This fixes an assert while testing with debug on.
Assert was due to static null pruning returning value
eval - futility_margin(depth, (ss-1)->futilityMoveCount)
That was sometimes higher than VALUE_INFINITE triggering
an assert at the caller site.
Because eval con be equal to ttValue and anyhow is read from
TT that can be corrupted in SMP case, we need to sanity
check it before to use.
bench: 4176431
Let TT clusters (16*4=64 bytes) to hold on a singe cache line.
This avoids the need for the double prefetch.
Original patches by Lucas and Jean-Francois that has also tested
on his AMD FX:
BIG HASHTABLE
./stockfish bench 1024 1 18 > /dev/null
Before:
1437642 nps
1426519 nps
1438493 nps
After:
1474482 nps
1476375 nps
1475877 nps
SMALL HASHTABLE
./stockfish bench 128 1 18 > /dev/null
Before:
1435207 nps
1435586 nps
1433741 nps
After:
1479143 nps
1471042 nps
1472286 nps
No functional change.
Crash is due to uninitialized ss->futilityMoveCount that
when happens to be negative, yields to an out of range
access in futility_margin().
Bug is subtle because it shows itself only in SMP case.
Indeed in single thread mode we only use the
Stack ss[MAX_PLY_PLUS_2];
Allocated at the begin of id_loop() and due to pure
(bad) luck, it happens that for all the MAX_PLY_PLUS_2
elements, ss[i].futilityMoveCount >= 0
Note that the patch does not prevent futilityMoveCount
to be overwritten after, for instance singular search
or null verification, but to keep things readable and
because the effect is almost unmeasurable, we here
prefer a slightly incorrect but simpler patch.
bench: 4311634
Instead of a pointer. This should fix the issue of
remaining with a stale pointer when for instance calling
IID, but also null search verification, singular search
and razoring where we call search with the same ss
pointer. In this case ss->ei is overwritten in the
search() call and upon returning remains stale.
This patch could have a performance hit because Eval::Info
is big (176 bytes) and during splitting we copy 4 ss entries.
On the good side, this patch is a clean solution.
Proposed by Gary.
No functional change.
Allow to use EvalInfo struct, populated by
evaluation(), in search.
In particular we allocate Eval::Info on the stack
and pass a pointer to this to evaluate().
Also add to Search::Stack a pointer to Eval::Info,
this allows to reference eval info of previous/next
nodes.
WARNING: Eval::Info is NOT initialized and is populated
by evaluate(), only if the latter is called, and this
does not happen in all the code paths, so care should be
taken when accessing this struct.
No functional change.
The idea is to fail high more easily in static
null test if in the parent node we are already
very deep in the move list, so the propability
to fail high there is very low.
[edit: I have slightly changed the functionality
moving
ss->futilityMoveCount = moveCount;
At the end of the pruning code, this should not affect
ELO in anyway, but makes code more natural and logic]
Test with SPRT is good at 15+0.05
LLR: 2.96 (-2.94,2.94)
Total: 50653 W: 10024 L: 9780 D: 30849
And at 60+0.05
LLR: 2.97 (-2.94,2.94)
Total: 40799 W: 7227 L: 6921 D: 26651
bench: 4530093
When we use sysconf(_SC_NPROCESSORS_ONLN) to get number of
cores, we have to include sysconf library that is unistd.h
Sometimes it happens to work just becuase unistd.h indirectly
included by some other libraries, but not always.
Reported and fixed by Eyal BD
No functional change.
The idea is to penalize a bishop in case of
its pawns are on the same colored squares.
Good at short 15"+0.05 TC
LLR: 2.95 (-2.94,2.94)
Total: 4252 W: 925 L: 806 D: 2521
And at longer 60"+0.05 TC
LLR: 2.95 (-2.94,2.94)
Total: 15006 W: 2743 L: 2564 D: 9699
bench: 5274705
Use an optinal argument instead of a template
parameter. Interestingly, not only is simpler,
but also faster, perhaps due to less L1 instruction
cache pressure because we don't duplicate the very
used SEE code path.
No functional change.
According to Jean-Paul this setup should be stronger
than default.
