]> git.sesse.net Git - stockfish/log
stockfish
3 years agoRevert futility pruning patches
Joost VandeVondele [Thu, 5 Aug 2021 14:34:37 +0000 (16:34 +0200)]
Revert futility pruning patches

reverts 09b6d28391cf582d99897360b225bcbbe38dd1c6 and
dbd7f602d3c7622df294f87d7239b5aaf31f695f that significantly impact mate
finding capabilities. For example on ChestUCI_23102018.epd, at 1M nodes,
the number of mates found is nearly reduced 2x without these depth conditions:

       sf6  2091
       sf7  2093
       sf8  2107
       sf9  2062
      sf10  2208
      sf11  2552
      sf12  2563
      sf13  2509
      sf14  2427
    master  1246
   patched  2467

(script for testing at https://github.com/official-stockfish/Stockfish/files/6936412/matecheck.zip)

closes https://github.com/official-stockfish/Stockfish/pull/3641

fixes https://github.com/official-stockfish/Stockfish/issues/3627

Bench: 5467570

3 years agoSEE simplification
VoyagerOne [Thu, 5 Aug 2021 12:50:24 +0000 (08:50 -0400)]
SEE simplification

Simplified SEE formula by removing std::min. Should also be easier to tune.

STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 22656 W: 1836 L: 1729 D: 19091
Ptnml(0-2): 54, 1426, 8267, 1521, 60
https://tests.stockfishchess.org/tests/view/610ae62f2a8a49ac5be79449

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 26248 W: 806 L: 744 D: 24698
Ptnml(0-2): 6, 668, 11715, 728, 7
https://tests.stockfishchess.org/tests/view/610b17ad2a8a49ac5be79466

closes https://github.com/official-stockfish/Stockfish/pull/3643

bench:  4915145

3 years agoUpdate default net to nn-46832cfbead3.nnue
SFisGOD [Wed, 4 Aug 2021 11:26:06 +0000 (19:26 +0800)]
Update default net to nn-46832cfbead3.nnue

SPSA 1: https://tests.stockfishchess.org/tests/view/6100e7f096b86d98abf6a832
Parameters: A total of 256 net weights and 8 net biases were tuned (output layer)
Base net: nn-56a5f1c4173a.nnue
New net: nn-ec3c8e029926.nnue

SPSA 2: https://tests.stockfishchess.org/tests/view/610733caafad2da4f4ae3da7
Parameters: A total of 256 net biases were tuned (hidden layer 2)
Base net: nn-ec3c8e029926.nnue
New net: nn-46832cfbead3.nnue

STC:
LLR: 2.98 (-2.94,2.94) <-0.50,2.50>
Total: 50520 W: 3953 L: 3765 D: 42802
Ptnml(0-2): 138, 3063, 18678, 3235, 146
https://tests.stockfishchess.org/tests/view/610a79692a8a49ac5be793f4

LTC:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 57256 W: 1723 L: 1566 D: 53967
Ptnml(0-2): 12, 1442, 25568, 1589, 17
https://tests.stockfishchess.org/tests/view/610ac5bb2a8a49ac5be79434

Closes https://github.com/official-stockfish/Stockfish/pull/3642

Bench: 5359314

3 years agoSimplify new cmh pruning thresholds by using directly a quadratic formula.
Stefan Geschwentner [Tue, 3 Aug 2021 14:32:48 +0000 (16:32 +0200)]
Simplify new cmh pruning thresholds by using directly a quadratic formula.

This decouples also the stat bonus updates from the threshold which creates less dependencies for tuning of stat bonus parameters.
Perhaps a further fine tuning of the now separated coefficients for constHist[0] and constHist[1] could give further gains.

STC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 78384 W: 6134 L: 6090 D: 66160
Ptnml(0-2): 207, 5013, 28705, 5063, 204
https://tests.stockfishchess.org/tests/view/6106d235afad2da4f4ae3d4b

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 38176 W: 1149 L: 1095 D: 35932
Ptnml(0-2): 6, 1000, 17030, 1038, 14
https://tests.stockfishchess.org/tests/view/6107a080afad2da4f4ae3def

closes https://github.com/official-stockfish/Stockfish/pull/3639

Bench: 5098146

3 years agoFutile pruning simplification
VoyagerOne [Mon, 2 Aug 2021 17:52:48 +0000 (13:52 -0400)]
Futile pruning simplification

Remove CMH conditions in futile pruning.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 93520 W: 7165 L: 7138 D: 79217
Ptnml(0-2): 222, 5923, 34427, 5982, 206
https://tests.stockfishchess.org/tests/view/61083104e50a153c346ef8df

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 59072 W: 1746 L: 1706 D: 55620
Ptnml(0-2): 13, 1562, 26353, 1588, 20
https://tests.stockfishchess.org/tests/view/610894f2e50a153c346ef913

closes https://github.com/official-stockfish/Stockfish/pull/3638

Bench: 5229673

3 years agoCMH Pruning Tweak
VoyagerOne [Sat, 31 Jul 2021 12:18:49 +0000 (08:18 -0400)]
CMH Pruning Tweak

replace CounterMovePruneThreshold by a depth dependent threshold

STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 35512 W: 2718 L: 2552 D: 30242
Ptnml(0-2): 66, 2138, 13194, 2280, 78
https://tests.stockfishchess.org/tests/view/6104442fafad2da4f4ae3b94

LTC:
LLR: 2.96 (-2.94,2.94) <0.50,3.50>
Total: 36536 W: 1150 L: 1019 D: 34367
Ptnml(0-2): 10, 920, 16278, 1049, 11
https://tests.stockfishchess.org/tests/view/6104b033afad2da4f4ae3bbc

closes https://github.com/official-stockfish/Stockfish/pull/3636

Bench: 5848718

3 years agoAvoid unnecessary stores in the affine transform
Tomasz Sobczyk [Tue, 27 Jul 2021 20:12:14 +0000 (22:12 +0200)]
Avoid unnecessary stores in the affine transform

This patch improves the codegen in the AffineTransform::forward function for architectures >=SSSE3. Current code works directly on memory and the compiler cannot see that the stores through outptr do not alias the loads through weights and input32. The solution implemented is to perform the affine transform with local variables as accumulators and only store the result to memory at the end. The number of accumulators required is OutputDimensions / OutputSimdWidth, which means that for the 1024->16 affine transform it requires 4 registers with SSSE3, 2 with AVX2, 1 with AVX512. It also cuts the number of stores required by NumRegs * 256 for each node evaluated. The local accumulators are expected to be assigned to registers, but even if this cannot be done in some case due to register pressure it will help the compiler to see that there is no aliasing between the loads and stores and may still result in better codegen.

See https://godbolt.org/z/59aTKbbYc for codegen comparison.

passed STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 140328 W: 10635 L: 10358 D: 119335
Ptnml(0-2): 302, 8339, 52636, 8554, 333

closes https://github.com/official-stockfish/Stockfish/pull/3634

No functional change

3 years agoUpdate default net to nn-56a5f1c4173a.nnue
SFisGOD [Tue, 27 Jul 2021 16:43:58 +0000 (00:43 +0800)]
Update default net to nn-56a5f1c4173a.nnue

SPSA 1: https://tests.stockfishchess.org/tests/view/60fd24efd8a6b65b2f3a796e
Parameters: A total of 256 net biases were tuned (hidden layer 2)
New best values: Half of the changes from the tuning run
New net: nn-5992d3ba79f3.nnue

SPSA 2: https://tests.stockfishchess.org/tests/view/60fec7d6d8a6b65b2f3a7aa2
Parameters: A total of 128 net biases were tuned (hidden layer 1)
New best values: Half of the changes from the tuning run
New net: nn-56a5f1c4173a.nnue

STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 140392 W: 10863 L: 10578 D: 118951
Ptnml(0-2): 347, 8754, 51718, 9021, 356
https://tests.stockfishchess.org/tests/view/610037e396b86d98abf6a79e

LTC:
LLR: 2.95 (-2.94,2.94) <0.50,3.50>
Total: 14216 W: 454 L: 355 D: 13407
Ptnml(0-2): 4, 323, 6356, 420, 5
https://tests.stockfishchess.org/tests/view/61019995afad2da4f4ae3a3c

