My Best-First-Approach on the GPU plays pretty bad for the amount of searched nodes...so i will switch to an Depth-First-Search for the next release.
I will couple 16 to 48 threads (one warp) to work on one chess position in parallel and will try to implement YBWC across 128 Warps. Still not sure if this will work.
One Warp on AMD HD7750 is capable to reach max. 40 Knps.
The current Best-First-Approch with Castle and EnPassant moves is online at github:
https://github.com/smatovic/Zeta/tree/zeta_cl_096890
...it makes 1.5 Mnps on Quad-Core-CPU and 1.0 Mnps on HD7750 GPU,
windows binaries are available.
--
Srdja
Zeta OpenCL Chess, switch from Best-First to Depth-First....
Moderators: hgm, Rebel, chrisw
-
smatovic
- Posts: 2658
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
-
smatovic
- Posts: 2658
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Zeta OpenCL Chess, switch from Best-First to Depth-First
followup:
coupling 16 threads together was an bad idea for Quiscence-Search, Performance falls < 20 Knps per Warp
But i found some bugs in my Move Generation....so back to Best-First.
--
Srdja
coupling 16 threads together was an bad idea for Quiscence-Search, Performance falls < 20 Knps per Warp
But i found some bugs in my Move Generation....so back to Best-First.
--
Srdja
-
jshriver
- Posts: 1342
- Joined: Wed Mar 08, 2006 9:41 pm
- Location: Morgantown, WV, USA
Re: Zeta OpenCL Chess, switch from Best-First to Depth-First
I really enjoy your posts and the progress keep it up.