New engine: Clover

[Damir]

Luca,

Thank you a lot for such a nice engine. I hope you continue in improving it.

[lucametehau]

After a long break, Clover 3.0 is here: https://github.com/lucametehau/CloverEn ... s/tag/v3.0. It is around 70 elo better in self-play, 40-50 against other engines. This version uses NNUE for evaluating positions. More details about the network on the release page.

As usual, tell me if you have any problems!

[Gabor Szots]

Hi Luca,

Using this command line:

c++ *.cpp tbprobe.c -DEVALFILE=\"$(Clover_323mil_e33_256.nn)\" -std=c++17 -flto -pthread -O3 -march=native -oClover_3.0-x64-SzG.exe

I get the following error message:

In file included from net.h:2,
from board.h:6,
from init.h:3,
from main.cpp:2:
net.h:11:13: error: expected ')' before 'EVALFILE'
11 | INCBIN(Net, EVALFILE);
| ^~~~~~~~
incbin.h:352:31: note: in definition of macro 'INCBIN'
352 | INCBIN_MACRO "\"" FILENAME "\"\n" \
| ^~~~~~~~

What am I doing wrong? With the makefile I also get error messages (see below) but at least it does compile.

In file included from tbprobe.c:137:
tbcore.c: In function 'void unmap_file(char*, uint64)':
tbcore.c:226:52: warning: format '%d' expects argument of type 'int', but argument 3 has type 'DWORD' {aka 'long unsigned int'} [-Wformat=]
226 | fprintf(stderr, "unmap failed, error code %d", GetLastError());
| ~^ ~~~~~~~~~~~~~~
| | |
| int DWORD {aka long unsigned int}
| %ld
tbcore.c:230:58: warning: format '%d' expects argument of type 'int', but argument 3 has type 'DWORD' {aka 'long unsigned int'} [-Wformat=]
230 | fprintf(stderr, "CloseHandle failed, error code %d", GetLastError());
| ~^ ~~~~~~~~~~~~~~
| | |
| int DWORD {aka long unsigned int}
| %ld

[lucametehau]

Hi Gabor.
Clover can't be compiled from command line, because of this bit

-DEVALFILE=\"Clover_323mil_e33_256.nn\"

which only works if compiled from makefile.

As for compiling using makefile, those "errors" are warnings which are generated by the Fathom code. I'll try resolving them, but in the mean time you can ignore them.

[Gabor Szots]

Hi Gabor.
Clover can't be compiled from command line, because of this bit

-DEVALFILE=\"Clover_323mil_e33_256.nn\"

which only works if compiled from makefile.

As for compiling using makefile, those "errors" are warnings which are generated by the Fathom code. I'll try resolving them, but in the mean time you can ignore them.

I'm surprised about the first part.

As for the makefile, it only generates one exe. I expected 5.

[lucametehau]

Well I tested Bit-Genie as well to see if it can be compiled by command line and it seems to have the same problem.

For compiling with makefile, if you want 5 exes to be created, use

make run

[Sylwy]

After a long break, Clover 3.0 is here: https://github.com/lucametehau/CloverEn ... s/tag/v3.0. It is around 70 elo better in self-play, 40-50 against other engines. This version uses NNUE for evaluating positions. More details about the network on the release page.

As usual, tell me if you have any problems!

Bravo, Luca ! I was sure you could do it!
A single bewilderment regarding the compilations you provided. The bmi2 version is slower than the avx2 version on my Haswell CPU. It is the only case among many engines!

[lucametehau]

Thanks Sylwy.

As for why the bmi2 is slower than the avx2, I have no idea. If you used the compiles from the release page, one cause might be that I bought a new laptop, with a new CPU (AMD Ryzen 7 5800H).

[Sylwy]

Thanks Sylwy.

As for why the bmi2 is slower than the avx2, I have no idea. If you used the compiles from the release page, one cause might be that I bought a new laptop, with a new CPU (AMD Ryzen 7 5800H).

Clear. Then the native one is faster.

[Gabor Szots]

For compiling with makefile, if you want 5 exes to be created, use

make run

Thanks. Odd that my native is slower than my BMI2.