Speed up your engine Part 3

[lauriet]

I don't really WANT to learn C. Far from it. But I am curious if a one to one translation would make an improvement in speed or do you need to use C tricks to improve the compiled code.

Is there anyone out there would like to do the conversion

[Dann Corbit]

Free Pascal is pretty close to C in speed.

Booot and Delfi along with some versions of Critter were written in Pascal and they are quite strong.

All three of these programs have source code available.

You can do bitboards in Pascal just as easily as C. The latest versions of 64 bit free pascal have all the important bit operation intrinsics.

If you are fluent in Pascal, I would leave the engine in Pascal unless you have a goal to learn the C programming language.

The other advantage to using C is that all the samples in the chess programming Wiki are either in C or C++.

[Dann Corbit]

Critter:
https://sourceforge.net/projects/critterchess/

Booot:
http://www.sdchess.ru/download_engines.htm

Delfi:
http://www.msbsoftware.it/delfi/source.htm

The easiest one to learn from is probably Critter, with Delfi second.
Booot is the strongest one, but it has all of its comments in Russian so you have to translate them one by one unless you speak the language.

[lauriet]

I have found a pascal version of tscp on the web and it does run faster than my program, so I guess the obvious conclusion is that I can improve my implementation.
I'm guessing the move gen in tscp is faster, so I will need to study whats going on.

I tried to profile my program with gprof, but got stuck and didn't go any further. Maybe I need to revisit that.

Watch this space !

[Stan Arts]

Nemeton is FreePascal.
While FPascal did not keep up with C in the optimization race it's not going to be worlds apart also simply depending on your own skill and sometimes looking at the assembler output can be a great help.

I like Pascal way more than C so when I have a choice I write in Pascal. The speedup is definitely not worth it to me and you'd have to think in the order of 10-20 Elo max.

That said learning C is pretty useful.

[lauriet]

Hi,

My program uses 0x88 board. Do you think the other mailbox types have any advantage over this ?
Is Nemeton open source ?
Can I download it ?

Laurie.

[Stan Arts]

Hi,

My program uses 0x88 board. Do you think the other mailbox types have any advantage over this ?
Is Nemeton open source ?
Can I download it ?

Laurie.

Yes,

Code: Select all

https://onedrive.live.com/redir?resid=17C791428DB77D82!636&authkey=!AOdXat1-hCCVUak&ithint=file%2crar 

I've never used 0x88 so can't say. For a first chessprogram making everything work as expected and understanding what's happening outweighs everything else.

Nemeton uses a 16x12 mailbox representation similar to 10x12 in the way I do off-the-board detection but 16 so I can bitshift the index.
Nemeton is not such a serious chessprogram, mostly an effort to write an essentially fast program which I had to get out of my system..

People always bringing up bitboards.. Likely won't even make movegeneration faster. Might find some speedup in attack or SEE stuff but that's a small gain for a hellish boardrepresentation.

[ZirconiumX]

Hi,

My program uses 0x88 board. Do you think the other mailbox types have any advantage over this ?
Is Nemeton open source ?
Can I download it ?

Laurie.

Yes,

Code: Select all

https://onedrive.live.com/redir?resid=17C791428DB77D82!636&authkey=!AOdXat1-hCCVUak&ithint=file%2crar 

I've never used 0x88 so can't say. For a first chessprogram making everything work as expected and understanding what's happening outweighs everything else.

Nemeton uses a 16x12 mailbox representation similar to 10x12 in the way I do off-the-board detection but 16 so I can bitshift the index.
Nemeton is not such a serious chessprogram, mostly an effort to write an essentially fast program which I had to get out of my system..

People always bringing up bitboards.. Likely won't even make movegeneration faster. Might find some speedup in attack or SEE stuff but that's a small gain for a hellish boardrepresentation.

Bitboards aren't "hellish" if done right. They're also fast, and can be very elegant.

How long does it take an array based program to calculate, say, a knight's mobility?

8 additions/subtractions for the vector, 8 register increments, 8 table lookups, assuming the compiler puts them in a temporary register, and 16 comparisons (8 for off board, 8 for avoiding counting own piece attack), plus 16(?) conditional branches. Assuming an incredibly optimistic RISC-style one instruction per clock, that's 56 clocks.

How long does it take a bitboard based program to calculate the same knight's mobility?

1 population count, a table lookup for the attacks for the knight square, a lookup for the friendly pieces, a NOT to get all squares which aren't friendly pieces, and an AND to get the attacks on the squares that aren't friendly pieces. Taking the same optimistic RISC chip, that's 5 instructions.

Or alternatively, the bitboard code is ~11 faster.

I'm not saying array based code is necessarily slow, if optimised well. But bitboards have a head start.

[Dann Corbit]

Biggest advantage of bitboards is in evaluation.
Magic bitboards are pretty fast, even in move generation.

[hgm]

How long does it take a bitboard based program to calculate the same knight's mobility?

That depends on how you implement it. When done incrementally the answer would be zero cycles, unless the last move happens to change it.