I am wondering whether there is any cusolver which can be used as a replacement for intel mkl pradiso.
I am dealing with the problem Ax=b, where “A” is sparse, symmetric and positive definite, and x and b are vectors which can hold multiple righthand sides/solutions. “A” is constant throughout the program but “Ax=b” is called in different parts of the program with different “x”'s and “b”'s.
So far I have used pardio which has a preparation, factorization and solving step, which can be called individually. Thus, preparation and the most expensive factorization can be called once at the beginning of the program, and the solver calls can be repeated anywhere throughout the program without much cost because the factors are reused.
With regard to cuda I have looked at “cusolverSpScsrlsvlu”, “cusolverSpScsrlsvqr” and “cusolverSpScsrlsvchol” and it appears to me that these function will do (redo) the factorization each time they are called which will produce a massive overhead. Is that correct?? And if so, is there any way to circumvent it??