What this patch does is:
* increase safety margin from 40ms to 60ms. It's worth noting that the previous
code not only used 60ms incompressible safety margin, but also an additional
buffer of 30ms for each "move to go".
* remove a whart, integrating the extra 10ms in Move Overhead value instead.
Additionally, this ensures that optimumtime doesn't become bigger than maximum
time after maximum time has been artificially discounted by 10ms. So it keeps
the code more logical.
Tested at 3 different time controls:
Standard 10+0.1
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 58008 W: 10674 L: 10617 D: 36717
Sudden death 16+0
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 59664 W: 10945 L: 10891 D: 37828
Tournament 40/10
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 16371 W: 3092 L: 2963 D: 10316
bench: 5479946
Reduces memory footprint by ~1.2MB (per thread).
Strong pressure: small but mesurable gain
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 258430 W: 46977 L: 45943 D: 165510
Low pressure: no regression
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 73542 W: 13058 L: 13026 D: 47458
Strong pressure + LTC: elo gain confirmed
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 31489 W: 4532 L: 4295 D: 22662
Tested for crashing on overflow and after 70K
games at STC we have only 4 time losses,
possible candidate for an overflow.
No functional change.
Previously, we had duplicated History:
- one with (piece,to) called History
- one with (from,to) called FromTo
Now that we have only one, rename it to History, which is the generally accepted
name in the chess programming litterature for this technique.
Also correct some comments that had not been updated since the introduction of CMH.
No functional change.
Tested using syzygy bench method:
- 2016 random positions ranging between 3 and 10 pieces
- each searched using bench at depth=10
Same node count (and no speed regression).
No functional change.
It was a bit of a hack, without intrinsic value, but rather compensating for the
fact that checks were mistuned.
STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 88308 W: 15553 L: 15545 D: 57210
LTC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 53115 W: 6741 L: 6662 D: 39712
bench 5468995
Allow to specifiy the log file name, this comes
handy in case of self-matches so that each SF
instance writes into a different log file.
No functional change.
Counter intuitively, make build ARCH=x86-32 does NOT produce a 32-bit compile
when running a 64-bit OS. Nor would ARCH=x86-64 produce a 64-bit compile when
running a 32-bit OS (assuming it compiled w/o errors).
No functional change
Resolves#621
lsb(b) and msb(b) are undefined when b == 0. This can lead to subtle bugs, where
the resulting code behaves differently on different configurations:
- It can be the home grown software LSB/MSB
- It can be the compiler generated software LSB/MSB (when using compiler
intrinsics without the right compiler flags to allow compiler to use hardware
LSB/MSB). Which of course depends on the compiler.
- It can be hardware LSB/MSB generated by the compiler.
- Not to mention that hardware LSB/MSB can return different value on different
hardware when b == 0.
No functional change
Resolves#610
There is no reason to compile 3 different copies of search(). PV nodes are on
the cold path, and PvNode is a template parameter, so there is no cost in
computing:
const bool RootNode = PvNode && (ss-1)->ply == 0;
And this simplifies code a tiny bit as well.
Speed impact is negligible on my machine (i7-3770k, linux 4.2, gcc 5.2):
nps +/-
test 2378605 3118
master 2383128 2793
diff -4523 2746
Bench: 7751425
No functional change.
Resolves#568
Comment is based on a misunderstanding of what unaligned memory access is. Here
is an article that explains it very clearly:
https://www.kernel.org/doc/Documentation/unaligned-memory-access.txt
No matter how we define TTEntry or TTCluster, there will never be any unaligned
memory access. This is because the complier knows the alignment rules, and does
the necessary adjustments to make sure unaligned memory access does not occur.
The issue being adressed here has nothing to do with unaligned memory access. It
is about cache performance. In order to achieve best cache performance:
- we prefetch the cacheline as soon as possible.
- we ensure that TT clusters do not spread across two cachelines. If they did,
we would need to prefetch 2 cachelines, which could hurt cache performance.
Therefore the true conditions to achieve this are:
1/ start adress of TT is cache line aligned. void TranspositionTable::resize()
enforces this.
