Solver updates
kuberry opened this issue · comments
Add "SVDA" option to solvers, where "A" is for approximate. Would require that basis being used knows it nullspace size. In CUDA, would use the call gesvdaStridedBatched() operator. If not known, then would revert to normal SVD operation.
Consider changing out LU factorization plus two triangular back solves with dgesv().
Opted for QR with pivoting.