There was actually some bugs in perft function. The total perft suite is actually: 32,362,126,796 nodes.
Time management is greatly simplified.
AiLevel was much improved.
Now AiLevel is much better. Some quick tests.
AiLevel 0 as random mover loses all games as random mover should.
Level 1/100 scores 2%.
10/100 scores 9%. etc.
Seems to work great now.
I wanted to check has there been any progress in terms of strength. It is faster and less buggy that's I know without testing.
I ran a little tournaments against older versions using Chess960 games. I test against Stockfish just to see there's no illegal moves/crashes.
Yes, the latest version is the strongest as it should be.
JohnWoe wrote: ↑Wed Nov 06, 2019 11:41 pm
I wanted to check has there been any progress in terms of strength. It is faster and less buggy that's I know without testing.
I ran a little tournaments against older versions using Chess960 games. I test against Stockfish just to see there's no illegal moves/crashes.
Yes, the latest version is the strongest as it should be.
Gabor Szots wrote: ↑Thu Nov 07, 2019 11:31 am
...
Still no chance for a Windows compile?
I created one for my architecture only, old quadcore (yorkfield), no popcount, SSE 4.1.
If you are interested I can add it here, (if talkchess has no hickups anymore for connection)
(connection problems since around 15 minutes from here)
Gabor Szots wrote: ↑Thu Nov 07, 2019 11:31 am
...
Still no chance for a Windows compile?
I created one for my architecture only, old quadcore (yorkfield), no popcount, SSE 4.1.
If you are interested I can add it here, (if talkchess has no hickups anymore for connection)
(connection problems since around 15 minutes from here)
Yes, I am interested. And sooner or later I may learn how to compile it myself.
Gabor Szots wrote: ↑Thu Nov 07, 2019 11:31 am
...
Still no chance for a Windows compile?
I created one for my architecture only, old quadcore (yorkfield), no popcount, SSE 4.1.
If you are interested I can add it here, (if talkchess has no hickups anymore for connection)
(connection problems since around 15 minutes from here)
Yes, I am interested. And sooner or later I may learn how to compile it myself.
Well, the quota for adding files is very low?
(140KB was too much already)
Sorry, the board attachment quota has been reached.
I have those other compilers installed on my laptop so I'll add Android + other builds later.
With this release I profiled and profiled Sapeli. And squeezed every NPS out I could. It is fast.
Some tuning to LMR. Slightly more aggressive pruning. Split search into smaller pieces.
1.65 searches a bit deeper than 1.64. But that wasn't the goal.
Source code polishing.
Strength wise on par with 1.64. Slightly stronger on my tests.
Compiled a few flavours for 1.65, with a big caveat, because I had never built versions before, I cannot run here myself,
which is a bit awkward.
Would be nice if someone could try the bmi or bmi2 version, if those run at all and with what speed.
The other two are are tested here and behave very similar at least on my old machine.