2/ TT cluster size should *divide* the cache line size. Currently, we pack 2
clusters per cache lines. It used to be 1 before "TT sardines". Does not matter
what the ratio is, all we want is to fit an integer number of clusters per cache
line.
No functional change.
Resolves#506
Indeed, if we use a depth >= DEPTH_MAX, we start having negative depth in the
TT (due to int8_t cast).
No functional change in single thread mode
Resolves#490
Rely on well defined behaviour for message passing, instead of volatile. Three
versions have been tested, to make sure this wouldn't cause a slowdown on any
platform.
v1: Sequentially consistent atomics
No mesurable regression, despite the extra memory barriers on x86. Even with 15
threads and extreme time pressure, both acting as a magnifying glass:
threads=15, tc=2+0.02
ELO: 2.59 +-3.4 (95%) LOS: 93.3%
Total: 18132 W: 4113 L: 3978 D: 10041
threads=7, tc=2+0.02
ELO: -1.64 +-3.6 (95%) LOS: 18.8%
Total: 16914 W: 4053 L: 4133 D: 8728
v2: Acquire/Release semantics
This version generates no extra barriers for x86 (on the hot path). As expected,
no regression either, under the same conditions:
threads=15, tc=2+0.02
ELO: 2.85 +-3.3 (95%) LOS: 95.4%
Total: 19661 W: 4640 L: 4479 D: 10542
threads=7, tc=2+0.02
ELO: 0.23 +-3.5 (95%) LOS: 55.1%
Total: 18108 W: 4326 L: 4314 D: 9468
As suggested by Joona, another test at LTC:
threads=15, tc=20+0.05
ELO: 0.64 +-2.6 (95%) LOS: 68.3%
Total: 20000 W: 3053 L: 3016 D: 13931
v3: Final version: SeqCst/Relaxed
threads=15, tc=10+0.1
ELO: 0.87 +-3.9 (95%) LOS: 67.1%
Total: 9541 W: 1478 L: 1454 D: 6609
Resolves#474
Align the behaviour with reductions. Initially castling moves had to be
treated differently, because the SEE did not handle them correctly. But now it
does.
STC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 83750 W: 15722 L: 15711 D: 52317
LTC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 97183 W: 15120 L: 15115 D: 66948
bench 7759837
Resolves#403
Instead of crafting a clever formula to calculate the array offset, simply use a
3 dimensional array. Remove the comment while at it, because now the code is
self-documenting.
No functional change.
Resolves#344
Introduce helper function Search::reset() which clears all kind of search
memory, in order to restore a deterministic search state.
Generalize TT.clear() into Search::reset() for the following use cases:
- bench: needed to guarantee deterministic bench (ie. if you call bench from
interactive command line twice in a row you get the same value).
- Clear Hash: restore clean search state, which is the purpose of this button.
- ucinewgame: ditto.
No functional change.
Resolves#346
Based on an idea and patch by VoyagerOne.
Small simplification, but was tedted for an ELO gain anyway.
STC:
LLR: 2.95 (-2.94,2.94) [-1.00,4.00]
Total: 5375 W: 1119 L: 977 D: 3279
LTC:
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 17893 W: 2984 L: 2792 D: 12117
bench 8322847
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
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
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.
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.
This optimization is aimed at old hardware only (withouth popcount), and even on
non popcount compile (ARCH=x86-64), it provides no mesurable speedup:
stat test master diff
mean 2,341,779 2,354,699 -12,920
stdev 12,910 14,770 18,150
speedup -0.55%
P(speedup>0) 23.8%
No functional change.
Resolves#187
Just like in Physics, the ratio of 2 things in the same unit, should be
without unit.
Example use case:
- Ratio of a Depth by a Depth (eg. ONE_PLY) should be an int.
- Ratio of a Value by a Value (eg. PawnValueEg) should be an int.
Remove a bunch of useless const while there.
No functional change.
