This patch removes some magic numbers in TT bit management and introduce proper
constants in the code, to improve documentation and ease further modifications.
No function change
Use TT memory functions to allocate memory for the NNUE weights. This
should provide a small speed-up on systems where large pages are not
automatically used, including Windows and some Linux distributions.
Further, since we now have a wrapper for std::aligned_alloc(), we can
simplify the TT memory management a bit:
- We no longer need to store separate pointers to the hash table and
its underlying memory allocation.
- We also get to merge the Linux-specific and default implementations
of aligned_ttmem_alloc().
Finally, we'll enable the VirtualAlloc code path with large page
support also for Win32.
STC: https://tests.stockfishchess.org/tests/view/5f66595823a84a47b9036fba
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 14896 W: 1854 L: 1686 D: 11356
Ptnml(0-2): 65, 1224, 4742, 1312, 105
closes https://github.com/official-stockfish/Stockfish/pull/3081
No functional change.
Fix the issue where a TT entry with key16==0 would always be reported
as a miss. Instead, we'll use depth8 to detect whether the TT entry is
occupied. In order to do that, we'll change DEPTH_OFFSET to -7
(depth8==0) to distinguish between an unoccupied entry and the
otherwise lowest possible depth, i.e., DEPTH_NONE (depth8==1).
To prevent a performance regression, we'll reorder the TT entry fields
by the access order of TranspositionTable::probe(). Memory in general
works fastest when accessed in sequential order. We'll also match the
store order in TTEntry::save() with the entry field order, and
re-order the 'if-or' expressions in TTEntry::save() from the cheapest
to the most expensive.
Finally, as we now have a proper TT entry occupancy test, we'll fix a
minor corner case with hashfull reporting. To reproduce:
- Use a big hash
- Either:
a. Start 31 very quick searches (this wraparounds generation to 0); or
b. Force generation of the first search to 0.
- go depth infinite
Before the fix, hashfull would incorrectly report nearly full hash
immediately after the search start, since
TranspositionTable::hashfull() used to consider only the entry
generation and not whether the entry was actually occupied.
STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 36848 W: 4091 L: 3898 D: 28859
Ptnml(0-2): 158, 2996, 11972, 3091, 207
https://tests.stockfishchess.org/tests/view/5f3f98d5dc02a01a0c2881f7
LTC:
LLR: 2.95 (-2.94,2.94) {0.25,1.25}
Total: 32280 W: 1828 L: 1653 D: 28799
Ptnml(0-2): 34, 1428, 13051, 1583, 44
https://tests.stockfishchess.org/tests/view/5f3fe77a87a5c3c63d8f5332
closes https://github.com/official-stockfish/Stockfish/pull/3048
Bench: 3760677
This patch ports the efficiently updatable neural network (NNUE) evaluation to Stockfish.
Both the NNUE and the classical evaluations are available, and can be used to
assign a value to a position that is later used in alpha-beta (PVS) search to find the
best move. The classical evaluation computes this value as a function of various chess
concepts, handcrafted by experts, tested and tuned using fishtest. The NNUE evaluation
computes this value with a neural network based on basic inputs. The network is optimized
and trained on the evalutions of millions of positions at moderate search depth.
The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward.
It can be evaluated efficiently on CPUs, and exploits the fact that only parts
of the neural network need to be updated after a typical chess move.
[The nodchip repository](https://github.com/nodchip/Stockfish) provides additional
tools to train and develop the NNUE networks.
This patch is the result of contributions of various authors, from various communities,
including: nodchip, ynasu87, yaneurao (initial port and NNUE authors), domschl, FireFather,
rqs, xXH4CKST3RXx, tttak, zz4032, joergoster, mstembera, nguyenpham, erbsenzaehler,
dorzechowski, and vondele.
This new evaluation needed various changes to fishtest and the corresponding infrastructure,
for which tomtor, ppigazzini, noobpwnftw, daylen, and vondele are gratefully acknowledged.
The first networks have been provided by gekkehenker and sergiovieri, with the latter
net (nn-97f742aaefcd.nnue) being the current default.
The evaluation function can be selected at run time with the `Use NNUE` (true/false) UCI option,
provided the `EvalFile` option points the the network file (depending on the GUI, with full path).
