And return on using TT as backing store for position
evaluations.
Tests (even on single thread) show eval cache was a regression.
In multi thread result should be even worst because eval cache
is a per-thread struct, while TT is shared.
After 4957 games at 15"+0.05 (single thread)
eval cache vs master 969 - 1093 - 2895 -9 ELO
So previous reported result of +18 ELO was probably due to an
issue in the testing framework (a bug in cutechess-cli) that
has been fixed in the meanwhile.
bench: 5386711
In case of null search at low depths returns a fail low
due to a threat then, rather than return beta-1 (to cause
a re-search at full depth in the parent node), we set a flag
threatExtension = true (false by default) that will cause
moves that prevent the threat to be extended of one ply
in the following search.
Idea and patch is by Lucas Braesch.
Lucas also did the tests:
1500 games in 5"+0.05":
SF_threatExtension vs SF_20121222: 366 - 331 - 803 [51.2%] LOS=90.8%
3000 games in 10"+0.1":
SF_threatExtension vs SF_20121222: 610 - 559 - 1831 [50.8%] LOS=93.2%
Tests confirmed by Gary after 10570 games,
ELO: 2.79 +- 99%: 8.72 95%: 6.63
LOS: 94.08%
Wins: 1523 Losses: 1438 Draws: 7607
And finally by me at 15"+0.05, single thread, 3824 games
threatExtension vs master 768 - 692 - 2364 +7 ELO
bench 4918443
This is difficult code becuase a bug here could lead
to very subtle crashes in case of SMP games where we
have TT move corruption due to concurrent access.
Anyhow I have fully verified te code throwing at it
random moves. It shoudl work.
No functional change.
Test by Joona prooves the new feature don't value 70 added lines of code.
Grand totals after 10040 games (crashes: 0) for tt_both
master_9edc7 - 6a93488_6a934: 1756 - 1688 - 6596 ELO +2 (+- 2.7)
Confirmed by test of Gary:
After 8680 games:
ELO: 0.80 +- 99%: 9.62 95%: 7.31
LOS: 65.38%
Wins: 1288 Losses: 1268 Draws: 6130
Thanks a lot to both for testing it !!!
bench 5149248
This is more complex than what I'd like but I
was unable to split in small chunks.
Here we add 2 slots to TTEntry (valueUpper and depthUpper)
so that sizeof(TTEntry) returns to the original 16 bytes
and we can pack exactly 4 entries in a 64 bytes cache line.
Now we save an upper bound score alongside a lower (exact)
score. The idea is to increase TT cut-offs rates becuase
there is now an higher probability for a node to use TT info.
This patch is highly experimental and probably needs further
steps as is hinted by an unrealistic bench number:
bench: 2022385
In almost all cases we already know in advance that
color_of() argument is different from NO_PIECE.
So avoid the check for NO_PIECE in color_of() and
test at caller site in the very few places where
this case could occur.
As a nice side effect, this patch fixes a (bogus)
warning under some versions of gcc.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Use an eval cache instead of TT to store node
position evaluations.
It is already an improvment and, because it frees
two TT entry slots, paves the way to extend TT to
store both upper and lower bounds.
After 4855 games, single thread, 15"+0.05
Mod vs Orig 1165 -920 - 2770 ELO +18
bench: 5149248
And document why this is an hard limit. It
seems for some (lucky) people 32 threads
are not enough.
No functional change.
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
In search(), after we evalute the position, in case there
isn't any TT entry we create one with just the evaluation
score.
This patches removes that code. The reason becuase the patch
deserves a single commit it is becuase introduces a (very small)
functional change due to the fact that the total number of
TT stores is less now and this slightly alters the TT hits
of our benchmark.
bench: 4983262
Rely fully on eval cache. Note that we still save eval
info to TT, this is not needed at this moment and will be
removed in future patches. We keep it so to have a "non
functional change" patch.
No functional change.
With this patch series we want to introduce a per-thread
evaluation cache to store node evaluation and do not
rely anymore on the TT table for this.
This patch just introduces the infrastructure.
No functional change.
In case of a RootNode or a SpNode move has
been already checked for legality so we can
skip a redundant check.
Spotted by Frank Genot.
No functional change.
In qsearch we should update the bestValue as we do
in case of futilityValue < beta, also when pruning
moves with non-positive see.
Spotted by Lucas Braesch
Bench: 5695710
Send the PV lines to GUI only once at the end of the
PV search loop or just in case of long searches.
We need to sync also sending of "currmove" info to
avoid sending info on current move without first
informing the GUI on the PV line we are searching on.
No functional change.
At this point we have already verified (value > alpha)
and this implies, in case of a non-PV node, where search
window size is zero, that value >= beta.
This is not so self-evident, so document the code with
an assert condition.
No functional change.
Let the caller to decide where to redirect (cout or cerr) the
ASCII representation of the position. Rename the function to
reflect this.
Renamed also from_fen() and to_fen() to set() and fen() respectively.
No functional change.
In case a PvNode node has a static evaluation above alpha but
no available moves we want to flag the node as BOUND_EXACT,
not as BOUND_UPPER as is currently.
