stockfish
10 years agoSort PV moves always in two steps
Marco Costalba [Tue, 2 Aug 2011 08:53:56 +0000 (09:53 +0100)]
Sort PV moves always in two steps

This should fix following issue:

Suppose the search with MultiPVIteration == 0 returns an exact score

move = Nxf4, score = 100

Now search with MultiPVIteration == 1 and get two scores

move = Qg8, score = 150
move = Ra1, score = 180

If we now reorder all the moves in one step we end up with

pv[0] = Ra1, pv[1] = Qg8

Instead reordering as the current patch we end up in:

pv[0] = Ra1, pv[1] = Nxf4

preserving the first searched move.

No functional change in single PV.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSmall cleanup of previous patches
Marco Costalba [Tue, 2 Aug 2011 06:50:14 +0000 (07:50 +0100)]
Small cleanup of previous patches

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoReimplement MultiPV mode
Joona Kiiski [Sun, 31 Jul 2011 22:15:11 +0000 (01:15 +0300)]
Reimplement MultiPV mode

No functional change

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoReimplement support for "searchmoves" option
Joona Kiiski [Sun, 31 Jul 2011 12:18:52 +0000 (15:18 +0300)]
Reimplement support for "searchmoves" option

No functional change

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRemove now unused RootMove.non_pv_score
Joona Kiiski [Sun, 31 Jul 2011 12:02:36 +0000 (15:02 +0300)]
Remove now unused RootMove.non_pv_score

No functional change

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoMake root search to use standard MovePicker.
Joona Kiiski [Sat, 30 Jul 2011 21:00:54 +0000 (00:00 +0300)]
Make root search to use standard MovePicker.

This patch temporarily breaks MultiPV and searchmove
features, but they will be re-implemented in future
patches.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDon't probe TT at RootNode
Joona Kiiski [Thu, 28 Jul 2011 21:02:50 +0000 (00:02 +0300)]
Don't probe TT at RootNode

In that case we should also update RootMoveList to avoid
bogus output

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix pretty_pv() output in Chess960
Marco Costalba [Tue, 26 Jul 2011 07:47:20 +0000 (09:47 +0200)]
Fix pretty_pv() output in Chess960

And move it to search.cpp

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix PV output in Chess960
Joona Kiiski [Sun, 24 Jul 2011 10:56:47 +0000 (11:56 +0100)]
Fix PV output in Chess960

We missed to set chess960 flag into the std::stringstream used to
setup the PV line.

Bug introduced with commit f803f33e6383fea1074aca2cf625983e148a6668
of 30/12/2010 when we started to print PV line into a std::stringstream
instead of directly into cout, where the chess960 flag is correctly set.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoTidy up benchmark.cpp
Marco Costalba [Sun, 24 Jul 2011 07:26:35 +0000 (08:26 +0100)]
Tidy up benchmark.cpp

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRename execute_uci_command() to uci_loop()
Marco Costalba [Sun, 24 Jul 2011 06:56:26 +0000 (07:56 +0100)]
Rename execute_uci_command() to uci_loop()

As a side effect now root position can be directly
allocated on the stack and doesn't need to be defined
static anymore.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUnhide the istringstream behind UCIParser
Marco Costalba [Sun, 24 Jul 2011 06:22:37 +0000 (07:22 +0100)]
Unhide the istringstream behind UCIParser

It is misnamed because it is not a parser, perhaps a
tokenizer, anyhow better call it for what it is, an
input string stream.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix startpos_ply_counter() regression
Marco Costalba [Sun, 24 Jul 2011 04:59:16 +0000 (05:59 +0100)]
Fix startpos_ply_counter() regression

Return the correct number of played plies at the end
of the setup moves. Currently it always returns 0 when
starting from start position, like in real games.

We fix this adding st->pliesFromNull that starts from 0
and is incremented after each setup move and is never
reset in the setup phase but only once search is started.

