Sparse matrix vector multiplication vectorization implementation method

Abstract

The invention discloses a sparse matrix vector multiplication vectorization implementation method, which comprises the following steps of: 1, respectively constructing a first matrix and a second matrix in a DDR (Double Data Rate), reading non-zero element data values in a sparse matrix to be calculated according to rows and storing the non-zero element data values in the first matrix according tocolumns, and storing corresponding column index values in the second matrix according to columns; 2, configuring a data buffer area in AM; 3, respectively transmitting the dense vector to be calculated and the data in the second matrix from the DDR to the GSM; 4, sequentially transmitting the data in the first matrix from the DDR to a data buffer area; 5, reading a column index value in the second matrix, reading a data value of the to-be-calculated dense vector according to the read column index value, and transmitting the data value to a data buffer area; and step 6, performing vectorization calculation on the data in the data buffer area. The method has the advantages of simple implementation operation, high resource utilization rate, high calculation efficiency, low hardware overheadand the like.

Description

technical field [0001] The invention relates to the technical field of multi-core vector processors, in particular to a method for implementing vectorization of sparse matrix-vector multiplication for multi-core vector processors. Background technique [0002] HPL (High Performance Linpack) is the most important benchmark test program for evaluating the performance of high-performance computer systems, and it is also the main evaluation index. What HPL solves is a large-scale dense linear equation system, which has good data locality, and is easy to develop parallelism and locality. However, with the continuous development of high-performance applications, a large number of scientific computing applications, such as large-scale marine weather forecasting, have application characteristics that do not match HPL. Therefore, since 2013, HPCG (HighPerformance Conjugate Gradient) has been used as a new benchmark test program to supplement the performance evaluation of high-perfor...

AI Technical Summary

Problems solved by technology

But using this type of method, on the one hand, maxnonzeros is not necessarily an integer multiple of the number of VPEs (for example, maxnonzeros=27, the number of VPEs=16), when maxnonzeros is not an integer multiple of the number of VPEs, it will lead to waste of computing resources; on the other hand , the multiplication and accumulation of vectors is carried out between each VPE, and the vector processor is required to support shuffling or reduction calculations, and the required hardware overhead (area, power consumption, etc.) is large

Images