Joona Kiiski [Tue, 23 Feb 2010 10:05:26 +0000 (12:05 +0200)]
Do not wait for threads falling asleep
I cannot see any reason to do this. Even this is not enough to fix
theoretical race case on Windows which doesn't seem to cause any
problems in practice anyhow
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 20 Feb 2010 22:56:02 +0000 (23:56 +0100)]
Revert "Recursive lock"
Joona says that sp_update_pv() does not pass the split point
boundaries, so there is no risk to corrupt data from another
split point. Also the race on thread_should_stop() is harmless
because of this.
So revert the patch and come back to single lock.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 20 Feb 2010 17:36:07 +0000 (18:36 +0100)]
Revert small state change optimization in idle_loop()
Joona says:
1. We should not be afraid of "AllThreadsShouldExit" flag.
Because when this is set to true we _must_not_ be searching (= All
splits must have been undone).
And if we are not searching it's impossible that some other thread
could give us work to do. So setting state to THREAD_AVAILABLE
doesn't do any harm. If you want to add check for this, you could do
it like this:
if (threads[threadID].state == THREAD_WORKISWAITING)
{
+ assert(!AllThreadsShouldExit)
threads[threadID].state = THREAD_SEARCHING;
2a. If waitSp->cpus == 0, setting state to THREAD_AVAILABLE makes
no harm either, because helpful master concept dictates that _only_
our own slave can book us. If we don't have any slaves, noone has the
right to book us.
2b. If point (2a) is not correct then your extra check only adds extra race:
In smp code checking for waitSp->cpus > 0 is not enough. It's possible that
our slave immediately exits and another thread
books us as a slave when our state is still
THREAD_AVAILABLE. So instead of adding extra level of security we have
just introduced extra race.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 20 Feb 2010 17:29:53 +0000 (18:29 +0100)]
Recursive lock all split point's chain
When we found a cut-off then lock all the split point chain,
not only current one to avoid races in case two threads running
on different split points where one is ancestor then the other,
find a beta cut-off at the same time, in this case we want only
one to call sp_update_pv().
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 20 Feb 2010 16:52:09 +0000 (17:52 +0100)]
Retire per-thread stopRequest flag
This is a per split-point request, not per-thread. When we find
a beta cut-off in current thread's split point or in or in some
ancestor of the current split point then threads should stop
immediately the search and return to idle_loop().
The check is done by thread_should_stop() that now looks only
at split point's chain.
No functional change and a good semplification.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Joona Kiiski [Fri, 12 Feb 2010 14:40:00 +0000 (16:40 +0200)]
Use zero null move margin when depth < 4 * OnePly
This is because when we are below 4 * OnePly, the null move
will directly jump to qsearch and if we are below beta,
our opponent is above beta and will get immediate
stand pat cut off.
So basically this patch is just optimizing away useless
evaluation calls. dbg_hit_on() runs show that this heuristic
is correct >99% of cases. Transposition table probably causes
some inaccurary?
After 1148 games on QUAD
mod-orig: 583 - 565 +5 elo
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sun, 14 Feb 2010 15:29:28 +0000 (16:29 +0100)]
Fix another setting of a flag out of lock protection
In this case is dangerous because in split() we reset the flag to
false, but if it was set due to a cut-off higher in the tree we
completely miss that and go on with the full search.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sun, 14 Feb 2010 15:04:53 +0000 (16:04 +0100)]
Rename flag 'stop' in 'stopRequest'
Instead of other flags this is not a state flag, i.e. does
not defines a state for the thread, but a request because
after we raise 'stopRequest' flag the corresponding thread is
not stopped, but continues to run for a while until it returns
from sp_search() in idle_loop.
It is important the name reflects this.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sun, 14 Feb 2010 12:31:05 +0000 (13:31 +0100)]
Fix 'stop' flag changed out of lock protection
This is the first nice effect of previous patch !
Because thread_should_stop() should be declared 'const' we
need to remove the setting of 'stop' flag to true that
turns out to be a bug because thread_should_stop() is called
outside from lock protection while 'stop' flag is a volatile
shared variable so cannot be changed when not in lock.