Resolves#128
Daniel Jose reported that it was an elo gain in his engine:
http://www.talkchess.com/forum/viewtopic.php?t=54290
STC: Hash=16
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 33067 W: 6670 L: 6571 D: 19826
LTC: Hash=64
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 41181 W: 7008 L: 6920 D: 27253
And another one to verify no regression with hash pressure:
STC: Hash=4
LLR: 2.96 (-2.94,2.94) [-4.00,0.00]
Total: 25085 W: 5059 L: 4991 D: 15035
Verified that qsearch does not explode after this patch (recapture threshold).
Bench 7962287
Resolves#112
16MB for 1" searches is more comensurate with the average use case.
And 16 is the default used by stockfish bench, so it makes sense to be
consistent, if only to have the same minimum memory requirement for using
SF and compiling it with PGO.
No functional change.
Resolves#109
* remove some erroneous comments, that were based on the ONE_PLY == 2.
* rename hd to d, because there's no more half-depth in SF.
* rescope variables into the for loops.
* reindent the for loops correctly.
* add a comment to explain the eval improving part (not so obvious to read
this code as array has 4 dimensions).
No functional change.
This should reduce search inconsistencies, and doesn't seem to have a measurable ELO Impact:
STC with Hash=16
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 49264 W: 10076 L: 10007 D: 29181
LTC with Hash=64
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 82149 W: 14044 L: 14023 D: 54082
Plus an extra test, to make sure it doesn't regress with strong hash pressure:
STC with Hash=4
LLR: 2.95 (-2.94,2.94) [-4.00,0.00]
Total: 21498 W: 4327 L: 4246 D: 12925
Bench: 7302735
Resolves#100
Idea is to apply king safety later in the endgame. Previously, we didn't
apply KS in a RR vs. Q ending for example, which causes poor play.
Now we calculate king attacks when the attacking side has a queen or more.
STC with 8moves_v3
LLR: 3.06 (-2.94,2.94) [0.00,4.00]
Total: 38481 W: 6228 L: 5952 D: 26301
LTC with 2moves_v1
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 51053 W: 8670 L: 8353 D: 34030
Bench: 7514010
Resolves#98
Clang 3.5 issues warning on constructs like: abs(f1 - f2). The thing is that
f1 and f2 are enum types, and the range given (all positive) allows the
compiler to choose an unsigned type (efficiency being one reason to prefer
unsigned arithmetic). If f1 < f2 are unsigned, then f1 - f2 wraps around zero
and the abs() becomes a no-op. It's the reinterpretation of the unsigned
result (large value) as a signed int that happens to give the correct result,
thanks to 2's complement. This is all tricky and dangerous!
In the spirit of the standard, we assume nothing on the signedness of enums,
and simply calculate the rank and file distances as:
- rank_dist(r1, r2) = r1 < r2 ? r2 - r1 : r1 - r2
- file_dist(f1, f2) = f1 < f2 ? f2 - f1 : f1 - f2
this logic can in fact be applied to any enum we may use, so for better
generality and to avoid code duplication, we use a template function diff()
here.
No functional change.
Resolves#95
This area has become obscure and tricky over the course of incremental
changes that did not respect the original's consistency and clarity. Now,
it's not clear why we use MAX_PLY = 120, or why we use MAX_PLY+6, among
other things.
This patch does the following:
* ID loop: depth ranges from 1 to MAX_PLY-1, and due to TT constraint (depth
must fit into an int8_t), MAX_PLY should be 128.
* stack[]: plies now range from 0 to MAX_PLY-1, hence stack[MAX_PLY+4],
because of the extra 2+2 padding elements (for ss-2 and ss+2). Document this
better, while we're at it.
* Enforce 0 <= ply < MAX_PLY:
- stop condition is now ss->ply >= MAX_PLY and not ss->ply > MAX_PLY.
- assert(ss->ply < MAX_PLY), before using ss+1 and ss+2.
- as a result, we don't need the artificial MAX_PLY+6 range. Instead we
can use MAX_PLY+4 and it's clear why (for ss-2 and ss+2).
* fix: extract_pv_from_tt() and insert_pv_in_tt() had no reason to use
MAX_PLY_PLUS_6, because the array is indexed by plies, so the range of
available plies applies (0..MAX_PLY before, and now 0..MAX_PLY-1).
