Parallel white noise generation on a GPU via cryptographic hash

Abstract

A good random number generator is essential for many graphics applications. As more such applications move onto parallel processing, it is vital that a good parallel random number generator be used. Unfortunately, most random number generators today are still sequential, exposing performance bottlenecks and denying random accessibility for parallel computations. Furthermore, popular parallel random number generators are still based off sequential methods and can exhibit statistical bias. In this paper, we propose a random number generator that maps well onto a parallel processor while possessing white noise distribution. Our generator is based on cryptographic hash functions whose statistical robustness has been examined under heavy scrutiny by cryptologists. We implement our generator as a GPU pixel program, allowing us to compute random numbers in parallel just like ordinary texture fetches: given a texture coordinate per pixel, instead of returning a texel as in ordinary texture fetches, our pixel program computes a random noise value based on this given texture coordinate. We demonstrate that our approach features the best quality, speed, and random accessibility for graphics applications. Copyright © 2008 by the Association for Computing Machinery, Inc.