Abstract: Using scientific computing to solve practical problems in the engineering field can often be transformed into solving large-scale linear equations. The most frequently called step in this process is sparse matrix-vector multiplication (SpMV). For the sparse quasi-diagonal matrix commonly-used in engineering, a blocked hybrid storage format BHDC, which combined the advantages of DIA and CSR, is proposed. In BHDC format, a sparse matrix was divided into several rows, and the diagonal dense areas and scattered points were stored respectively according to the threshold value. It not only made use of the good floating-point operation performance of DIA storage format, but also avoided the performance degradation caused by the sharp increase of diagonal lines through CSR storage format. Several sparse matrices were selected on CUDA platform for testing, and the time-space performance better than the above two storage formats and the time performance better than the non-blocked hybrid storage format HDC were obtained.