stockfish
5 years agoAssorted code style of latest commits
Marco Costalba [Sun, 29 Mar 2015 07:24:17 +0000 (09:24 +0200)]
Assorted code style of latest commits

No functional chnage.

5 years agoFix indentations for hanging pawns code
Joona Kiiski [Sun, 29 Mar 2015 07:54:25 +0000 (08:54 +0100)]
Fix indentations for hanging pawns code

No functional change

5 years agoRemove some difficult to understand C++11 constructs
lucasart [Fri, 27 Mar 2015 01:44:57 +0000 (09:44 +0800)]
Remove some difficult to understand C++11 constructs

Code like this is more a case of showing off one's C++ knowledge, rather than
using it adequately, IMHO.

**First loop (std::generate)**

Iterators are inadequate here, because they lose the key information which is
idx. As a result, we need to carry a redundant idx variable, and increment it
along the way. Very clumsy.
Usage of std::generate and a lambda function only obfuscate the code, which is
merely a simple and stupid loop over the elements of a vector.

**Second loop (std::accumulate)**

This code is thoroughlly incomprehensible. Restore the original, which was much
simpler to understand.

**Third loop (range based loop)**

Again, a range based loop is inadequate, because we lose idx! To resolve this
artificially created problem, the data model was made redundant (idx is a data
member of db[] elements!?), which is ugly and unjustified. A simple and stupid
for loop with idx does the job much better.

No functional change.

Resolves #313

5 years agoPSV3_1
Alain SAVARD [Sat, 28 Mar 2015 23:30:46 +0000 (07:30 +0800)]
PSV3_1

Small speed-up in pawn.cpp
Results for 10 tests for each version:

Base      Test      Diff
Mean    1435636   1445238   -9602
StDev   22576     23189     1848

p-value: 1
speedup: 0.007

No functional change

Resolves #295

5 years agoGive a reduced bonus for threats by hanging pawns
Ajith [Sat, 28 Mar 2015 23:27:43 +0000 (07:27 +0800)]
Give a reduced bonus for threats by hanging pawns

Passed STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 105539 W: 20389 L: 20001 D: 65149

and LTC:
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 9629 W: 1577 L: 1432 D: 6620

Bench: 7658627

Resolves #317

5 years agoRetire follow-up move heuristic
mbootsector [Sat, 28 Mar 2015 22:09:48 +0000 (22:09 +0000)]
Retire follow-up move heuristic

STC: http://tests.stockfishchess.org/tests/view/5501d0f30ebc5902160ec0fd
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 34891 W: 6904 L: 6808 D: 21179

LTC: http://tests.stockfishchess.org/tests/view/550328540ebc5902160ec133
LLR: 3.10 (-2.94,2.94) [-3.00,1.00]
Total: 182653 W: 29866 L: 29993 D: 122794

Bench: 8396161

Resolves #310

5 years agoUse CounterMoveHistory when calculating LMR for cut nodes
VoyagerOne [Sat, 28 Mar 2015 21:15:49 +0000 (21:15 +0000)]
Use CounterMoveHistory when calculating LMR for cut nodes

If the sum of CounterMoveHistory heuristic and History heuristic is below zero,
then reduce an extra ply in cut nodes

LTC:
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 6479 W: 1099 L: 967 D: 4413

Bench: 7773299

Resolves #315

5 years agoClean up previous patch
Marco Costalba [Wed, 25 Mar 2015 05:47:40 +0000 (06:47 +0100)]
Clean up previous patch

No functional change.

5 years agoTuned values for the pawn piece square table
joergoster [Tue, 24 Mar 2015 23:38:32 +0000 (07:38 +0800)]
Tuned values for the pawn piece square table

Quick measure at very fast tc:
ELO: 4.77 +-2.2 (95%) LOS: 100.0%
Total: 40124 W: 8711 L: 8160 D: 23253

LTC:
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 52284 W: 8880 L: 8559 D: 34845

Bench: 8865736

Resolves #311

5 years agoIntroduce a new counter move history penalty
VoyagerOne [Tue, 24 Mar 2015 23:00:31 +0000 (23:00 +0000)]
Introduce a new counter move history penalty

Extra penalty for TT move in previous ply when it gets refuted

STC:

LLR: 2.94 (-2.94,2.94) [-1.50,4.50]
Total: 31303 W: 6216 L: 6025 D: 19062

LTC:

LLR: 2.97 (-2.94,2.94) [0.00,6.00]
Total: 6950 W: 1189 L: 1054 D: 4707

Bench: 8191926

Resolves #309

5 years agoFully yielding locks, no spinning
Joona Kiiski [Sat, 14 Mar 2015 21:20:11 +0000 (21:20 +0000)]
Fully yielding locks, no spinning

7 threads:

ELO: 2.00 +-2.7 (95%) LOS: 92.4%
Total: 20000 W: 3276 L: 3161 D: 13563

There is no functional change in single thread mode

Resolves #304

5 years agoAllow Bitbases::init() to be called more than once
Marco Costalba [Sun, 22 Mar 2015 11:41:11 +0000 (12:41 +0100)]
Allow Bitbases::init() to be called more than once

Currently if we call it more than once, we crash.

