And update master->splitPointsCnt under lock
protection. Not stricly necessary because
single_bit() condition takes care of false
positives anyhow, but it is a bit tricky and
moving under lock is the most natural thing
to do to avoid races with "reparenting".
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
In Young Brothers Wait Concept (YBWC) available slaves are
booked by the split point master, then start to search below
the assigned split point and, once finished, return in idle
state waiting to be booked by another master.
This patch introduces "Active Reparenting" so that when a
slave finishes its job on the assigned split point, instead
of passively waiting to be booked, searches a suitable active
split point and reprents itselfs to that split point. Then
immediately starts to search below the split point in exactly
the same way of the others split point's slaves. This reduces
to zero the time waiting in idle loop and should increase
scalability especially whit many (8 or more) cores.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Unfortunatly accessing thread local variable
is much slower than object data (see previous
patch log msg), so we have to revert to old code
to avoid speed regression.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Much faster then pthread_getspecific() but still a
speed regression against the original code.
Following are the nps on a bench:
Position
454165
454838
455433
tls
441046
442767
442767
ms (Win)
450521
447510
451105
ms (pthread)
422115
422115
424276
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
So to avoid a crash when setting the moves in
UCI "position startpos moves ...." command.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
But use the newly introduced local storage
for this. A good code semplification and also
the correct way to go.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Use thread local storage to store a pointer to the thread we
are running on. This will allow to remove thread info from
Position class.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
A std::set (that is a rb_tree) seems really
overkill to store at most a handful of moves
and nothing in the common case.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
In particualr before to wake up main thread that
could take some random time. Until we don't reset
search time we are not able to correctly track
the elapsed search time and this can be dangerous
under extreme time pressure.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
In this case SF stop searching and goes sleeping
waiting for a stop / ponderhit before to return
best move. So when a "stop" arrives we need to wake
up the main thread again.
Another regression introduced by 3aa471f2a9,
hopefully the last one.
Thanks to Otello1984 to reporting this.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Fixes a not so rare crash (once every 100 games)
newly introduced. Unfortunatly I am still not
able to figure out why :-(
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
And add final touches to this long patch series.
All the series has been verified against regression with
20K games at fast TC.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
We cannot set do_sleep flag of main thread before
"bestmove" is sent to GUI, otherwise GUI could send
immediately the next "go" command that triggers
start_thinking() and because do_sleep is set UI
thread resets the flag to launch a new search. But
when shortly after main thread returns to main_loop()
flag is incorrectly reset and main thread goes to sleep
hanging the engine.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
We store pointers instead of Thread objects because
Thread is not copy-constructible nor copy-assignable
and default ones are not suitable. So we cannot store
directly in a std::vector.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Associate platform OS thread to the Thread class instead of
creating it from ThreadsManager.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Split the data allocation, now done (mostly once)
in read_uci_options(), from the wake up and sleeping
of the slave threads upon entering/exiting the search.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
A somewhat tricky function pointer cast allows us
to move the platform specifics to lock.h, the cast
is tricky because return type is not the same of the
casted function in Linux (for Windows return type is
a DWORD that is a long) but this should not be a
problem as long as the size is the same;
From: http://stackoverflow.com/questions/188839/function-pointer-cast-to-different-signature
"OpenSSL was only casting functions pointers to
other function types taking and returning the same
number of values of the same exact sizes, and this
(assuming you're not dealing with floating-point)
happens to be safe across all the platforms and
calling conventions I know of. However, anything
else is potentially unsafe."
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Introduce 'on change' actions that are triggered as soon as
an UCI option is changed by the GUI. This allows to set hash
size before to start the game, helpful especially on very fast
TC and big TT size.
As a side effect remove the 'button' type option, that now
is managed as a 'check' type.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
And introduce SPlitPoint bestMove to pass back the
best move after a split point.
This allow to define as const the search stack passed
to split.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
We shouldn't need lock protection to increment
splitPointsCnt and set curSplitPoint of masterThread.
