Low-Level Functional GPU Programming for Parallel Algorithms

We present a Functional Compute Language (FCL) for low-levelGPU programming. FCL is functional in style, which allows foreasy composition of program fragments and thus easy prototypingand a high degree of code reuse. In contrast with projects such asFuthark, Accelerate, Harlan, Nessie and Delite, the intention is notto develop a language providing fully automatic optimizations, butinstead to provide a platform that supports absolute control of theGPU computation and memory hierarchies. The developer is thusrequired to have an intimate knowledge of the target platform, as isalso required when using CUDA/OpenCL directly.FCL is heavily inspired by Obsidian. However, instead of relyingon a multi-staged meta-programming approach for kernel generationusing Haskell as meta-language, FCL is completely selfcontained,and we intend it to be suitable as an intermediate languagefor data-parallel languages, including data-parallel parts ofhigh-level array languages, such as R, Matlab, and APL.We present a type-system and a dynamic semantics suitablefor understanding the performance characteristics of both FCL andObsidian-style programs. Our aim is that FCL will be useful as aplatform for developing new parallel algorithms, as well as a target languagefor various code-generators targeting GPU hardware