Gabor Szots wrote: ↑Mon Feb 22, 2021 9:24 am
Thanks, it's OK now.
BTW Gabor, could you show, if the output is different for your hardware too, compared to the old official 1.51 version?
(Best again comparing both outputs for FEN and moves for the same end position - see start of thread)
Does the avx2+bmi2 run on your machine? That's the only popcount I have compiled.
Here are the outputs using your fixAVX compile. It seems it has really been fixed. The only problem is that the exe is slower than the original. I compiled it too but that is slower as well.
Gabor Szots wrote: ↑Mon Feb 22, 2021 1:12 pm
Here are the outputs using your fixAVX compile. It seems it has really been fixed. The only problem is that the exe is slower than the original. I compiled it too but that is slower as well.
...
position startpos moves g1f3 d7d5 d2d3 g8f6 g2g3 c8g4
go wtime 120000 btime 120000 movestogo 37
info depth 1 seldepth 3 score cp 65 hashfull 0 nodes 64 time 22 nps 64000 pv b1c3
info depth 2 seldepth 9 score cp 4 hashfull 0 nodes 544 time 31 nps 544000 pv b1c3 b8c6
info depth 3 seldepth 11 score cp 44 hashfull 0 nodes 2945 time 33 nps 2945000 pv b1c3 b8c6 c1e3
info depth 4 seldepth 19 score cp 7 hashfull 0 nodes 26275 time 53 nps 1313750 pv c1f4 d5d4 h2h3 g4e6
info depth 5 seldepth 23 score cp 33 hashfull 0 nodes 114550 time 113 nps 1975000 pv h2h3 g4f3 e2f3 b8c6 b1c3
info depth 6 seldepth 26 score cp 6 hashfull 0 nodes 392502 time 293 nps 2180566 pv b1c3 e7e6 c1f4 b8c6 h2h3 g4f5
info depth 7 seldepth 27 score cp 32 hashfull 0 nodes 2366298 time 1466 nps 2017304 pv f3e5 b8c6 e5g4 f6g4 b1c3 e7e5 c3b5
...
position fen rn1qkb1r/ppp1pppp/5n2/3p4/6b1/3P1NP1/PPP1PP1P/RNBQKB1R w KQkq - 1 4
go wtime 120000 btime 120000 movestogo 37
info depth 1 seldepth 3 score cp 31 hashfull 0 nodes 64 time 2 nps 64000 pv f3e5
info depth 2 seldepth 9 score cp -16 hashfull 0 nodes 574 time 2 nps 574000 pv f3e5 b8c6
info depth 3 seldepth 17 score cp 27 hashfull 0 nodes 5262 time 10 nps 5262000 pv h2h3 g4f3 e2f3
info depth 4 seldepth 17 score cp -16 hashfull 0 nodes 17906 time 22 nps 2238250 pv c1e3 b8c6 e3c5 b7b6
info depth 5 seldepth 23 score cp 29 hashfull 0 nodes 84315 time 62 nps 2107875 pv h2h3 g4f3 e2f3 b8c6 b1c3
info depth 6 seldepth 25 score cp -2 hashfull 0 nodes 360862 time 250 nps 2004788 pv f3e5 b8c6 c1f4 c6e5 f4e5 c7c6
info depth 7 seldepth 27 score cp 29 hashfull 0 nodes 1128135 time 641 nps 2960984 pv f3e5 b8c6 e5g4 f6g4 h2h3 g4e5 c1g5
bestmove f3e5
...
Please don't forget the shown nps is and was always wrong too. I don't remember anymore, if I had mentioned this already in this thread
or in PMs to Martin. Best calculate it yourself ;-)
I think the output looks better than before even on your hardware, but IMHO it is still quite different to the FEN output.
(for no popcount versions the difference is of course 'worlds like'. > +100 rating points under my conditions)
Martin assumes probably more bugs, but this needs too much time for checking.
I filed an issue already, but I don't think Stuart is looking anymore in the Pigeon repo since long.