It is an hack because startpos_ply_counter() will return
different values during the search and is correct only
at the beginning of the search.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse a circular buffer to keep track of setup states
Marco Costalba [Sat, 23 Jul 2011 11:52:49 +0000 (12:52 +0100)]
Use a circular buffer to keep track of setup states

This fixes a regression on real games due to the fact that
we have some mismatches:

    history[st->gamePly - i] != stp->key

when st->gamePly - i == 0,this is due to a nasty bug I have
introduced when using std::vector<> as StateInfo backup. The
point is that StateInfo keeps inside a pointer to the previous
StateInfo in a kind of linked list. But when std::vector<> is
resized reallocates a larger chunk of memory and moves the
data there so these pointers became stale.

This patch fixes the issue.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDon't update gamePly after each move
Marco Costalba [Wed, 20 Jul 2011 08:45:40 +0000 (10:45 +0200)]
Don't update gamePly after each move

We just need startup value to calculate available
thinking time. So remove from state.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse std::vector<Move> to store UCI search moves
Marco Costalba [Wed, 20 Jul 2011 08:21:41 +0000 (10:21 +0200)]
Use std::vector<Move> to store UCI search moves

Avoid the ugly and anyhow incorrect hard limit on the
maximum number of moves and allow to handle an arbitrary
number of moves to search.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse st->gamePly to store fullMoves
Marco Costalba [Wed, 20 Jul 2011 08:01:12 +0000 (10:01 +0200)]
Use st->gamePly to store fullMoves

This allow to retire do_setup_move() and also to simplify
draw detection logic becuase now we always have:

Min(st->rule50, st->gamePly) = st->rule50

This was already true when starting from starting position,
but now is true even when starting from a FEN string because
now we take in account fullmove number in counting gamePly so
that it is always.

st->rule50 <= st->gamePly

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire history[]
Marco Costalba [Mon, 18 Jul 2011 08:20:37 +0000 (10:20 +0200)]
Retire history[]

Use key saved in state instead.

No functional change (in real games) and no speed regression.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDo not exit early even when seeing mate
Joona Kiiski [Sat, 23 Jul 2011 21:18:28 +0000 (22:18 +0100)]
Do not exit early even when seeing mate

Fixes the reported KNNK ending problem:

http://talkchess.com/forum/viewtopic.php?t=39347

Joona says:

Now I finally had a time to take a look at on this issue.

I've reproduced the problem starting from this position:
1B6/1B2k3/P7/1P3p2/1K6/8/4b3/4b3 w - - 6 85

I made Stockfish play as white and Fruit as black.
I repeated test ten times and once SF was not able to deliver mate.

But I observed several times that SF had reported on last something like mate in 10.
However next time it played move with score mate in 15.

Easiest way to solve the problem is attached as a patch. I tested it several times and SF always
ended up playing the optimal move. Of course the downside is that now delivering mate
takes a bit longer, but IMO it's better to lose once in a while by time in sudden death
game than not being able to deliver simple mate with long time controls.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRevert previous patches due to bug
Marco Costalba [Wed, 20 Jul 2011 02:31:50 +0000 (03:31 +0100)]
Revert previous patches due to bug

We have a bug (possibly because of returning draw from
root move list), it is possible to see when looking at
games with a GUI, we can see rarely but consistently the
score return as #0 for many depths until it comes back to
normal values.

Revert patches until it is not fixed.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoMove the draw check also for qsearch
Joona Kiiski [Mon, 11 Jul 2011 23:00:30 +0000 (02:00 +0300)]
Move the draw check also for qsearch

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoMove draw checks right after doing the move
Joona Kiiski [Mon, 11 Jul 2011 21:09:50 +0000 (00:09 +0300)]
Move draw checks right after doing the move

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSmall touches in set_option()
Marco Costalba [Sun, 17 Jul 2011 11:26:50 +0000 (12:26 +0100)]
Small touches in set_option()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoValidate input UCI moves
Marco Costalba [Sun, 17 Jul 2011 09:22:08 +0000 (10:22 +0100)]
Validate input UCI moves

Running following command:

  position startpos moves e1e8

Makes SF to assert in debug mode in do_move() but to accept
bad input and continue in release mode where probably it is
going to crash little later.

So validate input before to feed do_move().

