> On Jun 11, 2019, at 10:29 AM, Zhang, Hong via petsc-users > <[email protected]> wrote: > > > Ian, > I would suggest start using AIJ format. > > I am Ian. I trying to implement a solver which involves a sparse symmetric > matrix A multiplied by a dense matrix X. And because of the nature of the > problem, the bandwidth of the matrix A would be kind of large.For A*X, I am > thinking using reverse Cuthill-Mckee algorithm to reduce the bandwidth. > > Are the following approach reasonable, or do you have a better advice? > > 1. Use MatGetOrdering to get a MATORDERINGRCM ordering, and MatPermute to > create a new with it. > SBAIJ may not support some orderings. Matrix ordering for sbaij matrix is > limited to symmetric ordering and requires restructuring the matrix. > > 2. What’s the difference by using MATAIJ and MATBAIJ in terms of the entry > insertion
If you can use MatSetValuesBlocked with BAIJ matrix then it is more efficient then calling it on AIJ. MatMult() on a vector is more efficient with BAIJ or SBAIJ than AIJ. But we haven't provided optimized code for MatMatMult() for BAIJ. it could be written and might improve the performance a little. I would recommend starting with AIJ get your code working doing everything you want correctly. Then measure performance using -log_view and if these computations dominate trying switching to BAIJ and ask for help, if needed to fill in the missing pieces Barry > and computation and MPI efficiency for a sparse-dense matrix multiplication? > Would it be better to use MATSBAIJ in terms of the computational efficiency? > SBAIJ stores upper triangular part of matrix, thus saves approximately half > memory of a sparse matrix. However, it requires more data communications for > y=A*x in general. in your case, C=A*X is a dense matrix, saving half of > storage for sparse A may not be worth trading communication cost and > functionalities. > > Hong
