Aumento do desempenho da support vector machine através de técnicas de paralelização de código em gpu

Abstract

The Support Vector Machine is a data classification method that has presented an increase in terms of its popularity. This increase is mainly due to the publication of LIBSVM, a library that implements Support Vector Machine which source code is available in a diversity of programming languages. Since its publication, several researches have tried to optimize its performance through parallelization of parts of its source code, mostly using the Graphics Processor Unit as a parallel processing device. Some of these work use the CUDA framework, limiting the solutions to the NVIDIA’s GPUs. This work proposes the parallelization of the Support Vector Machine through OpenCL framework, allowing the solution to be executed in a wider variety of GPUs. We choose training phase of the classificatiom process for the optimization. The final solution achieved a speedup of three in comparisson to the sequential version, and it is developed using two distinct GPU architectures.