Possible Board Presentation and Move Generation for GPUs?

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
smatovic
Posts: 952
Joined: Wed Mar 10, 2010 9:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic
Contact:

Re: Possible Board Presentation and Move Generation for GPUs

Post by smatovic » Sun Mar 20, 2011 12:24 pm

No idea with GPUs. Do they have 64-bit shifts with 128-bit SIMD vectors? If yes, I guess Kogge-Stone is the way to go on these architectures.
Yes, they have, i guess i have to get closer into Kogge-Stone, thanks Steffan and Gerd.

--
Srdja

smatovic
Posts: 952
Joined: Wed Mar 10, 2010 9:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic
Contact:

Re: Possible Board Presentation and Move Generation for GPUs

Post by smatovic » Tue Mar 22, 2011 3:35 pm

I'd strongly recommend you take a look a look at my method inspired by Kogge-Stone parallel prefix networks,
Great, i think i got it...

i could use 8 Threads/Processes in parallel, for each direction one, just have to care for the non-sliding pieces...thanks.

--
Srdja

bob
Posts: 20642
Joined: Mon Feb 27, 2006 6:30 pm
Location: Birmingham, AL

Re: Possible Board Presentation and Move Generation for GPUs

Post by bob » Tue Mar 22, 2011 5:24 pm

smatovic wrote:
I'd strongly recommend you take a look a look at my method inspired by Kogge-Stone parallel prefix networks,
Great, i think i got it...

i could use 8 Threads/Processes in parallel, for each direction one, just have to care for the non-sliding pieces...thanks.

--
Srdja
That's one approach. Or a GPU per square, ala' HiTech, which was somewhat based on Belle.

smatovic
Posts: 952
Joined: Wed Mar 10, 2010 9:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic
Contact:

Re: Possible Board Presentation and Move Generation for GPUs

Post by smatovic » Mon Mar 28, 2011 1:23 pm

That's one approach. Or a GPU per square, ala' HiTech, which was somewhat based on Belle.
Thanks for the advice,

i will try a KoggeStone Move Generator (with extension for all pieces) with 64x64 threads next.

--
Srdja

smatovic
Posts: 952
Joined: Wed Mar 10, 2010 9:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic
Contact:

Re: Possible Board Presentation and Move Generation for GPUs

Post by smatovic » Thu Jun 16, 2011 8:15 pm

For the record:

i have implemented a Kogge-Stone move generator with 8 Threads (8 directions) in parallel on a GPU but it could not outperform a Magic-Bitboard Move Generator with only one thread.

So i will use a Quad-Bitboard Board Representation with Magic Bitboards for sliders and precalculated AttackTables for non-sliders in future.

--
Srdja

PS: Zeta project is reopened again:
http://zeta-chess.blogspot.com/

Post Reply