Algorithmic Skeleton Framework for the Orchestration of GPU Computations

The Graphics Processing Unit (GPU) is gaining popularity as a co-processor to the Central Processing Unit (CPU), due to its ability to surpass the latter’s performance in certain application fields. Nonetheless, harnessing the GPU’s capabilities is a non-trivial exercise that requires good knowledge of parallel programming. Thus, providing ways to extract such computational power has become an emerging research topic. In this context, there have been several proposals in the field of GPGPU (Generalpurpose Computation on Graphics Processing Unit) development. However, most of these still offer a low-level abstraction of the GPU computing model, forcing the developer to adapt application computations in accordance with the SPMD model, as well as to orchestrate the low-level details of the execution. On the other hand, the higher-level approaches have limitations that prevent the full exploitation of GPUs when the purpose goes beyond the simple offloading of a kernel. To this extent, our proposal builds on the recent trend of applying the notion of algorithmic patterns (skeletons) to GPU computing. We propose Marrow, a high-level algorithmic skeleton framework that expands the set of skeletons currently available in this field. Marrow’s skeletons orchestrate the execution of OpenCL computations and introduce optimizations that overlap communication and computation, thus conjoining programming simplicity with performance gains in many application scenarios. Additionally, these skeletons can be combined (nested) to create more complex applications. We evaluated the proposed constructs by confronting them against the comparable skeleton libraries for GPGPU, as well as against hand-tuned OpenCL programs. The results are favourable, indicating that Marrow’s skeletons are both flexible and efficient in the context of GPU computing.