A Portable OpenCL-Based Approach for SVMs in GPU

Support Vector Machines (SVMs) is one of the most efficient methods for data classification in machine learning. Several efforts were dedicated towards improving its performance through source-code parallelization, particularly within the Graphics Processor Unit (GPU). Those studies make use of the well-known CUDA framework, which is provided by NVIDIA for its graphics cards. Nevertheless, the main disadvantage of CUDA-based solutions is that they are specific to NVIDIA cards, reducing the applicability of such solutions in heterogeneous environments. In this work, we propose the parallelization of SVMs through the OpenCL framework, which allows the generated solution to be portable to a wide range of GPU manufacturers. The proposed approach parallelizes the most costly steps that are performed when training SVMs. We show that the proposed solution achieves a significant speedup regarding the algorithm´s original version, and also that it outperforms the state-of-the-art CUDA-based approach in terms of computational performance in 11 out of the 12 datasets that were tested in this work.