And SPRT test seems to confirm it:
At fast TC 15"+0.05
ELO: 3.33 +-2.7 (95%) LOS: 99.2%
Total: 25866 W: 5461 L: 5213 D: 15192
At longer TC 60"+0.05
ELO: 7.27 +-5.0 (95%) LOS: 99.8%
Total: 6544 W: 1212 L: 1075 D: 4257
bench: 5473339
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Increasing depth limit to 10 plies seems stronger
after 16K games at 15"+0.05 (ELO: +3.56) and also
repeating the test at 60"+0.05 TC:
ELO: 2.08 +-3.1 (95%) LOS: 90.9%
Total: 16000 W: 2641 L: 2545 D: 10814
Moreover setting the limit to 12 is proved stronger
then limit set to 10 by direct SPRT test at 15"+0.05:
ELO: 2.56 +-2.0 (95%) LOS: 99.5%
Total: 46568 W: 9240 L: 8897 D: 28431
So we directly set the limit to 12, the strongest setup.
bench: 4361224
Master IID formula is depth / 2
Previous patch is depth - 4 * ONE_PLY
This one is the middle way:
(dept/2 + depth-4*ONE_PLY)/2 -> depth-2*ONE_PLY-depth/4
After 16000 games at 60+0.05 th 1
ELO: 4.08 +-3.1 (95%) LOS: 99.5%
Total: 16000 W: 2742 L: 2554 D: 10704
bench: 4781239
Same idea of 5af8179647
in qsearch() but applied to search()
After 15500 games at 15+0.05
ELO: 4.48 +-3.4 (95%) LOS: 99.5%
Total: 15500 W: 3061 L: 2861 D: 9578
bench: 4985829
This reverts commit a24da071f0
Seems a regression when tested against 2.3.1
With this patch, have after 20000 games at 60+0.05, we have
ELO: 13.42 +-4.8 (95%) LOS: 100.0%
Total: 20000 W: 3746 L: 2974 D: 13280
Instead with the patch reverted:
ELO: 16.62 +-4.8 (95%) LOS: 100.0%
Total: 20000 W: 3816 L: 2860 D: 13324
Although we are within error bounds here we take the conservative
approach of not introducing changes that are not proved stronger
It doesn't mean that the change shall be weaker, simply that we
don't want to take any risk.
No functional change.
This seems good at short TC controls.
After 10000 games at 20+0.05
ELO: 9.56 +-6.8 (95%) LOS: 100.0%
Total: 10000 W: 1949 L: 1674 D: 6377
Testing at long TC and regression testing is still
ongoing. So this is a bit speculative commit and
could be reverted in the future.
Also re-testing at long TC the SEE pruning in PV nodes
seems less effective (perhaps even a regression, but
still ongoing) so disabled for now.
bench: 4968764
This reverts commit 0d68b523a3.
After easy move semplification this machinery is not
needed anymore (because of we don't need to know if a
root move is a recapture)
No functional change.
Detect a move as easy only if it is the only one ;-)
or if is much better than remaining ones after we
have spent 20% of search time.
Tests are ongoing, but it seems this semplification
stands. Anyhow it is experimental for now and could
be reverted/improved with further work Gary is
testing right now.
No functional change.
After previous patch if split point master is
waiting for job and "Use Sleeping Threads" is
false (our condition for official releases) then
it will lock/unlock splitPoint mutex in a super
tight loop badly affecting performance.
Rewrite the code to lock only when we are about
to finish. Note that race condition on slavesMask
is anyhow fixed.
No functional change.
We detect an easy move as a recapture with an
high margin on the second best move.
Unfortunatly the recapture detection is broken
becuase we identify as a recapture any move that
follows an opponent's previous capture !
This patch fix the logic to correctly detect a
real re-capture.
No functional change.
kf = (kf == FILE_A) ? kf++ : ....
is tricky becuase kf is updated twice and it happens
to do the right thing just by accident.
Rewrite in a better way.
Spotted by pdimov
No functional change.
Give a bonus if a bishop can pin a piece or can
give a discovered check through an x-ray attack.