Note that this bugs fires ONLY when we use more then 2 threads,
so commonly only in a QUAD or OCTAL machine.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 13 Feb 2010 12:38:32 +0000 (13:38 +0100)]
Introduce ThreadsManager class
Main aim of this patch is to consolidate all the thread related stuff
behind a single class interface so to avoid messing with global flags
and having thread code scattered among non-thread related stuff.
Another advantage is that now access to thread's variables is
more controlled, in particular we can differentiate between
read and write accesses by the mean of different interfaces, it
is so simpler to understand how a function is related to threads.
Lastly this rewrite is the base for future code consolidations and
semplifications that are easier now that we have only one thread's
access point.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 13 Feb 2010 11:39:27 +0000 (12:39 +0100)]
Ensure function boundaries for threads state changes
Ensure threads are sleeping when leaving init_threads() and
the newly introduced put_threads_to_sleep().
Also ensure threads are not sleeping when leaving
wake_sleeping_threads().
As a side effect we now leave think() with all the threads
(but the main one) guaranteed to sleep. So when we enter
again in think(), after the opponent next move, we know
threads must be sleeping.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Fri, 12 Feb 2010 05:49:16 +0000 (06:49 +0100)]
Allow build on HP-UX 11.X
Patch from Richard Lloyd (slightly edited from me), following the list
of changes as described by the author:
src/Makefile:
- Added PREFIX and BINDIR for the install: rule.
- Added a "make hpux" line to the help: rule.
- Added "make test"/"make check" rule that runs the $(PGOBENCH) command.
- "make clean" now additionally removes core and bench.txt.
- Added an hpux: rule.
- Added an install: rule to mkdir $(BINDIR), copy $(EXE) to $(BINDIR) and
then strip it.
- "make strip" now ensures that $(EXE) is built first before trying to
strip it.
- Hide errors and output from the g++ command used by the .depend: rule and
then touch .depend in case g++ isn't available.
- Hide errors from the "include .depend" in case .depend doesn't exist
(e.g. directly after a "make clean").
src/book.cpp and src/book.h:
- HP-UX's aCC really didn't like the const keywords used for the
Book::file_name() definitions, so they were removed. I checked that this
didn't affect a Linux build and it was still fine.
src/misc.cpp:
- HP-UX uses <sys/pstat.h> and pstat_getdynamic() to determine the number of
CPU cores, so added conditional code for that (if pstat_getdynamic() fails,
set the number of cores to 1).
src/tt.cpp:
- <xmmintrin.h> and _mm_prefetch() seem highly specific to the Intel x86(_64)
and gcc platforms - neither exist in HP-UX, so conditionally avoid that
code in HP-UX's case. Perhaps some sort of define is needed here
such as -DHAS_MM_PREFETCH that could be #ifdef'ed for instead?
Even after these changes, it's more convenient for HP-UX users to edit the
default: rule in the Makefile to run "$(MAKE) hpux" before they build
stockfish, but that's not a big deal if they're warned about that first (the
same applies to all other builds other than the standard "$(MAKE) gcc" one).
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Wed, 10 Feb 2010 20:27:44 +0000 (21:27 +0100)]
Fix a couple of new MSVC 2010 warnings
Compiler complains because in Book we have a d'tor but not
copy c'tor and assignement operator (warning C4511 and C4512),
note that after adding them (just declared) you now need also
default c'tor !
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Mon, 8 Feb 2010 09:48:58 +0000 (10:48 +0100)]
Retire EvalInfo* in SearchStack
It is an hidden bug waiting to fire. The main problem is
that ss[ply] is overwritten by search() and qsearch() called
from IID and razoring, so that we cannot hold a pointer to a
local EvalInfo variable.
For instance if we go razoring then we overwrite the pointer
with the address of a variable local to qsearch(), when we return
from qsearch() variable goes out of scope and now ss[ply].evalInfo
holds a stale pointer !
