Implementing an embedded GPU language by combining translation and generation

Dynamic languages typically allow programs to be written at a very high level of abstraction. But their dynamic nature makes it very hard to compile such languages, meaning that a price has to be paid in terms of performance. However under certain restricted conditions compilation is possible. In this paper we describe how a domain specific language for image processing in Python can be compiled for execution on high speed graphics processing units. Previous work on similar problems have used either translative or generative compilation methods, each of which has its limitations. We propose a strategy which combine these two methods thereby achieving the benefits of both.