Seer

[Werner]

Seer 1.0 znver1 compile on my PC:


Seer 1.0 skylake compile (installed on my PC):


Seer 1.0 znver2 compile on my PC (AMD nostalgia ):

PS: I wonder, why in Arena I can see the whole analysis only after the thinking time is over , and in command prompt it works fine?

[connor_mcmonigle]

...
PS: I wonder, why in Arena I can see the whole analysis only after the thinking time is over , and in command prompt it works fine?

I'm not entirely sure. It's likely I'm just doing something dumb. My guess is that it's a consequence of my running the text interface on a separate thread from my search threads. The text interface thread checks for new info from the search threads. If new information is available, it presents it to the console, building a new PV string from the TT. This causes depths to be skipped occasionally and inconsistent node counts. This is just all around bad and I plan to change it to be more consistent with other engines soon (likely by having my iterative deepening loop take a callback).

[connor_mcmonigle]

Unrelated observation: The AVX2 compiles (skylake, znver1, znver2) behave differently than the non-AVX2 compiles as a consequence of, from what I can tell, FMA instructions being slightly more accurate in the case of AVX2. The differences are substantial enough to cause differing PVs and evaluations, though unlikely to influence playing strength.

[carldaman]

...
Again the appeal to engine authors - please create a separate thread if introducing a new engine or important updates. The new engines 2020 thread is intended for testers. It would be interesting to see whether AMD's PEXT beats the native C PEXT implementation, most use magic bitboards as fallback, specially if people demand 32-bit binaries
...

Sorry to clutter the new engines 2020 thread. I was uncertain as to whether creating a new thread or posting in the existing new engines thread was preferred. In any case, thanks for your interest in my engine.

No need to say sorry! You did not clutter this thread, which I introduced the same way, as the one for 2019.

Gerd is wrong in the way it would be intended for testers only, it was also created, because of easier doing research
in the future, instead of crawling through 1000s of single posts in the future.
(I am also announcing my compilations of programs sometimes, which are not updated with binaries or just source only)

Ok, but it's no fun to scan through 50++ pages with lot of engine sub-trees.

The CCW engine list is practically dead already, if I once should close my chronology too, other people could create
something reliable new, much easier.

May be I will revive the CPW engine releases page.

OTH Gerd is right, when he asks for additional threads from authors of new programs by themselves.

Thanks

Thanks for thinking of reviving the engines list page, if you can do it. If there were several capable and responsible volunteers allowed to update it, then it would be unlikely that it would be abandoned, as is the case if just one person can make updates. A single individual can suddenly be overwhelmed by outside factors and leave everything nonessential aside.

[Graham Banks]

Seer will not run properly on my old i5. I'm using ChessGUI.
It doesn't display any info for close to 30 seconds and then makes a move that has around depth 7.

[connor_mcmonigle]

Seer will not run properly on my old i5. I'm using ChessGUI.
It doesn't display any info for close to 30 seconds and then makes a move that has around depth 7.

Hmm. Did you download the save.bin weights file from the release page and set the "Weights" UCI option to the absolute path to the weights file? If you could give the name of your i5 (is it the ivy bridge part you mentioned earlier?), that would be helpful as well. Thanks for your time

[Graham Banks]

Seer will not run properly on my old i5. I'm using ChessGUI.
It doesn't display any info for close to 30 seconds and then makes a move that has around depth 7.

Hmm. Did you download the save.bin weights file from the release page and set the "Weights" UCI option to the absolute path to the weights file? If you could give the name of your i5 (is it the ivy bridge part you mentioned earlier?), that would be helpful as well. Thanks for your time

Lynnfield.

[connor_mcmonigle]

Lynnfield.

I see. Thanks Graham. Both the seer_generic and seer_core2 compiles should work in theory...
Have you had any success running it on other machines? I'd also recommend trying another GUI. I've verified my compiles work in Arena and CuteChess on Windows. I tried to use ChessGUI on Windows with Seer just to see if it was possibly an issue specific to that GUI, but ChessGUI would consistently crash before I could even get a chance to install my engine. If I was to take a wild guess as to why it's not working for you, I would guess ChessGUI is failing to set the "Weights" UCI parameter or you don't have the "Weights" parameter configured correctly.

[Werner]

I test NN engines only on my Skylake pcs - not on the older ivybridge. Here is a result of a short testmatch made with 40/10:
1: Naum 4.6 x64 1CPU 14,5/20 111111110110=1====01
2: Seer 1.0NN x64 1CPU 5,5/20 000000001001=0====10
so I think this version is between Deep Shredder 12 and Koivisto 2.0 (very small sample).

[Guenther]

...
Thanks for the binaries. I still would like to create a core2 compile (if you have time for a makefile instead of cmake), may be I can squeeze a bit
more out of Seer for my old quadcore. What flags are used for the generic one currently?
...

Sure. I can add a core2 compile in a little bit. For the makefile, I'm prioritizing exploring a new NNUE idea involving taking the game phase into consideration right now. I'm getting similar accuracy during training with a model which should be twice as fast.

In any case, the performance you're seeing is .... disappointing :( Slower than Leela on a medium to high end GPU.
If you want to compile using MINGW64 on Windows and can't get CMake configured, Seer should compile with:

(from the build directory)

Code: Select all

g++ ../src/seer.cc -o seer -march=native --std=c++17 -O3 -fopenmp -fconstexpr-ops-limit=1000000000 -I ../include/ -static 

...

Thanks! I will try it out today. I am not very optimistic though, because the core2 binary you released now, just adds around 4% in speed.
Obviously missing popcount and/or newer instructions makes a huge difference here (note that this hardware is from 2009).

For better comparison, vanilla SF12 (nnue integrated) official release has around 300kn/s on this machine.