Seems good after 24000 games at 15"+0.05 (single thread):
ELO: 12.30 +- 99%: 5.79 95%: 4.40 LOS: 100.00%
Total: 24000 W: 4931 L: 4082 D: 14987
bench: 4917064
Rename startPosPly to gamePly and increment/decrement
the variable after each do/undo move. This adds a little
overhead in do_move() but we will need to have the
game ply during the search for the next patches now
under test.
Currently we don't increment gamePly in do_null_move()
becuase it is not needed at the moment. Could change
in the future.
As a nice side effect we can now remove an hack in
startpos_ply_counter().
No functional change.
Another trick, along the same lines of previous
patch. This time we first check positions with
white side to move that, becuase we start with
pawn on rank 7, are easily classified as wins,
then black ones.
Number of cycles reduced to 15 !
Becuase now it is faster we can remove a lot of
code to detect theoretical draws. We will calculate
them anyhow, although a bit slower, but the speed
up trick more than compensates it.
Verified that generated bitbases match original ones.
No functional change.
Change the way the index is coded so that
now looping from 0 to IndexMax generates
the pawns from RANK_7 down to RANK2.
Becuase positions with pawns at RANK_7
are easily classified as wins/draws, this
small trick allows to reduce the number
of needed iterations from 30 down to 26!
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Rename to insertion_sort so to avoid confusion
with std::sort, also move it to movepicker.cpp
and use the bit slower std::stable_sort in
search.cpp where it is used in not performance
critical paths.
No functional change.
We can encode the ClusterSize directly in the
hashMask, this allows to skip the left shift.
There is no real change, but bench number is now
different because instead of using the lowest order
bits of the key to index the start of the cluster,
now we don't use the last two lsb bits that are
always set to zero (cluster size is 4). So for
instance, if 10 bits are used to index the cluster,
instead of bits [9..0] now we use bits [11..2].
This changes the positions that end up in the same
cluster affecting TT hits and so bench is different.
Also some renaming while there.
bench: 5383795
Do a 32bit bitwise 'and' instead of a 64bit
subtract and bitwise 'and'.
This is possible because even in the biggest
hash table case (8GB) the number of entries
is 2^29 so storable in an unsigned int.
No functional change.
Save the current active position in each Thread
instead of keeping a centralized array in struct
SplitPoint.
This allow to skip a memset() call at each split.
No functional change.
Obsolete renmant of when position was directly
passed to the search instead of being copied
for the main thread as is now.
From Jundery.
No functional change.
The syntax splitPoints() should force the compiler to
value-initialize the array and because there is no
user defined c'tor it falls back on zero-initialization.
Unfortunatly this is broken in MSVC compilers, because
value initialization for non-POD types is not supported,
so left splitPoints un-initialized and add in split()
initialization of slavesPositions, that is the only
member not already set at split time.
This fixes an assert under MSVC when running with
more than one thread.
Spotted and reported by Jundery.
No functional change.
To return a pointer to the available
thread instead of a bool. This allows
to simplify the core loop in split().
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
This function "returns" two values: bestValue and bestMove
Instead of returning one and passing as pointer the other
be consistent and pass as pointers both.
No functional change.
Currently a ttMove is reduced with ss->reduction = DEPTH_ZERO,
so it is actually not reduced (as it should be), but the
trick works just becuase it happens that ttMove is the first
to be tried and
reduction(depth, 1)
Always returns zero. So explicitly forbid reduction of ttMove
in the LMR condition. This is much clear and self-documented.
No functional change.
Surprisingly this rare case was not considered
when scoring a capture.
Also take in account that in the promotion case
we gain a new piece (typically a queen) but we
lose the promoting pawn.
These small issues were present since Glaurung times!
Found while browsing DiscoCheck sources
bench: 5400063
Handling of History and Gains is almost the same, with
the exception of the update logic, so unify both
classes under a single Stats struct.
No functional change.
And handle the castle directly in do/undo_move().
This allow to greatly simplify the code.
Here the beast is the nasty Chess960 that is
really tricky to get it right because could be
that 'from' and 'to' squares are the same or
that king's 'to' square is rook's 'from' square.
Anyhow should work: verified on all Chess960
starting positions.
No functional and no speed change also in Chess960.
Use a more traditional approach, along the same lines
of do_move().
It is true that we copy more in do_null_move(), but we
save the work in undo_null_move(). Speed test shows the
new code to be even a bit faster.