Closes #3633

Bench: 4801359

3 years agoUpdate default net to nn-26abeed38351.nnue
SFisGOD [Sun, 25 Jul 2021 11:43:25 +0000 (19:43 +0800)]
Update default net to nn-26abeed38351.nnue

SPSA: https://tests.stockfishchess.org/tests/view/60fba335d8a6b65b2f3a7891

New best values: Half of the changes from the tuning run.
Setting: nodestime=300 with 10+0.1 (approximate real TC is 2.5 seconds)
The rest is the same as described in #3593

The change from nodestime=600 to 300 was suggested by gekkehenker to prevent time losses for some slow workers
SFisGOD@94cd757#commitcomment-53324840

STC:
LLR: 2.96 (-2.94,2.94) <-0.50,2.50>
Total: 67448 W: 5241 L: 5036 D: 57171
Ptnml(0-2): 151, 4198, 24827, 4391, 157
https://tests.stockfishchess.org/tests/view/60fd50f2d8a6b65b2f3a798e

LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 48752 W: 1504 L: 1358 D: 45890
Ptnml(0-2): 13, 1226, 21754, 1368, 15
https://tests.stockfishchess.org/tests/view/60fd7bb2d8a6b65b2f3a79a9

Closes https://github.com/official-stockfish/Stockfish/pull/3630

Bench:  5124774

3 years agoSimplification in LMR
Giacomo Lorenzetti [Sat, 24 Jul 2021 20:03:29 +0000 (22:03 +0200)]
Simplification in LMR

This commit removes the `!captureOrPromotion` condition from ttCapture reduction and from good/bad history reduction (similar to #3619).

passed STC:
https://tests.stockfishchess.org/tests/view/60fc734ad8a6b65b2f3a7922
LLR: 2.97 (-2.94,2.94) <-2.50,0.50>
Total: 48680 W: 3855 L: 3776 D: 41049
Ptnml(0-2): 118, 3145, 17744, 3206, 127

passed LTC:
https://tests.stockfishchess.org/tests/view/60fce7d5d8a6b65b2f3a794c
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 86528 W: 2471 L: 2450 D: 81607
Ptnml(0-2): 28, 2203, 38777, 2232, 24

closes https://github.com/official-stockfish/Stockfish/pull/3629

Bench: 4951406

3 years agoUpdate the default net to nn-76a8a7ffb820.nnue.
MichaelB7 [Sat, 24 Jul 2021 12:42:00 +0000 (08:42 -0400)]
Update the default net to nn-76a8a7ffb820.nnue.

combined work by Serio Vieri, Michael Byrne, and Jonathan D (aka SFisGod) based on top of previous developments, by restarts from good nets.

Sergio generated the net https://tests.stockfishchess.org/api/nn/nn-d8609abe8caf.nnue:

The initial net nn-d8609abe8caf.nnue is trained by generating around 16B of training data from the last master net nn-9e3c6298299a.nnue, then trained, continuing from the master net, with lambda=0.2 and sampling ratio of 1. Starting with LR=2e-3, dropping LR with a factor of 0.5 until it reaches LR=5e-4. in_scaling is set to 361. No other significant changes made to the pytorch trainer.

Training data gen command (generates in chunks of 200k positions):

generate_training_data min_depth 9 max_depth 11 count 200000 random_move_count 10 random_move_max_ply 80 random_multi_pv 12 random_multi_pv_diff 100 random_multi_pv_depth 8 write_min_ply 10 eval_limit 1500 book noob_3moves.epd output_file_name gendata/$(date +"%Y%m%d-%H%M")_${HOSTNAME}.binpack

PyTorch trainer command (Note that this only trains for 20 epochs, repeatedly train until convergence):

python train.py --features "HalfKAv2^" --max_epochs 20 --smart-fen-skipping --random-fen-skipping 500 --batch-size 8192 --default_root_dir $dir --seed $RANDOM --threads 4 --num-workers 32 --gpus $gpuids --track_grad_norm 2 --gradient_clip_val 0.05 --lambda 0.2 --log_every_n_steps 50 $resumeopt $data $val

See https://github.com/sergiovieri/Stockfish/tree/tools_mod/rl for the scripts used to generate data.

Based on that Michael generated nn-76a8a7ffb820.nnue in the following way:

The net being submitted was trained with the pytorch trainer: https://github.com/glinscott/nnue-pytorch

python train.py i:/bin/all.binpack i:/bin/all.binpack --gpus 1 --threads 4 --num-workers 30 --batch-size 16384 --progress_bar_refresh_rate 30 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^ --auto_lr_find True --lambda=1.0 --max_epochs=240 --seed %random%%random% --default_root_dir exp/run_109 --resume-from-model ./pt/nn-d8609abe8caf.pt

This run is thus started from Segio Vieri's net nn-d8609abe8caf.nnue

all.binpack equaled 4 parts Wrong_NNUE_2.binpack https://drive.google.com/file/d/1seGNOqcVdvK_vPNq98j-zV3XPE5zWAeq/view?usp=sharing plus two parts of Training_Data.binpack https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing
Each set was concatenated together - making one large Wrong_NNUE 2 binpack and one large Training so the were approximately equal in size. They were then interleaved together. The idea was to give Wrong_NNUE.binpack closer to equal weighting with the Training_Data binpack

model.py modifications:
loss = torch.pow(torch.abs(p - q), 2.6).mean()
LR = 8.0e-5 calculated as follows: 1.5e-3*(.992^360) - the idea here was to take a highly trained net and just use all.binpack as a finishing micro refinement touch for the last 2 Elo or so. This net was discovered on the 59th epoch.
optimizer = ranger.Ranger(train_params, betas=(.90, 0.999), eps=1.0e-7, gc_loc=False, use_gc=False)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.992)
For this micro optimization, I had set the period to "5" in train.py. This changes the checkpoint output so that every 5th checkpoint file is created

The final touches were to adjust the NNUE scale, as was done by Jonathan in tests running at the same time.

passed LTC
https://tests.stockfishchess.org/tests/view/60fa45aed8a6b65b2f3a77a4
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 53040 W: 1732 L: 1575 D: 49733
Ptnml(0-2): 14, 1432, 23474, 1583, 17

passed STC
https://tests.stockfishchess.org/tests/view/60f9fee2d8a6b65b2f3a7775
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 37928 W: 3178 L: 3001 D: 31749
Ptnml(0-2): 100, 2446, 13695, 2623, 100.

closes https://github.com/official-stockfish/Stockfish/pull/3626

Bench: 5169957

3 years agoApply good/bad history reduction also when inCheck
Giacomo Lorenzetti [Sun, 18 Jul 2021 18:14:11 +0000 (20:14 +0200)]
Apply good/bad history reduction also when inCheck

Main idea is that, in some cases, 'in check' situations are not so different from 'not in check' ones.
Trying to use piece count in order to select only a few 'in check' situations have failed LTC testing.
It could be interesting to apply one of those ideas in other parts of the search function.

passed STC:
https://tests.stockfishchess.org/tests/view/60f1b68dd1189bed71812d40
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 53472 W: 4078 L: 4008 D: 45386
Ptnml(0-2): 127, 3297, 19795, 3413, 104

passed LTC:
https://tests.stockfishchess.org/tests/view/60f291e6d1189bed71812de3
LLR: 2.92 (-2.94,2.94) <-2.50,0.50>
Total: 89712 W: 2651 L: 2632 D: 84429
Ptnml(0-2): 60, 2261, 40188, 2294, 53

closes https://github.com/official-stockfish/Stockfish/pull/3619

Bench: 5185789

3 years agoSimplify lowply-history scoring logic
pb00067 [Thu, 15 Jul 2021 18:56:21 +0000 (20:56 +0200)]
Simplify lowply-history scoring logic

STC:
https://tests.stockfishchess.org/tests/view/60eee559d1189bed71812b16
LLR: 2.97 (-2.94,2.94) <-2.50,0.50>
Total: 33976 W: 2523 L: 2431 D: 29022
Ptnml(0-2): 66, 2030, 12730, 2070, 92

LTC:
https://tests.stockfishchess.org/tests/view/60eefa12d1189bed71812b24
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 107240 W: 3053 L: 3046 D: 101141
Ptnml(0-2): 56, 2668, 48154, 2697, 45

closes https://github.com/official-stockfish/Stockfish/pull/3616

bench: 5199177

3 years agoPrune illegal moves in qsearch earlier
Vizvezdenec [Sun, 18 Jul 2021 10:51:14 +0000 (13:51 +0300)]
Prune illegal moves in qsearch earlier

The main idea is that illegal moves influencing search or
qsearch obviously can't be any sort of good. The only reason
why initially legality checks for search and qsearch were done
after they actually can influence some heuristics is because
legality check is expensive computationally. Eventually in
search it was moved to the place where it makes sure that
illegal moves can't influence search.

