Getting the hash near full hurts strength https://github.com/official-stockfish/Stockfish/wiki/Useful-data#elo-cost-of-small-hash
So maybe try to avoid long thinks if under hash pressure. And increase typical time usage to compensate.
#Factor hashfull into TM
shell sonnet
shell sonnet
Test or commit link?
Also it sounds hard to get a passer on standard tests. IIRC there is very little hash pressure with default hash size.
drowsy folio
Iirc default hash pressure at stc is like sub 200
Which is where performance shows no penalty
digital girder
Which is like why they failed
drowsy folio
Yep, basically a no op
high remnant
Create a test with lower hash then
shell sonnet
Yeah, I would be tempted to argue for a non-regression test for the standard hash and larger hash but require a gainer on low hash. But I think you would expect at least one yellow for 5 tests...
For hashfull(-2-3) I don't think this method is quite right. If there is a bonus for low hashfull then even if we get nonreg on standard hash we might regress on 4x hash. And if there is only penalty for high hashfull then that shifts time usage towards endgame. In my mind TM is "right" for low hash pressure.
I like the idea of lastHashfull though. I feel like this has potential for the tm version. Maybe some sort of rolling average from previous moves. I think the longest searches need to be penalized and the the rest need to be increased to not shift time to the endgame.
Adding links for future:
https://tests.stockfishchess.org/tests/view/642d98c377ff3301150dfb3e
https://tests.stockfishchess.org/tests/view/642c4c0c77ff3301150da16e
https://tests.stockfishchess.org/tests/view/642c507e77ff3301150da267
https://tests.stockfishchess.org/tests/view/642c513377ff3301150da287
https://tests.stockfishchess.org/tests/view/642d99e577ff3301150dfc45
https://tests.stockfishchess.org/tests/view/642d9b9177ff3301150dfd24