This is not a real problem, because this function is
indeed called just once. Nevertheless with this small fix,
that gets rid of a hidden 'static' variable, we cleanly
resolve the issue.

While there, fix also ThreadPool::exit to return in a
consistent state. Now all the init() functions but
UCI::init() are reentrant and can be called multiple
times.

No functional change.

5 years agoDouble magics generation speed
Marco Costalba [Sun, 22 Mar 2015 10:27:32 +0000 (11:27 +0100)]
Double magics generation speed

Profiling shows that resetting attacks table after
a failed candidate magic attempt is the biggest
time consumer, so rewrite the logic avoiding the
memset()

Magics init for rook+bishop goes from 200msecs to
under 100msec.

No functional change.

5 years agoGet rid of nativeThread
Marco Costalba [Sat, 21 Mar 2015 10:50:14 +0000 (11:50 +0100)]
Get rid of nativeThread

No functional change.

5 years agoUse only one ConditionVariable to sync UI
Marco Costalba [Fri, 20 Mar 2015 16:58:43 +0000 (17:58 +0100)]
Use only one ConditionVariable to sync UI

To sync UI with main thread it is enough a single
condition variable because here we have a single
producer / single consumer design pattern.

Two condition variables are strictly needed just for
many producers / many consumers case.

Note that this is possible because now we don't send to
sleep idle threads anymore while searching, so that now
only UI can wake up the main thread and we can use the
same ConditionVariable for both threads.

The natural consequence is to retire wait_for_think_finished()
and move all the logic under MainThread class, yielding the
rename of teh function to join()

No functional change.

5 years agoFix comment
lucasart [Wed, 18 Mar 2015 14:10:13 +0000 (22:10 +0800)]
Fix comment

We always probe, but we do not prune at PV nodes.

No functional change.

Resolves #300

5 years agoRetire ConditionVariable
Marco Costalba [Wed, 18 Mar 2015 07:25:15 +0000 (08:25 +0100)]
Retire ConditionVariable

Now that we use spinlocks everywhere and don't put
threads to sleep while idle, we can use the slower
(but no more in hot path) std::condition_variable_any
instead of our homwgrown ConditionVariable struct.

Verified fo rno regression at STC with 7 threads:
ELO: -0.66 +-2.7 (95%) LOS: 31.8%
Total: 20000 W: 3210 L: 3248 D: 13542

No functional change

5 years agoTuned mobility with another SPSA run
joergoster [Fri, 20 Mar 2015 00:08:21 +0000 (08:08 +0800)]
Tuned mobility with another SPSA run

Further improved mobility values after another SPSA session, 50k
iterations.

Elo measure at very fast 9+0.05":
ELO: 3.40 +-2.2 (95%) LOS: 99.9%
Total: 40000 W: 8434 L: 8042 D: 23524

and LTC SPRT[0, 4]:
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 11052 W: 1874 L: 1687 D: 7491

Bench: 8226843

Resolves #301

5 years agoFix the comment for Position::is_draw()
joergoster [Tue, 17 Mar 2015 16:42:29 +0000 (17:42 +0100)]
Fix the comment for Position::is_draw()

We no longer check for insufficient material.

No functional change

Resolves #299

5 years agoconnected should be bool, not Bitboard
lucasart [Tue, 17 Mar 2015 14:01:07 +0000 (22:01 +0800)]
connected should be bool, not Bitboard

There's no reason to define it as a Bitboard, so for consistency, use bool.

This is even a speedup on my machine: i7-3770k, using gcc 4.9.1 (linux):

    stat        test     master    diff
    mean   2,341,338  2,327,998  13,134
    stdev     15,765     14,717   5,405

    speedup       0.56%
    P(speedup>0) 100.0%

No functional change.

Resolves #298

5 years agoReformat FastMove
Marco Costalba [Wed, 18 Mar 2015 07:12:59 +0000 (08:12 +0100)]
Reformat FastMove

Align to SF coding style.

Verified no regression:

LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 55938 W: 10893 L: 10835 D: 34210

No functional change.

5 years agoSimplify nosleep logic
Marco Costalba [Wed, 18 Mar 2015 07:01:50 +0000 (08:01 +0100)]
Simplify nosleep logic

Avoid redundant 'while' conditions. It is enough to
check them in the outer loop.

