Stockfish 210720 with NNUE

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Carlos777
Posts: 693
Joined: Sun Dec 13, 2009 5:09 pm

Re: Stockfish 210720 with NNUE

Post by Carlos777 » Thu Jul 23, 2020 9:35 pm

Thank you!
Most of popcount engines compiles work on my old Phenom II x4, but this does not :(
Could you make a compile for old PCs like mine?

Thanks in advance.

PS: I am using this compile: stockfish.sse2.halfkp_256x2-32-32.nnue-learn.2020-07-19 (I think I downloaded it from nodchip)

kranium wrote:
Thu Jul 23, 2020 12:24 am
Hi all-

I got a nice speed increase (~8-10%) for these aio (all-in-one) builds:
https://github.com/FireFather/sf-nnue/r ... 07-23-2020

peter
Posts: 2199
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Stockfish 210720 with NNUE

Post by peter » Thu Jul 23, 2020 9:54 pm

kranium wrote:
Thu Jul 23, 2020 11:20 am
Tried aio popcnt- compile with nnue-GUI 1.4.
Worked well to the point to validation with 24 threads chosen, task manager confirms 24 threads used, yet learning itself runs with single core only, in config 24 threads are noted, task manager shows one thread only.
390Mb data.bin will take very long single- threaded, won't it?
Peter.

User avatar
kranium
Posts: 1922
Joined: Thu May 29, 2008 8:43 am

Re: Stockfish 210720 with NNUE

Post by kranium » Thu Jul 23, 2020 10:21 pm

peter wrote:
Thu Jul 23, 2020 9:54 pm
kranium wrote:
Thu Jul 23, 2020 11:20 am
Tried aio popcnt- compile with nnue-GUI 1.4.
Worked well to the point to validation with 24 threads chosen, task manager confirms 24 threads used, yet learning itself runs with single core only, in config 24 threads are noted, task manager shows one thread only.
390Mb data.bin will take very long single- threaded, won't it?
The learning/training runs in cycles...for ex: I've experienced the CPU maxing out for a short time then retreating (over and over) until nn.bin is built.
If task manager indicates 24 threads used, I think it's accurate and the process is busy.
Generating the validation data generally takes less time than generating the training data (which normally take many hours, a day or even more)

You might just have to be patient and let it finish...

peter
Posts: 2199
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Stockfish 210720 with NNUE

Post by peter » Thu Jul 23, 2020 10:53 pm

kranium wrote:
Thu Jul 23, 2020 10:21 pm
The learning/training runs in cycles...for ex: I've experienced the CPU maxing out for a short time then retreating (over and over) until nn.bin is built.
If task manager indicates 24 threads used, I think it's accurate and the process is busy.
You got me wrong, Norman, task manager indicated 24 threads in use only as long as process to end of validation longed, training itself now is already running longer than validation was (about 4 hours with gensfen in all) for quite a time (about 5 hours after end of validation)
As long as learning goes on now, always only 1 thread is busy according to task manager. Of course difference is to be heard as for van running much less too.
Peter.

User avatar
kranium
Posts: 1922
Joined: Thu May 29, 2008 8:43 am

Re: Stockfish 210720 with NNUE

Post by kranium » Fri Jul 24, 2020 1:40 am


User avatar
Ovyron
Posts: 4410
Joined: Tue Jul 03, 2007 2:30 am

Re: Stockfish 210720 with NNUE

Post by Ovyron » Fri Jul 24, 2020 3:30 am

kranium wrote:
Fri Jul 24, 2020 1:40 am
https://github.com/FireFather/sf-nnue/r ... 07-24-2020
CB/Fritz compatibility ?!
Hi Kranium, when trying to compile with the command I use to compile regular stockfish I get an error:

make profile-build ARCH=x86-64 COMP=mingw

Code: Select all

make[2]: Entering directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
g++ -Wall -Wcast-qual -fno-exceptions -std=c++17 -fprofile-generate  -Wextra -Ws
hadow -DNDEBUG -O3 -DIS_64BIT -msse -DUSE_SSE2   -c -o benchmark.o benchmark.cpp

g++.exe: error: unrecognized command line option '-std=c++17'
make[2]: *** [benchmark.o] Error 1
make[2]: Leaving directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
make[1]: *** [gcc-profile-make] Error 2
make[1]: Leaving directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
make: *** [profile-build] Error 2

zullil
Posts: 6442
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: Stockfish 210720 with NNUE

