-- Improvement to work-sharing in lazy smp implementation, motivated by
ABDADA discussion by Daniel Shawul on talkchess.com, but using a data
structure rather than hash table to control this
-- Addition of a pawn hash move to pawn hash table
-- Changed from one to two layers of history based reductions
-- Use null move to detected threatened pieces when apparently far
ahead but null move is refuted by a capture
-- Various improvements to King Safety
-- Aging history table between moves rather than initializing each time
-- Using piece-ID rather than from-square to index history and reply
tables
-- Padding hash table entries to give identical sizes on 32 and 64
bit systems
-- Some parameter adjustments with CLOP
-- Various minor bug fixes
Hi Daniel
I am flattered that my thread inspired you but you should give yourself more credit because no body will Well you are the one who started those lazy smp threads which got me thinking there might be something we are missing. Indeed ABDADA was a real surprise for me with its performance. If I knew about that performance long time ago, it would have been my first choice. I hope you got similar results with it too. Please make a post in programmers forum when you can. It would be interesting to read about the difference in implementation details. So often a hash table is replaced with linked-list data structure in other areas of heuristic search like MCTS. I find use of hashtables there a bit awkward ,but ABDADA iseems to require some sort of hashing to coordinate work sharing anyway so i am curious what data structure you use.
Daniel
Thanks, I will write something up in awhile (heading for vacation in a few hours, so it will have to wait until I am back). The scheme is embarrasingly simple, so it might not be that interesting. It can't handle transpositions, but it does avoid having to make the move to check whether it is being worked on.
EXchess v7.11 is now available from my website...
....
Hi Dan,
do you expect a remarkable difference between 64bit and 32bit version?
Thnx (for the engine and your answer... )
They should play the same moves, but the 32 bit version is slower. I measure about 20% slower than the 64 bit version. How much this affects elo probably depends on time control, but I am guessing around 20-30 elo less.
maybe you remember how I annoyed you with problems running ExChess on the ARM architecture.
Recently we had in this thread a discussion about an issue with Crafty 23.6 on ARM. Bob Hyatt found, that the problem looked like "char x" defaults to unsigned and Jim Ablett knew a solution by adding the compiler switch "-fsigned-char".
Log story short: compiled with this switch, ExChess runs fine on ARM.
The hash move in the pawn hash table is basically a killer move that worked in that pawn structure. I am writing from my iPhone, so I can't write much about it, but I can say more when I am back from traveling in a few days.
maybe you remember how I annoyed you with problems running ExChess on the ARM architecture.
Recently we had in this thread a discussion about an issue with Crafty 23.6 on ARM. Bob Hyatt found, that the problem looked like "char x" defaults to unsigned and Jim Ablett knew a solution by adding the compiler switch "-fsigned-char".
Log story short: compiled with this switch, ExChess runs fine on ARM.
Thanks,
Max
Thanks to you and Jim for figuring this out! I never would have guessed that without an ARM system to develop on.
Well you are the one who started those lazy smp threads which got me thinking there might be something we are missing. Indeed ABDADA was a real surprise for me with its performance. If I knew about that performance long time ago, it would have been my first choice. I hope you got similar results with it too. Please make a post in programmers forum when you can. It would be interesting to read about the difference in implementation details. So often a hash table is replaced with linked-list data structure in other areas of heuristic search like MCTS. I find use of hashtables there a bit awkward ,but ABDADA iseems to require some sort of hashing to coordinate work sharing anyway so i am curious what data structure you use.