Suggestion by Yakovlev Vadim.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRemove some useless include
Marco Costalba [Sat, 16 Jul 2011 21:59:14 +0000 (22:59 +0100)]
Remove some useless include

Spotted by Rein Halbersma.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDon't need to assert for pos.is_ok() when position is constant
Marco Costalba [Sat, 16 Jul 2011 09:42:27 +0000 (10:42 +0100)]
Don't need to assert for pos.is_ok() when position is constant

It's only necessary to do the checking at the end of every non-const
member (including the constructors and from_fen()) of class Position.
Once the post-condition of every modifier guarantees the class invariant,
we don't need to verify sanity of the position as preconditions for outside
callers such as movegen, search etc. For non-class types such as Move and
Square we still need to assert of course.

Suggested by Rein Halbersma.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoChange hidden checkers API
Marco Costalba [Sat, 16 Jul 2011 08:56:34 +0000 (09:56 +0100)]
Change hidden checkers API

After previous patch is no more needed to pass
the color, becuase it is always the side to move.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSimplified discovered check connected_moves()
Marco Costalba [Sat, 16 Jul 2011 08:36:03 +0000 (09:36 +0100)]
Simplified discovered check connected_moves()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSimplify and micro-optimize hidden_checkers()
Marco Costalba [Fri, 15 Jul 2011 21:24:18 +0000 (22:24 +0100)]
Simplify and micro-optimize hidden_checkers()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoNo need to declare default Position c'tor
Marco Costalba [Fri, 15 Jul 2011 20:51:19 +0000 (21:51 +0100)]
No need to declare default Position c'tor

Pointed out by Rein Halbersma.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix a warning under MSVC
Marco Costalba [Fri, 15 Jul 2011 11:20:31 +0000 (12:20 +0100)]
Fix a warning under MSVC

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix parametrized direction in pawns generation
Marco Costalba [Fri, 15 Jul 2011 09:40:09 +0000 (11:40 +0200)]
Fix parametrized direction in pawns generation

It worked by accident because we always called both directions,
but definition was wrong.

Functional change due to different generation order, but
perft numbers are the correct ones.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoCleanup handling of Delta enums
Marco Costalba [Fri, 15 Jul 2011 09:19:49 +0000 (11:19 +0200)]
Cleanup handling of Delta enums

Ispired by Rein's code.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse std library to sort moves
Marco Costalba [Thu, 14 Jul 2011 11:12:49 +0000 (12:12 +0100)]
Use std library to sort moves

Functional change due to the fact that now pick_best() is
stable, but should be no change in strenght.

Example code and ideas by Rein Halbersma.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix icc's "unreachable code" warning
Marco Costalba [Mon, 11 Jul 2011 10:17:50 +0000 (11:17 +0100)]
Fix icc's "unreachable code" warning

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoTeach to_fen() about Halfmove and Fullmove number
Marco Costalba [Sun, 3 Jul 2011 10:32:34 +0000 (11:32 +0100)]
Teach to_fen() about Halfmove and Fullmove number

And also fix the last '/' at the end of the piece placement field.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse MoveList also in Position::move_is_pl_slow()
Marco Costalba [Sun, 3 Jul 2011 09:38:20 +0000 (10:38 +0100)]
Use MoveList also in Position::move_is_pl_slow()

And rename it in Position::move_is_legal()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire move_is_short_castle() and move_is_long_castle()
Marco Costalba [Sun, 3 Jul 2011 09:08:09 +0000 (10:08 +0100)]
Retire move_is_short_castle() and move_is_long_castle()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRestore startpos_ply_counter() instead of full_moves()
Marco Costalba [Sat, 2 Jul 2011 15:15:20 +0000 (16:15 +0100)]
Restore startpos_ply_counter() instead of full_moves()

And pass correct currentPly to TimeManager::init().