Quick tested for no regression 10K games at 4 threads
ELO: -1.32 +-3.9 (95%) LOS: 25.6%
Total: 10000 W: 1653 L: 1691 D: 6656

No functional change.

5 years agoFix a bogus use of mutex
Marco Costalba [Tue, 17 Mar 2015 06:31:45 +0000 (07:31 +0100)]
Fix a bogus use of mutex

Spinlock must be used instead.

Tested for no regression at 15+0.05 th 4:

LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 25928 W: 4303 L: 4190 D: 17435

No functional change.

Resolves #297

5 years agoRe-arrange history update code
Marco Costalba [Sun, 15 Mar 2015 09:05:57 +0000 (10:05 +0100)]
Re-arrange history update code

Unify the quites moves loop for both cases,
the compiler optimizes away the

 if (is_ok((ss-1)->currentMove))

inside loop, so that the result is same
speed as original.

No functional change.

5 years agoUse acquire() and release() for spinlocks
Marco Costalba [Sun, 15 Mar 2015 06:39:10 +0000 (07:39 +0100)]
Use acquire() and release() for spinlocks

It is more idiomatick than lock() and unlock()

No functional change.

5 years agoFix dependency generation for MacOSX
Vince Negri [Sat, 14 Mar 2015 14:14:37 +0000 (14:14 +0000)]
Fix dependency generation for MacOSX

No functional change

Resolves #290

5 years agoFix dependency generation for C++11
Joona Kiiski [Sat, 14 Mar 2015 21:12:13 +0000 (21:12 +0000)]
Fix dependency generation for C++11

No functional change

Resolves #291

5 years agoDo not sleep, but yield
Joona Kiiski [Sun, 15 Mar 2015 19:45:30 +0000 (19:45 +0000)]
Do not sleep, but yield

During the search, do not block on condition variable, but instead use std::this_thread::yield().

Clear gain with 16 threads. Again results vary highly depending on hardware, but on average it's a clear gain.

ELO: 12.17 +-4.3 (95%) LOS: 100.0%
Total: 7998 W: 1407 L: 1127 D: 5464

There is no functional change in single thread mode

Resolves #294

5 years agoNew values for Mobility and Outposts.
joergoster [Sat, 14 Mar 2015 21:35:52 +0000 (21:35 +0000)]
New values for Mobility and Outposts.

Both are the result of a SPSA tuning session with a custom book, 50k iterations each.

After an additional tuning session of the mobility values, tuning the delta values, with following result.

40k games at 9+0.05:
ELO: 4.13 +-2.2 (95%) LOS: 100.0%
Total: 40000 W: 8581 L: 8106 D: 23313

and LTC
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 36518 W: 6049 L: 5782 D: 24687

Bench: 8567402

Resolves #284

5 years agoLink with -static in mingw
Marco Costalba [Sat, 14 Mar 2015 07:49:50 +0000 (08:49 +0100)]
Link with -static in mingw

Fixes reported startup error about missing libwinpthread-1.dll
when the dll is not in the path.

The current -static-xxxx flags, introduced with:

https://github.com/official-stockfish/Stockfish/commit/373503f4a9a990054b5

Only take in account standard libraries, but not thread
library.

No functional change.

Resolves #289

5 years agoIntroduce yielding spin locks
Joona Kiiski [Thu, 12 Mar 2015 20:36:32 +0000 (20:36 +0000)]
Introduce yielding spin locks

Idea and original implementation by Stephane Nicolet

7 threads 15+0.05
ELO: 3.54 +-2.9 (95%) LOS: 99.2%
Total: 17971 W: 2976 L: 2793 D: 12202

There is no functional change in single thread mode

5 years agoRemove check for gcc version from Makefile.
Joona Kiiski [Thu, 12 Mar 2015 22:05:18 +0000 (22:05 +0000)]
Remove check for gcc version from Makefile.

This check is obsolete.
very old gcc versions can't compile c++11 code.

No functional change

Resolves #285

5 years agoNew easy move implementation
mstembera [Thu, 12 Mar 2015 19:49:30 +0000 (19:49 +0000)]
New easy move implementation

Spend much less time in positions where one move is much better than all other alternatives.
We carry forward pv stability information from the previous search to identify such positions.
It's based on my old InstaMove idea but with two significant improvements.

1) Much better instability detection inside the search itself.
2) When it's time to make a FastMove we no longer make it instantly but still spend at least 10% of normal time verifying it.

Credit to Gull for the inspiration.
BIG thanks to Gary because this would not work without accurate PV!