The behaviour was recently introduced with patch d471c49700
of 3/10/2012
Spotted by Hongzhi Cheng.
bench: 5558464
Both Lucas re-test and Jean-Francois confirrm it
is a regression.
Here Jean-Francois's results after 3600 games :
Score of 96d3b1c92b vs 3b87314: 690 - 729 - 2181 [0.495] 3600
ELO: -3.86 +- 99%: 14.94 95%: 11.35
LOS: 15.03%
Wins: 690 Losses: 729 Draws: 2181 Total: 3600
Bench: 5404066
Don't prune and eventually extend check moves of type
DISCO_CHECK (pun intended, Lucas will understand :-) ).
Patch from Lucas Braesch that has also tested it:
Result: 879-661-2137, score=52.96%, LOS=100.00% (time 10"+0.1")
I have started a verification test right now.
bench: 6004966
From Jean-Francois's
Final result after 5000 games :
Score of c581b7e vs a878312: 1163 - 970 - 2867 [0.519] 5000
ELO: 13.35 +- 99%: 12.71 95%: 9.65
LOS: 100.00%
Wins: 1163 Losses: 970 Draws: 2867 Total: 5000
From me
After 3266 games at 20"+0,05
Score of c581b7e vs a878312: 612 - 607 - 2047
So no regression at longer TC and perhaps a little gain at
fast TC.
In this case we try a rather drastic
approach: we simply don't futility prune
in qsearch when arriving from a null move.
So we save evaluating and also save to mess
with eval margins at all because margin is used
only in futility.
Also accuracy should not be affected, actually it
improves because we don't prune anything anymore.
bench: 5404066
Performs well at very short TC of 40/4+0.05 (courtesy of Jean-Francois):
Wins: 2503 Losses: 2146 Draws: 5581 Total: 10230 +12 ELO
But is poor at longer TC of 20"+0.05
Wins: 321 Losses: 373 Draws: 1141 Total: 1808 -10 ELO
The patch was clearly a tradoff between speed and accuracy and
the most interesting part of it are test results that can be
commented as follows:
- A short TC is very sensible to any speed increase
- A longer TC is more sensible to accuracy and less to speed
So a patch that does not change speed is suitable to be tested at
short TC, while a speed/accuracy compromise patch is IMO better to
be tested at longer TC to verify loss of accuracy can be tolerated.
In this case the revert is only temporary. We will come back again
once we will be able to preserve the evaluation margin.
bench: 5809010
Reuse the evaluation of the parent with inverted sign and
set margin to zero (this is an hack!).
This is done only in qsearch where almost 15% of calls are
from a null move. In normal search the number of nodes where
(ss-1)->currentMove == MOVE_NULL
is almost zero and so there is no need of using this trick.
The big advantage of this patch is a speed-up due to skipped
evaluate() calls, that are very costly.
Functionality is of course affected and we will need to proper
test it later. For now we just register a 3-4% speed up.
Suggested by Hongzhi Cheng.
bench: 5051328
When testing if a move blocks the threat path there is no
reason to require the threat to be a slider. Indeed threat
can be a double pawn push like in this example:
r1bq1rk1/ppp1np1p/4n1p1/3p4/3P2Q1/2P1B3/PPBN2PP/R4RK1 w - - 0 16
Where white's move Rf6 blocks the threat f5.
As a nice side effect we can retire the now useless helper
piece_is_slider().
This patch kicks in only very rare cases, indeed the bench is
still the same!
bench: 5809010
There is no guarantee that split() consumes all the node's
moves. Indeed split() can return without performing any job
for instance because MAX_SPLITPOINTS_PER_THREAD is reached
or becuase no available threads are found (this latter case
is much more common).
So search must continue in those cases and we cannot force
exiting from move's loop.
Bug introduced by 1ac417edb8 of 5/10/2012
Spotted by Frank Genot.
No functional change.
If a previous move attacks the king (with the piece
of the threat move removed) then must be a discovered
check, otherwise it means that first move gave check
and we were not able to do a null move.
Also renamed stuff to better document the function's
context.
No functional change.
When testing if a piece is moving through the squares
vacated by a previous move there is no reason to require
the piece to be a slider, indeed we can have a double
pawn push like in this example:
r1q2rk1/2p1bppp/2Pp4/pN5b/Q1P1p3/4B2P/PP1R1PP1/1K5R w - - 3 18
Where black's move f5 is connected to previous move Be7 that
frees the path.
Or we can have a castle move:
r1bqkb1r/pppp1ppp/2n1pn2/1B6/4P3/2N2N2/PPPP1PPP/R1BQK2R b KQkq - 5 1
Where a previous move Bb5 allows the white to castle king side.
This time patch is mine ;-)
new bench: 5809010
We send to GUI multi-pv info after each cycle,
not just once at the end of the PV loop. This is
because at high depths a single root search can
be very slow and we want to update the gui as
soon as we have a new PV score.
Idea is good but implementation is broken because
sort() takes as arguments a pointer to the first
element and one past the last element.
So fix the bug and rename sort arguments to better
reflect their meaning.
Another hit by Hongzhi Cheng. Impressive!
No functional change.
