Philox in Mojo 🔥

This repository contains an implementation of the Philox Random Number Generator in Mojo programming language. Philox is a counter-based random number generator (RNG) that provides high-quality random numbers with excellent statistical properties.

Features

• Stateless generators
• Statefull CPU streams
• Parallel CPU fill functions
• Reusable GPU kernel
• GPU fill functions

Usage

Basic Example

from random import seed, random_ui64
from philox.streams import Stream64F

fn main():
    seed()
    var seed1 = random_ui64(0, UInt64.MAX)
    var seed2 = random_ui64(0, UInt64.MAX)

    # Create a generator with specific seeds
    var generator = Stream64F(seed1, seed2)
    
    # Generate a SIMD vector of 4 random numbers
    var random_quad = generator.next()
    print(random_quad)
    
    # Fill a buffer with random numbers
    var random_array = InlineArray[Float64, 42](uninitialized = True)
    generator.fill(random_array.unsafe_ptr(), len(random_array))
    print(random_array[0])

Benchmarks

CPU benchmark is against built-in rand function. Number of samples: 10 000 000

$ magic run cpu
Philox 4x32:
CPU time: 7.2638 ms
CPU time: 614.2812 ms
Philox 4x64:
CPU time: 10.3108 ms
CPU time: 593.7460 ms

GPU benchmark measures only kernel time, without memory copy to host. Number of samples: 1 000 000 000

$ magic run gpu
Philox 4x32:
Kernel time: 14.9248 ms
Philox 4x64:
Kernel time: 50.7695 ms

cuRANDDx: CUDA kernels (same algorithm, Philox4-10)

$ ./benchmark_curanddx.sh 
Performance Statistics:
Average time: 14.92 ms
Standard deviation: 0.13 ms

Testing

Tests include statistics verification for mean, variance, histogram uniformity, and chi-square test

$ magic run test
Testing Time: 2.187s

Total Discovered Tests: 5

Passed : 5 (100.00%)
Failed : 0 (0.00%)
Skipped: 0 (0.00%)