Tested with debug compile, using MAX_PLY=16, and running bench at depth 17,
using 1 and 7 threads. No assert() fired. Feel free to submit to more severe
crash-tests, if you can think of any.
No functional change.
SSE4.2 has nothing to do with POPCNT. We must dispell this myth, because
Stockfish is a reference and many will copy this mistake if they see it in Stockfish:
* SSE is an SIMD instruction set, relative to vectorization (using special 128-bit registers).
* POPCNT/LZCNT work on normal registers (eg. AL, AX, EAX, RAX).
The confusion comes from the fact that, in the Intel product line, it just
so happens that SSE4.2 and POPCNT/LZCNT came out at the same time. But this
is not true for AMD. For example, all AMD Pheniom II have SSE3 but no
POPCNT/LZCNT, and that is why the modern compile uses -msse3 -popcnt and not -msse4.2.
No functional change.
Resolves#86
Objects that are only accessible at file-scope should be put in the anonymous namespace.
This is what the C++ standard recommends, rather than using static, which is really C-style and results in static linkage.
Stockfish already does this throughout the code. So let's weed out the few exceptions,
because... they have no reason to be exceptional.
No functional change.
Resolves#84
Make even more clear what are the terms that
contribute to evaluate connected pawns, and
completely separate them from the weights
that are now fully looked up in a table.
For future tuning makes sense to init the table with
a formula instead of hard-code it. This allows to
reduce problem space cardinality and makes tuning
easier.
And fix a MSVC warning while there:
warning C4804: '>>' : unsafe use of type 'bool' in operation
No functional change.
These two notions are very correlated. Since connected has the most
generality, it makes sense to generalize it to encompass what is
covered by candidate.
STC:
LLR: 4.03 (-2.94,2.94) [-3.00,1.00]
Total: 11970 W: 2577 L: 2379 D: 7014
LTC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 13194 W: 2389 L: 2255 D: 8550
bench 7328585
Now that half plies have been removed from the engine, we can encode
TT depth into an int8_t.
Range is -128 to +127, so it goes still further than the previous
limit of 121 plies (with ONE_PLY == 2 where depth - DEPTH_NONE was
encoded as an uint8_t).
No functional change.
Resolved#60
Now "Write Search Log" will pring moves in UCI format, consistent with all the rest. This functionality is
not aimed at end-users anyway. It's hardly useful at all, in fact. Also, pretty-printing SAN moves is
something that better belongs in the GUI than in the engine.
No functional change.
Where they better belong.
Also, this removes '#include <string>' from types.h, which reduces the amount of code to compile (every
translation unit includes types.h).
No functional change.
So that one can redirect cout to /dev/null and only print print cerr in the terminal (for more accurate speed
tests).
Suggested by Marco.
No functional change.
The compiler tries to cast Options["Hash"] into a string, using:
Option::operator std::string() const {
assert(type == "string");
return currentValue;
}
And, as expected, the assert() fails.
std::to_string() would be the right solution, but it's C++11. And using a stringstream is too much code to
achieve so little. Let's keep it the way it was: hardcoded (ie. default hash defined in two places).
No functional change.
First, remove some dead code (function never called with a Move argument).
Then, remove printing of legal moves, which does not belong here. Let's keep commands orthogonal and minimal:
- the "d" command should display the board, nothing more, or less.
- "perft 1" will display the list of legal moves.
No functional change.
Stockfish allocates the default hash (32MB) in main(), before entering UCI::loop(). If there is not enough
memory, the program will crash even before UCI::loop() is entered and the GUI is given a change to specify a
lower Hash value.
This defective design could be resolved by doing a lazy allocation upon "isready" command, as the UCI protocol
guarantees that "isready" will be sent at least once before any search. But it's a bit cumbersome when using
Stockfish "manually" to have to remember to type "isready" everytime.
So leave the current design, but reduce the default hash to 16MB instread of 32MB. In order to perform such
quick searches (depth=13), there is no reason to use so much Hash anyway. Another benefit is to introduce a
bit of hash pressure in bench, which increases chances to detect rare bugs related to TT replacement, for
example.