This patch shows that the same can be done for qsearch + it
passed STC with elo-gaining bounds + it removes 3 lines of code
because one no longer needs to increment/decrement movecount
on illegal moves.

passed STC with elo-gaining bounds
https://tests.stockfishchess.org/tests/view/60f20aefd1189bed71812da0
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 61512 W: 4688 L: 4492 D: 52332
Ptnml(0-2): 139, 3730, 22848, 3874, 165

The same version functionally but with moving condition ever earlier
passed LTC with simplification bounds.
https://tests.stockfishchess.org/tests/view/60f292cad1189bed71812de9
LLR: 2.98 (-2.94,2.94) <-2.50,0.50>
Total: 60944 W: 1724 L: 1685 D: 57535
Ptnml(0-2): 11, 1556, 27298, 1597, 10

closes https://github.com/official-stockfish/Stockfish/pull/3618

bench 4709569

3 years agoAdd macOS and windows to CI
Liam Keegan [Wed, 21 Jul 2021 07:33:13 +0000 (09:33 +0200)]
Add macOS and windows to CI

- macOS
  - system clang
  - gcc
- windows / msys2
  - mingw 64-bit gcc
  - mingw 32-bit gcc
- minor code fixes to get new CI jobs to pass
  - code: suppress unused-parameter warning on 32-bit windows
  - Makefile: if arch=any on macos, don't specify arch at all

fixes https://github.com/official-stockfish/Stockfish/issues/2958

closes https://github.com/official-stockfish/Stockfish/pull/3623

No functional change

3 years agoDon't save excluded move eval in TT
VoyagerOne [Mon, 12 Jul 2021 18:44:29 +0000 (14:44 -0400)]
Don't save excluded move eval in TT

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 17544 W: 1384 L: 1236 D: 14924
Ptnml(0-2): 37, 1031, 6499, 1157, 48
https://tests.stockfishchess.org/tests/view/60ec8d9bd1189bed71812999

LTC:
LLR: 2.95 (-2.94,2.94) <0.50,3.50>
Total: 26136 W: 823 L: 707 D: 24606
Ptnml(0-2): 6, 643, 11656, 755, 8
https://tests.stockfishchess.org/tests/view/60ecb11ed1189bed718129ba

closes https://github.com/official-stockfish/Stockfish/pull/3614

Bench: 5505251

3 years agoRemove second futility pruning depth limit
Vizvezdenec [Sat, 10 Jul 2021 21:09:15 +0000 (00:09 +0300)]
Remove second futility pruning depth limit

This patch removes futility pruning lmrDepth limit for futility pruning at parent nodes.
Since it's already capped by margin that is a function of lmrDepth there is no need to extra cap it with lmrDepth.

passed STC
https://tests.stockfishchess.org/tests/view/60e9b5dfd1189bed71812777
LLR: 2.97 (-2.94,2.94) <-2.50,0.50>
Total: 14872 W: 1264 L: 1145 D: 12463
Ptnml(0-2): 37, 942, 5369, 1041, 47

passed LTC
https://tests.stockfishchess.org/tests/view/60e9c635d1189bed71812790
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 40336 W: 1280 L: 1225 D: 37831
Ptnml(0-2): 24, 1057, 17960, 1094, 33

closes https://github.com/official-stockfish/Stockfish/pull/3612

bench: 5064969

3 years agoSEE: simplify stm variable initialization
pb00067 [Wed, 7 Jul 2021 12:32:54 +0000 (14:32 +0200)]
SEE: simplify stm variable initialization

Pull #3458 removed the only usage of pos.see_ge() moving pieces that
don't belong to the side to move, so we can simplify this, adding an assert.

closes https://github.com/official-stockfish/Stockfish/pull/3607

No functional change

3 years agoRemove futility pruning depth limit
Vizvezdenec [Tue, 6 Jul 2021 17:44:50 +0000 (20:44 +0300)]
Remove futility pruning depth limit

This patch removes futility pruning depth limit for child node futility pruning.
In current master it was double capped by depth and by futility margin, which is also a function of depth, which didn't make much sense.

passed STC
https://tests.stockfishchess.org/tests/view/60e2418f9ea99d7c2d693e64
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 116168 W: 9100 L: 9097 D: 97971
Ptnml(0-2): 319, 7496, 42476, 7449, 344

passed LTC
https://tests.stockfishchess.org/tests/view/60e3374f9ea99d7c2d693f20
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 43304 W: 1282 L: 1231 D: 40791
Ptnml(0-2): 8, 1126, 19335, 1173, 10

closes https://github.com/official-stockfish/Stockfish/pull/3606

bench 4965493

3 years agoUpdate default net to nn-9e3c6298299a.nnue
SFisGOD [Fri, 2 Jul 2021 22:13:13 +0000 (06:13 +0800)]
Update default net to nn-9e3c6298299a.nnue

Optimization of nn-956480d8378f.nnue using SPSA
https://tests.stockfishchess.org/tests/view/60da2bf63beab81350ac9fe7

Same method as described in PR #3593

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 17792 W: 1525 L: 1372 D: 14895
Ptnml(0-2): 28, 1156, 6401, 1257, 54
https://tests.stockfishchess.org/tests/view/60deffc59ea99d7c2d693c19

LTC:
LLR: 2.96 (-2.94,2.94) <0.50,3.50>
Total: 36544 W: 1245 L: 1109 D: 34190
Ptnml(0-2): 12, 988, 16139, 1118, 15
https://tests.stockfishchess.org/tests/view/60df11339ea99d7c2d693c22

closes https://github.com/official-stockfish/Stockfish/pull/3601

Bench: 4687476

3 years agoAllow passing RTLIB=compiler-rt to make
Paul Mulders [Tue, 29 Jun 2021 09:13:54 +0000 (11:13 +0200)]
Allow passing RTLIB=compiler-rt to make

Not all linux users will have libatomic installed.
When using clang as the system compiler with compiler-rt as the default
runtime library instead of libgcc, atomic builtins may be provided by compiler-rt.
This change allows such users to pass RTLIB=compiler-rt to make sure
the build doesn't error out on the missing (unnecessary) libatomic.

closes https://github.com/official-stockfish/Stockfish/pull/3597

No functional change

3 years agono cut node reduction for killer moves.
candirufish [Thu, 1 Jul 2021 17:51:41 +0000 (19:51 +0200)]
no cut node reduction for killer moves.

stc:
LLR: 2.95 (-2.94,2.94) <-0.50,2.50>
Total: 44344 W: 3474 L: 3294 D: 37576
Ptnml(0-2): 117, 2710, 16338, 2890, 117
https://tests.stockfishchess.org/tests/view/60d8ea673beab81350ac9eb8

ltc:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 82600 W: 2638 L: 2441 D: 77521
Ptnml(0-2): 38, 2147, 36749, 2312, 54
https://tests.stockfishchess.org/tests/view/60d9048f3beab81350ac9eed

closes https://github.com/official-stockfish/Stockfish/pull/3600

Bench: 5160239

3 years agoSimplify lazy_skip.
xoto10 [Wed, 30 Jun 2021 08:22:59 +0000 (09:22 +0100)]
Simplify lazy_skip.

Small speedup by removing operations in lazy_skip.

