This rewrites in part the message passing part, using in place gather, and collecting, rather than merging, the data of all threads.
neutral with a single thread per rank:
Score of new-2mpi-1t vs old-2mpi-1t: 789 - 787 - 2615 [0.500] 4191
Elo difference: 0.17 +/- 6.44
likely progress with multiple threads per rank:
Score of new-2mpi-36t vs old-2mpi-36t: 76 - 53 - 471 [0.519] 600
Elo difference: 13.32 +/- 12.85
This generalizes exchange of signals between the ranks using a non-blocking all-reduce. It is now used for the stop signal and the node count, but should be easily generalizable (TB hits, and ponder still missing). It avoids having long-lived outstanding non-blocking collectives (removes an early posted Ibarrier). A bit too short a test, but not worse than before:
Score of new-r4-1t vs old-r4-1t: 459 - 401 - 1505 [0.512] 2365
Elo difference: 8.52 +/- 8.43
use a counter to track available elements.
Some elo gain, on 4 ranks:
Score of old-r4-1t vs new-r4-1t: 422 - 508 - 1694 [0.484] 2624
Elo difference: -11.39 +/- 7.90
avoid saving to TT the part of the receive buffer that actually originates from the same rank.
Now, on 1 mpi rank, we have the same bench as the non-mpi code on 1 thread.
since the logic for saving moves in the sendbuffer and the associated rehashing is expensive, only do it for TT stores of sufficient depth.
quite some gain in local testing with 4 ranks against the previous version.
Elo difference: 288.84 +/- 21.98
This starts to make the branch useful, but for on-node runs, difference remains to the standard threading.
In the original code, the position key stored in the TT is used to probe&store TT entries after message passing. Since we only store part of the bits in the TT, this leads to incorrect rehashing. This is fixed in this patch storing also the full key in the send buffers, and using that for hashing after message arrival.
Short testing with 4 ranks (old vs new) shows this is effective:
Score of mpiold vs mpinew: 84 - 275 - 265 [0.347] 624
Elo difference: -109.87 +/- 20.88
This reverts my earlier change that only the root node gets to output best move after fixing problem with MPI_Allreduce by our custom operator(BestMoveOp). This function is not commutable and we must ensure that its output is consistent among all nodes.
Previous behavior was to wait on all nodes to finish their search on their own TM and aggregate to root node via a blocking MPI_Allreduce call. This seems to be problematic.
In this commit a proper non-blocking signalling barrier was implemented to use TM from root node to control the cluster search, and disable TM on all non-root nodes.
Also includes some cosmetic fix to the nodes/NPS display.
Based on Peter Österlund's "Lazy Cluster" algorithm,
but with some simplifications.
To compile, point COMPCXX to the MPI C++ compiler wrapper (mpicxx).
I tried to improve the Readme because many people in my local
chess club do not understand some of the UCO options properly.
Starting point of this was Cfish's Readme by Ronald de Man,
some internet resources and the Stockfish code itself.
Closes https://github.com/official-stockfish/Stockfish/pull/1898
Initial commit by user @erbsenzaehler, with help from users
Adrian Petrescu, @alayan-stk-2 and Elvin Liu.
No functional change
Co-Authored-By: Alayan-stk-2 <alayan-stk-2@users.noreply.github.com>
Co-Authored-By: Adrian Petrescu <apetresc@gmail.com>
Co-Authored-By: Elvin Liu <solarlight2@users.noreply.github.com>
Recent tests by @xoto10, @Vizvezdenec, and myself seemed to hint that Elo could
be gained by expanding the number of cases where king safety is applied. Several
users (@Spliffjiffer, @Vizvezdenec) have anticipated benefits specifically in
evaluation of tactics. It appears that we actually do not need to restrict the
cases in which we initialize and evaluate king safety at all: initializing and
evaluating it in every position appears roughly Elo-neutral at STC and possibly
a substantial Elo gain at LTC.
Any explanation for this scaling is, at this point, conjecture. Assuming it is
not due to chance, my hypothesis is that initialization of king safety in all
positions is a mild slowdown, offset by an Elo gain of evaluating king safety
in all positions. At STC this produces Elo gains and losses that offset each
other, while at longer time control the slowdown is much less important, leaving
only the Elo gain. It probably helps SF to explore king attacks much earlier in
search with high numbers of enemy pieces concentrating but not essentially attacking
king ring.
