I believe the performance reported in their paper is circumstantial. It’s not that much faster when I tried it, and not worth the horrible macro syntax.