STC 10+0.1 :
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 55088 W: 4553 L: 4482 D: 46053
Ptnml(0-2): 163, 3546, 20045, 3637, 153
https://tests.stockfishchess.org/tests/view/60daa2cb3beab81350aca04d

LTC 60+0.6 :
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 46136 W: 1457 L: 1407 D: 43272
Ptnml(0-2): 10, 1282, 20442, 1316, 18
https://tests.stockfishchess.org/tests/view/60db0e753beab81350aca08e

closes https://github.com/official-stockfish/Stockfish/pull/3599

Bench 5122403

3 years agoSimplify format_cp_aligned_dot()
Stéphane Nicolet [Wed, 23 Jun 2021 07:55:42 +0000 (09:55 +0200)]
Simplify format_cp_aligned_dot()

closes https://github.com/official-stockfish/Stockfish/pull/3583

No functional change

3 years agoRestore development version
Joost VandeVondele [Sat, 3 Jul 2021 07:20:06 +0000 (09:20 +0200)]
Restore development version

No functional change

3 years agoStockfish 14
Joost VandeVondele [Mon, 28 Jun 2021 19:46:04 +0000 (21:46 +0200)]
Stockfish 14

Official release version of Stockfish 14

Bench: 4770936

---

Today, we have the pleasure to announce Stockfish 14.

As usual, downloads will be freely available at https://stockfishchess.org

The engine is now significantly stronger than just a few months ago,
and wins four times more game pairs than it loses against the previous
release version [0]. Stockfish 14 is now at least 400 Elo ahead of
Stockfish 7, a top engine in 2016 [1]. During the last five years,
Stockfish has thus gained about 80 Elo per year.

Stockfish 14 evaluates positions more accurately than Stockfish 13 as
a result of two major steps forward in defining and training the
efficiently updatable neural network (NNUE) that provides the evaluation
for positions.

First, the collaboration with the Leela Chess Zero team - announced
previously [2] - has come to fruition. The LCZero team has provided a
collection of billions of positions evaluated by Leela that we have
combined with billions of positions evaluated by Stockfish to train the
NNUE net that powers Stockfish 14. The fact that we could use and combine
these datasets freely was essential for the progress made and demonstrates
the power of open source and open data [3].

Second, the architecture of the NNUE network was significantly updated:
the new network is not only larger, but more importantly, it deals better
with large material imbalances and can specialize for multiple phases of
the game [4]. A new project, kick-started by Gary Linscott and
Tomasz Sobczyk, led to a GPU accelerated net trainer written in
pytorch.[5] This tool allows for training high-quality nets in a couple
of hours.

Finally, this release features some search refinements, minor bug
fixes and additional improvements. For example, Stockfish is now about
90 Elo stronger for chess960 (Fischer random chess) at short time control.

The Stockfish project builds on a thriving community of enthusiasts
(thanks everybody!) that contribute their expertise, time, and resources
to build a free and open-source chess engine that is robust, widely
available, and very strong. We invite our chess fans to join the fishtest
testing framework and programmers to contribute to the project on
github [6].

Stay safe and enjoy chess!

The Stockfish team

[0] https://tests.stockfishchess.org/tests/view/60dae5363beab81350aca077
[1] https://nextchessmove.com/dev-builds
[2] https://stockfishchess.org/blog/2021/stockfish-13/
[3] https://lczero.org/blog/2021/06/the-importance-of-open-data/
[4] https://github.com/official-stockfish/Stockfish/commit/e8d64af1
[5] https://github.com/glinscott/nnue-pytorch/
[6] https://stockfishchess.org/get-involved/

3 years agoUpdate Top CPU Contributors
Brad Knox [Tue, 29 Jun 2021 06:40:16 +0000 (01:40 -0500)]
Update Top CPU Contributors

closes https://github.com/official-stockfish/Stockfish/pull/3595

No functional change

3 years agoUpdate default net to nn-3475407dc199.nnue
SFisGOD [Mon, 28 Jun 2021 06:58:51 +0000 (14:58 +0800)]
Update default net to nn-3475407dc199.nnue

Optimization of eight subnetwork output layers of Michael's nn-190f102a22c3.nnue using SPSA
https://tests.stockfishchess.org/tests/view/60d5510642a522cc50282ef3

Parameters: A total of 256 net weights and 8 net biases were tuned
New best values: The raw values at the end of the tuning run were used (800k games, 5 seconds TC)
Settings: default ck value and SPSA A is 30,000 (3.75% of the total number of games)

STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 29064 W: 2435 L: 2269 D: 24360
Ptnml(0-2): 72, 1857, 10505, 2029, 69
https://tests.stockfishchess.org/tests/view/60d8ea123beab81350ac9eb6

LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 61848 W: 2055 L: 1884 D: 57909
Ptnml(0-2): 18, 1708, 27310, 1861, 27
https://tests.stockfishchess.org/tests/view/60d8f0393beab81350ac9ec6

closes https://github.com/official-stockfish/Stockfish/pull/3593

Bench: 4770936

3 years agoMake net nn-956480d8378f.nnue the default
MichaelB7 [Sun, 27 Jun 2021 15:26:09 +0000 (11:26 -0400)]
Make net nn-956480d8378f.nnue the default

Trained with the pytorch trainer: https://github.com/glinscott/nnue-pytorch

python train.py i:/bin/all.binpack i:/bin/all.binpack --gpus 1 --threads 4 --num-workers 30 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^ --lambda=1.0 --max_epochs=440 --seed %random%%random% --default_root_dir exp/run_18 --resume-from-model ./pt/nn-75980ca503c6.pt

This run is thus started from a previous master net.

all.binpack equaled 4 parts Wrong_NNUE_2.binpack https://drive.google.com/file/d/1seGNOqcVdvK_vPNq98j-zV3XPE5zWAeq/view?usp=sharing plus two parts of Training_Data.binpack https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing
Each set was concatenated together - making one large Wrong_NNUE 2 binpack and one large Training so the were approximately equal in size. They were then interleaved together. The idea was to give Wrong_NNUE.binpack closer to equal weighting with the Training_Data binpack

passed STC:
https://tests.stockfishchess.org/tests/view/60d0c0a7a8ec07dc34c072b2
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 18440 W: 1693 L: 1531 D: 15216
Ptnml(0-2): 67, 1225, 6464, 1407, 57

passed LTC:
https://tests.stockfishchess.org/tests/view/60d762793beab81350ac9d72
LLR: 2.98 (-2.94,2.94) <0.50,3.50>
Total: 93120 W: 3152 L: 2933 D: 87035
Ptnml(0-2): 48, 2581, 41076, 2814, 41

passed LTC (rebased branch to current master):
https://tests.stockfishchess.org/tests/view/60d85eeb3beab81350ac9e2b
LLR: 2.96 (-2.94,2.94) <0.50,3.50>
Total: 42688 W: 1347 L: 1206 D: 40135
Ptnml(0-2): 14, 1097, 18981, 1238, 14.

closes https://github.com/official-stockfish/Stockfish/pull/3592

Bench: 4906727

3 years agoUpdate WDL model for NNUE
Joost VandeVondele [Wed, 23 Jun 2021 05:23:21 +0000 (07:23 +0200)]
Update WDL model for NNUE

This updates the WDL model based on the LTC statistics in June this year (10M games),
so from pre-NNUE to NNUE based results.

(for old results see, https://github.com/official-stockfish/Stockfish/pull/2778)

As before the fit by the model to the data is quite good.

closes https://github.com/official-stockfish/Stockfish/pull/3582

No functional change

3 years agoSimplify Reductions Initialization
bmc4 [Tue, 22 Jun 2021 22:33:14 +0000 (19:33 -0300)]
Simplify Reductions Initialization

passed

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 45032 W: 3600 L: 3518 D: 37914
Ptnml(0-2): 111, 2893, 16435, 2957, 120
https://tests.stockfishchess.org/tests/view/60d2655d40925195e7a6c527

LTC:
LLR: 3.00 (-2.94,2.94) <-2.50,0.50>
Total: 25728 W: 786 L: 722 D: 24220
Ptnml(0-2): 5, 650, 11494, 706, 9
https://tests.stockfishchess.org/tests/view/60d2b14240925195e7a6c577

closes https://github.com/official-stockfish/Stockfish/pull/3584

bench: 4602977

3 years agoDetect fortresses a little bit quicker
Stéphane Nicolet [Tue, 22 Jun 2021 07:08:37 +0000 (09:08 +0200)]
Detect fortresses a little bit quicker

In the so-called "hybrid" method of evaluation of current master, we use the
classical eval (because of its speed) instead of the NNUE eval when the classical
material balance approximation hints that the position is "winning enough" to
rely on the classical eval.

