Streaming Irregular Arrays
In the new draft paper Streaming Irregular Arrays, we extend the high-performance array library Accelerate for Haskell with support for irregular streams of multidimensional arrays. We discuss why this is useful to work with large datasets on memory-constrained devices, such as GPUs, and how it widens the range of applications for which we achieve portable parallelism with Accelerate. The paper explains how we compile these programs and what runtime mechanisms we need for execution. Moreover, it includes a range of benchmarks to evaluate the performance of the resulting code.