Static Code Analysis

[jsgroby]

Did any of you developers use a static code analysis tool to take a look at your code and see if there were any areas where you could optimize it? I am not super happy with my move generator times compared to some of the more stronger engines out there and am trying to determine the best method of analyzing where there might be a bottleneck, etc.

Any thoughts or consideration in this matter is greatly appreciated.

Thanks
Jeff

[mar]

Static analysis is used to detect bugs, if you want to improve performance you want profiling tools instead.
As for move generator, I wouldn't worry much about it (unless it's a factor of 100
Of course being faster here or there is always good as it accumulates in the end.
Mine is 2+ times slower than top engines for example. You will only spend a fraction of total time generating moves.
Real work has to be done elsewhere to make an engine strong.

[jsgroby]

Static analysis is used to detect bugs, if you want to improve performance you want profiling tools instead.
As for move generator, I wouldn't worry much about it (unless it's a factor of 100
Of course being faster here or there is always good as it accumulates in the end.
Mine is 2+ times slower than top engines for example. You will only spend a fraction of total time generating moves.
Real work has to be done elsewhere to make an engine strong.

I stand corrected, profiling tools are what I am asking about, thank you Martin.

Basically wanting to know if I am having bottlenecks and slowdowns in loops, etc...

Right now my engine is taking about 10 seconds to get to perft 6 on FEN entries that Stockfish and other engines are hitting in about 2-3 seconds.

[mar]

Right now my engine is taking about 10 seconds to get to perft 6 on FEN entries that Stockfish and other engines are hitting in about 2-3 seconds.

It's not that bad. Do you do bulk counting of leaf nodes? This can give a nice speed up.
Also if you use staged move generator having a full legal one for perft gives a small speed up as well.
EDIT: move generator is only part of the story, it's also important how fast you can make/unmake moves
For example I do incremental update of psqs and material key which can be omitted in perft. Also (unless you hash perft) you don't have to update hash key(s).
Those can give a small speed up too.

[jsgroby]

Right now my engine is taking about 10 seconds to get to perft 6 on FEN entries that Stockfish and other engines are hitting in about 2-3 seconds.

It's not that bad. Do you do bulk counting of leaf nodes? This can give a nice speed up.
Also if you use staged move generator having a full legal one for perft gives a small speed up as well.
EDIT: move generator is only part of the story, it's also important how fast you can make/unmake moves
For example I do incremental update of psqs and material key which can be omitted in perft. Also (unless you hash perft) you don't have to update hash key(s).
Those can give a small speed up too.

Thanks Martin. You have given me some ideas where to start poking around. I am examining my make move and take move functions and doing some overall testing with my perft function to see what impact it has.

I did try some compiler optimizations and got the following results:

No compiler optimization resulted in 25059ms for a perft 6

-O2 compiler optimization resulted in 10421ms for a perft 6

-O3 compiler optimization resulted in 10111ms for a perft 6

-O3 -ffast-math optimization resulted in 10149ms for a perft 6

Position used was starting FEN,

rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1

[zullil]

-O3 compiler optimization resulted in 10111ms for a perft 6

-O3 -ffast-math optimization resulted in 10149ms for a perft 6

Position used was starting FEN,

rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1

My guess is -ffast-math might be helpful for code with many floating point ops, which you probably don't have.

As Martin said, this is not really a key issue in the big picture. But you might try letting the compiler do some profile-guided optimization for you. Assuming you are using gcc, you can follow the steps in my model below:

Code: Select all

ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -fprofile-generate -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 8069 ms
ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -fprofile-use -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 5442 ms

For comparison, here's without the PGO:

Code: Select all

ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 6394 ms

[jsgroby]

-O3 compiler optimization resulted in 10111ms for a perft 6

-O3 -ffast-math optimization resulted in 10149ms for a perft 6

Position used was starting FEN,

rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1

My guess is -ffast-math might be helpful for code with many floating point ops, which you probably don't have.

As Martin said, this is not really a key issue in the big picture. But you might try letting the compiler do some profile-guided optimization for you. Assuming you are using gcc, you can follow the steps in my model below:

Code: Select all

ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -fprofile-generate -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 8069 ms
ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -fprofile-use -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 5442 ms

For comparison, here's without the PGO:

Code: Select all

ProcyonLeo: ~/Documents/Chess/Kirby] gcc -Wall -O3 -o perft Perft.c
ProcyonLeo: ~/Documents/Chess/Kirby] ./perft
FEN string = rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -
Depth = 6
Leaf nodes = 119060324
Time taken = 6394 ms

Thanks Louis. I will indeed try that.

Jeff

[jdart]

I have used oprofile (http://oprofile.sourceforge.net/about/), which works on both AMD & Intel systems, and VTune (https://software.intel.com/en-us/intel- ... plifier-xe), which is Intel only (there is a free version for non-commercial use on Linux: it is costly for Windows, although there is a free trial). These are very helpful although they are complex tools. You can use them for just the basics (finding code execution hotspots) but anything beyond that will require a good knowledge of modern CPU and PC system architecture.

--Jon