This trade-off idea between speed and accuracy works well in general, but in
some fortress positions the classical eval is just bad. So in shuffling branches
of the search tree, we (slowly) increase the thresehold so that eventually we
don't trust classical anymore and switch to NNUE evaluation.

This patch increases that threshold faster, so that we switch to NNUE quicker
in shuffling branches. Idea is to incite Stockfish to spend less time in fortresses
lines in the search tree, and spend more time searching the critical lines.

passed STC:
LLR: 2.96 (-2.94,2.94) <-0.50,2.50>
Total: 47872 W: 3908 L: 3720 D: 40244
Ptnml(0-2): 122, 3053, 17419, 3199, 143
https://tests.stockfishchess.org/tests/view/60cef34b457376eb8bcab79d

passed LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 73616 W: 2326 L: 2143 D: 69147
Ptnml(0-2): 21, 1940, 32705, 2119, 23
https://tests.stockfishchess.org/tests/view/60cf6d842114332881e73528

Retested at LTC against lastest master:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 18264 W: 642 L: 532 D: 17090
Ptnml(0-2): 6, 479, 8055, 583, 9
https://tests.stockfishchess.org/tests/view/60d18cd540925195e7a6c351

closes https://github.com/official-stockfish/Stockfish/pull/3578

Bench: 5139233

3 years agoMake net nn-190f102a22c3.nnue the default net.
MichaelB7 [Mon, 21 Jun 2021 12:10:35 +0000 (08:10 -0400)]
Make net nn-190f102a22c3.nnue the default net.

Trained with the pytorch trainer: https://github.com/glinscott/nnue-pytorch

python train.py i:/bin/all.binpack i:/bin/all.binpack --gpus 1 --threads 4 --num-workers 30 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^ --lambda=1.0 --max_epochs=440 --seed %random%%random% --default_root_dir exp/run_17 --resume-from-model ./pt/nn-75980ca503c6.pt

This run is thus started from the previous master net.

all.binpack equaled 4 parts Wrong_NNUE_2.binpack https://drive.google.com/file/d/1seGNOqcVdvK_vPNq98j-zV3XPE5zWAeq/view?usp=sharing plus two parts of Training_Data.binpack https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing
Each set was concatenated together - making one large Wrong_NNUE 2 binpack and one large Training so the were approximately equal in size. They were then interleaved together. The idea was to give Wrong_NNUE.binpack closer to equal weighting with the Training_Data binpack

passed LTC
https://tests.stockfishchess.org/tests/view/60d09f52b4c17000d679517f
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 32184 W: 1100 L: 970 D: 30114
Ptnml(0-2): 10, 878, 14193, 994, 17

passed STC
https://tests.stockfishchess.org/tests/view/60d086c02114332881e7368e
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 11360 W: 1056 L: 906 D: 9398
Ptnml(0-2): 25, 735, 4026, 853, 41

closes https://github.com/official-stockfish/Stockfish/pull/3576

Bench: 4631244

3 years agoFix build error on OSX
Joost VandeVondele [Mon, 21 Jun 2021 06:23:50 +0000 (08:23 +0200)]
Fix build error on OSX

directly use integer version for cp calculation.

fixes https://github.com/official-stockfish/Stockfish/issues/3573

closes https://github.com/official-stockfish/Stockfish/pull/3574

No functional change

3 years agoRemove the Contempt UCI option
Stéphane Nicolet [Wed, 16 Jun 2021 05:23:26 +0000 (07:23 +0200)]
Remove the Contempt UCI option

This patch removes the UCI option for setting Contempt in classical evaluation.

It is exactly equivalent to using Contempt=0 for the UCI contempt value and keeping
the dynamic part in the algo (renaming this dynamic part `trend` to better describe
what it does). We have tried quite hard to implement a working Contempt feature for
NNUE but nothing really worked, so it is probably time to give up.

Interested chess fans wishing to keep playing with the UCI option for Contempt and
use it with the classical eval are urged to download the version tagged "SF_Classical"
of Stockfish (dated 31 July 2020), as it was the last version where our search
algorithm was tuned for the classical eval and is probably our strongest classical
player ever: https://github.com/official-stockfish/Stockfish/tags

Passed STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 72904 W: 6228 L: 6175 D: 60501
Ptnml(0-2): 221, 5006, 25971, 5007, 247
https://tests.stockfishchess.org/tests/view/60c98bf9457376eb8bcab18d

Passed LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 45168 W: 1601 L: 1547 D: 42020
Ptnml(0-2): 38, 1331, 19786, 1397, 32
https://tests.stockfishchess.org/tests/view/60c9c7fa457376eb8bcab1bb

closes https://github.com/official-stockfish/Stockfish/pull/3575

Bench: 4947716

3 years agoKeep more pawns and pieces when attacking
Stéphane Nicolet [Sun, 20 Jun 2021 08:29:20 +0000 (10:29 +0200)]
Keep more pawns and pieces when attacking

This patch increase the weight of pawns and pieces from 28 to 32
in the scaling formula we apply to the output of the NNUE pure eval.

Increasing this gradient for pawns and pieces means that Stockfish
will try a little harder to keep material when she has the advantage,
and try a little bit harder to escape into an endgame when she is
under pressure.

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 53168 W: 4371 L: 4177 D: 44620
Ptnml(0-2): 160, 3389, 19283, 3601, 151
https://tests.stockfishchess.org/tests/view/60cefd1d457376eb8bcab7ab

LTC:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 10888 W: 386 L: 288 D: 10214
Ptnml(0-2): 3, 260, 4821, 356, 4
https://tests.stockfishchess.org/tests/view/60cf709d2114332881e7352b

closes https://github.com/official-stockfish/Stockfish/pull/3571

Bench: 4965430

3 years agoMake net nn-75980ca503c6.nnue the default.
MichaelB7 [Sat, 19 Jun 2021 13:57:09 +0000 (09:57 -0400)]
Make net nn-75980ca503c6.nnue the default.

trained with the Python command

c:\nnue>python train.py i:/bin/all.binpack i:/bin/all.binpack --gpus 1 --threads 4 --num-workers 30 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^ --lambda=1.0 --max_epochs=440 --seed %random%%random% --default_root_dir exp/run_10 --resume-from-model ./pt/nn-3b20abec10c1.pt
`
all.binpack equaled 4 parts Wrong_NNUE_2.binpack https://drive.google.com/file/d/1seGNOqcVdvK_vPNq98j-zV3XPE5zWAeq/view?usp=sharing plus two parts of Training_Data.binpack https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing
Each set was concatenated together - making one large Wrong_NNUE 2 binpack and one large Training so the were approximately equal in size. They were then interleaved together. The idea was to give Wrong_NNUE.binpack closer to equal weighting with the Training_Data binpack .

Net nn-3b20abec10c1.nnue was chosen as the --resume-from-model with the idea that through learning, the manually hex edited values will be learned and will not need to be manually adjusted going forward. They would also be fine tuned by the learning process.

passed STC:
https://tests.stockfishchess.org/tests/view/60cdf91e457376eb8bcab66f
LLR: 2.95 (-2.94,2.94) <-0.50,2.50>
Total: 18256 W: 1639 L: 1479 D: 15138
Ptnml(0-2): 59, 1179, 6505, 1313, 72

passed LTC:
https://tests.stockfishchess.org/tests/view/60ce2166457376eb8bcab6e1
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 18792 W: 654 L: 542 D: 17596
Ptnml(0-2): 9, 490, 8291, 592, 14

closes https://github.com/official-stockfish/Stockfish/pull/3570

Bench: 5020972

3 years agoChange trace with NNUE eval support
Tomasz Sobczyk [Thu, 17 Jun 2021 10:36:06 +0000 (12:36 +0200)]
Change trace with NNUE eval support

This patch adds some more output to the `eval` command. It adds a board display
with estimated piece values (method is remove-piece, evaluate, put-piece), and
splits the NNUE evaluation with (psqt,layers) for each bucket for the NNUE net.

