Page 1 of 2
Laser 1.6 Release
Posted: Sat Jul 14, 2018 9:28 pm
by jeffreyan11
Laser 1.6 has now been released and is available at
https://github.com/jeffreyan11/uci-ches ... s/tag/v1.6
Self-play regressions:
+151 in 12000 games at 3+0.02
+151 in 12000 games at 15+0.05
Estimated CCRL elo is 3180
Major changes:
Search and evaluation tweaks
SMP rewrite
Time management rewrite
Re: Laser 1.6 Release
Posted: Sat Jul 14, 2018 9:52 pm
by Guenther
Thanks Jeffrey. That selfplay improvement looks gorgeous.
Re: Laser 1.6 Release
Posted: Sat Jul 14, 2018 10:26 pm
by AndrewGrant
Thanks, Jeffrey.
My testing says your last few time control changes have a big impact at STC
Re: Laser 1.6 Release
Posted: Sat Jul 14, 2018 10:26 pm
by CMCanavessi
Great release, awesome improvement over 1.5!
Will definitely get to the Elite League and stay there this season... hopefully.
Re: Laser 1.6 Release
Posted: Sat Jul 14, 2018 11:00 pm
by Graham Banks
Thanks Jeffrey. Fantastic progress!
Re: Laser 1.6 Release
Posted: Sun Jul 15, 2018 10:08 am
by Dokterchen
Awesome! Thanks Jeffrey
Re: Laser 1.6 Release
Posted: Sun Jul 15, 2018 11:05 am
by Dariusz
Good news! Thank You Jeffrey
Re: Laser 1.6 Release
Posted: Sun Jul 15, 2018 12:04 pm
by Damir
Awesome work. Thx Jeffrey.
Re: Laser 1.6 Release
Posted: Sun Jul 15, 2018 4:59 pm
by CMCanavessi
jeffreyan11 wrote: ↑Sat Jul 14, 2018 9:28 pm
Laser 1.6 has now been released and is available at
https://github.com/jeffreyan11/uci-ches ... s/tag/v1.6
Self-play regressions:
+151 in 12000 games at 3+0.02
+151 in 12000 games at 15+0.05
Estimated CCRL elo is 3180
Major changes:
Search and evaluation tweaks
SMP rewrite
Time management rewrite
Hi Jeffrey, I've benchmarked all 3 versions and something looks suspicious with the BMI2 and the Popcount version. They have almost the same speed, which on a Ryzen processor is kind of nonsense due to the BMI2 bug, so either the BMI2 is not using BMI2, or the Popcount version is also using BMI2 and running slower than it should. The Nopopcount version is the slower of them all, as expected. Usually with other engines, the Popcount version is much faster. Maybe you can check that, or others can also test and see if they find the same?
I'll post benchmarks later.
Re: Laser 1.6 Release
Posted: Sun Jul 15, 2018 7:15 pm
by AndrewGrant
CMCanavessi wrote: ↑Sun Jul 15, 2018 4:59 pm
jeffreyan11 wrote: ↑Sat Jul 14, 2018 9:28 pm
Laser 1.6 has now been released and is available at
https://github.com/jeffreyan11/uci-ches ... s/tag/v1.6
Self-play regressions:
+151 in 12000 games at 3+0.02
+151 in 12000 games at 15+0.05
Estimated CCRL elo is 3180
Major changes:
Search and evaluation tweaks
SMP rewrite
Time management rewrite
Hi Jeffrey, I've benchmarked all 3 versions and something looks suspicious with the BMI2 and the Popcount version. They have almost the same speed, which on a Ryzen processor is kind of nonsense due to the BMI2 bug, so either the BMI2 is not using BMI2, or the Popcount version is also using BMI2 and running slower than it should. The Nopopcount version is the slower of them all, as expected. Usually with other engines, the Popcount version is much faster. Maybe you can check that, or others can also test and see if they find the same?
I'll post benchmarks later.
Laser does not have BMI2/PEXT support in the sense that most expect. Usually this implies replacing the multiplication shift and step of the magic bitboard computation with a single pext instruction. Jeffrey has not done this.
Here is this BMI2 support patch
, which just assumes you have haswell architecture, which is still a speedup I presume.