Sparse Matrix Selection for CSR-Based SpMV Using Deep Learning
Document Type
Conference Proceeding
Publication Date
12-1-2019
Abstract
CSR (Compressed Sparse Row) is the most popular and widely used sparse matrix representation format for Sparse Matrix-Vector Multiplication (SpMV), which is a key operation in many scientific and engineering applications. However, considering different matrix features and the given GPUs, CSR-based SpMV on some sparse matrices does not always have better performance than that of SpMV based on other sparse matrix formats. In this paper, we explore deep learning techniques and present a methodology to select the proper sparse matrices for CSR-based SpMV on NVIDIA GPUs. To address the challenge of this matrix selection problem, we convert it to a matrix classification problem, then address this classification problem by using the Convolutional Neural Networks (CNN). The effectiveness of our proposed methodology has been demonstrated by our experimental evaluations performed on NVIDIA GPUs.
Publication Title
2019 IEEE 5th International Conference on Computer and Communications, ICCC 2019
First Page Number
2097
Last Page Number
2101
DOI
10.1109/ICCC47050.2019.9064309
Recommended Citation
Guo, Ping and Zhang, Changjiang, "Sparse Matrix Selection for CSR-Based SpMV Using Deep Learning" (2019). Kean Publications. 1298.
https://digitalcommons.kean.edu/keanpublications/1298