Example:

```

./stockfish
position fen 3Qb1k1/1r2ppb1/pN1n2q1/Pp1Pp1Pr/4P2p/4BP2/4B1R1/1R5K b - - 11 40
eval

 Contributing terms for the classical eval:
+------------+-------------+-------------+-------------+
|    Term    |    White    |    Black    |    Total    |
|            |   MG    EG  |   MG    EG  |   MG    EG  |
+------------+-------------+-------------+-------------+
|   Material |  ----  ---- |  ----  ---- | -0.73 -1.55 |
|  Imbalance |  ----  ---- |  ----  ---- | -0.21 -0.17 |
|      Pawns |  0.35 -0.00 |  0.19 -0.26 |  0.16  0.25 |
|    Knights |  0.04 -0.08 |  0.12 -0.01 | -0.08 -0.07 |
|    Bishops | -0.34 -0.87 | -0.17 -0.61 | -0.17 -0.26 |
|      Rooks |  0.12  0.00 |  0.08  0.00 |  0.04  0.00 |
|     Queens |  0.00  0.00 | -0.27 -0.07 |  0.27  0.07 |
|   Mobility |  0.84  1.76 |  0.01  0.66 |  0.83  1.10 |
|King safety | -0.99 -0.17 | -0.72 -0.10 | -0.27 -0.07 |
|    Threats |  0.27  0.27 |  0.73  0.86 | -0.46 -0.59 |
|     Passed |  0.00  0.00 |  0.79  0.82 | -0.79 -0.82 |
|      Space |  0.61  0.00 |  0.24  0.00 |  0.37  0.00 |
|   Winnable |  ----  ---- |  ----  ---- |  0.00 -0.03 |
+------------+-------------+-------------+-------------+
|      Total |  ----  ---- |  ----  ---- | -1.03 -2.14 |
+------------+-------------+-------------+-------------+

 NNUE derived piece values:
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |   Q   |   b   |       |   k   |       |
|       |       |       | +12.4 | -1.62 |       |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |   r   |       |       |   p   |   p   |   b   |       |
|       | -3.89 |       |       | -0.84 | -1.19 | -3.32 |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   p   |   N   |       |   n   |       |       |   q   |       |
| -1.81 | +3.71 |       | -4.82 |       |       | -5.04 |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|   P   |   p   |       |   P   |   p   |       |   P   |   r   |
| +1.16 | -0.91 |       | +0.55 | +0.12 |       | +0.50 | -4.02 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |   P   |       |       |   p   |
|       |       |       |       | +2.33 |       |       | +1.17 |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |   B   |   P   |       |       |
|       |       |       |       | +4.79 | +1.54 |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |       |       |       |   B   |       |   R   |       |
|       |       |       |       | +4.54 |       | +6.03 |       |
+-------+-------+-------+-------+-------+-------+-------+-------+
|       |   R   |       |       |       |       |       |   K   |
|       | +4.81 |       |       |       |       |       |       |
+-------+-------+-------+-------+-------+-------+-------+-------+

 NNUE network contributions (Black to move)
+------------+------------+------------+------------+
|   Bucket   |  Material  | Positional |   Total    |
|            |   (PSQT)   |  (Layers)  |            |
+------------+------------+------------+------------+
|  0         |  +  0.32   |  -  1.46   |  -  1.13   |
|  1         |  +  0.25   |  -  0.68   |  -  0.43   |
|  2         |  +  0.46   |  -  1.72   |  -  1.25   |
|  3         |  +  0.55   |  -  1.80   |  -  1.25   |
|  4         |  +  0.48   |  -  1.77   |  -  1.29   |
|  5         |  +  0.40   |  -  2.00   |  -  1.60   |
|  6         |  +  0.57   |  -  2.12   |  -  1.54   | <-- this bucket is used
|  7         |  +  3.38   |  -  2.00   |  +  1.37   |
+------------+------------+------------+------------+

Classical evaluation   -1.00 (white side)
NNUE evaluation        +1.54 (white side)
Final evaluation       +2.38 (white side) [with scaled NNUE, hybrid, ...]

```

Also renames the export_net() function to save_eval() while there.

closes https://github.com/official-stockfish/Stockfish/pull/3562

No functional change

3 years agoFix for Cygwin's environment build-profile (fixed)
proukornew [Fri, 18 Jun 2021 21:52:46 +0000 (00:52 +0300)]
Fix for Cygwin's environment build-profile (fixed)

The Cygwin environment has two g++ compilers, each with a different problem
for compiling  Stockfish at the moment:

(a) g++.exe : full posix build compiler, linked to cygwin dll.

    => This one has a problem embedding the net.

(b) x86_64-w64-mingw32-g++.exe : native Windows build compiler.

    => This one manages to embed the net, but has a problem related to libgcov
       when we use the profile-build target of Stockfish.

This patch solves the problem for compiler (b), so that our recommended command line
if you want to build an optimized version of Stockfish on Cygwin becomes something
like the following (you can change the ARCH value to whatever you want, but note
the COMP and CXX variables pointing at the right compiler):

```
   make -j profile-build ARCH=x86-64-modern COMP=mingw CXX=x86_64-w64-mingw32-c++.exe
```

closes https://github.com/official-stockfish/Stockfish/pull/3569

No functional change

3 years agoMake net nn-50144f835024.nnue the default
Joost VandeVondele [Fri, 18 Jun 2021 21:50:01 +0000 (23:50 +0200)]
Make net nn-50144f835024.nnue the default

trained with the Python command

c:\nnue>python train.py i:/bin/all.binpack i:/bin/all.binpack --gpus 1 --threads 4 --num-workers 30 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^ --lambda=1.0 --max_epochs=440 --seed %random%%random% --default_root_dir exp/run_8 --resume-from-model ./pt/nn-6ad41a9207d0.pt
`
all.binpack equaled 4 parts Wrong_NNUE_2.binpack https://drive.google.com/file/d/1seGNOqcVdvK_vPNq98j-zV3XPE5zWAeq/view?usp=sharing plus two parts of Training_Data.binpack https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing
Each set was concatenated together - make one large Wrong_NNUE 2 binpack and one large Training_Data of approximate size. They were then interleaved together. The idea was to give Wrong_NNUE.binpack closer to equal weighting with the Training _Data binpack .

nn-6ad41a9207d0.pt was derived from a net vondele ran which passed STC quickly,
but faltered in LTC. https://tests.stockfishchess.org/tests/view/60cba666457376eb8bcab443

STC:
LLR: 2.95 (-2.94,2.94) <-0.50,2.50>
Total: 18792 W: 2068 L: 1889 D: 14835
Ptnml(0-2): 82, 1480, 6117, 1611, 106
https://tests.stockfishchess.org/tests/view/60ccda8b457376eb8bcab568

LTC:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 11376 W: 574 L: 454 D: 10348
Ptnml(0-2): 4, 412, 4747, 510, 15
https://tests.stockfishchess.org/tests/view/60ccf952457376eb8bcab58d

closes https://github.com/official-stockfish/Stockfish/pull/3568

Bench: 4900906

3 years agoAdd basic github workflow
Tomasz Sobczyk [Fri, 18 Jun 2021 10:03:03 +0000 (12:03 +0200)]
Add basic github workflow

move to github actions to replace travis CI.

First version, testing on linux using gcc and clang.
gcc build with sanitizers and valgrind.

