s417-lama / lxm_random

C++17 header-only library for the LXM random number generator

LXM Random Number Generator in C++

This C++17 header-only library lxm_random is an (unofficial) implementation of the LXM random number generator, which was presented in the following paper:

Guy L. Steele Jr. and Sebastiano Vigna. "LXM: better splittable pseudorandom number generators (and almost as fast)" in ACM OOPSLA '21.

LXM is splittable, which means that a random number generator (RNG) can be splitted into two (apparently) independent streams of RNGs. This split operation is particularly useful for parallel computing; an RNG can be split into multiple child RNGs to be passed to child threads, even recursively. Therefore, the threads can independently generate random numbers in a deterministic manner, without interacting with other threads.

Usage

The C++ header lxm_random/lxm_random.hpp is included in the include/ directory.

Minimal example:

#include <iostream>
#include "lxm_random/lxm_random.hpp"

int main() {
  // Initialize an RNG with a random seed
  uint64_t seed = 42;
  lxm_random::lxm_random rng(seed);

  // Generate a random number
  std::cout << rng() << std::endl;
}

See ./samples/demo.cpp for more practical usage.

Split Operation

How to split random number generators and pass them to child threads:

// Create a root random number generator with a seed
lxm_random::lxm_random rng(seed);

std::vector<std::thread> threads;
for (int i = 0; i < n_threads; i++) {
  // Fork a new thread and pass a new random number generator to it
  // by splitting the root random number generator
  threads.emplace_back([=, rng = rng.split()]() mutable {
    ...
    rng(); // Generate a random number independently from other threads
    ...
  });
}

Alternative approach:

// Create a root random number generator with a seed
lxm_random::lxm_random rng(seed);

std::vector<std::thread> threads;
for (int i = 0; i < n_threads; i++) {
  // Fork a new thread
  threads.emplace_back([=]() mutable {
    // Split the copied random number generator by giving a unique thread ID
    // so that threads with different IDs will have independent streams
    rng = rng.split(i);
    ...
    rng(); // Generate a random number independently from other threads
    ...
  });
}

Notes

• Currently, only L64X128MixRandom variant is implemented