In this case it's not tricks like those done in BLAS kernels. However, there are some intricacies about the algorithms that are chosen as noted in the paper. Th

In this case it's not tricks like those done in BLAS kernels. However, there are some intricacies about the algorithms that are chosen as noted in the paper. That said, most the difference here is simply from the difference in the high level details of how the parallelism is designed, not necessarily low level bit hacking optimizations. We leave those for another day.