Abstract

A collection of rng_engine modules that implement the Romu family of random number generators by Mark Overton. While the Romu generators are mostly fast on out-of-order CPUs because they can be "zero latency" by fitting into otherwise unused execution slots, they may also be useful on GPU. In particular, the 32-bit variants run very fast.

As usual, the seeding, splitting, and joining functions are not part of the upstream Romu algorithm, but are simply best-effort implementations that try to mix in a little different entropy. Caveat emptor.

Synopsis

open import "lib/github.com/diku-dk/cpprandom/random"
module romu_quad: rng_engine with int.t = u64
module romu_trio: rng_engine with int.t = u64
module romu_duo: rng_engine with int.t = u64
module romu_duo_jr: rng_engine with int.t = u64
module romu_quad32: rng_engine with int.t = u32
module romu_trio32: rng_engine with int.t = u32
module romu_mono32: rng_engine with int.t = u16

Description

↑module romu_quad

More robust than anyone could need, but uses more registers than romu_trio.

  • Est. capacity: 2⁹⁰ bytes
  • State size: 256 bits
↑module romu_trio

Great for general purpose work, including huge jobs.

  • Est. capacity: 2⁷⁵ bytes
  • State size: 192 bits
↑module romu_duo

Might be faster than romu_trio due to using fewer registers, but might struggle with massive jobs.

  • Est. capacity: 2⁶¹ bytes
  • State size: 128 bits
↑module romu_duo_jr

The fastest generator using 64-bit arithmetic, but not suited for huge jobs.

  • Est. capacity: 2⁵¹ bytes
  • State size: 128 bits
↑module romu_quad32

32-bit arithmetic: Good for general purpose use.

  • Est. capacity: 2⁶¹ bytes
  • State size: 128
↑module romu_trio32

32-bit arithmetic: Good for general purpose use, except for huge jobs.

  • Est. capacity: 2⁵³ bytes
  • State size = 96 bits.
↑module romu_mono32

32-bit arithmetic: Suitable only up to 2²⁶ output-values. Outputs 16-bit numbers. Fixed period of (2³²)-47.

  • Capacity: 2²⁷ bytes
  • State size; 32 bits