12 releases (7 breaking)

Uses new Rust 2021

0.8.2 Nov 22, 2022
0.8.0 Oct 26, 2022
0.5.1 Jul 13, 2022

#147 in Algorithms

Download history 124/week @ 2022-08-16 273/week @ 2022-08-23 78/week @ 2022-08-30 126/week @ 2022-09-06 103/week @ 2022-09-13 26/week @ 2022-09-20 117/week @ 2022-09-27 40/week @ 2022-10-04 30/week @ 2022-10-11 68/week @ 2022-10-18 186/week @ 2022-10-25 135/week @ 2022-11-01 228/week @ 2022-11-08 191/week @ 2022-11-15 167/week @ 2022-11-22 132/week @ 2022-11-29

735 downloads per month
Used in 2 crates (via bevy_turborand)

Apache-2.0 OR MIT

105KB
2K SLoC

turborand

CI License Cargo Documentation

Fast random number generators.

turborand's internal implementations use Wyrand, a simple and fast generator but not cryptographically secure, and also ChaCha8, a cryptographically secure generator tuned to 8 rounds of the ChaCha algorithm in order to increase throughput considerably without sacrificing too much security, as per the recommendations set out in the Too Much Crypto paper.

Examples

use turborand::prelude::*;

let rand = Rng::new();

if rand.bool() {
    println!("Success! :D");
} else {
    println!("Failure... :(");
}

Sample a value from a list:

use turborand::prelude::*;

let rand = Rng::new();

let values = [1, 2, 3, 4, 5];

let value = rand.sample(&values);

Generate a vector with random values:

use turborand::prelude::*;
use std::iter::repeat_with;

let rand = Rng::new();

let values: Vec<_> = repeat_with(|| rand.f32()).take(10).collect();

Performance

Wyrand is a pretty fast PRNG, and is a good choice when speed is needed while still having decent statistical properties. Currently, the turborand implementation benches extremely well against similar rand algorithms. Below is a chart of the fill_bytes method performance, tested on Windows 10 x64 on an AMD Ryzen 1700 clocked at 3.7Ghz with 32GB RAM at 3066Mhz.

fill_bytes benchmark

For filling 2048 byte array buffers, turborand's Rng is able to do so in around 170-180ns, whereas SmallRng does it between 260-268ns, and Pcg64Mcg (the fastest PCG impl on 64bit systems) does it in 305-312ns.

u64 gen benchmark

For generating unbound u64 values, turborand and fastrand are equal in performance, which is expected given they both implement the Wyrand algorithm, consistently performing around 820-830ps for generating a u64 value. SmallRng performs around 1.16ns, while Pcg64Mcg is at 1.35ns.

Migration from 0.5 to 0.6

Version 0.6 introduces a major reworking of the crate, with code reorganised and also exposed more granularly via features. First things to note:

  • All major exports for the crate are now in the prelude module. Top level only exports the new traits for turborand.
  • Rng is now split into Rng and AtomicRng, no more top level generics that required exporting internal traits. State trait is now made private and no longer available to be implemented, as this was an internal implementation detail for WyRand.
  • All previous methods for Rng are now implemented in TurboCore, GenCore, SeededCore and TurboRand traits. These are part of the prelude so as long as they are included, all existing methods will work as expected.
  • Rng is now under a feature flag, wyrand. This is enabled by default however, unless default-features = false is applied on the dependency declaration in Cargo.toml.
  • Yeet the rng!, atomic_rng! macros, as these are no longer needed to manage the generics spam that has since been refactored out. Instead, use ::new(), ::default() or ::with_seed(seed) methods instead.

Migration from 0.6 to 0.7

Version 0.7 hasn't changed much except that the internals module is now fully private (so the State traits and CellState/AtomicState structs are no longer public). They are not accessible from the prelude any more. The removal of these from the public API thus constitutes a breaking change, leading to a new major version.

Also, the serialisation format of ChaChaRng has changed, so 0.7 is not compatible with older serialised structs. The plus side is also a flatter serialised format for ChaChaRng. Also, ChaChaRng is no longer backed by a Vec for caching generated entropy, now preferring to use an aligned array for better random number generation at the slight cost of initialisation/cloning performance and increased struct size. This means that the single heap allocation ChaChaRng needed is now reduced to zero.

Migration from 0.7 to 0.8

Version 0.8 seperates the old Clone behaviour into two: standard Clone which maintains the original state and clones it to the new instance as is (and so both old and new equal to each other), and ForkableCore which mutates the state of the original to fork a new instance with a random state generated from the original. Previous usage of .clone() now should make use of .fork() instead. Cloning now should be used where preserving the state of the original to the cloned instance is required.

License

Licensed under either of

at your option.

Dependencies

~0–370KB