The performance of the NNUE evaluation relative to the classical evaluation depends somewhat on
the hardware, and is expected to improve quickly, but is currently on > 80 Elo on fishtest:
60000 @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f28fe6ea5abc164f05e4c4c
ELO: 92.77 +-2.1 (95%) LOS: 100.0%
Total: 60000 W: 24193 L: 8543 D: 27264
Ptnml(0-2): 609, 3850, 9708, 10948, 4885
40000 @ 20+0.2 th 8
https://tests.stockfishchess.org/tests/view/5f290229a5abc164f05e4c58
ELO: 89.47 +-2.0 (95%) LOS: 100.0%
Total: 40000 W: 12756 L: 2677 D: 24567
Ptnml(0-2): 74, 1583, 8550, 7776, 2017
At the same time, the impact on the classical evaluation remains minimal, causing no significant
regression:
sprt @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f2906a2a5abc164f05e4c5b
LLR: 2.94 (-2.94,2.94) {-6.00,-4.00}
Total: 34936 W: 6502 L: 6825 D: 21609
Ptnml(0-2): 571, 4082, 8434, 3861, 520
sprt @ 60+0.6 th 1
https://tests.stockfishchess.org/tests/view/5f2906cfa5abc164f05e4c5d
LLR: 2.93 (-2.94,2.94) {-6.00,-4.00}
Total: 10088 W: 1232 L: 1265 D: 7591
Ptnml(0-2): 49, 914, 3170, 843, 68
The needed networks can be found at https://tests.stockfishchess.org/nns
It is recommended to use the default one as indicated by the `EvalFile` UCI option.
Guidelines for testing new nets can be found at
https://github.com/glinscott/fishtest/wiki/Creating-my-first-test#nnue-net-tests
Integration has been discussed in various issues:
https://github.com/official-stockfish/Stockfish/issues/2823https://github.com/official-stockfish/Stockfish/issues/2728
The integration branch will be closed after the merge:
https://github.com/official-stockfish/Stockfish/pull/2825https://github.com/official-stockfish/Stockfish/tree/nnue-player-wip
closes https://github.com/official-stockfish/Stockfish/pull/2912
This will be an exciting time for computer chess, looking forward to seeing the evolution of
this approach.
Bench: 4746616
Conceptually group hash clusters into super clusters of 256 clusters.
This scheme allows us to use hash sizes up to 32 TB
(= 2^32 super clusters = 2^40 clusters).
Use 48 bits of the Zobrist key to choose the cluster index. We use 8
extra bits to mitigate the quantization error for very large hashes when
scaling the hash key to cluster index.
The hash index computation is organized to be compatible with the existing
scheme for power-of-two hash sizes up to 128 GB.
Fixes https://github.com/official-stockfish/Stockfish/issues/1349
closes https://github.com/official-stockfish/Stockfish/pull/2722
Passed non-regression STC:
LLR: 2.93 (-2.94,2.94) {-1.50,0.50}
Total: 37976 W: 7336 L: 7211 D: 23429
Ptnml(0-2): 578, 4295, 9149, 4356, 610
https://tests.stockfishchess.org/tests/view/5edcbaaef29b40b0fc95abc5
No functional change.
for users that set the needed privilige "Lock Pages in Memory"
large pages will be automatically enabled (see Readme.md).
This expert setting might improve speed, 5% - 30%, depending
on the hardware, the number of threads and hash size. More for
large hashes, large number of threads and NUMA. If the operating
system can not allocate large pages (easier after a reboot), default
allocation is used automatically. The engine log provides details.
closes https://github.com/official-stockfish/Stockfish/pull/2656
fixes https://github.com/official-stockfish/Stockfish/issues/2619
No functional change
Align the TT allocation by 2M to make it huge page friendly and advise the
kernel to use huge pages.
Benchmarks on my i7-8700K (6C/12T) box: (3 runs per bench per config)
vanilla (nps) hugepages (nps) avg
==================================================================================
bench | 3012490 3024364 3036331 3071052 3067544 3071052 +1.5%
bench 16 12 20 | 19237932 19050166 19085315 19266346 19207025 19548758 +1.1%
bench 16384 12 20 | 18182313 18371581 18336838 19381275 19738012 19620225 +7.0%
On my box, huge pages have a significant perf impact when using a big
hash size. They also speed up TT initialization big time:
vanilla (s) huge pages (s) speed-up
=======================================================================
time stockfish bench 16384 1 1 | 5.37 1.48 3.6x
In practice, huge pages with auto-defrag may always be enabled in the
system, in which case this patch has no effect. This
depends on the values in /sys/kernel/mm/transparent_hugepage/enabled
and /sys/kernel/mm/transparent_hugepage/defrag.
closes https://github.com/official-stockfish/Stockfish/pull/2463
No functional change
Simplification that eliminates ONE_PLY, based on a suggestion in the forum that
support for fractional plies has never been used, and @mcostalba's openness to
the idea of eliminating it. We lose a little bit of type safety by making Depth
an integer, but in return we simplify the code in search.cpp quite significantly.
