Moderators: hgm, Rebel, chrisw
that's not reality of course, you also keep referring in postings here on what worked for perft in hashtable and what didn't. Would be bad if another guy is doing that.hgm wrote:He is not measuring performance, but correctness...
Maybe switch language, speed increase of factor 4 or so?Felpo wrote:yes,
I'm just measuring the correctness: my performance now are reasonable ( in term of move generation ) for a C# program as mine: I can't beat the HG Mueller qperft, neither the "emptran" shown in the Steven Edwards post, but is competitive in term of speed with some other engines I found, even in C++.
This obsession for testing correctness derive from my previous engine: I was unexperienced, and I created a slow and buggy move generator, I had a TT as well, but who can trust without a solid testing fundation ? So I ended in avery poor job, in an untestable search and a very uncompetitive engine.
I did not find any other way to probe if TT is working correct than challenge it on a perft session.
This appear to be tricky as well, because a big testing with a lot of position, but ending at perft 6 was perfectly working, to discover errors I had to go down to perft 9, and I'm still not sure on what will happen at perft 10
As I learned from previous thread over this forum, the move generation speed is not influencing too much the overall performance of the engine: Achieving more cut-off is for sure better than a 4*speed in the move generation process. And Actually the main topic in this particular thread is not really speed, but how to probe if the TT behave correctly ( even if as a side effect they improve the perft speed too ).diep wrote:Maybe switch language, speed increase of factor 4 or so?Felpo wrote:yes,
I'm just measuring the correctness: my performance now are reasonable ( in term of move generation ) for a C# program as mine: I can't beat the HG Mueller qperft, neither the "emptran" shown in the Steven Edwards post, but is competitive in term of speed with some other engines I found, even in C++.
This obsession for testing correctness derive from my previous engine: I was unexperienced, and I created a slow and buggy move generator, I had a TT as well, but who can trust without a solid testing fundation ? So I ended in avery poor job, in an untestable search and a very uncompetitive engine.
I did not find any other way to probe if TT is working correct than challenge it on a perft session.
This appear to be tricky as well, because a big testing with a lot of position, but ending at perft 6 was perfectly working, to discover errors I had to go down to perft 9, and I'm still not sure on what will happen at perft 10
C is a lot faster than the object oriented languages practical spoken.
Vincent
A hashtable that keeps your search 100% intact might not be so efficient in an engine which goal is not to count how many moves each subtree has, yet focuses upon playing the best move insteadFelpo wrote:As I learned from previous thread over this forum, the move generation speed is not influencing too much the overall performance of the engine: Achieving more cut-off is for sure better than a 4*speed in the move generation process. And Actually the main topic in this particular thread is not really speed, but how to probe if the TT behave correctly ( even if as a side effect they improve the perft speed too ).diep wrote:Maybe switch language, speed increase of factor 4 or so?Felpo wrote:yes,
I'm just measuring the correctness: my performance now are reasonable ( in term of move generation ) for a C# program as mine: I can't beat the HG Mueller qperft, neither the "emptran" shown in the Steven Edwards post, but is competitive in term of speed with some other engines I found, even in C++.
This obsession for testing correctness derive from my previous engine: I was unexperienced, and I created a slow and buggy move generator, I had a TT as well, but who can trust without a solid testing fundation ? So I ended in avery poor job, in an untestable search and a very uncompetitive engine.
I did not find any other way to probe if TT is working correct than challenge it on a perft session.
This appear to be tricky as well, because a big testing with a lot of position, but ending at perft 6 was perfectly working, to discover errors I had to go down to perft 9, and I'm still not sure on what will happen at perft 10
C is a lot faster than the object oriented languages practical spoken.
Vincent
I'm interested on submit more precalculated test to my move generation subsystem, I used perftestsuite.epd, that I found looking at Roce chess engine, It contains some insidious positions as this one:sje wrote:Note that you will have to test to depth ten or greater from the initial position to trigger black queenside castling.
I forget who first mentioned the idea, but another test strategy would be to take a well distributed problem suite like WAC (300 positions) and run perft N on each and sum the results.
Never mind about where to find it, I suppose it is:Felpo wrote:I'm interested on submit more precalculated test to my move generation subsystem, I used perftestsuite.epd, that I found looking at Roce chess engine, It contains some insidious positions as this one:sje wrote:Note that you will have to test to depth ten or greater from the initial position to trigger black queenside castling.
I forget who first mentioned the idea, but another test strategy would be to take a well distributed problem suite like WAC (300 positions) and run perft N on each and sum the results.
r3k2r/p1ppqpb1/bn2pnp1/3PN3/1p2P3/2N2Q1p/PPPBBPPP/R3K2R w KQkq - 0 1
and many others. Where can I found the WAC you are pointing me ?
Thanks