I'm getting ~40% speedup on my machine using tevador's x25519 implementation for view tag scanning compared to cryptonote view tag scanning (i.e. 1) compute DH derivation, 2) compute view tag). A small amount of that is from using blake2b for view tags instead of keccak (<5%).

plus it should allow significant speedup for mobile devices since they don't have any assembler optimized code currently

very cool

So what would the total speedup be using x25519 compared to no view tags?

not that much, since there are a bunch of steps after view tags

but the view tag filter handles 99.6% of enotes on-chain

oh you mean compared to pre-view tags? about 55% per enote on my machine