Reading / storing of structs is not atomic, unless they happen to fit in a 64-bit machine word.

When hashing, you should be prepared to handle the situation where you sometimes get faulty information. Even with a 64-bit key this occasionally happens through key collisions. Often enough to make the ...

Just gained the motivation to do my first implementation of parallel search but first I tried to do parallel perft with hashing in order to test some ideas and the correctness of some code. In this test I have stumbled with one problem. Very rarely, when multiple threads try to store and lookup the ...

Shoot... that is where ALL of my hash-related bugs have shown up. :) (when hashing is enabled.)

hehehe :)

I just found the problem.. It was pretty simple. I was calculating the node type(flag of the transposition table in the code) after adjusting the best score to be stored in the transposition ...

I think I see the issue.


if(val > alpha){
alpha = val;
if(val >= beta)
break;
update_pv(it, ply);
}


Here, if the value >= beta, you exit the loop and go on to store this alpha in the transposition table, but you don't update the PV, meaning the PV could have erroneous moves after the ...

My point was not that you needed the - 1 or + 1, but that you have to carefully look. For example, why not do this:

when you store a mate score, display the tree and print the actual score and the stored score. When you hit on a position with a mate score, display the position, the table score ...

Note we don't match exactly. Here is mine again:

if (val > 32000)
val -= ply - 1;
else if (val < -32000)
val += ply - 1;

Note the val += ply -1 or val =- ply - 1. Perhaps that -1 is what you are missing?

Subtracting one ply doesn't do anything. This is the scores I'm getting for the last ...

I am hoping you meant

val = - alphabeta(-beta, -alpha, etc)???

If so, that should not be a problem. I would re-factor your code so that you only have one call to alphabet(), not one for in check and one for not in check, which is a good way to introduce bugs that are hard to find...

yes, its ...

I do exactly the same thing. On storing:

if (value > 32000)
value += ply - 1;
else if (value < -32000)
value -= ply - 1;

In Crafty, any score > 32000 or < -32000 has to be mate, there is no way to have a 320 pawn advantage of deficit.

On probing:

if (val > 32000)
val -= ply - 1;
else if ...

Mate scores are special. They represent mate in N move from the ROOT position. When you store a mate score internally inside the tree, you have to adjust it to mate in N minus something from the current position. Simple example.

the mate sequence looks like this:

P1 - P2 - P3 - P4 - P5 <mate in 3 ...

I'm having some trouble understanding how I can store and then retrieve the mate scores from the transposition table. I couldn't find in the website chessprogramming.wikispaces.com any useful info about this matter.

For example in this test position:
3k4/1R3p2/2p3pp/8/8/pR6/p7/4K3 w - - 0 1
There ...