Abstract: Sampling is important for a variety of graphics applications include rendering, imaging, and geometry processing. However, producing sample sets with desired efficiency and blue noise statistics has been a major challenge, as existing methods are either sequential with limited speed, or are parallel but only through pre-computed datasets and thus fall short in producing samples with blue noise statistics. We present a Poisson disk sampling algorithm that runs in parallel and produces all samples on the fly with desired blue noise properties. Our main idea is to subdivide the sample domain into grid cells and we draw samples concurrently from multiple cells that are sufficiently far apart so that their samples cannot conflict one another. We present a parallel implementation of our algorithm running on a GPU with constant cost per sample and constant number of computation passes for a target number of samples. Our algorithm also works in arbitrary dimension, and allows adaptive sampling from a user-specified importance field. Furthermore, our algorithm is simple and easy to implement, and runs faster than existing techniques.