Post by zullil » Fri Jul 24, 2020 6:25 am

Ovyron wrote:
Fri Jul 24, 2020 3:30 am
kranium wrote:
Fri Jul 24, 2020 1:40 am
https://github.com/FireFather/sf-nnue/r ... 07-24-2020
CB/Fritz compatibility ?!
Hi Kranium, when trying to compile with the command I use to compile regular stockfish I get an error:

make profile-build ARCH=x86-64 COMP=mingw

Code: Select all

make[2]: Entering directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
g++ -Wall -Wcast-qual -fno-exceptions -std=c++17 -fprofile-generate  -Wextra -Ws
hadow -DNDEBUG -O3 -DIS_64BIT -msse -DUSE_SSE2   -c -o benchmark.o benchmark.cpp

g++.exe: error: unrecognized command line option '-std=c++17'
make[2]: *** [benchmark.o] Error 1
make[2]: Leaving directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
make[1]: *** [gcc-profile-make] Error 2
make[1]: Leaving directory `/c/Chess/S/N/sf-nnue-07-24-2020/src'
make: *** [profile-build] Error 2
What version of g++? Yours must be quite old.

What do you get if you enter the following at a command line?

g++.exe -v

User avatar
M ANSARI
Posts: 3500
Joined: Thu Mar 16, 2006 6:10 pm

Re: Stockfish 210720 with NNUE

Post by M ANSARI » Fri Jul 24, 2020 7:16 am

Actually I had something really weird happen with the CB GUI ... finally one of the engines seemed to load without any exception errors ... but then the evaluation kept showing 0.00 in positions where it is obviously a huge advantage for one side ... then when mate shows up it shows a mate score (I guess the GUI EGTB takes over). I think I'll wait a few more days to let the kinks smoothed out but it is a very exciting engine and going through some of the games it is really a game changer!

peter
Posts: 2199
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Stockfish 210720 with NNUE

Post by peter » Fri Jul 24, 2020 8:11 am

M ANSARI wrote:
Fri Jul 24, 2020 7:16 am
Actually I had something really weird happen with the CB GUI ... finally one of the engines seemed to load without any exception errors ... but then the evaluation kept showing 0.00 in positions where it is obviously a huge advantage for one side ... then when mate shows up it shows a mate score (I guess the GUI EGTB takes over). I think I'll wait a few more days to let the kinks smoothed out but it is a very exciting engine and going through some of the games it is really a game changer!
The 0.00- syndrome I know from a case of mine in which I tried to have a 256x- Binary run a 384- net. If nn.bin isn't found by the engine at all, it could be the same behaviour. Have a look at the bigness of the net, it's file- name and the path to it given at UCI- options.

Binaries latest but one from Norman (popcnt) work well for me with chessbase- GUI 12 and with Fritz 14. Didn't try the most recent ones.

BTW the all in one- binary together with nn-gui 1.4 did finish training of a new 20Mb- net overnight (gensfen and validation were togehter about 4 hours), even if learning itself seemed to work single- threaded only, till end of validation 24 threads were used for sure.

Even if still message "running" was there, something like "eval save file folder final finished" was to be read too and the CPU wasn't working anymore, so I guess it might be ok., nn.bin in folder final (7th of 7, 0-5 and "final") 20Mb big.

Thanks again, Norman, seems to have worked, even if I still think learning after validation uses 1 thread only.
Peter.

User avatar
Ovyron
Posts: 4410
Joined: Tue Jul 03, 2007 2:30 am

Re: Stockfish 210720 with NNUE

Post by Ovyron » Fri Jul 24, 2020 6:51 pm

zullil wrote:
Fri Jul 24, 2020 6:25 am
What version of g++? Yours must be quite old.

What do you get if you enter the following at a command line?

g++.exe -v
Yeah, it's from 2014...

I get:

gcc version 4.9.2 (x86_64-posix-seh-rev1, Built by MinGW-W64 project)

However, this compiles Stockfish dev just fine. Do I need to upgrade or can I make changes to source so it compiles with this one?

Post Reply