new version: rofChade 3.1

[Guenther]

The SSE2 version works here, thanks.
The SSSE3 version has probably popcnt included too, that's why it fails too run on my old quadcore I guess.

That was a stupid mistake. Hardware popcount support somehow feels much older, but actually it was introduced with SSE4.2...

I've created yet another SSSE3 version with software popcount compile. This version is again downloadable from rofchade.nl (maybe a refresh of the page is needed. Hopefully they all work now.

Ronald

Thanks again Ronald.
Now we have some numbers. As expected by me the SSSE3 one is around 50% faster. (at least from startpos)

Code: Select all

rofChade 3.1 SSE2
...
position startpos
go infinite
info score cp 32 depth 10 seldepth 13 nodes 42722 nps 227244 time 188 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 d2d4 c5d4 d1d4
info score cp 28 depth 11 seldepth 16 nodes 79070 nps 229188 time 345 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 g1f3
info score cp 30 depth 12 seldepth 16 nodes 110942 nps 242761 time 457 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 g1f3 g8f6 d2d4 c5d4 d1d4 e7e6 a2a3 b8c6
info score cp 32 depth 13 seldepth 20 nodes 200898 nps 251436 time 799 tbhits 0 pv e2e4 c7c5 g1f3 d7d6 d2d4 c5d4 f3d4 e7e6 f1e2 g8f6 b1c3
info score cp 24 depth 14 seldepth 23 nodes 413918 nps 255347 time 1621 tbhits 0 pv e2e4 c7c5 g1f3 d7d6 d2d4 c5d4 f3d4 g8f6 b1c3 a7a6
info score cp 22 depth 15 seldepth 27 nodes 913534 nps 259085 time 3526 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3 e7e6 f1e2 g8f6 e1g1 d7d5 e4d5 f6d5 c3d5 d8d5 f1e1
info score cp 36 depth 16 seldepth 26 nodes 1647226 nps 261796 time 6292 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3 e7e6 d2d4 c5d4 f3d4 g8f6 d1d3 f8b4 a2a3 b4a5 c1e3 f6g4
info score cp 44 depth 17 seldepth 29 nodes 2781745 nps 264399 time 10521 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3
info score cp 30 depth 18 seldepth 29 nodes 4816622 nps 267783 time 17987 tbhits 0 pv e2e4 c7c5 b1c3 b8c6 g1f3 e7e5 f1c4 f8e7 d2d3 d7d6 c3d5 g8f6 d5f6 e7f6 c2c3
info score cp 23 depth 19 seldepth 33 nodes 12963207 nps 269242 time 48147 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 b1c3 g8f6 f1b5 c6d4 b5c4 d4f3 d1f3 c7c6 c4b3 f8c5 e1g1 e8g8 d2d3 a7a5 f3g3

Code: Select all

rofChade 3.1 SSSE3
...
position startpos
go infinite
info score cp 32 depth 10 seldepth 13 nodes 42722 nps 328630 time 130 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 d2d4 c5d4 d1d4
info score cp 28 depth 11 seldepth 16 nodes 79070 nps 349867 time 226 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 g1f3
info score cp 30 depth 12 seldepth 16 nodes 110942 nps 369806 time 300 tbhits 0 pv e2e4 c7c5 b1c3 d7d6 g1f3 g8f6 d2d4 c5d4 d1d4 e7e6 a2a3 b8c6
info score cp 32 depth 13 seldepth 20 nodes 200898 nps 389337 time 516 tbhits 0 pv e2e4 c7c5 g1f3 d7d6 d2d4 c5d4 f3d4 e7e6 f1e2 g8f6 b1c3
info score cp 24 depth 14 seldepth 23 nodes 413918 nps 401083 time 1032 tbhits 0 pv e2e4 c7c5 g1f3 d7d6 d2d4 c5d4 f3d4 g8f6 b1c3 a7a6
info score cp 22 depth 15 seldepth 27 nodes 913534 nps 402260 time 2271 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3 e7e6 f1e2 g8f6 e1g1 d7d5 e4d5 f6d5 c3d5 d8d5 f1e1
info score cp 36 depth 16 seldepth 26 nodes 1647226 nps 405820 time 4059 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3 e7e6 d2d4 c5d4 f3d4 g8f6 d1d3 f8b4 a2a3 b4a5 c1e3 f6g4
info score cp 44 depth 17 seldepth 29 nodes 2781745 nps 409682 time 6790 tbhits 0 pv e2e4 c7c5 g1f3 b8c6 b1c3
info score cp 30 depth 18 seldepth 29 nodes 4816622 nps 410064 time 11746 tbhits 0 pv e2e4 c7c5 b1c3 b8c6 g1f3 e7e5 f1c4 f8e7 d2d3 d7d6 c3d5 g8f6 d5f6 e7f6 c2c3
info score cp 23 depth 19 seldepth 33 nodes 12963207 nps 413328 time 31363 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 b1c3 g8f6 f1b5 c6d4 b5c4 d4f3 d1f3 c7c6 c4b3 f8c5 e1g1 e8g8 d2d3 a7a5 f3g3

BTW how much is the difference between AVX2 and AVX512 for your nn size/architecture in Rofchade?

Guenther

[Ronald]

That' s great!

Concerning AVX512, I don't own an AVX512 machine myself, I only have a lousy Threadripper 3990X..

Fortunately Ipman did a few tests with different compiles. On his Intel machine the AVX512 was around 19% faster on the opening position than the BMI2AVX2 version. On his AMD zen4 the AVX512VNNI version was around 10% faster than the BMI2AVX2 version.

[Frank Quisinsky]

Hi Ronald,

a small bug ...
Saw it today ...

Constellation:
rofChade 3.1 three moves before game is won with mate ... my games are without resign mode.
KRQ-K ... mate in 3 moves (syzygybases-5).

OK, rofchade gave the queen and set mate in 10 moves with KR-K.
So, move average goes higher, game need longer ... mate distance should be checked!!

Not very important but if you have time for it ...

Best and congratulation for v3.1, clearly improved.



Frank