Functional programming is intuitively a great fit for massively parallel programming, and Blelloch's seminal work on NESL appeared to show exactly how to transform nested data parallelism into efficient flat parallelism. So how come we are not all using functional languages to program our GPUs? In this lecture we will discuss the gap between simply expressing a lot of parallelism, and actually obtaining good performance on real hardware. I will present the design and implementation of the Futhark programming language, which has been carefully restricted (compared to NESL) to permit the construction of an aggressively optimising compiler. We will look at how close Futhark gets to the dream of automatically transforming high-level hardware-agnostic functional parallel code into efficient low-level code, and how to tailor one's parallel programming style to the restrictions imposed by Futhark.

Recommended reading (in decreasing order of relevance):

You are required to read this paper, which outlines the basis of the Futhark implementation strategy (Links to an external site.)
The Futhark User's Guide and language reference (for use when doing the lab) (Links to an external site.)
Parallel Programming in Futhark ("the Futhark book"), although incomplete, is likely the simplest introduction (Links to an external site.)
This paper shows a simple example of Futhark's multi-versioned code generation technique (Links to an external site.)
For the particularly motivated, the first part of my thesis (Links to an external site.)
The slides.pdf