Towards High-Performance Code Generation for Multi-GPU Clusters Based on a Domain-Specific Language for Algorithmic Skeletons

In earlier work, we defined a domain-specific language (DSL) with the aim to provide an easy-to-use approach for programming multi-core and multi-GPU clusters. The DSL incorporates the idea of utilizing algorithmic skeletons, which are well-known patterns for parallel programming, such as map and reduce. Based on the chosen skeleton, a user-defined function can be applied to a data structure in parallel with the main advantage that the user does not have to worry about implementation details. So far, we had only implemented a generator for multi-core clusters and in this paper we present and evaluate two prototypes of generators for multi-GPU clusters, which are based on OpenACC and CUDA. We have evaluated the approach with four benchmark applications. The results show that the generation approach leads to execution times, which are on par with an alternative library implementation.