Comparing Different Programming Approaches for SpMV-Operations on GPUs

There exist various different high- and low-level approaches for GPU programming. These include the newer directive based OpenACC programming model, Nvidia’s programming platform CUDA and existing libraries like cuSPARSE with a fixed functionality. This work compares the attained performance and development effort of different approaches based on the example of implementing the SpMV operation, which is an important and performance critical building block in many application fields. We show that the main differences in development effort using CUDA and OpenACC are related to the memory management and the thread mapping