This is not a functional change, although it obviously changes the bench.
bench 7461879
Instead of defining it both in ucioption.cpp and benchmark.cpp. Obviously changing the default Hash will
change the bench as a result.
No functional change.
The main purpose of perft is to help debugging. But without the breakdown in sum of perft(N-1), it is a
completely useless debugging tool.
So perft now displays the breakdown, and divide is therefore removed.
No functional change.
Despite being neutral at STC, it turned out to be regressive at LTC:
40k games at LTC with Hash=8
ELO: -2.06 +-1.9 (95%) LOS: 1.4%
Total: 39720 W: 5740 L: 5976 D: 28004
40k games at LTC with Hash=128
ELO: -2.69 +-1.9 (95%) LOS: 0.2%
Total: 39149 W: 5702 L: 6005 D: 27442
bench 7477963
Also raise the admissible bounds to (-100,100), as there is no reason to prevent users from using high
values if they want to.
Does not regress in self play:
ELO: 0.10 +-2.0 (95%) LOS: 53.7%
Total: 40000 W: 7084 L: 7073 D: 25843
master vs SF 3
ELO: 182.86 +-2.7 (95%) LOS: 100.0%
Total: 40000 W: 21843 L: 2541 D: 15616
Contempt = 20 vs SF 3
ELO: 189.25 +-2.8 (95%) LOS: 100.0%
Total: 40000 W: 22721 L: 2859 D: 14420
Diff is therefore 6.4 +/- 3.9 elo against a 180-190 elo weaker engine, which is significantly positive,
as expected. This elo difference is likely understated, because of FishTest aggressive draw adjudication
though.
We could push Contempt further, but after 20cp, it would get in the way of FishTest draw adjudication
rule, and is likely to reduce the testing throughput as a result.
bench 8198667
Another attempt at retiring current asymmetric
king evaluation and use a much simpler symmetric
one. As a good side effect we can avoid recalculating
eval after a null move.
Tested in no-regression mode and passed
STC
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 21580 W: 3752 L: 3632 D: 14196
LTC
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 18253 W: 2593 L: 2469 D: 13191
And a LTC regression test against SF DD to
verify we don't have regression against
weaker engines due to some kind of 'contempt'
effect:
ELO: 54.69 +-2.1 (95%) LOS: 100.0%
Total: 40000 W: 11072 L: 4827 D: 24101
bench: 8205159
Retire current asymmetric king evaluation
and use a much simpler symmetric one.
As a side effect retire the infamous
'Aggressiveness' and 'Cowardice' UCI
options.
Tested in no-regression mode,
Passed both STC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 33855 W: 5863 L: 5764 D: 22228
And LTC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 40571 W: 5852 L: 5760 D: 28959
bench: 8321835
This is more consistent with what other engines are doing.
Often people thinks that SF's scores are overblown. In the
end, it just boils down to the arbitrary way of rescaling them.
No functional change.
Small simplification.
Passed SPRT(-3,1) both at STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 17051 W: 3132 L: 3005 D: 10914
and LTC:
LLR: 4.55 (-2.94,2.94) [-3.00,1.00]
Total: 24890 W: 3842 L: 3646 D: 17402
The rationale behind this is that I've never managed to add a
Queen on 7th rank bonus in DiscoCheck, because it never showed
to be positive (evne slightly) in testing. The only thing that
worked is Rook on 7th rank.
In terms of SF code, it seemed natural to group it with QueenOnPawn
as well as those are done together. I know you're against groupping
in general, but when it comes to non regression test, you are being
more conservative by groupping. If the group passes SPRT(-3,1) it's
safer to commit, than test every component in SPRT(-3,1) and end up
with the risk of commiting several -1 elo regression instead of just
one -1 elo regression.
In chess terms, perhaps it's just easier to manouver a Queen (which
can more also diagonaly) than a Rook. Therefore you can let the search
do its job without needing eval ad-hoc terms to guide it. For the Rook
which takes more moves to manouver such eval terms can be (marginally)
useful.
bench: 7473314
Tested with SPRT in simplification mode [-4.00,0.00],
this ensures that the patch is (very probably) not
a regression.
