Eelco de Groot wrote:This is a bugfix version of the 2.3.1, not a lot of line changes but some I think were bugs so those changes I think can be important.
There was a suspected bug in the bad bishop initialization and a certain bug in the end of the pawn evaluation. Not a bug but a code change is that singular extensions are not forced to go through LMR reduction but can go directly to full search depth. The node is expected to fail high so I don't think skipping the reduced search is bring much instability, there are some more (sometimes False-)Fail Lows visible now (there were hardly any before) probably due to this but the benefit should be that a Fail Low is now certainly from a full search and therefore in some instances a bit earlier (which is what you want, trust me :geek: )

In spite of fixes, there may be major bugs left and testing is still very minimal... I do hope this is still a bit better than the 2.3.1 now that some of the bugs are fixed, hence the 2.3.2 designation. No new read.me file yet, it is copied from the last archive like most other files. The slightly changed files are thread.h, thread.cpp, pawns.cpp, evaluate.cpp and search.cpp. All of the rest is still very much like Stockfish 2.3.1 + some changes from the GitHub master, all those were already in Rainbow Serpent 2.3.1.

One search change of the "hollowPV" condition should do nothing at all, but I have not yet verified that it does nothing so I left it in for now :) .