Anyhow because this code is very tricky and prone to
races bound the change in a single patch.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
But only when needed, after a split point. This behaviour
does not apply when useSleepingThreads is false, becuase
in this case threads are not woken up at split points so
must be already running.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
This method belongs to Thread, not to ThreadsManager.
Reshuffle stuff in thread.cpp while there.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Release split point lock before to wake up
master thread. This seems to increase speed
in case "sleeping threads" are used:
After 7792 games with 4 threads at very fast TC (2"+0.05)
Mod vs Orig 1722 - 1627 - 4443 ELO +4 (+- 5.1)
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
When allocating a slave we set both is_searching
and splitPoint under lock protection.
Unfortunatly the order in which the variables are
set is not defined. This article was very clarifying:
http://software.intel.com/en-us/blogs/2007/11/30/volatile-almost-useless-for-multi-threaded-programming/
So when in idle loop we test for is_searching and then
access splitPoint, it could happen that splitPoint is still
not updated leading to a possible crash.
Fix the race lock protecting splitPoint access.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Apart from some renaming the biggest change
is the retire of split_point_finished()
replaced by slavesMask flags. As a side
effect we now take also split point lock
when allocation available threads.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Instead of Posix threads. This seems to fix time
losses of the gcc compiled version for Windows.
The patch replaces the MSVC specific _MSC_VER flag
with _WIN32 and _WIN64 that are defined both by
MSVC and mingw-gcc.
Workaround found by Jim Ablett.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
pass references (Windows style) instead of
pointers (Posix style) as function arguments.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
And directly pass RootMoves instead of SearchMoves
to main thread. A class declaration is better suited
in a header and slims a bit the fatty search.cpp
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
We just need to verify if a legal move is among the
SearchMoves, so we don't need a vector for this.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
New gcc 4.7 complains about casting a volatile pointer
to void* so assign the variables directly.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
We try hard not to lose on time even under extreme
time pressure. We achieve this through 3 different but
coordinated steps:
1) Increase max frequency of timer events
2) Quickly return after a stop signal
3) Take in account timer resolution
With these SF played under LittleBlitzer at 1"+0.02 and 3"+0
without losing on time even one game.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Before creating main thread we set its do_sleep flag to true,
then thread is created and it will go to sleep in main_loop()
after resetting do_sleep.
But if after the setting of do_sleep and before its resetting
the UI thread calls start_thinking() it will not wait on:
if (!asyncMode)
while (!main.do_sleep)
cond_wait(&sleepCond, &main.sleepLock);
as it should but will immediately return before the main thread has
started the search. This very rare race show itself during bench,
when the first position is erroneusly skipped so that bench node count
results of 5309038 instead of the correct 5457475.
The patch is somewhat tricky, but is simple and it works!
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
It seems it yields to missing wake-up events with the
result of SF loosing on time as reported by many people.
So revert the patch and use a more robust approach: assume
there can be spurious wake ups events and make the code to
work also in those cases.
While debugging I found that WaitForSingleObject() had wrong
parameter 0 instead of INFINITE yielding to a crash while
exiting under Windows, strangely unnoticed til now.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Currently after a 'quit' command UI thread raises stop
signal, exits from uci_loop() and calls Threads.exit()
while the search threads are still active.
In Threads.exit() main thread is asked to terminate, but
if it is parked in idle_loop() it will exit and free its
resources (in particular the shared Movepicker object) while
sibling slaves are still active and this leads to a crash.
The fix is to let the UI thread always wait for main thread
to finish the search before to return from uci_loop().
Found by Valgrind when running with 8 threads.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Greatly improves the usage. User defined conversions
are a novelity for SF, another amazing C++ facility
at work !
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Make a better use of C++ operators overloading to
streamline the APIs.
Also sync polyglot.ini file while there.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Diretcly use the underlying std::vector<Move> and the
STL algorithms. Also a bit of cleanup while there.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