This restores old behaviour, in particular now black has
a different timing than white becuase is no more:

currentPly = 2 * fullMoveNumber;

but becomes

2 * (fullMoves - 1) + int(sideToMove == BLACK)

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoIntroduce and use struct MoveList
Marco Costalba [Sat, 2 Jul 2011 12:33:06 +0000 (13:33 +0100)]
Introduce and use struct MoveList

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRemove MSVC debug window hack in bench
Marco Costalba [Sat, 2 Jul 2011 10:52:22 +0000 (11:52 +0100)]
Remove MSVC debug window hack in bench

Interference with profile-build under mingw

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRearrange structs to avoid internal padding
Marco Costalba [Sat, 2 Jul 2011 10:01:12 +0000 (11:01 +0100)]
Rearrange structs to avoid internal padding

Found with gcc -Wpadded gcc option.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRevert PHQ-2
Marco Costalba [Sat, 2 Jul 2011 08:45:41 +0000 (09:45 +0100)]
Revert PHQ-2

No clear advantage again standard, possibly we will
retest before to release.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSmall touches to do/undo_castle_move()
Marco Costalba [Thu, 30 Jun 2011 14:28:58 +0000 (16:28 +0200)]
Small touches to do/undo_castle_move()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoBetter document generate_castle_moves()
Marco Costalba [Thu, 30 Jun 2011 08:29:58 +0000 (10:29 +0200)]
Better document generate_castle_moves()

No functional change also in Chess960

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSmall tweaks to search()
Marco Costalba [Sat, 25 Jun 2011 13:31:57 +0000 (14:31 +0100)]
Small tweaks to search()

No functional change also in faked split mode

To be sure verified in real games with 4 threads TC 2"+0.1
After 11125 games 2497 - 2469 - 6159 ELO +0 (+- 4.4)

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRemove duplicated enum Phase definition
Marco Costalba [Wed, 29 Jun 2011 16:40:33 +0000 (17:40 +0100)]
Remove duplicated enum Phase definition

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire find_checkers()
Marco Costalba [Tue, 28 Jun 2011 17:01:51 +0000 (18:01 +0100)]
Retire find_checkers()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire square_is_weak()
Marco Costalba [Tue, 28 Jun 2011 14:25:44 +0000 (16:25 +0200)]
Retire square_is_weak()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire one piece_list() overload
Marco Costalba [Tue, 28 Jun 2011 14:16:43 +0000 (16:16 +0200)]
Retire one piece_list() overload

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRename pieces_of_color() in pieces()
Marco Costalba [Tue, 28 Jun 2011 13:36:39 +0000 (15:36 +0200)]
Rename pieces_of_color() in pieces()

To be uniform with other overloads.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire redundant square_is_occupied()
Marco Costalba [Tue, 28 Jun 2011 13:29:58 +0000 (15:29 +0200)]
Retire redundant square_is_occupied()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoChange Position::pst() signature
Marco Costalba [Tue, 28 Jun 2011 12:59:33 +0000 (14:59 +0200)]
Change Position::pst() signature

To be more clear what is the underlying table.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix Shredder-FEN regression in from_fen()
Marco Costalba [Tue, 28 Jun 2011 07:35:51 +0000 (09:35 +0200)]
Fix Shredder-FEN regression in from_fen()

Fix also an incredible 3% speed regression by an almost
never called function. I guess this is due to mingw very low
quality standard libraries implementation.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoGreatly simplify castling rights handling
Marco Costalba [Mon, 27 Jun 2011 15:06:15 +0000 (16:06 +0100)]
Greatly simplify castling rights handling

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoAssume input FEN string is correct in from_fen()
Marco Costalba [Mon, 27 Jun 2011 09:07:57 +0000 (10:07 +0100)]
Assume input FEN string is correct in from_fen()

And also tolerate a 0 value for full move number.

Revert BUG_41 patch, now we set initial King file only
if a castling is possible, so we don't need the fix
anymore in case of correct FEN.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix a bug in Position::is_ok()
Marco Costalba [Sun, 26 Jun 2011 10:42:58 +0000 (11:42 +0100)]
Fix a bug in Position::is_ok()

If we cannot castle castleRightsMask[] could be not valid,
for instance when king initial file is FILE_A as queen rook.

In this case castleRightsMask[] at initialQRFile is different
from the expected (ALL_CASTLES ^ WHITE_OOO).

