Dec 10, 2024

This is with measured avx512 hashing. On this machine shani always beats sse4 and avx2 but avx512 always beats shani (until the set size precludes avx512 at less than 16 blocks). I also generated a baseline, which is about 20% more efficient than our unoptimized implementation (because we use buffer expansion). This however gives is a greater advantage (5-30%) in the vectorizations, primarily for non-Merkle.