oxicuda-rand

GPU-accelerated random number generation -- pure Rust cuRAND equivalent.

Part of the OxiCUDA project.

Overview

oxicuda-rand provides GPU-accelerated random number generation implemented entirely in Rust, targeting feature parity with NVIDIA's cuRAND. It includes three PRNG engines, four distribution transforms, and Sobol quasi-random sequences, all generating numbers in parallel on the GPU via PTX kernels.

Each engine is counter-based or state-based, producing statistically independent streams across GPU threads. Philox-4x32-10 is the default (matching cuRAND), offering an excellent balance of speed and quality. XORWOW provides fast generation for Monte Carlo workloads. MRG32k3a delivers the highest statistical quality for applications that demand it.

The high-level RngGenerator API dispatches to any engine and automatically tracks per-thread offsets to ensure non-overlapping sequences across successive generation calls. Sobol quasi-random sequences support up to 20 dimensions with Gray-code generation for low-discrepancy sampling in Monte Carlo integration.

Engines

Engine	Module	Description
Philox-4x32-10	`engines::philox`	Counter-based, cuRAND default, 2^128 period
XORWOW	`engines::xorwow`	XORshift + Weyl sequence, fast
MRG32k3a	`engines::mrg32k3a`	Combined MRG, highest statistical quality

Distributions

Distribution	Module	Method
Uniform	`distributions::uniform`	[0, 1) in f32/f64
Normal	`distributions::normal`	Box-Muller transform
Log-normal	`distributions::log_normal`	exp(Normal(mu, sigma))
Poisson	`distributions::poisson`	Knuth (small lambda), normal approx (large)

Quasi-Random

Sequence	Module	Description
Sobol	`quasi::sobol`	Gray-code Sobol, up to 20 dimensions

Quick Start

use oxicuda_rand::prelude::*;

// With a GPU context:
// let mut gen = RngGenerator::new(&ctx, RngEngine::Philox, 42)?;
//
// // Generate 1M uniform f32 values on the GPU
// let uniform = gen.generate_uniform_f32(1_000_000)?;
//
// // Generate normal-distributed f64 values
// let normal = gen.generate_normal_f64(1_000_000, 0.0, 1.0)?;
//
// // Sobol quasi-random sequence (5 dimensions)
// let sobol = SobolGenerator::new(&ctx, 5)?;
// let points = sobol.generate(10_000)?;

Feature Flags

Feature	Description
`f16`	Half-precision (fp16) RNG support
`gpu-tests`	Enable GPU integration tests

Status

Metric	Value
Version	0.1.3
Tests passing	270
Release date	2026-04-17

oxicuda-rand 0.1.3