high performance computing on graphics processing units: hgpu.org

Sparse matrix-vector product (SpMV) is a central operation in many iterative methods for solving linear systems and as such is an attractive candidate for acceleration on the GPU. However, the performance of the SpMV kernel can vary depending both on the target architecture as well as on the sparsity pattern of the matrix. Thus, to achieve optimal performance, the implementation might need to be adjusted for each particular matrix and architecture, which can be achieved through dynamic autotuning. The goal of this thesis is to implement dynamic autotuning of the SpMV kernel on the GPU for the DIA and ELL sparse matrix formats using the KTT autotuning framework and integrate it into the CUSP library. As part of the thesis, a set of possible tuning parameters for both formats is identified and they are implemented into optimized GPU kernels. The proposed tuning parameters and their implementation is evaluated by comparing the autotuned kernels with the original CUSP kernels on a set of representative matrices and by examining the contribution of autotuning. The results show that the autotuned kernels can reach up to about 2 times better performance compared to a fixed implementation.