I have a partial implementation that is still missing some pre-processing to improve performance. Currently, it fails to beat the JAMA code, hence my patch on
MATH-611. I don't believe that the performance of my code can be improved to beat JAMA, though I am willing to continue development slow time.
If I understand correctly, the algorithm in question is not necessarily the best SVD algorithm available, it merely pushes the boundaries of what Jacobi methods for SVD are capable of achieving compared to, say, bi- or tri-diagonalisation methods.
I would certainly not regard this as 'Major' issue anymore. I could pursue development further if it is considered useful to have another implementation within ACM (raising a new issue later if required to submit).