Passed both short TC
LLR: 2.95 (-2.94,2.94) [-4.00,0.00]
Total: 27543 W: 4278 L: 4209 D: 19056
And long TC
LLR: 2.95 (-2.94,2.94) [-4.00,0.00]
Total: 39483 W: 7325 L: 7305 D: 24853
bench: 8347121
Hopefully this patch makes the code more:
* Self-documenting: Null search is always a zero window search,
because it is testing for a fail high. It should never be done
on a full window! The current code only works because we don't
do it at PV nodes, and therefore (alpha, beta) = (beta-1, beta):
that's the kind of "clever" trick we should avoid.
* Idiot-proof: If we want to enable null search at PV nodes, all we
need to do now is comment out the !PvNode condition. It's that simple!
In theory, null search should not be done at PV nodes, because PV nodes
should never fail high. But in practice, they DO fail high, because of
aspiration windows, and search inconsistencies, for example. So it makes
sense to keep that flexibility in the code.
No functional change.
Use ralpha instead of rbeta
* rbeta is confusing people. It took THREE attempts to code razoring
at PV nodes correctly in a recent test, because of the rbeta trick.
Unnecessary tricks should be avoided.
* The more correct and self-documenting way of doing this, is to say
that we use a zero window around alpha-margin, not beta-margin.
The fact that, because we only do it at PV nodes, alpha happens to be
beta-1 and that the current stuff with rbeta works, may be correct,
but is confusing.
Remove the misleading and partially erroneous comment about returning
v + margin:
* comments should explain what the code does, not what it could have done.
* this comment is partially wrong in saying that v+margin is "logical",
and that it is "surprising" that is doesn't work.
From a theoretical perspective, at least 3 ways of doing this are equally
defendable:
1/ fail hard: return alpha: The most conservative. We bet that the search
will fail low, but we don't know by how much and don't want to take risks.
2/ aggressive fail soft: return v (what the current code does). This
corresponds to normal fail soft, with the added assumption that we don't
care about the reduction effect (see below point 3/)
3/ conservative fail soft: return v + margin. If the reduced search (qsearch)
gives us a score <= v, we bet that the non reduced search will give us a
score <= v + margin.
* Saying that 2/ is "logical" implies that 1/ and 3/ are not, which is
arguably wrong. Besides, experimental results tell us that 2/ beats 3/,
and that's not something we can argue against: experimental results are
the only trusted metric.
* Also, with the benefit of hindsight, I don't think the fact that 2/ is
better than 3/ is surprising at all. The point is that it is YOUR turn to
move, and you are assuming that by NOT playing (and letting the opponent
capture your hanging pieces in QS) you cannot generally GAIN razor_margin(depth).
No functional change.
A great simplification that shows no regression
and it seems even a bit scalable.
Tested with fixed number of games:
Short TC
ELO: 0.60 +-2.1 (95%) LOS: 71.1%
Total: 39554 W: 7477 L: 7409 D: 24668
Long TC
ELO: 2.97 +-2.0 (95%) LOS: 99.8%
Total: 36424 W: 5894 L: 5583 D: 24947
bench: 8184352
We already update killers so it is natural to extend to
history and counter move too.
Passed both short TC
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 52690 W: 9955 L: 9712 D: 33023
And long TC
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 5555 W: 935 L: 808 D: 3812
bench: 7876473
1/ eval margin and gains removed:
16bit are now free on TT entries, due to the removal of eval margin. may be useful
in the future :) gains removed: use instead by Value(128). search() and qsearch()
are now consistent in this regard.
2/ futility_margin()
linear formula instead of complex (log(depth), movecount) formula.
3/ unify pre & post futility pruning
pre futility pruning used depth < 7 plies, while post futility pruning used
depth < 4 plies. Now it's always depth < 7.
Tested with fixed number of games both at short TC:
ELO: 0.82 +-2.1 (95%) LOS: 77.3%
Total: 40000 W: 7939 L: 7845 D: 24216
And long TC
ELO: 0.59 +-2.0 (95%) LOS: 71.9%
Total: 40000 W: 6876 L: 6808 D: 26316
bench 7243575
1/ eval margin and gains removed:
- gains removed by Value(128): search() and qsearch() now behave consistently!