20K
ELO: 8.22 +-3.0 (95%) LOS: 100.0%
Total: 20000 W: 4203 L: 3730 D: 12067

STC
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 23266 W: 4662 L: 4492 D: 14112

LTC
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 12470 W: 2091 L: 1931 D: 8448

Resolves #283

5 years agoIntroduce Counter Move History tables
Stefan Geschwentner [Thu, 12 Mar 2015 07:29:57 +0000 (07:29 +0000)]
Introduce Counter Move History tables

Introduce a counter move history table which additionally is indexed by the last move's piece and target square.
For quiet move ordering use now the sum of standard and counter move history table.

STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 4747 W: 1005 L: 885 D: 2857

LTC:
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 5726 W: 1001 L: 872 D: 3853

Because of reported low NPS on multi core test
STC (7 threads):
ELO: 7.26 +-3.3 (95%) LOS: 100.0%
Total: 14937 W: 2710 L: 2398 D: 9829

Bench: 7725341

Resolves #282

5 years agoUse thread specific mutexes instead of a global one.
Joona Kiiski [Wed, 11 Mar 2015 21:50:41 +0000 (21:50 +0000)]
Use thread specific mutexes instead of a global one.

This is necessary to improve the scalability with high number of cores.

There is no functional change in a single thread mode.

Resolves #281

5 years agoRetire spinlocks
Marco Costalba [Tue, 10 Mar 2015 20:50:45 +0000 (21:50 +0100)]
Retire spinlocks

Use Mutex instead.

This is in preparaation for merging with master branch,
where we stilll don't have spinlocks.

Eventually spinlocks will be readded in some future
patch, once c++11 has been merged.

No functional change.

5 years agoCleanup thread_win.h
Marco Costalba [Tue, 10 Mar 2015 16:13:52 +0000 (17:13 +0100)]
Cleanup thread_win.h

No functional change.

5 years agoDisable spinlocks
Marco Costalba [Tue, 10 Mar 2015 11:47:49 +0000 (12:47 +0100)]
Disable spinlocks

To allow testing on fishtest.

No functional change.

5 years agoAdd thread_win32.h header
Marco Costalba [Tue, 10 Mar 2015 11:42:40 +0000 (12:42 +0100)]
Add thread_win32.h header

Workaround slow std::thread implementation in mingw
and gcc for Windows with our own old low level thread
functions.

No functional change.

5 years agoSync with master
Marco Costalba [Sat, 7 Mar 2015 09:52:55 +0000 (10:52 +0100)]
Sync with master

bench: 8285241

5 years agoSync with master
Marco Costalba [Sat, 7 Mar 2015 09:41:07 +0000 (10:41 +0100)]
Sync with master

No functional change.

5 years agoTransform minKingPawnDistance into a local variable
hxim [Sat, 7 Mar 2015 09:03:19 +0000 (09:03 +0000)]
Transform minKingPawnDistance into a local variable

minKingPawnDistance is used only as local variable in one place so we don't need it to be part of "Pawns::Entry" structure.

No functional change.

Resolves #277

5 years agoUpdate Makefile for Mac OS X compilation
Stéphane Nicolet [Mon, 2 Mar 2015 21:01:19 +0000 (22:01 +0100)]
Update Makefile for Mac OS X compilation

This change in the Makefile restores the possibility to compile
Stockfish on Mac OS X 10.9 and 10.10 after the C++11 has been merged.

To use the default (fastest) settings, compile with:

make build ARCH=x86-64-modern

To test the clang settings, compile with

make build ARCH=x86-64-modern COMP=clang

Beware that the clang settings may provide a slightly slower (6%)
executable.

Backported from master.

No functional change

Resolves #275

5 years agoRe-enable spinlocks
Marco Costalba [Sat, 7 Mar 2015 07:36:23 +0000 (08:36 +0100)]
Re-enable spinlocks

For branch C++11, that doe snot run on fishtest,
there is no need of this kludge, let only master
have it.

No functional change.

5 years agoRevert C++11 merge
Joona Kiiski [Sat, 7 Mar 2015 07:38:22 +0000 (07:38 +0000)]
Revert C++11 merge

Restore the state of repo back to commit 'Simplify pawn code a bit' (1e6d21dbb6)

No functional change

5 years agoUpdate Makefile for Mac OS X compilation
Stéphane Nicolet [Mon, 2 Mar 2015 21:01:19 +0000 (22:01 +0100)]
Update Makefile for Mac OS X compilation

This change in the Makefile restores the possibility to compile
Stockfish on Mac OS X 10.9 and 10.10 after the C++11 has been merged.

To use the default (fastest) settings, compile with:

make build ARCH=x86-64-modern

To test the clang settings, compile with

make build ARCH=x86-64-modern COMP=clang

Beware that the clang settings may provide a slightly slower (6%)
executable.