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRename type_of_piece() and color_of_piece()
Marco Costalba [Sun, 26 Jun 2011 09:52:42 +0000 (10:52 +0100)]
Rename type_of_piece() and color_of_piece()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire Position::color_of_piece_on()
Marco Costalba [Sun, 26 Jun 2011 09:37:13 +0000 (10:37 +0100)]
Retire Position::color_of_piece_on()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire Position::type_of_piece_on()
Marco Costalba [Sun, 26 Jun 2011 09:28:54 +0000 (10:28 +0100)]
Retire Position::type_of_piece_on()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire seeValues[] and move PieceValue[] out of Position
Marco Costalba [Sun, 26 Jun 2011 09:16:31 +0000 (10:16 +0100)]
Retire seeValues[] and move PieceValue[] out of Position

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRestore user weights to 100
Marco Costalba [Fri, 24 Jun 2011 09:22:26 +0000 (10:22 +0100)]
Restore user weights to 100

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoPHQ settings for King and Mobility
Marco Costalba [Sat, 18 Jun 2011 22:48:42 +0000 (23:48 +0100)]
PHQ settings for King and Mobility

See:
http://talkchess.com/forum/viewtopic.php?t=39327

After 8130 games on QUAD at 20"+0.1
1342 - 1359 - 5429 ELO +0 (+- 4.4)

Tried also version with just king settings changed:
After 5932 games 962 - 1052 - 3918 ELO -5 (+- 5.2)

And with just mobility settings changed:
After 4114 games 618 - 619  - 2877 ELO +0 (+- 5.9)

Frank has tested only 1200 games, but at longer TC and
against many engines, so because PHQ results are not worst
than other combination and not worst than original let's
commit his version.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix move_is_capture() to detect capture promotions
Marco Costalba [Wed, 22 Jun 2011 22:04:57 +0000 (23:04 +0100)]
Fix move_is_capture() to detect capture promotions

We miss to account as a capture a promotion capture !

Incredible bug in this critical function that is here
since a long time (b50921fd5c3e1753 of 21/10/2009 !!)

This patch fixes the bug and readds the faster
move_is_capture_or_promotion() that slightly increases
overall speed.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRewrite how uci info is sent to GUI
Marco Costalba [Wed, 22 Jun 2011 06:03:28 +0000 (07:03 +0100)]
Rewrite how uci info is sent to GUI

It is now much more modular than before and also we
always send the seldepth when we send the depth, this
avoids to make seldepth disappearing from GUI at the
start of a new iteration.

Print also fails high/low pv lines at high enough
search depths.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSimplify sliding_attacks()
Marco Costalba [Tue, 21 Jun 2011 17:48:14 +0000 (18:48 +0100)]
Simplify sliding_attacks()

Easy, almost trivial simplification, I don't understand
how I missed this before !!

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire ksq from CheckInfo
Marco Costalba [Tue, 21 Jun 2011 06:52:35 +0000 (08:52 +0200)]
Retire ksq from CheckInfo

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse CheckInfo to store pinned bitboard
Marco Costalba [Mon, 20 Jun 2011 07:46:31 +0000 (09:46 +0200)]
Use CheckInfo to store pinned bitboard

This trivial change gives an impressive 2,5% speedup !!!!

Also retire one unused move_gives_check() overload.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoOmit mate distance pruning at root
Marco Costalba [Sat, 18 Jun 2011 16:10:29 +0000 (17:10 +0100)]
Omit mate distance pruning at root

Restore original behaviour, before root unification and
remove a now useless ugly hack for alpha in multi-pv case.

No functional change

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse an array index instead of an iterator in root list
Marco Costalba [Sat, 18 Jun 2011 06:43:25 +0000 (07:43 +0100)]
Use an array index instead of an iterator in root list

It is not correct to use an iterator stick on a vector that
is sorted becuase iterator is invalidated in general case.

It happens to work by accident because iterators are implemented
as pointers and so they behave in the same (correct) way then
using array indices, but the latters are the correct thing to use.

