A Parallel Error Diffusion Implementation On A Gpu

In this paper, we investigate the suitability of the GPU for a parallel implementation of the pinwheel error diffusion. We demonstrate a high-performance GPU implementation by efficiently parallelizing and unrolling the image processing algorithm. Our GPU implementation achieves a 10 - 30x speedup over a two-threaded CPU error diffusion implementation with comparable image quality. We have conducted experiments to study the performance and quality tradeoffs for differences in image block sizes. We also present a performance analysis at assembly level to understand the performance bottlenecks.