2/ futility_margin()
- testing showed that there is no added value in this weird (log(depth), movecount)
formula, and a much simpler linear formula is just as good. In fact, it is most
likely better, as it is not yet optimally tuned.
- the new simplified formula also means we get rid of FutilityMargins[], its
initialization code, and more importantly ss->futilityMoveCount, and the hacky
code that updates it throughout the search().
- the current formula gives negative futility margins, and there is a hidden interaction
between the move coutn pruning formula and the futility margin one: what happens is
that MCP is supposed to be triggered before we use the non-sensical negative futility
margins.
3/ unify pre & post futility pruning
- pre futility pruning (what SF calls value based pruning) used depth < 7 plies,
while post futility pruning (what SF calls static null move pruning) used depth < 4 plies.
- also the condition depth < 7 in pre futility pruning was not obvious, and it seemd
to be depth < 16 (futility_margin() returns an infinite value when depth >= 7).
Tested with fixed number of games both at short TC:
ELO: 0.82 +-2.1 (95%) LOS: 77.3%
Total: 40000 W: 7939 L: 7845 D: 24216
And long TC
ELO: 0.59 +-2.0 (95%) LOS: 71.9%
Total: 40000 W: 6876 L: 6808 D: 26316
bench: 10206576
Rationale:
- Speed of double and float is about the same (not on the hot path anyway)
- Double makes code prettier (no need to write 1.0f, just 1.0)
- Only practical advantage of float is to use less memory, but since we never
store large arrays of double, we don't care.
No functional change.
Unify extensions between PV and not PV nodes
and remove all but check extensions.
This is a simplification so tested at fixed number
of games where proved to not regress.
About 45k games at 15+0.05
ELO: 1.23 +-2.0 (95%) LOS: 88.5%
Total: 45643 W: 9107 L: 8946 D: 27590
About 45k games at 60+0.05
ELO: 1.07 +-1.8 (95%) LOS: 87.8%
Total: 46786 W: 7728 L: 7584 D: 31474
bench: 3172206
Union of
- LMR >= 3 plies from Gary tests.stockfishchess.org/tests/view/522522960ebc595d328fcafd
- allows() tweak from Reuven tests.stockfishchess.org/tests/view/5225fa1c0ebc595d328fcb53
Both passed Step I and failed Step II.
Instead this union passed both short TC:
LLR: 2.95 (-2.94,2.94)
Total: 14525 W: 3063 L: 2874 D: 8588
And long TC
LLR: 2.94 (-2.94,2.94)
Total: 31075 W: 5566 L: 5308 D: 20201
bench: 4238160
This patch is actually the sum of two contributions that
have been tested independently:
1) Pruning of negative SEE moves in PV
After 10000 games at 20+0.05
ELO: 5.18 +-7 (95%) LOS: 99.2%
Total: 10000 W: 1952 L: 1803 D: 6245
2) Remove of bestValue > VALUE_MATED_IN_MAX_PLY condition
After 23000 games at 20+0.05
ELO: 1.63 +-4 (95%) LOS: 88.1%
Total: 23000 W: 4232 L: 4124 D: 14644
The whole patch as been re-tested at long TC with positive results:
After 10000 games at 60+0.05
ELO: 4.31 +-7 (95%) LOS: 98.3%
Total: 10000 W: 1765 L: 1641 D: 6594
Great code simplification: - instead do not futility
prune threat refutations. allows_move() is therefore removed.
4000 games at 50,000 nodes/move:
1085-989-1926 [51.2%] LOS=98.3%
4000 games in 10"+0.1"
756-751-2493 [50.1%] LOS=55.1%
EDIT: I have retested the patch of Lucas in a slightly different form
(without pruning in PvNode) and test mre or less confirms that
60 lines of code are totally unuseful:
After 6195 games at 15"+0.05"
1333 - 1325 - 3537 ELO 0
bench 5140990
lucasart