Because we are not looking for troubles we go through the
safe route and we remove it entirely.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Joona Kiiski [Fri, 5 Feb 2010 16:05:48 +0000 (18:05 +0200)]
Copy 4 SearchStack items in split()
In search routines we use information from previous ply
and init killers two plies ahead.
So for me it seems correct to copy 4 searchstack items
in split:
ply - 1, ply, ply + 1, ply + 2
Because
a) we do not split at root (ply == 0)
b) ply < PLY_MAX and SearchStack size is PLY_MAX_PLUS_2
there should be no risk of underflows or overflows
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Fri, 5 Feb 2010 16:37:33 +0000 (17:37 +0100)]
Remove sorting optimization for many zeroes
With negative history we don't have anymore a
lot of zeroes to score, so just split moves in
positives and non-positives sets.
Speed up is almost zero, we cannot test speed directly
because node count changed due to reorder, but I have
verified sorting is correct. With a profiler I have
seen we gain a little in sort_moves() and lose a little
in insertion_sort(), so the net effect is almost zero,
but code is simpler.
No real change, just move reordering.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Fri, 5 Feb 2010 07:19:39 +0000 (08:19 +0100)]
Be sure negative see evasions are at the bottom
Because H.move_ordering_score() can return negative values
some negative see moves could be searched before non-negative
see moves with negative history.
This patch restores proper ordering.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Mon, 1 Feb 2010 13:06:59 +0000 (14:06 +0100)]
Check bounds in set_option_value()
Normally it's up to the GUI to check for option's limits,
but we could receive the new value directly from the user
by teminal window. So let's check the bounds anyway.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Sat, 30 Jan 2010 18:18:27 +0000 (19:18 +0100)]
Add hardware POPCNT support for gcc
With new target 'make gcc-popcnt' it is now
possible to compile with enabled hardware POPCNT
support also with gcc. Until now was possible only
for Intel and MSVC compilers.
When this instruction is supported by CPU, for instance
on Intel i7 or i5 family, produced binary is a bit faster.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Produces following build error after 'make clean', 'make icc' under Mandriva
with icc version 11.0
Makefile:306: .depend: No such file or directory
In file included from tt.cpp:28:
/usr/lib/gcc/i586-manbo-linux-gnu/4.3.2/include/xmmintrin.h:35:3: error: #error "SSE instruction set not enabled"
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Thu, 28 Jan 2010 23:16:00 +0000 (00:16 +0100)]
Retire captures pruning
Futility captures alone does not seem an improvment.
Perhaps is a combination of stand pat + futility that is winning,
so revert for now and continue testing starting from a standard
base until we find the correct receipe.
After 999 games at 1+0
Mod vs Orig +231 0506 -201 +10 ELO
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Joona Kiiski [Tue, 26 Jan 2010 18:22:51 +0000 (20:22 +0200)]
Do not initialize RootPosition at startup
Initializing high-level object at startup is very dangerous,
because low-level snippets are not yet initialized.
For example Position's constructor calls find_checkers() which
calls attackers_to() which depends on various global bitboard arrays
which are not yet initialized. I think we are lucky not to crash.
RootPosition.from_fen(StartPosition); is called immediately after
all initializations are made at uci_main_loop() which is the
correct behaviour
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Wed, 27 Jan 2010 18:46:57 +0000 (19:46 +0100)]
Aspiration window rewrite
Joona new aspiration window. Main idea is to always
research aspiration fail highs/low at the same
ply and use much smaller aspiration window than previously.
Testing result is very positive.
1CPU:
953-1149
4CPU:
545 - 656
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Marco Costalba [Wed, 27 Jan 2010 16:29:34 +0000 (17:29 +0100)]
Simplify locking in sp_search and sp_search_pv
Avoid to take the lock two times in a tight sequence, the first
in get_next_move() and the second to update sp->moves.
Do all with one lock and so retire the now useless locked version
of get_next_move().
Also fix some theorical race due to comparison sp->bestValue < sp->beta
is done out of lock protection. Finally fix another (harmless but time
waster) race that coudl occur because thread_should_stop() is also
called outside of lock protection.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
projects / stockfish / log