Mixed precision iterative refinement for symmetric positive definite matrix

I am trying to perform mixed precision iterative refinement on tensor core. The matrix that I have is symmetric positive definite. I am able to use the gesv solver cusolverDnIRSXgesv(). I have gone though the paper by Haidar et.al. and was wondering if I can do something similar for my positive definite matrix. I also wanted to understand the method a little better. Is it possible to have access to the cusolver source code so that I can go through the implementation and get the finer details.


Dear user,
Thank you very much for your interest and for contacting us related to this subject.
Unfortunately the cuSolver source code cannot be shared.
the method was described well in the paper you mention it.
The issue of symmetric positive definite is that (depending on the matrix), the lower precision can loose positiveness. However there is some work done by a previous colleagues for symmetric positive matrices available here https://icl.utk.edu/files/publications/2020/icl-utk-1389-2020.pdf and its implementation in MAGMA.
Suggestion: in case your matrices loose positiveness, you can use TF32 instead of FP16. TF32 showed better accuracy with similar performance for many problem.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.