high performance computing on graphics processing units: hgpu.org

Infinite impulse response (IIR) or recursive filters, are essential for image processing because they turn expensive large-footprint convolutions into operations that have a constant cost per pixel regardless of kernel size. However, their recursive nature constrains the order in which pixels can be computed, severely limiting both parallelism within a filter and memory locality across multiple filters. Prior research has developed algorithms that can compute IIR filters with image tiles. Using a divide-and-recombine strategy inspired by parallel prefix sum, they expose greater parallelism and exploit producer-consumer locality in pipelines of IIR filters over multidimensional images. While the principles are simple, it is hard, given a recursive filter, to derive a corresponding tile-parallel algorithm, and even harder to implement and debug it. We show that parallel and locality-aware implementations of IIR filter pipelines can be obtained through program transformations, which we mechanize through a domain-specific compiler. We show that the composition of a small set of transformations suffices to cover the space of possible strategies. We also demonstrate that the tiled implementations can be automatically scheduled in hardwarespecific manners using a small set of generic heuristics. The programmer specifies the basic recursive filters, and the choice of transformation requires only a few lines of code. Our compiler then generates high-performance implementations that are an order of magnitude faster than standard GPU implementations, and outperform hand tuned tiled implementations of specialized algorithms which require orders of magnitude more programming effort-a few lines of code instead of a few thousand lines per pipeline.