Damage 15.3

Discussion about development of draughts in the time of computer and Internet.
Post Reply
BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Wed May 27, 2020 14:17

Could you run a test with 1 core and search depth 26, and post a picture.

1 core , change in ini file.
in the engine type:
level depth=26
go

I believe that I suffer much from the PEXT implementation.
But at least now the engine recognizes an AMD processor.
So I will work the next days in replacing the PEXT (think Ed did the same).
Will post a new version and will let you know.

Before you run a match with the new version (think in the weekend im ready with this one), I hope you can repeat this test (so depth=26 with 1 core), so you should see a speed increase.
Hope then we will go below the 20 ELO difference, but still Scan and Kingsrow are in another league.

After the PEXT is solved, I will work on the opening book.

Bert

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Wed May 27, 2020 14:35

I understand and thank you, Bert. I am asking for what you are asking for.
Attachments
Damage.jpg
Damage.jpg (165.68 KiB) Viewed 1932 times

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Wed May 27, 2020 15:04

Thanks , think you posted a similar picture in the past.
So not a perft but a normal 26 ply search, with 1 core, so i know the base speed with 1 core and this implementation on your AMD machine.

Bert

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Fri May 29, 2020 09:21

I posted yesterday a new Engine version.
This version detects if the processor is AMD or Intel, and in case of AMD a specific PEXT implementation/emulation is used.
A first test on the latest Threadripper (with 32 cores), reveals a speed of 200+ MNodes/seconds.
Maybe this value could be further increased with Lazy SMP,

Bert

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Sat May 30, 2020 11:33

Thank you Bert.

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Sat May 30, 2020 14:52

Krzysztof, your welcome.

I'm interested to see the results on you computer when Damage plays against Kingsrow, when Damage uses all available 16 cores.
Both a match of 1 min / 75 moves as 2 min / 75 moves.

Bert

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Sat May 30, 2020 15:33

Ok.

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Sat May 30, 2020 20:19

1 minute / 75 moves

Damage 15.3 x64 - Kingsrow 1.61 x64 0 wins, 5 losses, 153 draws, 0 unknowns
Attachments
dxpgames.pdn
1 minute / 75 moves - 16 threads
(151.87 KiB) Downloaded 31 times

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Sun May 31, 2020 06:34

2 minute / 75 moves

Damage 15.3 x64 - Kingsrow 1.61 x64 0 wins, 4 losses, 154 draws, 0 unknowns
Attachments
dxpgames.pdn
2 minute / 75 moves - 16 cores
(148.24 KiB) Downloaded 33 times

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Sun May 31, 2020 11:37

Krzysztof, thanks for your support.
I analyzed the 4 lost games from the 2 min match, see analysis graph below.
game analysis.PNG
game analysis.PNG (6.36 KiB) Viewed 1678 times
The red line is game 53, here you see a sudden drop in evaluation (between move 25 - 30), so definitely something went wrong, so here i need to dig deeper to understand.

Green line is game 80, quite typical behavior, Damage has most likely a worse position, which is does not understand, and from move 20 you see the effects.

Basically the same pattern you see in game 135 (blue line), and game 149 (yellow line), in both cases one enters into positions with king(s) on both sides each. I recognize that my evaluation does not really know how to handle these.

So still an awful lot of work to do, both Scan as Kingsrow put the bar extremely high.
Compliments to Fabien and Ed.

Bert

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Sun May 31, 2020 23:03

I uploaded an improved Engine.
The only difference is better performance on an AMD processor.
Further no changes.
Test results with a 32 core Threadripper to be expected soon.

Bert

https://www.dropbox.com/sh/pxqa6zc8bu5o ... WOqua?dl=0

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Mon Jun 01, 2020 07:03

Let me understand if there are any changes since the last download damageengine153.exe

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Mon Jun 01, 2020 17:33

Krzysztof, no changes only a slighter faster emulation of the PEXT instruction, in case the program runs on an AMD processor.
I expect no boost in real performance, so no need for testing.
On a 32 core Threadripper the maximum Damage speed measured was 240 Mega Nodes/second.

Bert

Krzysztof Grzelak
Posts: 927
Joined: Thu Jun 20, 2013 17:16
Real name: Krzysztof Grzelak

Re: Damage 15.3

Post by Krzysztof Grzelak » Mon Jun 01, 2020 17:43

And I played one match.

1 minute / 75 moves.

Damage 15.3 x64 - Kingsrow 1.61 x64 1 wins, 4 losses, 153 draws, 0 unknowns
Attachments
dxpgames.pdn
Damage - Kingsrow - 16cores
(150.97 KiB) Downloaded 33 times

BertTuyt
Posts: 1435
Joined: Wed Sep 01, 2004 19:42

Re: Damage 15.3

Post by BertTuyt » Mon Jun 01, 2020 18:48

Krzysztof thanks for testing.
I will now more focus on the opening book.

Bert

Post Reply