r2 = rand_pcg::Pcg64::new(seed, Rand64::DEFAULT_INC); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand(), r2.next_u64()); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand64::new_inc(seed, inc); |
| let mut r2 = rand_pcg::Pcg64::new(seed, inc); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand(), r2.next_u64()); |
| } |
| } |
| } |
| */ |
| |
| // Test vs. random-fast-rng, which I will call rfr |
| // rfr's float conversion uses yet a different algorithm |
| // than ours, so we can't really check that. |
| #[test] |
| fn test_rand32_vs_rfr() { |
| use random_fast_rng as rfr; |
| use rfr::Random; |
| { |
| let seed = 54321; |
| let mut r1 = Rand32::new(seed); |
| let mut r2 = rfr::FastRng::seed(seed, Rand32::DEFAULT_INC); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.get_u32()); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand32::new_inc(seed, inc); |
| let mut r2 = rfr::FastRng::seed(seed, inc); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.get_u32()); |
| } |
| } |
| } |
| |
| /// Make sure that saving a RNG state and restoring |
| /// it works. |
| /// See https://todo.sr.ht/~icefox/oorandom/1 |
| #[test] |
| fn test_save_restore() { |
| { |
| let seed = 54321; |
| let mut r1 = Rand32::new(seed); |
| let s1 = r1.state(); |
| let mut r2 = Rand32::from_state(s1); |
| assert_eq!(r1, r2); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.rand_u32()); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand32::new_inc(seed, inc); |
| let s1 = r1.state(); |
| let mut r2 = Rand32::from_state(s1); |
| assert_eq!(r1, r2); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u32(), r2.rand_u32()); |
| } |
| } |
| |
| { |
| let seed = 54321; |
| let mut r1 = Rand64::new(seed); |
| let s1 = r1.state(); |
| let mut r2 = Rand64::from_state(s1); |
| assert_eq!(r1, r2); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u64(), r2.rand_u64()); |
| } |
| } |
| |
| { |
| let seed = 3141592653; |
| let inc = 0xDEADBEEF; |
| let mut r1 = Rand64::new_inc(seed, inc); |
| let s1 = r1.state(); |
| let mut r2 = Rand64::from_state(s1); |
| assert_eq!(r1, r2); |
| for _ in 0..1000 { |
| assert_eq!(r1.rand_u64(), r2.rand_u64()); |
| } |
| } |
| } |
| } |