Abstract

Array fusion is a vital optimization in functional array programming. With array fusion, functions operating on arrays can be written in a high-level, compositional style, yet still “fuse” down to a single, efficient loop, with no need to allocate intermediary arrays. This produces data-parallel code, but often does not leave any opportunities for task-parallelism. Array fissioning goes in the opposite direction: single functions are “fizzed” into two or more chunks. Fission is compatible with fusion, and allows for functional array programs to be implicitly distributed across multiple devices.

In this talk, I will discuss different functional array representations and describe how to implement fusion and fissioning.

The presentation slides are available here: http://vollmerm.github.io/FizzFuseArrays/slides.pdf