No functional change
------------------------------------------
The argument favoring eliminating ONE_PLY:
* The term “ONE_PLY” comes up in a lot of forum posts (474 to date)
https://groups.google.com/forum/?fromgroups=#!searchin/fishcooking/ONE_PLY%7Csort:relevance
* There is occasionally a commit that breaks invariance of the code
with respect to ONE_PLY
https://groups.google.com/forum/?fromgroups=#!searchin/fishcooking/ONE_PLY%7Csort:date/fishcooking/ZIPdYj6k0fk/KdNGcPWeBgAJ
* To prevent such commits, there is a Travis CI hack that doubles ONE_PLY
and rechecks bench
* Sustaining ONE_PLY has, alas, not resulted in any improvements to the
engine, despite many individuals testing many experiments over 5 years.
The strongest argument in favor of preserving ONE_PLY comes from @locutus:
“If we use par example ONE_PLY=256 the parameter space is increases by the
factor 256. So it seems very unlikely that the optimal setting is in the
subspace of ONE_PLY=1.”
There is a strong theoretical impediment to fractional depth systems: the
transposition table uses depth to determine when a stored result is good
enough to supply an answer for a current search. If you have fractional
depths, then different pathways to the position can be at fractionally
different depths.
In the end, there are three separate times when a proposal to remove ONE_PLY
was defeated by the suggestion to “give it a few more months.” So… it seems
like time to remove this distraction from the community.
See the pull request here:
https://github.com/official-stockfish/Stockfish/pull/2289
High rootDepths, selDepths and generally searches are increasingly
common with long time control games, analysis, and improving hardware.
In this case, depths of MAX_DEPTH/MAX_PLY (128) can be reached,
and the search tree is truncated.
In principle MAX_PLY can be easily increased, except for a technicality
of storing depths in a signed 8 bit int in the TT. This patch increases
MAX_PLY by storing the depth in an unsigned 8 bit, after shifting by the
most negative depth stored in TT (DEPTH_NONE).
No regression at STC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 42235 W: 9565 L: 9484 D: 23186
http://tests.stockfishchess.org/tests/view/5cdb35360ebc5925cf0595e1
Verified to reach high depths on
k1b5/1p1p4/pP1Pp3/K2pPp2/1P1p1P2/3P1P2/5P2/8 w - -
info depth 142 seldepth 154 multipv 1 score cp 537 nodes 26740713110 ...
No bench change.
Introducing new concept, saving principal lines into the transposition table
to generate a "critical search tree" which we can reuse later for intelligent
pruning/extension decisions.
For instance in this patch we just reduce reduction for these lines. But a lot
of other ideas are possible.
To go further : tune some parameters, how to add or remove lines from the
critical search tree, how to use these lines in search choices, etc.
STC :
LLR: 2.94 (-2.94,2.94) [0.50,4.50]
Total: 59761 W: 13321 L: 12863 D: 33577 +2.23 ELO
http://tests.stockfishchess.org/tests/view/5c34da5d0ebc596a450c53d3
LTC :
LLR: 2.96 (-2.94,2.94) [0.00,3.50]
Total: 26826 W: 4439 L: 4191 D: 18196 +2.9 ELO
http://tests.stockfishchess.org/tests/view/5c35ceb00ebc596a450c65b2
Special thanks to Miguel Lahoz for his help in transposition table in/out.
Bench: 3399866
* 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.
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.
The patch was tested for correctness by running bench with and
without the change against current master, and the tablebase hit
numbers were found to be identical in both cases. See the pull
request comments for details:
https://github.com/official-stockfish/Stockfish/pull/1826
No functional change.
Preparation commit for the upcoming Stockfish 10 version, giving a chance to catch last minute feature bugs and evaluation regression during the one-week code freeze period. Also changing the copyright dates to include 2019.
No functional change
Enable numa machinery only for STRICTLY MORE than 8 threads. Reason for this
change is that nowadays SMP tests are always done with 8 threads. That is a
problem for multi-socket Windows machines running on fishtest.
No functional change
Storing unconditionally the current generation and bound is equivalent to master.
Part of the condition was added as a speed optimization in #429.