No functional change

3 years agoUpdate default net to nn-aa9d7eeb397e.nnue
SFisGOD [Fri, 18 Jun 2021 19:09:20 +0000 (03:09 +0800)]
Update default net to nn-aa9d7eeb397e.nnue

Optimization of vondele's nn-33c9d39e5eb6.nnue using SPSA
https://tests.stockfishchess.org/tests/view/60ca68be457376eb8bcab28b
Setting: ck values are default based on how large the parameters are
The new values for this net are the raw values at the end of the tuning (80k games)

The significant changes are in buckets 1 and 2 (5-12 pieces) so the main difference is in playing endgames if we compare it to nn-33c9. There is also change in bucket 7 (29-32 pieces) but not as substantial as the changes in buckets 1 and 2. If we interpret the changes based on an experiment a few months ago, this new net plays more optimistically during endgames and less optimistically during openings.

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 49504 W: 4246 L: 4053 D: 41205
Ptnml(0-2): 140, 3282, 17749, 3407, 174
https://tests.stockfishchess.org/tests/view/60cbd752457376eb8bcab478

LTC:
LLR: 2.95 (-2.94,2.94) <0.50,3.50>
Total: 88720 W: 4926 L: 4651 D: 79143
Ptnml(0-2): 105, 4048, 35793, 4295, 119
https://tests.stockfishchess.org/tests/view/60cc7828457376eb8bcab4fa

closes https://github.com/official-stockfish/Stockfish/pull/3566

Bench: 4758885

3 years agoNew default net nn-3b20abec10c1.nnue
ap [Thu, 17 Jun 2021 23:43:58 +0000 (01:43 +0200)]
New default net nn-3b20abec10c1.nnue

This net was created by @pleomati, who manually edited with an hex editor
10 values randomly chosen in the LCSFNet10 net (nn-6ad41a9207d0.nnue) to
create this one. The LCSFNet10 net was trained by Joost VandeVondele from
a dataset combining Stockfish games and Leela games (16x10^9 positions from
SF self-play at depth 9, and 6.3x10^9 positions from Leela games, so overall
72% of Stockfish positions and 28% of Leela positions).

passed STC 10+0.1:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 50888 W: 5881 L: 5654 D: 39353
Ptnml(0-2): 281, 4290, 16085, 4497, 291
https://tests.stockfishchess.org/tests/view/60cbfa68457376eb8bcab49a

passed LTC 60+0.6:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 25480 W: 1498 L: 1338 D: 22644
Ptnml(0-2): 36, 1155, 10193, 1325, 31
https://tests.stockfishchess.org/tests/view/60cc4af8457376eb8bcab4d4

closes https://github.com/official-stockfish/Stockfish/pull/3564

Bench: 4904930

3 years agoRevert "Fix for Cygwin's environment build-profile"
Stéphane Nicolet [Thu, 17 Jun 2021 16:09:42 +0000 (18:09 +0200)]
Revert "Fix for Cygwin's environment build-profile"

This reverts commit "Fix for Cygwin's environment build-profile", as it was
giving errors for "make clean" on some Windows environments. See comments in
https://github.com/official-stockfish/Stockfish/commit/68bf362ea2385a641be9f5ed9ce2acdf55a1ecf1

Possibly somebody can propose a solution that would fix Cygwin builds and
not break on other system too, stay tuned! :-)

No functional change

3 years agoSimplify reduction when best move doesn't change frequently.
bmc4 [Tue, 15 Jun 2021 23:56:09 +0000 (20:56 -0300)]
Simplify reduction when best move doesn't change frequently.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 40400 W: 3468 L: 3377 D: 33555
Ptnml(0-2): 134, 2734, 14388, 2795, 149
https://tests.stockfishchess.org/tests/view/60c93e5a457376eb8bcab15f

LTC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 34200 W: 1190 L: 1128 D: 31882
Ptnml(0-2): 22, 998, 15001, 1054, 25
https://tests.stockfishchess.org/tests/view/60c96a1a457376eb8bcab180

closes https://github.com/official-stockfish/Stockfish/pull/3559

bench: 5629669

3 years agoFix for Cygwin's environment build-profile
proukornew [Thu, 13 May 2021 21:49:28 +0000 (00:49 +0300)]
Fix for Cygwin's environment build-profile

The Cygwin environment has two g++ compilers, each with a different problem
for compiling  Stockfish at the moment:

(a) g++.exe : full posix build compiler, linked to cygwin dll.

    => This one has a problem embedding the net.

(b) x86_64-w64-mingw32-g++.exe : native Windows build compiler.

    => This one manages to embed the net, but has a problem related to libgcov
       when we use the profile-build target of Stockfish.

This patch solves the problem for compiler (b), so that our recommended command line
if you want to build an optimized version of Stockfish on Cygwin becomes something
like the following (you can change the ARCH value to whatever you want, but note
the COMP and CXX variables pointing at the right compiler):

```
   make -j profile-build ARCH=x86-64-modern COMP=mingw CXX=x86_64-w64-mingw32-c++.exe
```

closes https://github.com/official-stockfish/Stockfish/pull/3463

No functional change

3 years agoNew default net nn-33c9d39e5eb6.nnue
Joost VandeVondele [Tue, 15 Jun 2021 10:49:23 +0000 (12:49 +0200)]
New default net nn-33c9d39e5eb6.nnue

As the previous net, this net is trained on Leela games as provided by borg.
See also https://lczero.org/blog/2021/06/the-importance-of-open-data/

The particular data set, which is a mix of T60 and T74 data, is now available as a single binpack:
https://drive.google.com/file/d/1RFkQES3DpsiJqsOtUshENtzPfFgUmEff/view?usp=sharing

The training command was:
python train.py ../../training_data_pylon.binpack ../../training_data_pylon.binpack --gpus 1 --threads 2 --num-workers 2 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 10 --features=HalfKAv2^   --lambda=1.0  --max_epochs=440 --seed $RANDOM --default_root_dir exp/run_2

passed STC:
https://tests.stockfishchess.org/tests/view/60c887cb457376eb8bcab054
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 12792 W: 1483 L: 1311 D: 9998
Ptnml(0-2): 62, 989, 4131, 1143, 71

passed LTC:
https://tests.stockfishchess.org/tests/view/60c8e5c4457376eb8bcab0f0
LLR: 2.95 (-2.94,2.94) <0.50,3.50>
Total: 11272 W: 601 L: 477 D: 10194
Ptnml(0-2): 9, 421, 4657, 535, 14

also had strong LTC performance against another strong net of the series:
https://tests.stockfishchess.org/tests/view/60c8c40d457376eb8bcab0c6

closes https://github.com/official-stockfish/Stockfish/pull/3557

Bench: 5032320

3 years agoFix a rare case of wrong TB ranking
J. Oster [Mon, 14 Jun 2021 15:28:30 +0000 (17:28 +0200)]
Fix a rare case of wrong TB ranking
of a root move leading to a 3-fold repetition.
With this small fix a draw ranking and thus a draw score is being applied.
This works for both, ranking by dtz or wdl tables.

Fixes https://github.com/official-stockfish/Stockfish/issues/3542

(No functional change without TBs.)
Bench: 4877339

3 years agoReduce the number of accumulator states
Tomasz Sobczyk [Sat, 12 Jun 2021 18:45:14 +0000 (20:45 +0200)]
Reduce the number of accumulator states

Reduce from 3 to 2. Make the intent of the states clearer.

STC: https://tests.stockfishchess.org/tests/view/60c50111457376eb8bcaad03
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 61888 W: 5007 L: 4944 D: 51937
Ptnml(0-2): 164, 3947, 22649, 4030, 154

LTC: https://tests.stockfishchess.org/tests/view/60c52b1c457376eb8bcaad2c
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 20248 W: 688 L: 618 D: 18942
Ptnml(0-2): 7, 551, 8946, 605, 15

closes https://github.com/official-stockfish/Stockfish/pull/3548

No functional change.

3 years agoUpdate default net to nn-8e47cf062333.nnue
JWmer [Sun, 13 Jun 2021 21:48:32 +0000 (23:48 +0200)]
Update default net to nn-8e47cf062333.nnue

This net is the result of training on data used by the Leela project. More precisely,
we shuffled T60 and T74 data kindly provided by borg (for different Tnn, the data is
a result of Leela selfplay with differently sized Leela nets).

The data is available at vondele's google drive:
https://drive.google.com/drive/folders/1mftuzYdl9o6tBaceR3d_VBQIrgKJsFpl.

