Sparse Matrix Multiplication using CUDA and Mex Interface
College of Arts and Sciences, Florida State University
Florida State University, master project, 2012
@article{deshpande2012sparse,
title={Sparse Matrix Multiplication using CUDA and Mex Interface},
author={Deshpande, A.},
year={2012}
}
In recent years, the development in the architecture of graphics processing units (GPUs) has revolutionized the area of high performance computing by offering massive parallelism and performance improvements in many applications including matrix algebra. While it is possible to harness the power of GPUs for dense matrix computations, sparse matrix computations are still complex since mapping the sparsity of individual matrices at thread level granularity is a challenging task. In this report, we explain the use of mex interface of MATLAB and its collaboration with CUDA [4] (Compute Unified Device Architecture from Nvidia) to perform the sparse matrix computation. We explain how we have exploited various CUDA template libraries (such as Thrust and CUSP) for performing essential matrix transformation and present a custom multiplication CUDA kernel (function) that was implemented as a part of deliverable of this project.
December 20, 2012 by hgpu
