A CSR storage format-based spmv implementation method and device and medium

Abstract

The application discloses a CSR storage format-based SpMV implementation method and device and a medium, wherein the method comprises the following steps: dividing the SpMV into three segment operations, the three segment operations being a value operation, a product operation and a summation operation respectively, and each segment operation corresponding to a task; dividing each task into a plurality of task blocks; dividing all threads into a plurality of vectors, one vector comprising vector_size threads, and the vector being an execution unit of the task block; designing two launch queues and one counter to record the ready state of the task block; each vector acquiring and executing the task block in the ready state, and updating the ready state of the task block after execution; and when all the task blocks are executed, the SpMV operation is completed. The application divides the value and the product into task blocks of a fixed size, and improves the data utilization rate of GPU memory access. The application can be widely applied to the technical field of high-performance numerical calculation.

Images