Rust crate implementing the high performance splitmix64, xoroshiro128+, xorshift128+, and xorshift1024* PRNGs. Derived from their respective public-domain C implementations. See COPYRIGHT for details.
Please see http://xoroshiro.di.unimi.it for an overview of the PRNGs and their preferred use cases. For parallel simulations, xorshift1024* is recommended, otherwise xoroshiro128+. splitmix64 is convenient for seeding PRNG states.
[dependencies]
xorshift = "0.1"
extern crate xorshift;
extern crate time;
extern crate xorshift;
use time::precise_time_ns;
use xorshift::{Rand, Rng, SeedableRng, SplitMix64, Xoroshiro128, Xorshift128, Xorshift1024};
fn main() {
// Use the high-resolution performance counter for seeding
let now = precise_time_ns();
// Manually seed a Xorshift128+ PRNG
let states = [now, now];
let mut rng: Xorshift128 = SeedableRng::from_seed(&states[..]);
println!("Xorshift128+ random u64: {}", rng.next_u64());
// Use a SplitMix64 PRNG to seed a Xoroshiro128+ PRNG
let mut sm: SplitMix64 = SeedableRng::from_seed(now);
let mut rng: Xoroshiro128 = Rand::rand(&mut sm);
println!("Xoroshiro128+ random u64: {}", rng.next_u64());
let mut rng: Xorshift1024 = Rand::rand(&mut sm);
println!("Xorshift1024* random u64: {}", rng.next_u64());
// Generate 20 random u32s
let vals = rng.gen_iter::<u32>().take(20).collect::<Vec<u32>>();
println!("Xorshift1024* random u32: {:?}", vals);
// Generate 50 random u64s
let vals = rng.gen_iter::<u64>().take(50).collect::<Vec<u64>>();
println!("Xorshift1024* random u64: {:?}", vals);
}
Applications with little parallelism, should use the Xoroshiro128+ generator.
For large scale parallel computations, use Xorshift1024*. Either use the
thread_rng()
function to create generators with the same seed but incremented
jump states or explicitly use the jump function to forward generator
state.
extern crate xorshift;
use std::thread;
use xorshift::{Rng, Xorshift1024};
use xorshift::thread_rng;
fn main() {
let mut threads = Vec::new();
for i in 0..17 {
threads.push(thread::spawn(move || {
let mut r: Xorshift1024 = thread_rng();
println!("Thread: {}, random u64: {}", i, r.next_u64());
}));
}
for child in threads {
let _ = child.join();
}
}
extern crate time;
extern crate xorshift;
use std::thread;
use time::precise_time_ns;
use xorshift::{Rand, Rng, RngJump, SeedableRng, SplitMix64, Xorshift1024};
fn main() {
// Use the high-resolution performance counter for seeding
let now = precise_time_ns();
let mut sm: SplitMix64 = SeedableRng::from_seed(now);
let rng: Xorshift1024 = Rand::rand(&mut sm);
let mut threads = Vec::new();
for i in 0..17 {
threads.push(thread::spawn(move || {
let mut r = rng;
r.jump(i);
println!("Thread: {}, random u64: {}", i, r.next_u64());
}));
}
for child in threads {
let _ = child.join();
}
}
Written by Alexander Stocko and released in the Public Domain using the Creative Commons Zero Declaration.