TP_PASSES directly controls the accuracy of the decomposition. As a guide:
- 3-5 passes typically suffice for well-conditioned matrices.
- More passes may be needed for ill-conditioned matrices or when high accuracy of
Uis required. - Fewer passes reduce latency and resource use at the cost of accuracy.
Singular values in the output S vector are sorted in descending order, with the columns of U and V permuted consistently. V is orthonormal by construction regardless of pass count; the orthonormality of U improves with each additional pass.