As I’ve explained in previous posts, a perft routine is a great way to debug a move generator. If you can correctly calculate the leaf node count for a variety of positions you are (almost) certain to have eradicated all the bug in your move generator.
But what about the quiescent search? Unlike the regular search, the quiescent search doesn’t look at all moves. And I’m not aware of any specialized quiescent perft test.
At a minimum the quiescent search generates all of the possible capture moves: and in most cases it also generates checks (at least for the first ply). How can we write these move generation routines with a high level of certainty they are bug free?
When I write Monarch I created a perft routine but didn’t create any test routines for the ‘generate_captures’ or ‘generate_checks’ routines. With Maverick I’m more determined to try to write a test routine for each procedure. So I wrote three new move generators:
As the name suggests, the first routine generates all captures except en-passant. Captures which promote the pawn are limited only to queen promotions.
The second routine, generates checking moves which are not captures. It ignores pawn promotions, en-passant moves and castling move which give check.
The third routine generates all the other moves. When Maverick finally plays chess I don’t expect to use this routine at all. However, all three routines together generate the same set of moves as the regular move generator. This means I can test them using a version of the perft routine which calls all three specialized move generators for each position. It worked well and I was able to trap a couple of bugs.
What really amazed me was the speed of this adjusted perft routine – it was only about 10% slower than the regular routine. This is all down to the bitboard structure. With a letterbox structure I’m sure there would have been a much wider gap between the normal perft and this specialized move generator version.