RNG Collection

A small, high-quality collection of non-cryptographic pseudo-random number generators (PRNGs) in C++. All have been tested and passed PractRand at 128 GB.

Contents

• Random Number Generator Classes

  • class wyrand
  • class wy256
  • class SplitMix64
  • class xoshiro256plusplus

• Utility Functions

  • uint64_t mul128(uint64_t a, uint64_t b, uint64_t& hi) - portable fast 64x64 -> 128 multiply
  • uint64_t mix_fold(uint64_t a, uint64_t b) - rapid-hash-style mixing function
  • uint64_t mix_fold_protected(uint64_t a, uint64_t b) - rapid-hash-style mixing function
  • uint64_t mix_fold_protected_modified(uint64_t a, uint64_t b) - rapid-hash-style mixing function
  • uint8_t add_carry(uint8_t c, uint64_t a, uint64_t b, uint64_t* out) - portable add_carry function
  • uint64_t nasam(uint64_t x) - Pelle Evensen's NASAM mixer
  • uint64_t pelican(uint64_t state) - Pelican mixing function
  • uint64_t mx3(uint64_t x) – John Maiga's mixing function.
  • uint64_t better_rand_device() – Improvement of rand_device, used to create random seeds.

Usage

#include "RNG.h"
#include <iostream>
int main() {
    uint64_t seed = RNG::Primitives::better_rand_device();
    RNG::SplitMix64 seeder(seed);
    RNG::wyrand wyrand_gen(seeder());
    RNG::wy256 wy256_gen(seeder());
    RNG::xoshiro256plusplus xoshiro_gen(seeder());

    // Advance 2^35 steps forward in the wy256 sequence.
    // This also works with RNG::SplitMix64 and RNG::wyrand.
    // See also the jump() and long_jump() functions in each class.
    wy256_gen.discard(1ull << 35);

    std::cout << "wyrand_gen() = " << wyrand_gen() << "\n";
    std::cout << "wy256_gen() = " << wy256_gen() << "\n";
    std::cout << "xoshiro_gen() = " << xoshiro_gen() << "\n";

    // Create 1024 instances of the xoshiro256plusplus generator, placing them in a vector.
    // Each generator is initialized with the same seed, but separated with a call to long_jump()
    // to ensure they provide independent streams of random numbers.
    seed = 123456789;
    size_t n_generators = 1024;
    auto generators = RNG::xoshiro256plusplus::factory(n_generators, seed);
}

Discard and Jump Functions

The following routines are available for skipping forward in the sequence of random numbers. All methods are O(1).

- SplitMix64
	- discard(n) - equivalent to n calls to operator()
	- jump() - advance 2^32 steps
	- long_jump() - advance 2^48 steps
- wyrand
	- discard(n) - equivalent to n calls to operator()
	- jump() - advance 2^32 steps
	- long_jump() - advance 2^48 steps
- wy256
	- discard(n) - equivalent to n calls to operator()
	- jump() - advance 2^128 steps
	- long_jump() - advance 2^192 steps

- xoshiro256plusplus
	- Note: no arbitrary discard(n) function
	- jump() - advance 2^128 steps
	- long_jump() - advance 2^192 steps

Random Number Generation Functions

All generator classes include the following methods for generating random numbers.

	- operator()(): Returns the next random number in the sequence.
	- draw64(): Alias for operator()(), returns a 64 bit random number.
	- draw32(): Returns the upper 32 bits of the next random number.
	- uni(): Returns a double in the range [0.0, 1.0) using IEEE 754 bit manipulation for maximum performance.
	- uniform(u64 limit): returns an unbiased random number in the range [0, limit) using rejection sampling.
	- uniform(u64 lo, u64 hi): returns an unbiased random number in the range [lo, hi].

Requirements

C++20 or later x86/x64 for __rdtsc() (the current implementation) , , , and x86 intrinsics (<x86intrin.h> or <intrin.h>)

Portability Note

An effort has been made to make this at least somewhat portable. Comments/suggestions welcome.

Installation

Just drop the header files into your project. No external dependencies beyond the standard library and intrinsics.

License

MIT

Contributing

Suggestions for:

• Better mixing strategies
• Portable fallbacks (ARM, PowerPC, etc.)
• Integration helpers for popular PRNGs
• Benchmarks

are welcome!