This repo previously contained a number of experiments comparing BLAKE2b and BLAKE2s. Eventually I figured out that BLAKE2s is faster in essentially all cases. Now the repo contains some new experiments:

• Using hash_many on only leaves (bao_standard) vs using it on parents also (bao_parallel_parents).
• Using a very large chunk size (bao_large_chunks). This is a baseline for what performance should be if parent node overhead disappears entirely.
• Using a tree of arity > 2 (bao_nary).

Mostly unrelated to the above, it also contains some benchmarks for different SIMD algorithms for loading transposed message words.

Run these benchmarks with cargo +nightly bench.