No functional change.
We have only one call place so inline its content.
BTW, function is already declared as FORCE_INLINE.
Also some small refactoring while there.
No functional change.
When a thread is allocated a bit is set in slavesMask.
This bit corresponds to the thread's index field that,
because it happens to be the position in the threads
array, eventually it is equal to the loop index 'i'.
But instead of relying on this 'coincidence', explicitly
use the 'idx' field so to clarify slavesMask usage.
Backported from c++11 branch.
No functional change.
Intel Compiler has 'invented' this pearl:
warning #1476: field uses tail padding of a base class
Just becuase we have subclassed MainThread and added
the field 'bool thinking'.
Pure nosense. Silence the warning.
No functional change.
Fix again TimerThread::idle_loop() to prevent a
theoretical race with 'exit' flag in ~Thread().
Indeed in Thread d'tor we raise 'exit' and then
call notify() that is lock protected, so we
have to check again for 'exit' before going to
sleep in idle_loop().
Also same change in Thread::idle_loop() where we
now check for 'exit' before to go to sleep.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Silly logic bug introduced in dda7de17e7
Timer thread, when msec = 0, instead of going
to sleep, calls check_time() in an endless loop.
Spotted and reported by snino64 due to abnormally
high CPU usage.
No functional change.
Subclass MainThread and TimerThread and declare
idle_loop() virtual. This allow us to cleanly
remove a good bunch of hacks, relying on C++
polymorphism to do the job.
No functional change.
Rename it is_finished and use it only in main
thread to signal search is finished. This allows
us to simplify the complex SMP logic.
Ultra tricky patch: deep test is required under
wide conditions like pondering on and option
"Use Sleeping Threads" set to false.
No functional change.
This reverts commit 869c924410
I misunderstood here. Actually it can happen that
thread is created but still not entered idle_loop
and at the same time start_searching() is called.
Becuase 'do_sleep' is set start_searching() will
set it to false and start the search, but when,
at last, the thread enters idle_loop(), resets
the flag and goes to sleep: not what we want.
Revert the hack waiting for a better solution
in the next patches.
No functional change.
Finally we can now merge the 'ponderhit' case with
'stop' and 'quit'.
The patches have been done step by step to help debugging
becuase this is really tricky code.
No functional change.
Reset Limits.ponder only if search continue, but if
we are going to stop the search there is no need
(and is also confusing) to clear the 'ponder' flag.
This mimics the behaviour upon rceiving 'stop' when
pondering.
No functional change.
In SAN notation when two pieces of the same type can
move to a given destination square, a disambiguation
additional info (like starting file) shall be added
to the SAN move.
If one of the two pieces is pinned, the corresponding
move _could_ be illegal and in this case disambiguation
is not needed. But to be pinned alone it is not enough
to deduce that the move is illegal, for instance in this
position:
R3rk2/2r6/8/8/8/8/8/K7 b - - 0 1
The move Rc8 is ambiguous although the rook in e8 is pinned
and the correct SAN notation should be Rcc8.
No functional change.
Don't wait for the search to finish after a 'stop'
command, but keep processing the GUI input if any.
Also explicitly wake up the main thread (that could be
sleeping) after a 'stop' or 'quit' command and do not
rely on wait_for_search_finished() doing it for us.
This patch cleans up the code and functions's definitions,
but it is risky and needs a good test under different
conditions to be sure it does not introduces hungs up.
No functional change.
Revert patch c581b7ea36
Seems a regression after testing from Gary:
ELO: 7.24 +- 99%: 17.03 95%: 12.93
LOS: 97.86%
Wins: 439 Losses: 381 Draws: 1962
And mine:
After 5410 games at 15"+0.05
Wins: 936 Losses: 1141 Draws: 3333 ELO -13
Moreover we know that there is a regression in the range
of patches which include the fromNull patch.
Probably this is not the only regression since 2.3.1 and
perhaps the idea under fromNull is good, but at the moment,
while in deep regression hunting, better to be on the safe
side and revert it entirely.
My guess on why this is a regression is that using the
negated evaluation of previous ply in case of null search
fails to take in account the king safety asymmetry between
the two colors. This is of course just a guess.
bench 5503830
Marco Costalba