No functional change

Resolves #275

5 years agoDisable spinlocks
Marco Costalba [Mon, 2 Mar 2015 07:11:39 +0000 (08:11 +0100)]
Disable spinlocks

Now that c++11 branch has been merged in master,
disable unconditionally the spinlocks and use mutex
instead. This will allow to run fishtest even on HT
machines withouth changes.

In the future we will reintorduce spinlocks, once
we will have took care of fishtest.

No functional change.

5 years agoAllow to disable spinlocks
Marco Costalba [Sun, 1 Mar 2015 16:12:09 +0000 (17:12 +0100)]
Allow to disable spinlocks

And use mutex instead. You may never want to do this.
It is a workaround to run c++11 on fishtest where many
machiens have HTenabled and this can be a problem when
number of cores set is higher than number of physical cores.

To disable spinlocks, just compile with -DNO_SPINLOCK flag

No functional change.

5 years agoRename available_to()
Marco Costalba [Sun, 1 Mar 2015 11:13:14 +0000 (12:13 +0100)]
Rename available_to()

Change this API to be more natural and simple.

Inspired by a patch by Joona.

No functional change.

5 years agoSync with master
Marco Costalba [Sat, 28 Feb 2015 19:21:38 +0000 (20:21 +0100)]
Sync with master

bench: 8285241

5 years agoSimplify pawn code a bit
Stéphane Nicolet [Sat, 28 Feb 2015 17:00:52 +0000 (01:00 +0800)]
Simplify pawn code a bit

Simplify a bit the number of bitwise operators used to calculate the
pawn evaluation in pawns.cpp

No functional change.

Resolves #269

5 years agoRaise penalty for knight attacked by pawn
Stéphane Nicolet [Sat, 28 Feb 2015 16:58:32 +0000 (00:58 +0800)]
Raise penalty for knight attacked by pawn

Raise a bit the penalty for knight attacked by pawn.

STC:
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 27744 W: 5563 L: 5380 D: 16801

LTC:
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 87571 W: 14747 L: 14314 D: 58510

Bench: 8285241

Resolves #270

5 years agoRetire apply_weight()
Marco Costalba [Fri, 27 Feb 2015 08:52:56 +0000 (09:52 +0100)]
Retire apply_weight()

Use the more natural operator*() instead.

No functional change.

5 years agoSync with master
Marco Costalba [Thu, 26 Feb 2015 19:49:41 +0000 (20:49 +0100)]
Sync with master

bench: 8069601

5 years agoNormalize twice supported pawns
Marco Costalba [Thu, 26 Feb 2015 17:56:16 +0000 (01:56 +0800)]
Normalize twice supported pawns

Align codying style to current conventions and move
formula for twice supported pawns to Pawns::init()
where it should be.

No functional change.

5 years agoApex Pawns
Alain SAVARD [Thu, 26 Feb 2015 17:50:35 +0000 (01:50 +0800)]
Apex Pawns

Pawns which are supported already have a bonus. Apex are pawns which are
supported twice.
This patch gives an additional 50% bonus for them.

STC
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 6549 W: 1333 L: 1209 D: 4007

LTC
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 18002 W: 3037 L: 2850 D: 12115

Bench: 8069601

Resolves #267

5 years agoDon't assume the type of Time::point
Marco Costalba [Tue, 24 Feb 2015 11:24:53 +0000 (12:24 +0100)]
Don't assume the type of Time::point

But instead use the proper definition. Also
rewrite chrono functions while there.

No functional change.

5 years agoSmall tweaks in pawns.cpp
Marco Costalba [Tue, 24 Feb 2015 10:33:40 +0000 (11:33 +0100)]
Small tweaks in pawns.cpp

No functional change.

5 years agoImprove spinlock implementation
Marco Costalba [Mon, 23 Feb 2015 18:22:37 +0000 (19:22 +0100)]
Improve spinlock implementation

Calling lock.test_and_set() in a tight loop creates expensive
memory synchronizations among processors and penalize other
running threads. So syncronize only only once at the beginning
with fetch_sub() and then loop on a simple load() that puts much
less pressure on the system.

Reported about 2-3% speed up on various systems.

Patch by Ronald de Man.

No functional change.

5 years agoUse spinlock instead of mutex for Threads and SplitPoint
Marco Costalba [Sun, 22 Feb 2015 13:59:55 +0000 (14:59 +0100)]
Use spinlock instead of mutex for Threads and SplitPoint

It is reported to be defenitly faster with increasing
number of threads, we go from a +3.5% with 4 threads
to a +15% with 16 threads.

The only drawback is that now when testing with more
threads than physical available cores, the speed slows
down to a crawl. This is expected and was similar at what
we had setting the old sleepingThreads to false.

No functional change.

