swAHSpMV: an adaptive and heterogeneous optimization of SpMV for Hypre on SW26010-pro processor

In the fields of computational science and engineering, Sparse Matrix-Vector Multiplication (SpMV) is a core numerical operation, particularly crucial when dealing with large-scale and complex systems. This paper proposes an adaptive and heterogeneous optimization method named swAHSpMV, which aims to enhance the performance of the SpMV kernel for the Hypre library on the SW26010-Pro processor. The adaptive dynamic data loading method is adopted to adjust the data loading amount to effectively utilize the slave core LDM resources. The SIMD operation is implemented by manually unrolling loops and rearranging data to improve computational efficiency. The alternate buffer optimization method is used to reduce the waiting time and improve processor utilization by using two buffers to alternately perform calculation and data transmission. Experimental results show that the proposed algorithm achieves an average speedup of 21.88x on many-core processors. Compared to the SpMV kernel in the Sunway xMath2.0 library, an average speedup of 4.97x is obtained. The multi-node parallel version can be extended to 128 core groups on the Sunway heterogeneous platform and shows good scalability.