SkyNet wrote: ↑24/02/2023, 21:11
Joachim26 wrote: ↑24/02/2023, 10:58
Thanks for the two links, sources doesn't matter, since Sopel, I, and others now exactly could reproduce the code you have used for compiling. Everything's fine with GPLv3 now
.
It's interesting that you and Chessman got regression soon after applying only few patches. When I started I expected that this would happen sooner or later also to my dev builds, after skipping the update to NNUEv3 net architecture.
So I did a lot of testing, after applying 1, 2, or max 3 commits, and tweaked the code a bit, if there was some obvious regression. After about 100 commits I'm still quite sure, that the released (win) builds slowly but steadily get stronger.
Concerning the Android builds: I don't make tournaments but I check from time to time the bench on Termux (and compare it with the win build) and do engine selfplay games in Droidfish. Sopel did an impressively great job concerning the NEON support in CFish but missed one(!) #define and didn't know about that till yesterday.
Yep, i forgot to mention if you begin to examine the Sopel's patches from the beginning you'll see that one (very simple) patch is missing, it my be the one that you mentioned.
Regarding your updates, i would suggest to remove the patch with "persistent ttsearch - persistent ttdeserialize", it is useless for the current network that you use. Also remove the Robert Houdart's patch, it is useless.
Regarding your tests, i would suggest you to test your updated version vs stable version eg your update vs Cf 13 or SF 14.
You see the the thing is that SF team testing SF master vs new update using thousand of games to see the effect of the specific patch, but in general it doesn't mean that the new patch will be good in general. You can found the regression/progression tests on other sites, and sometimes the results of development versions are not good against the stable versions.
Anyway, if you need any help you can PM me and
when have a free time i''l help you with everything i can.
Thanks for your offer to help me. At the moment you are already doing this by performing the first Android tournaments with CFNN.armv8. As I mentioned earlier, I'm quite curious about the results, which then can be compared with the Windows tournaments.
Above there are several things which could be discussed. It's just to late to do this now. But one point I want to clearify immediately:
CFNN does show the exact nps output since I have, certainly, corrected it in position.c:
...
pos->sideToMove = !pos->sideToMove;
pos->nodes++;
set_check_info(pos);
...
I guess that Leon didn't change it (his build "
seems" to be faster), but I'm to lacy now to search his sources.
Fun fact of the story is, that I prepared a '"funny" Pascal-style' PR on Sep 2, 2021 (after FTv1.0 release) to correct the nps output, but Sopel just rejected it back then.
https://github.com/Sopel97/FatTitz/commit/cc85ccff23aad1d2391781bcbecc34d6cf7731ac
Greetings from here and me to Poland, Tomasz, have a nice weekend
Edit: Here his reaction, today more funny for me than my PR
https://github.com/Sopel97/FatTitz/pull/8