A GPU Accelerated BiConjugate Gradient Stabilized Solver for Speeding-up Large Scale Model Evaluation
The Bucharest University of Economic Studies, Faculty of Management, Romania
International Journal of Economic Practices and Theories, Vol. 3, No. 3, 2013
@article{voicu2013gpu,
title={A GPU Accelerated BiConjugate Gradient Stabilized Solver for Speeding-up Large Scale Model Evaluation},
author={Voicu, Alexandru},
journal={International Journal of Economic Practices and Theories},
volume={3},
number={3},
pages={186–191},
year={2013}
}
Solving linear systems remains a key activity in of economics modelling, therefore making fast and accurate methods for computing solutions highly desirable. In this paper, a proof of concept C++ AMP implementation of an iterative method for solving linear systems, BiConjugate Gradient Stabilized (henceforth BiCGSTAB), is presented. The method relies on matrix and vector operations, which can benefit from parallel implementations. The work contained herein details the process of arriving at a moderately parallel implementation and a widely parallel implementation. Furthermore, the construction of two typical sparse data containers in a C++ AMP friendly manner is fleshed out. The implementation is evaluated by solving a number of large-scale linear systems to an exact or epsilon-exact solution.
September 14, 2013 by hgpu