Also better document the code.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRemove useless bestValue = alpha assignement
Marco Costalba [Wed, 15 Jun 2011 11:31:27 +0000 (12:31 +0100)]
Remove useless bestValue = alpha assignement

It is a fossil from the root_search() era, no more
needed today.

Spotted by Onno

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoTry only recaptures in qsearch if depth is very low
Marco Costalba [Sun, 12 Jun 2011 06:25:51 +0000 (07:25 +0100)]
Try only recaptures in qsearch if depth is very low

This avoids search explosion in qsearch for some
patological cases like:

r1n1n1b1/1P1P1P1P/1N1N1N2/2RnQrRq/2pKp3/3BNQbQ/k7/4Bq2 w - - 0 1

After 9078 games 20"+0.1 QUAD:
Mod vs Orig 1413 - 1319 - 6346 ELO +3 (+- 4)

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSpeed up kpk initialization
Marco Costalba [Wed, 15 Jun 2011 07:09:04 +0000 (09:09 +0200)]
Speed up kpk initialization

The trick is to classify more position at first cycle,
so to reduce following work. Speed up is of about 50% !

Also some cleanup while there.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse Carry-Rippler trick to speed up magics
Marco Costalba [Tue, 14 Jun 2011 17:40:52 +0000 (18:40 +0100)]
Use Carry-Rippler trick to speed up magics

Nice trick discovered on:

http://chessprogramming.wikispaces.com/Traversing+Subsets+of+a+Set

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire init_piece_square_tables()
Marco Costalba [Mon, 13 Jun 2011 14:59:03 +0000 (16:59 +0200)]
Retire init_piece_square_tables()

Merge in init_zobrist() and rename the latter.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoIncrease LMR limit by one ply
Marco Costalba [Thu, 2 Jun 2011 07:35:35 +0000 (08:35 +0100)]
Increase LMR limit by one ply

Seems there is no regression so prefer to prune less.

After 8278 games
Mod vs Orig 1246 - 1265 - 5767 +0 ELO (+- 4.2)

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoBetter self-document init_zobrist()
Marco Costalba [Sun, 12 Jun 2011 10:46:24 +0000 (11:46 +0100)]
Better self-document init_zobrist()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix initialization of BSFTable[]
Marco Costalba [Sun, 12 Jun 2011 09:49:34 +0000 (10:49 +0100)]
Fix initialization of BSFTable[]

We should start from i = 0, it works by accident because
static storage BSFTable[] is init to zero by default.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDon't update_gains() in qsearch
Marco Costalba [Sun, 12 Jun 2011 06:09:30 +0000 (07:09 +0100)]
Don't update_gains() in qsearch

Is almost unuseful becuase captures are skipped.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFix score_captures() for the case of capture promotions
Marco Costalba [Sat, 11 Jun 2011 21:11:17 +0000 (22:11 +0100)]
Fix score_captures() for the case of capture promotions

In case we have more than one promotion move, prefer
the one that captures the biggest piece.

Almost no functional change, anyhow I don't expect any
ELO change, it is just the correct thing to do.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire some unused functions
Marco Costalba [Sat, 11 Jun 2011 14:50:14 +0000 (15:50 +0100)]
Retire some unused functions

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoMicro-optimize castling handling in do_move()
Marco Costalba [Sat, 11 Jun 2011 11:41:20 +0000 (12:41 +0100)]
Micro-optimize castling handling in do_move()

And better self-document the code.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRetire PieceLetters struct
Marco Costalba [Fri, 10 Jun 2011 13:50:39 +0000 (15:50 +0200)]
Retire PieceLetters struct

Use a much simpler std::string instead.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoBetter document move_to_san()
Marco Costalba [Fri, 10 Jun 2011 06:49:12 +0000 (08:49 +0200)]
Better document move_to_san()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoRevert refinedValue in ProbCut
Marco Costalba [Thu, 9 Jun 2011 20:57:42 +0000 (21:57 +0100)]
Revert refinedValue in ProbCut

It seems much worst in number of nodes seacrhed to reach
the depth and anyhow does not give any advantage to the
Onno's oroginal one.

