git-annex using BLAKE3 is significantly slower than b3sum. The latter uses BLAKE3's multithreading, which is not available in the haskell library.
https://github.com/k0001/hs-blake3/issues/9
On a 4 gb file:
git-annex fsck:
1.25user 2.36system 0:07.81elapsed 46%CPU (0avgtext+0avgdata 57676maxresident)k
5539480inputs+208outputs (2major+9161minor)pagefaults 0swaps
b3sum:
1.60user 3.03system 0:02.98elapsed 155%CPU (0avgtext+0avgdata 3158216maxresident)k
7013984inputs+0outputs (15345major+57763minor)pagefaults 0swaps
b3sum --num-threads 1:
1.24user 1.62system 0:07.31elapsed 39%CPU (0avgtext+0avgdata 2326392maxresident)k
5902456inputs+0outputs (3262major+43283minor)pagefaults 0swaps
A massive amount of performance is being left on the table.
Although even so, BLAKE3 is still signficantly faster than SHA2, eg on the same file with SHA256SUM, git-annex fsck:
11.18user 2.11system 0:13.41elapsed 99%CPU (0avgtext+0avgdata 53792maxresident)k
7816912inputs+288outputs (2major+8615minor)pagefaults 0swaps
Using -J with enough concurrent files might approach b3sum's speed, but often a user wants to add or fsck 1 file as fast as possible.
Supporting BLAKE3 multithreading would not be trivial. It needs mmaped file access, and the C++ library oneTBB, which would need to be somehow linked into the haskell library.
Calling out to b3sum when in PATH would be easier.. --Joey