These are chat archives for ujh/iomrascalai

24th
Mar 2015
iopq
@iopq
Mar 24 2015 02:19
How did you measure the iterations per second in the threaded version?
iopq
@iopq
Mar 24 2015 03:48
25% better performance or thereabouts
in my branch right now
iopq
@iopq
Mar 24 2015 06:09
even further performance improvements
Urban Hafner
@ujh
Mar 24 2015 06:19
:+1:
I'm currently on my way to work so I can't really check the code (I only have a phone). I'll take a closer look later today.
Urban Hafner
@ujh
Mar 24 2015 06:36
I already made a few comments. I'll test the speed up on the reference machine (i.e. my laptop :P) later. And as I don't like those incline things at all I'll see if they even make a difference.
iopq
@iopq
Mar 24 2015 06:46
probably don't make a difference now as much because I don't use the same functions
Urban Hafner
@ujh
Mar 24 2015 07:44
So the performance of the multi threaded playouts: I didn’t really measure those as that’s a bit hard to do (right now). I just looked at the debug output (how many simulations were run) of the first move on a 9x9 board using the default time limits.
And sorry about being such a PITA about the code. I think I’ll just merge it now.
Urban Hafner
@ujh
Mar 24 2015 12:10
BTW, it may be of interest to you to see the performance of the playouts over time. I’ve kept records on the speed of the benchmarks over time here: https://www.icloud.com/numbers/AwBUCAESEA6puplN_B3iG6utIDWz0z0aKbkFBY5jdPSFukaEuQZ2KoozroDiZ2GyHjV6mrO8hIMkAtogEuldWQiCMCUCAQEEIKUIwP55STKkis5HQ1pXtzRjJ6SF0lIbuAny9YFCCIAB#Iomrascálaí
iopq
@iopq
Mar 24 2015 18:01
There were faster playouts on 19x19 before? I wonder why. They're very old versions, I can't test them.
iopq
@iopq
Mar 24 2015 18:51
oh because of the overflow bug
Urban Hafner
@ujh
Mar 24 2015 19:11
Exactly. 19 x 19 x 3 is too big for u8 so they were cut short.