Nov 25, 2024

This approach also allows us to vectorize K additions, because they are applied across the fully scheduled buffer. In the typical shani algorithm (using rotating variables as opposed to a message buffer), this vectorization is limited to 4 lanes (4 concurrent additions). This is however independent of the actual shani implementation. By separating the implementation in this manner we can take advantage of 4/8/16 lane vectorization for K additions - mixing avx512 and shani.