MANet: An Architecture Adaptive Method for Sparse Matrix Format Selection

The proliferation of modern computer architectures brings a great challenge to sparse matrix-vector multiplication (SpMV), which is widely used in scientific computing and artificial intelligence. Providing a suitable sparse matrix format for SpMV is crucial to achieve high performance by enhance data locality and cache performance. However, for different architectures, the best sparse matrix format varies. In this paper, we propose a novel architecture adaptive sparse matrix format selection method, MANet, to select proper format to optimize performance of SpMV. This method transforms a sparse matrix into a high-dimensional image, with the matrix sparseness feature and architecture feature combined as inputs. To evaluate the effectiveness of this method, we generated a dataset that includes various scientific problems and architectures with augmentation. Results show that MANet improves sparse matrix selection accuracy by 6% compared to previous works and can achieve a speedup of up to 230% compared to methods with a fixed format. When adapting to an architecture that is not presented in the training, it can still provide 88% selection accuracy and 14% higher than the previous approaches, without further training.