c - Solving sparse linear systems in CUDA using LU factorization -

the current matlab based c implementation takes around 6ms solving ax=b, a banded sparse matrix band-width 3 of dimensions 780 x 780.

now looking use cublas/cusparse find faster solution. need solve 1440 of such equations in loop.

i tried using pcg based method slow , output not matching.

is there direct solution using cublas/cusparse solving ax=b?

if problem can converted tri-diagonal problem, can use cusparsexgtsvstridedbatch multiple problems without using loop. have use cusparse_v2.h instead of cusparse.h work.

if problem can not converted tri-diagonal problem, can use routines cula solve problem. more information regarding can read in blog post. commercial library. may not best suited band of matrix 3 bands only.