Here the branch is fully eliminated.
passed STC single-threaded:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 73515 W: 16378 L: 16359 D: 40778
http://tests.stockfishchess.org/tests/view/5b2fc38c0ebc5902b2e57fd5
passed STC multi-threaded:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 63725 W: 12916 L: 12874 D: 37935
http://tests.stockfishchess.org/tests/view/5b307b8f0ebc5902b2e5895f
The multithreaded test was run after a plausible suggestion by @mstembera that the effect of this could be larger with many cores. The result seems to indicate this doesn't really matter on the 8core architecture abundantly available on fishtest.
No functional change
Makes sure the potential benefit of first touch does not depend on
the order of the UCI commands Threads and Hash, by reallocating the
hash if a Threads is issued. The cost is zeroing the TT once more
than needed. In case the prefered order (first Threads than Hash)
is employed, this amounts to zeroing the default sized TT (16Mb),
which is essentially instantaneous.
Follow up for https://github.com/official-stockfish/Stockfish/pull/1601
where additional data and discussion is available.
Closes https://github.com/official-stockfish/Stockfish/pull/1620
No functional change.
Stockfish currently takes a while to clear the TT when using larger hash sizes.
On one machine with 128 GB hash it takes about 50 seconds with a single thread,
allowing it to use all allocated cores brought that time down to 4 seconds on
some Linux systems. The patch was further tested on Windows and refined with
NUMA binding of the hash initializing threads (we refer to pull request #1601
for the complete discussion and the speed measurements).
Closes https://github.com/official-stockfish/Stockfish/pull/1601
No functional change
as discussed in issue #1349, the way pages are allocated with calloc might imply some overhead on first write.
This overhead can be large and slow down the first search after a TT resize significantly, especially for large TT.
Using an explicit clear of the TT on resize fixes this problem.
Not implemented, but possibly useful for large TT, is to do this zero-ing using all search threads. Not only would this be faster, it could also lead to a more favorable memory allocation on numa systems with a first touch policy.
No functional change.
For efficiency reasons current master only allows for transposition table sizes that are N = 2^k in size, the index computation can be done efficiently as (hash % N) can be written instead as (hash & 2^k - 1). On a typical computer (with 4, 8... etc Gb of RAM), this implies roughly half the RAM is left unused in analysis.
This issue was mentioned on fishcooking by Mindbreaker:
http://tests.stockfishchess.org/tests/view/5a3587de0ebc590ccbb8be04
Recently a neat trick was proposed to map a hash into the range [0,N[ more efficiently than (hash % N) for general N, nearly as efficiently as (hash % 2^k):
https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
namely computing (hash * N / 2^32) for 32 bit hashes. This patch implements this trick and now allows for general hash sizes. Note that for N = 2^k this just amounts to using a different subset of bits from the hash. Master will use the lower k bits, this trick will use the upper k bits (of the 32 bit hash).
There is no slowdown as measured with [-3, 1] test:
http://tests.stockfishchess.org/tests/view/5a3587de0ebc590ccbb8be04
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 128498 W: 23332 L: 23395 D: 81771
There are two (smaller) caveats:
1) the patch is implemented for a 32 bit hash (so that a 64 bit multiply can be used), this effectively limits the number of clusters that can be used to 2^32 or to 128Gb of transpostion table. That's a change in the maximum allowed TT size, which could bother those using 256Gb or more regularly.
2) Already in master, an excluded move is hashed into the position key in rather simple way, essentially only affecting the lower 16 bits of the key. This is OK in master, since bits 0-15 end up in the index, but not in the new scheme, which picks the higher bits. This is 'fixed' by shifting the excluded move a few bits up. Eventually a better hashing scheme seems wise.
Despite these two caveats, I think this is a nice improvement in usability.
Bench: 5346341
Rename shift_bb() to shift(), and DELTA_S to SOUTH, etc.
to improve code readability, especially in evaluate.cpp
when they are used together:
old b = shift_bb<DELTA_S>(pos.pieces(PAWN))
new b = shift<SOUTH>(pos.pieces(PAWN))
While there fix some small code style issues.
No functional change.
This non-functional change patch is a deep work to allow SF to be independent
from the actual value of ONE_PLY (currently set to 1). I have verified SF is
now independent for ONE_PLY values 1, 2, 4, 8, 16, 32 and 256.
This patch gives consistency to search code and enables future work, opening
the door to safely tweaking the ONE_PLY value for any reason.
Verified for no speed regression at STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 95643 W: 17728 L: 17737 D: 60178
No functional change.
Only refresh TT entry when it's really necessary.
This should give a small speed boost for some machines.
And it's a risk-free change.
No functional change.
Resolves#429