So revert by now and perhaps readd when we find something
clearly better.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse double rotate for magic generation
Marco Costalba [Thu, 9 Jun 2011 17:09:19 +0000 (18:09 +0100)]
Use double rotate for magic generation

Allow to choose among 4096 instances of pseudo-random
sequences instead of the previous 64 so the probability
to find a better sequence increases and actually we have
a much better 64 bit case and we can also use the 64 bit
version of pick_magic() also for 32 bits and althoug sub
optimal, because now we can have more choices results are
even slightly better also for 32 bit.

Use also a faster submask().

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoUse refinedValue in ProbCut condition
Marco Costalba [Wed, 8 Jun 2011 17:03:26 +0000 (18:03 +0100)]
Use refinedValue in ProbCut condition

After 12613 games at 20"+0.1 on QUAD
Mod vs Orig 1870 - 1863 - 8880 ELO +0 (+- 3.3)

So no performance change but it is a code semplification
and also is more easy to understand.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSmall pick_magic() touches
Marco Costalba [Wed, 8 Jun 2011 06:04:13 +0000 (07:04 +0100)]
Small pick_magic() touches

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoFind magics on the fly
Marco Costalba [Tue, 7 Jun 2011 13:12:07 +0000 (15:12 +0200)]
Find magics on the fly

Good result for 32 bit case where computation is very fast,
still not satisfying on 64 bit case where the magics seem
a bit harder to get.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoPartially revert previous patches
Marco Costalba [Mon, 6 Jun 2011 07:58:59 +0000 (09:58 +0200)]
Partially revert previous patches

Due to a -2% speed penalty. This patch takes the best
of the previous series without the regression due to
introduction of Magic struct.

Speedup against previous revision is of almost 3% !!!!

No functional change both in 32 and 64 bits.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSkip offset calculation in slider attacks
Marco Costalba [Sun, 5 Jun 2011 13:41:59 +0000 (14:41 +0100)]
Skip offset calculation in slider attacks

Another small simplification and micro optimization.

No functional change in both 32 and 64 bits.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoGet rid of Shift[] tables
Marco Costalba [Sun, 5 Jun 2011 11:09:55 +0000 (12:09 +0100)]
Get rid of Shift[] tables

We can calculate them counting the masks bits.

Also small tweak to sliding_attacks()

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoBetter name and document magic botboard stuff
Marco Costalba [Sun, 5 Jun 2011 08:04:57 +0000 (09:04 +0100)]
Better name and document magic botboard stuff

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoTry to keep memory access in the same cache line
Marco Costalba [Sat, 4 Jun 2011 15:36:52 +0000 (16:36 +0100)]
Try to keep memory access in the same cache line

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSkip draw by repetition check in qsearch
Marco Costalba [Sat, 4 Jun 2011 10:29:54 +0000 (11:29 +0100)]
Skip draw by repetition check in qsearch

Cut in half the time spent in pos.draw() that accounts
for a whopping 1% of total time !

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoMove bitboards initializations under one function
Marco Costalba [Sat, 4 Jun 2011 09:16:44 +0000 (10:16 +0100)]
Move bitboards initializations under one function

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoCalculate Bit Scan tables at initialization
Marco Costalba [Sat, 4 Jun 2011 07:54:57 +0000 (08:54 +0100)]
Calculate Bit Scan tables at initialization

Instead of hard-coding them.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoDo not sort negative non captures at low depth
Marco Costalba [Sat, 4 Jun 2011 05:25:16 +0000 (06:25 +0100)]
Do not sort negative non captures at low depth

Speedup of the whole 3 patch series is of 2,5% !!

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoSplit non capture in two sets when ordering
Marco Costalba [Fri, 3 Jun 2011 21:10:23 +0000 (22:10 +0100)]
Split non capture in two sets when ordering

But keep same ordering. This patch is prerequisite
for future work.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
10 years agoInline extension()
Marco Costalba [Fri, 3 Jun 2011 12:13:28 +0000 (13:13 +0100)]
Inline extension()

Signed-off-by: Marco Costalba <mcostalba@gmail.com>