5 years agoIntroduce Spinlock class
Marco Costalba [Sun, 22 Feb 2015 13:40:46 +0000 (14:40 +0100)]
Introduce Spinlock class

Initialization is more complex than what I'd like due
to MSVC compatibility that for some reason does not like:

std::atomic_flag lock = ATOMIC_FLAG_INIT;

No functional change.

5 years agoSync with master
Marco Costalba [Mon, 23 Feb 2015 12:35:04 +0000 (13:35 +0100)]
Sync with master

bench: 8253813

5 years agoFix build under OS X
Marco Costalba [Sun, 22 Feb 2015 15:30:53 +0000 (16:30 +0100)]
Fix build under OS X

Reported by Vince Negri

No functional change.

5 years agoSync with master
Marco Costalba [Sun, 22 Feb 2015 13:19:12 +0000 (14:19 +0100)]
Sync with master

bench: 8253813

5 years agoUse only 'level' as late join metric
Marco Costalba [Thu, 19 Feb 2015 13:14:20 +0000 (14:14 +0100)]
Use only 'level' as late join metric

It seems other metric are useless, this allow us
to simplify the code and to prune useless stuff.

STC 20K games 4 threads
ELO: -0.76 +-2.8 (95%) LOS: 29.9%
Total: 20000 W: 3477 L: 3521 D: 13002

STC 10K games 16 threads
ELO: 1.36 +-3.9 (95%) LOS: 75.0%
Total: 10000 W: 1690 L: 1651 D: 6659

bench: 8253813

5 years agoFurther refine SMP code
Marco Costalba [Sat, 21 Feb 2015 10:33:03 +0000 (11:33 +0100)]
Further refine SMP code

Backported from C++11 branch:

https://github.com/official-stockfish/Stockfish/commit/7ff965eebfbc17d2b
https://github.com/official-stockfish/Stockfish/commit/e74c2df907d5336d3d2b

Fully verified it is equivalent to master (see log msg
of individual commits for details).

No functional change.

5 years agoUse sp->master instead of bestThread
Marco Costalba [Sat, 21 Feb 2015 09:40:59 +0000 (10:40 +0100)]
Use sp->master instead of bestThread

Verified with:

dbg_hit_on(th != sp->master);

It is 100% equivalent on more than 200K hits.

No functional change.

5 years agoFix comment for kingAdjacentZoneAttacksCount
Stéphane Nicolet [Thu, 19 Feb 2015 20:33:07 +0000 (21:33 +0100)]
Fix comment for kingAdjacentZoneAttacksCount

The comment for kingAdjacentZoneAttacksCount[] was bogus, using
reversed semantics for color.

No functional change

Resolves #262

5 years agoMobile phalanxes
snicolet [Fri, 20 Feb 2015 19:56:57 +0000 (19:56 +0000)]
Mobile phalanxes

Try to create mobile phalanxes

STC:
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 52393 W: 10912 L: 10656 D: 30825

LTC:
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 30398 W: 5315 L: 5063 D: 20020

Bench: 8253813

Resolves #261

5 years agoImprove comments in SMP code
Marco Costalba [Fri, 20 Feb 2015 11:13:09 +0000 (12:13 +0100)]
Improve comments in SMP code

No functional change.

5 years agoUse range-based-for in late join
Marco Costalba [Fri, 20 Feb 2015 09:49:51 +0000 (10:49 +0100)]
Use range-based-for in late join

No functional change.

5 years agoSync with master
Marco Costalba [Fri, 20 Feb 2015 09:36:45 +0000 (10:36 +0100)]
Sync with master

bench: 7911944

5 years agoClarify we don't late join with only 2 threads
Marco Costalba [Thu, 19 Feb 2015 22:12:59 +0000 (23:12 +0100)]
Clarify we don't late join with only 2 threads

Thanks to Gary for pointing this out.

No functional change.

5 years agoUse size_t consistently across thread code
Marco Costalba [Thu, 19 Feb 2015 09:43:28 +0000 (10:43 +0100)]
Use size_t consistently across thread code

No functional change.

5 years agoRetire redundant sp->slavesCount field
Marco Costalba [Thu, 19 Feb 2015 09:27:24 +0000 (10:27 +0100)]
Retire redundant sp->slavesCount field

It should be used slavesMask.count() instead.

Verified 100% equivalent when sp->allSlavesSearching:

dbg_hit_on(sp->allSlavesSearching, sp->slavesCount != sp->slavesMask.count());

No functional change.

5 years agoFix a warning under MSVC
Marco Costalba [Thu, 19 Feb 2015 09:18:24 +0000 (10:18 +0100)]
Fix a warning under MSVC

Assignment of size_t to int.

No functional change.