Thanks to @xoto10 and @Vizvezdenec for helping run my LTC!
Closes https://github.com/official-stockfish/Stockfish/pull/1906
STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 35432 W: 7815 L: 7721 D: 19896
http://tests.stockfishchess.org/tests/view/5c24779d0ebc5902ba131b26
LTC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 12887 W: 2217 L: 2084 D: 8586
http://tests.stockfishchess.org/tests/view/5c25049a0ebc5902ba132586
Bench: 3163951
------------------
How to continue from there?
* Next step will be to tune all the king danger terms once more after that :-)
When iterating through 'SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX' structure, do not use structure member beyond known size.
API is guaranteed to provide us at lease one element upon successful, and no element in the structure can have a zero size.
No functional change.
This pull request fixes a rare crashing bug on Windows inside our NUMA code, first
reported by Dann Corbit in the following forum thread (thanks!):
https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/gA6aoMEuOwg
The fix is to not use structure member beyond known size when iterating through
'SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX' structure. We note that the Microsoft
API is guaranteed to provide us at least one element upon successful, and no
element in the structure can have a zero size.
No functional change.
The `&& (ss-1)->killers[0] ` conditions are there seemingly to protect
accessing ss-5.
This is unneeded and not so intuitive (as the killer is checked for equality
with currentMove, and that one is non-zero once we're high enough in the stack,
this protects access to ss-5). We can just extend the stack from ss-4 to ss-5,
so we can call update_continuation_histories(ss-1, ..) always in search.
This goes a bit further than #1881 and addresses a comment in #1878.
passed STC:
http://tests.stockfishchess.org/tests/view/5c1aa8d50ebc5902ba127ad0
LLR: 3.12 (-2.94,2.94) [-3.00,1.00]
Total: 53515 W: 11734 L: 11666 D: 30115
passed LTC:
http://tests.stockfishchess.org/tests/view/5c1b272c0ebc5902ba12858d
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 140176 W: 23123 L: 23192 D: 93861
Bench: 3451321
Even when playing without endgame table bases, this particular endgame should
be a win 100% of the time when Stockfish is given a KRBK position, assuming
there are enough moves remaining in the FEN to finish the game without hitting
the 50 move rule.
PROBLEM: The issue with master here is that the PushClose difference per square
is 20, however, the difference in squares for the PushToCorners array is usually
less. Thus, the engine prefers to move the kings closer together rather than pushing
the weak king to the correct corner.
What happens is if the weak king is in a safe corner, SF still prefers pushing the
kings together. Occasionally, the strong king traps the weak king in the safe corner.
It takes a while for SF to figure it out, but often draws the game by the 50 move rule
(on shorter time controls).
This patch increases the PushToCorners values to correct this problem. We also added
an assert to catch any overflow problem if anybody would want to increase the array
values again in the future.
It was tested in a couple of matches starting with random KRBK positions and showed
increased winning rates, see https://github.com/official-stockfish/Stockfish/pull/1877
No functional change
* Update our continuous integration machinery
Ubuntu 16.04 can now be used with travis. Updating all the other stuff
when there.
Invoking the lld linker seems to save 5 minutes with clang on linux.
No functional change.
* fix
* Use a bit less code to calculate hashfull(). Change post increment to preincrement as is more standard
in the rest of stockfish. Scale result so we report 100% instead of 99.9% when completely full.
No functional change.
Although this is a compile-time constant, we stick the castlingSide into a CastlingRight, then pull it out again. This seems unecessarily complex.
No functional change.
We do not need to change the winnerKSq variable, so we can simplify
a little bit the logic of the code by changing only the loserKSq
variable when it is necessary. Also consolidate and clarify comments.
See the pull request thread for a proof that the code is correct:
https://github.com/official-stockfish/Stockfish/pull/1854
No functional change
~stronglyProtected is quite similar to ~attackedBy[Them][PAWN] & ~attackedBy2[Them],
the only difference appears to be that the former includes squares attacked twice
by both sides. The resulting logic is simpler, and the change appears to be at least
Elo-neutral at both STC and LTC.
STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 35924 W: 7978 L: 7885 D: 20061
http://tests.stockfishchess.org/tests/view/5c14a5c00ebc5902ba11ed72
LTC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 37078 W: 6125 L: 6030 D: 24923
http://tests.stockfishchess.org/tests/view/5c14ae880ebc5902ba11eed8
Bench: 3646542
this patch fixes a rare but reproducible segfault observed playing a
multi-threaded match, it is discussed somewhat in fishcooking.
From the core file, it could be observed that the issue was in qsearch, namely:
````
ss->pv[0] = MOVE_NONE;
````
and the backtrace shows the it arrives there via razoring, called from the rootNode:
````
(gdb) bt
alpha=-19, beta=682, depth=DEPTH_ZERO) at search.cpp:1247
````
Indeed, ss->pv can indeed by a nullptr at the rootNode. However, why is the
segfault so rare ?
The reason is that the condition that guards razoring:
````
(depth < 2 * ONE_PLY && eval <= alpha - RazorMargin)
````
is almost never true, since at the root alpha for depth < 5 is -VALUE_INFINITE.
Nevertheless with the new failHigh scheme, this is not guaranteed, and rootDepth > 5,
can still result in a depth < 2 search at the rootNode. If now another thread,
via the hash, writes a new low eval to the rootPos qsearch can be entered.
Rare but not unseen... I assume that some of the crashes in fishtest recently
might be due to this.
Closes https://github.com/official-stockfish/Stockfish/pull/1860
No functional change
As noticed in the forum, a crash in extract_ponder_from_tt could result
if hash size is set before the ponder move is printed. While it is arguably
a GUI issue (but it got me on the cli), it is easy to avoid this issue.
Closes https://github.com/official-stockfish/Stockfish/pull/1856
No functional change.
On November 30th, @xoto10 experimented with removing this threshold,
but the simplification barely failed LTC. I was inspired to try various
[0, 4] tweaks to increase its value, which would narrow the effects of
this threshold without removing it entirely. Various values repeatedly
led to Elo gains at both STC and LTC, most of which were insufficient
to pass.
After a couple of weeks, I tried again to find an Elo-gaining tweak
but noticed that I could raise the threshold higher and higher without
regression. I decided to try removing it entirely--forgetting that
@xoto10 had already attempted this. However, this now performs much
better at both STC and LTC, producing a STC Elo gain and also potentially
a smaller LTC one.
The reason appears to be a recent change in master (e8ffca3) near
this code, which interacts with this patch. This simplification
governs the conditions under which that patch's effects are applied.
Something non-obvious about that change has significantly improved
the performance of this simplification.
I recognize and thank @xoto10, who originally had this idea. Since
I ran several LTCs recently (to determine whether to open this PR,
or one for a related [0, 4]), I would also like to acknowledge the
other developers and CPU donors for their patience. Thank you all!
STC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 13445 W: 3000 L: 2862 D: 7583
http://tests.stockfishchess.org/tests/view/5c11f01b0ebc5902ba11a6b8
LTC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 33868 W: 5663 L: 5563 D: 22642
http://tests.stockfishchess.org/tests/view/5c11ffe90ebc5902ba11a8a9
Closes https://github.com/official-stockfish/Stockfish/pull/1870
Bench: 3343286
STC:
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 13323 W: 3015 L: 2818 D: 7490
http://tests.stockfishchess.org/tests/view/5c00a2520ebc5902bcedd41b
LTC:
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 52294 W: 9093 L: 8756 D:34445
http://tests.stockfishchess.org/tests/view/5c00b2c40ebc5902bcedd596
Some obvious followups to this are to further tune this PSQT, or
try 8x8 for other pieces. As of now I don't plan on trying this
for other pieces as I think the majority of the ELO it brings is
for pawns and kings.
Looking at the new values, the differences between kingside and
queenside are quite significant. I am very hopeful that this a
llows SF to understand and plan pawn structures even better than
it already does. Cheers!
Closes https://github.com/official-stockfish/Stockfish/pull/1839
Bench: 3569243
I've gone through the RENAME/REFORMATTING thread and changed everything I could find, plus a few more. With this, let's close the previous issue and open another.
No functional change.
This reverts commit 33d9548218 ,
which crashed in DEBUG mode because of the following assert in position.h
````
Assertion failed: (is_ok(m)), function capture, file ./position.h, line 369.
````
No functional change
