Functional programming for nested data parallelism on GPUs

Recent advances in general purpose GPU computing technology allow new data parallel kernel jobs to be dispatched dynamically during kernel execution. This enables significantly more expressive programming using nested data parallelism (NDP), where the restrictive need for flat data structures and computation has been lifted. Functional programming is fundamentally well suited for expressing data parallel computation. Expensive flattening and fusion transformations have so far been necessary in vectorizing irregular data structures and computation, causing functional NDP to be inefficient. With hardware supported nesting on massively parallel GPUs, high-abstraction functional nested data parallelism can now be explored further. This paper introduces nested data parallelism and briefly makes a case for using functional languages to program GPUs for general purpose data parallel computing.