5 years agoAdd a couple of asserts to late join
Marco Costalba [Thu, 19 Feb 2015 09:08:29 +0000 (10:08 +0100)]
Add a couple of asserts to late join

Document and clarify that we cannot rejoin on ourselves
and that we never late join if we are master and all
slaves have finished, inded in this case we exit idle_loop.

No functional change.

5 years agoRemove useless condition in late join
Marco Costalba [Thu, 19 Feb 2015 08:51:17 +0000 (09:51 +0100)]
Remove useless condition in late join

In case of Threads.size() == 2 we have that sp->allSlavesSearching
is always false (because we have finished our search), bestSp is
always NULL and we never late join, so there is no need to special
case here.

Tested with dbg_hit_on(sp && sp->allSlavesSearching) and
verified it never fires.

No functional change.

5 years agoCompute SplitPoint::spLevel on the fly
Marco Costalba [Tue, 17 Feb 2015 09:10:58 +0000 (10:10 +0100)]
Compute SplitPoint::spLevel on the fly

And retire a redundant field. This is important also
from a concept point of view becuase we want to keep
SMP structures as simple as possible with the only
strictly necessary data.

Verified with

dbg_hit_on(sp->spLevel != level)

that the values are 100% the same out of more 50K samples.

No functional change.

5 years agoSimplify attackUnits formula
Marco Costalba [Tue, 17 Feb 2015 07:23:35 +0000 (08:23 +0100)]
Simplify attackUnits formula

Use '/ 8' instead of '* 31 / 256'

Passed STC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 55077 W: 10999 L: 10940 D: 33138

And LTC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 14751 W: 2530 L: 2400 D: 9821

bench: 7911944

5 years agoImprove smp performance for high number of threads
Joona Kiiski [Sat, 14 Feb 2015 20:46:00 +0000 (20:46 +0000)]
Improve smp performance for high number of threads

Balance threads between split points.

There are huge differences between different machines and autopurging makes it very difficult to measure the improvement in fishtest, but the following was recorded for 16 threads at 15+0.05:

    For Bravone (1000 games): 0 ELO
    For Glinscott (1000 games): +20 ELO
    For bKingUs (1000 games): +50 ELO
    For fastGM (1500 games): +50 ELO

The change was regression for no one, and a big improvement for some, so it should be fine to commit it.
Also for 8 threads at 15+0.05 we measured a statistically significant improvement:
ELO: 6.19 +-3.9 (95%) LOS: 99.9%
Total: 10325 W: 1824 L: 1640 D: 6861

Finally it was verified that there was no (significant) regression for

4 threads:
ELO: 0.09 +-2.8 (95%) LOS: 52.4%
Total: 19908 W: 3422 L: 3417 D: 13069

2 threads:
ELO: 0.38 +-3.0 (95%) LOS: 60.0%
Total: 19044 W: 3480 L: 3459 D: 12105

1 thread:
ELO: -1.27 +-2.1 (95%) LOS: 12.3%
Total: 40000 W: 7829 L: 7975 D: 24196

Resolves #258

5 years agoCompute checkers from scratch
lucasart [Mon, 16 Feb 2015 01:34:26 +0000 (09:34 +0800)]
Compute checkers from scratch

This micro-optimization only complicates the code and provides no benefit.
Removing it is even a speedup on my machine (i7-3770k, linux, gcc 4.9.1):

stat        test     master    diff
mean   2,403,118  2,390,904  12,214
stdev     12,043     10,620   3,677

speedup       0.51%
P(speedup>0) 100.0%

No functional change.

5 years agoRetire one do_move() overload
Marco Costalba [Sun, 15 Feb 2015 11:20:47 +0000 (12:20 +0100)]
Retire one do_move() overload

After Lucas patch it is almost useless.

No functional change.

5 years agoCompute checkers from scratch
lucasart [Sun, 15 Feb 2015 07:49:20 +0000 (15:49 +0800)]
Compute checkers from scratch

This micro-optimization only complicates the code and provides no benefit.
Removing it is even a speedup on my machine (i7-3770k, linux, gcc 4.9.1):

stat        test     master    diff
mean   2,403,118  2,390,904  12,214
stdev     12,043     10,620   3,677

speedup       0.51%
P(speedup>0) 100.0%

No functional change.

5 years agoRevert "Delayed killers checking"
Marco Costalba [Sun, 15 Feb 2015 10:24:34 +0000 (11:24 +0100)]
Revert "Delayed killers checking"

It seems a slowdown when run with fishbench.

No functional change.

5 years agoFurther simplify KingDanger init
Marco Costalba [Sat, 14 Feb 2015 14:55:11 +0000 (15:55 +0100)]
Further simplify KingDanger init

And remove a tale whitespace while there.

No functional change.

