Page 19 of 25
Re: A complete 2000 lines of code engine
Posted: Fri Dec 13, 2019 8:06 pm
by xr_a_y
Minic 1.19 is now released
I shall be a little stronger due to some history (included CMH) tuning
Some UCI fixes about pondering, but probably still not ok
Re: A complete 2000 lines of code engine
Posted: Wed Dec 18, 2019 11:32 am
by xr_a_y
The given figure is Minic SMP scaling in elo from 1 to 8 threads on a i7-9700K using only 256Mb hash.
What are your opinion on it ?
Code: Select all
1 minic_dev_8_threads 94 24 303 63.2% 61.1%
2 minic_dev_7_threads 70 22 303 59.9% 65.7%
3 minic_dev_6_threads 66 24 303 59.4% 61.4%
4 minic_dev_5_threads 25 22 302 53.6% 66.9%
5 minic_dev_4_threads 10 24 302 51.5% 63.9%
6 minic_dev_3_threads -24 23 302 46.5% 64.6%
7 minic_dev_2_threads -74 25 302 39.6% 57.9%
8 minic_dev -180 28 303 26.2% 45.9%
Re: A complete 2000 lines of code engine
Posted: Wed Dec 18, 2019 11:40 am
by Guenther
xr_a_y wrote: ↑Wed Dec 18, 2019 11:32 am
The given figure is Minic SMP scaling in elo from 1 to 8 threads on a i7-9700K using only 256Mb hash.
What are your opinion on it ?
Code: Select all
1 minic_dev_8_threads 94 24 303 63.2% 61.1%
2 minic_dev_7_threads 70 22 303 59.9% 65.7%
3 minic_dev_6_threads 66 24 303 59.4% 61.4%
4 minic_dev_5_threads 25 22 302 53.6% 66.9%
5 minic_dev_4_threads 10 24 302 51.5% 63.9%
6 minic_dev_3_threads -24 23 302 46.5% 64.6%
7 minic_dev_2_threads -74 25 302 39.6% 57.9%
8 minic_dev -180 28 303 26.2% 45.9%
Is the third column the average error bar?
Re: A complete 2000 lines of code engine
Posted: Wed Dec 18, 2019 1:19 pm
by CMCanavessi
To me it looks like elo, error, games played, % of points, % of draws.
Re: A complete 2000 lines of code engine
Posted: Wed Dec 18, 2019 1:32 pm
by Alayan
xr_a_y wrote: ↑Wed Dec 18, 2019 11:32 am
The given figure is Minic SMP scaling in elo from 1 to 8 threads on a i7-9700K using only 256Mb hash.
What are your opinion on it ?
This data is not very useful, because what really matters is how single-core with T time compare to N threads with T/N time (keeping the total hardware time the same). Giving the same time to different threads count creates an extra dependency on how much strength Minic gains with extra time at the bullet TC you used.
The ideal measure would be "time to elo" curves, but that requires more effort to get that T - 1 vs T/N - N.
Re: A complete 2000 lines of code engine
Posted: Wed Dec 18, 2019 6:32 pm
by xr_a_y
CMCanavessi wrote: ↑Wed Dec 18, 2019 1:19 pm
To me it looks like elo, error, games played, % of points, % of draws.
Exactly, cutechess output, sorry for the missed header.
Re: A complete 2000 lines of code engine
Posted: Sat Dec 28, 2019 10:18 am
by xr_a_y
I've just released Minic 1.21 with some fixes (UCI support shall be better/good ?).
This is the version that may be used for TCEC in a few days
Re: A complete 2000 lines of code engine
Posted: Sun Dec 29, 2019 8:16 pm
by xr_a_y
Late patch for TCEC : Minic 1.22 is available
Not well tested versus other engines
, mainly intuition ... so I keep fingers crossed ...
Re: A complete 2000 lines of code engine
Posted: Mon Dec 30, 2019 5:45 pm
by CMCanavessi
Did you fix ponder or it's still on the to-do list?
Re: A complete 2000 lines of code engine
Posted: Mon Dec 30, 2019 6:01 pm
by xr_a_y
CMCanavessi wrote: ↑Mon Dec 30, 2019 5:45 pm
Did you fix ponder or it's still on the to-do list?
Ponder UCI option is here.
"Ponder move" is displayed but I'm still not sure how to handle ponderhit.
I see other engines going from "infinite mode" to "timed mode" and continue to search, but I don't get that.
So for now I just stop search on ponderhit and wait for another go command.
I feel silly about that.
After reading
http://talkchess.com/forum3/viewtopic.p ... 9&start=50, I am even more lost.