ACC-SVM: Accelerating SVM on GPUs using OpenACC

GPUs have been successfully applied in scientific computing in the last decade. Many machine learning algorithms have also used GPUs to accelerate their computations. This includes the Support Vector Machine (SVM) which is a classical machine learning algorithm that has been successfully used in many applications such as text classification and image recognition. There have been many open-source CUDA SVM implementations. However, CUDA versions of SVM are not portable and difficult to maintain or redesign. Porting SVM to a directive-based portable model like OpenACC will make it possible to target multiple accelerators. In this paper, we use OpenACC programming model to parallelize SVM and produce ACC-SVM. Since the high-level programming model simplifies the programming by sacrificing performance, there is a performance gap between the OpenACC and the CUDA SVM versions. In order to improve the performance of ACC-SVM, we apply our auto-tuning framework to decrease the gap between the CUDA and the OpenACC performance results. The performance difference between the optimized ACC-SVM resulting from the auto-tuner and the CUDA SVM is 15.58% for the kernels code and 7.87% with respect to the whole application. For many applications this loss of performance is more than made up for by the convenience and portability of the high-level approach.