5 years agoFurther simplify KingDanger init
Marco Costalba [Sat, 14 Feb 2015 14:55:11 +0000 (15:55 +0100)]
Further simplify KingDanger init

And remove a tale whitespace while there.

No functional change.

5 years agoSync with master
Marco Costalba [Sat, 14 Feb 2015 14:29:21 +0000 (15:29 +0100)]
Sync with master

Bench: 7369224

5 years agoSmall bonus for all safe pawn pushes
snicolet [Fri, 13 Feb 2015 21:33:00 +0000 (21:33 +0000)]
Small bonus for all safe pawn pushes

Pawn flexibility: add a small bonus for all safe pawn pushes

STC:
LLR: 2.70 (-2.94,2.94) [-1.50,4.50]
Total: 18233 W: 3705 L: 3557 D: 10971

LTC:
LLR: 2.97 (-2.94,2.94) [0.00,6.00]
Total: 17684 W: 3042 L: 2854 D: 11788

Bench: 7369224

Resolves #253

5 years agoReformat tracing functions
Marco Costalba [Fri, 13 Feb 2015 10:54:46 +0000 (11:54 +0100)]
Reformat tracing functions

No functional change.

5 years agoFix KingDanger[] array initialization
hxim [Mon, 9 Feb 2015 18:12:04 +0000 (19:12 +0100)]
Fix KingDanger[] array initialization

Use integer arithmetic instead of floating point arithmetic.
Floating point arithmetic was causing different results for some 32-bit compiles

No functional change

Resolves #249
Resolves #250

5 years agoSync with master
Marco Costalba [Sun, 8 Feb 2015 20:31:25 +0000 (21:31 +0100)]
Sync with master

bench: 7699138

5 years agoPawn Center Bind Bonus
Joona Kiiski [Sun, 8 Feb 2015 19:20:39 +0000 (19:20 +0000)]
Pawn Center Bind Bonus

Bonus for two pawns controlling the same central square

STC:

LLR: 3.14 (-2.94,2.94) [-1.50,4.50]
Total: 15974 W: 3291 L: 3133 D: 9550

LTC:

LLR: 3.24 (-2.94,2.94) [0.00,6.00]
Total: 10449 W: 1837 L: 1674 D: 6938

Idea from Lyudmil Tsvetkov.

Bench: 7699138

Resolves #248

5 years agoShuffle put_piece() and friends signatures
Marco Costalba [Sun, 8 Feb 2015 12:21:50 +0000 (13:21 +0100)]
Shuffle put_piece() and friends signatures

It is more consistent with the others member functions.

No functional change.

5 years agoSmall tweaks in do_move and friends
Marco Costalba [Sat, 7 Feb 2015 18:34:24 +0000 (19:34 +0100)]
Small tweaks in do_move and friends

Also remove useless StateCopySize64 optimization:
compiler uses SSE movups instruction anyhow and
does not need this trick (verified with fishbench).

No functional change.

5 years agoAvoid casting to char* in prefetch()
Marco Costalba [Sat, 7 Feb 2015 18:13:41 +0000 (19:13 +0100)]
Avoid casting to char* in prefetch()

Funny enough, gcc __builtin_prefetch() expects
already a void*, instead Windows's _mm_prefetch()
requires a char*.

The patch allows to remove ugly casts from caller
sites.

No functional change.

5 years agoRewrite pos_is_ok()
Marco Costalba [Sat, 7 Feb 2015 13:34:35 +0000 (14:34 +0100)]
Rewrite pos_is_ok()

No functional change.

5 years agoMicro-optimize SEE
Marco Costalba [Sat, 7 Feb 2015 11:07:43 +0000 (12:07 +0100)]
Micro-optimize SEE

Results for 10 tests for each version (gcc 4.8.3 on mingw):

            Base      Test      Diff
    Mean    1502447   1507917   -5470
    StDev   3119      1364      4153

p-value: 0,906
speedup: 0,004

Results for 10 tests for each version (MSVC 2013):

            Base      Test      Diff
    Mean    1400899   1403713   -2814
    StDev   1273      2804      2700

p-value: 0,851
speedup: 0,002

No functional change.

5 years agoRename dbg_hit_on_c() to dbg_hit_on()
Marco Costalba [Sat, 7 Feb 2015 10:15:38 +0000 (11:15 +0100)]
Rename dbg_hit_on_c() to dbg_hit_on()

Use an overload instead of a new named function.

I have found this handier and easier when adding
some quick debug code.

No functional change.

5 years agoRename dbg_hit_on_c() to dbg_hit_on()
Marco Costalba [Sat, 7 Feb 2015 10:15:38 +0000 (11:15 +0100)]
Rename dbg_hit_on_c() to dbg_hit_on()

Use an overload instead of a new named function.

I have found this handier and easier when adding
some quick debug code.

No functional change.