The Leela data comes in small chunks of .binpack files. To shuffle them, we simply
used a small python script to randomly rename the files, and then concatenated them
using `cat`. As validation data we picked a file of T60 data. We will further investigate
T74 data.

The training for the NNUE architecture used 200 epochs with the Python trainer from
the Stockfish project. Unlike the previous run we tried with this data, this run does
not have adjusted scaling — not because we didn't want to, but because we forgot.
However, this training randomly skips 40% more positions than previous run. The loss
was very spiky and decreased slower than it does usually.

Training loss: https://github.com/official-stockfish/images/blob/main/training-loss-8e47cf062333.png
Validation loss: https://github.com/official-stockfish/images/blob/main/validation-loss-8e47cf062333.png

This is the exact training command:
python train.py --smart-fen-skipping --random-fen-skipping 14 --batch-size 16384 --threads 4 --num-workers 4 --gpus 1 trainingdata\training_data.binpack validationdata\val.binpack

---

10k STC result:
ELO: 3.61 +-3.3 (95%) LOS: 98.4%
Total: 10000 W: 1241 L: 1137 D: 7622
Ptnml(0-2): 68, 841, 3086, 929, 76
https://tests.stockfishchess.org/tests/view/60c67e50457376eb8bcaae70

10k LTC result:
ELO: 2.71 +-2.4 (95%) LOS: 98.8%
Total: 10000 W: 659 L: 581 D: 8760
Ptnml(0-2): 22, 485, 3900, 579, 14
https://tests.stockfishchess.org/tests/view/60c69deb457376eb8bcaae98

Passed LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 9648 W: 685 L: 545 D: 8418
Ptnml(0-2): 22, 448, 3740, 596, 18
https://tests.stockfishchess.org/tests/view/60c6d41c457376eb8bcaaecf

---

closes https://github.com/official-stockfish/Stockfish/pull/3550

Bench: 4877339

3 years agoRegister count for feature transformer
Tomasz Sobczyk [Thu, 10 Jun 2021 15:43:42 +0000 (17:43 +0200)]
Register count for feature transformer

Compute optimal register count for feature transformer accumulation dynamically.
This also introduces a change where AVX512 would only use 8 registers instead of 16
(now possible due to a 2x increase in feature transformer size).

closes https://github.com/official-stockfish/Stockfish/pull/3543

No functional change

3 years agoDo less LMR extensions
Vizvezdenec [Sat, 29 May 2021 03:39:14 +0000 (06:39 +0300)]
Do less LMR extensions

This patch restricts LMR extensions (of non-transposition table moves) from being
used when the transposition table move was extended by two plies via singular
extension. This may serve to limit search explosions in certain positions.

This makes a lot of sense because the precondition for the tt-move to have been
singular extended by two plies is that the result of the alternate search (with
excluded the tt-move) has been a hard fail low: it is natural to later search less
for non tt-moves in this situation.

The current state of depth/extensions/reductions management is getting quite tricky
in our search algo, see https://github.com/official-stockfish/Stockfish/pull/3546#issuecomment-860174549
for some discussion. Suggestions welcome!

Passed STC
https://tests.stockfishchess.org/tests/view/60c3f293457376eb8bcaac8d
LLR: 2.95 (-2.94,2.94) <-0.50,2.50>
Total: 117984 W: 9698 L: 9430 D: 98856
Ptnml(0-2): 315, 7708, 42703, 7926, 340

passed LTC
https://tests.stockfishchess.org/tests/view/60c46ea5457376eb8bcaacc7
LLR: 2.97 (-2.94,2.94) <0.50,3.50>
Total: 11280 W: 401 L: 302 D: 10577
Ptnml(0-2): 2, 271, 4998, 364, 5

closes https://github.com/official-stockfish/Stockfish/pull/3546

Bench: 4709974

3 years agoClarify use of UCI options
Stéphane Nicolet [Sun, 13 Jun 2021 07:59:34 +0000 (09:59 +0200)]
Clarify use of UCI options

Update README.md to clarify use of UCI options

closes https://github.com/official-stockfish/Stockfish/pull/3540

No functional change

3 years agoRead NNUE net faster
Tomasz Sobczyk [Wed, 9 Jun 2021 09:21:55 +0000 (11:21 +0200)]
Read NNUE net faster

Load feature transformer weights in bulk on little-endian machines.
This is in particular useful to test new nets with c-chess-cli,
see https://github.com/lucasart/c-chess-cli/issues/44

```
$ time ./stockfish.exe uci

Before : 0m0.914s
After  : 0m0.483s
```

No functional change

3 years agoLimit double extensions
Joost VandeVondele [Wed, 9 Jun 2021 21:23:13 +0000 (23:23 +0200)]
Limit double extensions

Double extensions can lead to search explosions, for specific positions.
Currently, however, these double extensions are worth about 10Elo and cannot
be removed. This patch instead limits the number of double extensions given
to a maximum of 3.

This fixes https://github.com/official-stockfish/Stockfish/issues/3532
where the following testcase was shown to be problematic:

```
uci
setoption name Hash value 4
setoption name Contempt value 0
ucinewgame
position fen 8/Pk6/8/1p6/8/P1K5/8/6B1 w - - 37 130
go depth 20
```

passed STC:
https://tests.stockfishchess.org/tests/view/60c13161457376eb8bcaaa0f
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 73256 W: 6114 L: 6062 D: 61080
Ptnml(0-2): 222, 4912, 26306, 4968, 220

passed LTC:
https://tests.stockfishchess.org/tests/view/60c196fb457376eb8bcaaa6b
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 166440 W: 5559 L: 5594 D: 155287
Ptnml(0-2): 106, 4921, 73197, 4894, 102

closes https://github.com/official-stockfish/Stockfish/pull/3544

Bench: 5067605

3 years agoSimplify promotion move generator
bmc4 [Mon, 7 Jun 2021 18:47:37 +0000 (15:47 -0300)]
Simplify promotion move generator

This patch removes Knight promotion checks from Captures. As a consequence,
it also removes this underpromotion from qsearch.

STC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 37776 W: 3113 L: 3023 D: 31640
Ptnml(0-2): 103, 2419, 13755, 2507, 104
https://tests.stockfishchess.org/tests/view/60be6a06457376eb8bcaa775

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 39760 W: 1257 L: 1203 D: 37300
Ptnml(0-2): 11, 1079, 17646, 1133, 11
https://tests.stockfishchess.org/tests/view/60beb972457376eb8bcaa7c5

closes https://github.com/official-stockfish/Stockfish/pull/3536

Bench: 5530620

3 years agoReduce in LMR reduction on PvNode
bmc4 [Sun, 6 Jun 2021 16:31:57 +0000 (13:31 -0300)]
Reduce in LMR reduction on PvNode

reduce reduction in LMR by 1 on PvNode.

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 266080 W: 22438 L: 21996 D: 221646
Ptnml(0-2): 774, 17874, 95376, 18168, 848
https://tests.stockfishchess.org/tests/view/60bc0661457376eb8bcaa4bb

LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 20144 W: 698 L: 587 D: 18859
Ptnml(0-2): 2, 529, 8906, 626, 9
https://tests.stockfishchess.org/tests/view/60bcc3f2457376eb8bcaa58d

closes https://github.com/official-stockfish/Stockfish/pull/3534

bench: 5173012

3 years agoMakefile: Extend sanitize support
Guy Vreuls [Thu, 3 Jun 2021 14:46:05 +0000 (16:46 +0200)]
Makefile: Extend sanitize support

Enable compiling with multiple sanitizers at once.

Syntax:
make build ARCH=x86-64-avx512 debug=on sanitize="address undefined"

closes https://github.com/official-stockfish/Stockfish/pull/3524

No functional change.

3 years agoEnhance CI to error on leaks
Joost VandeVondele [Thu, 3 Jun 2021 17:18:24 +0000 (19:18 +0200)]
Enhance CI to error on leaks

Add flags to valgrind in our Continuous Integration scripts,
to error on memory leaks.

closes https://github.com/official-stockfish/Stockfish/pull/3525

No functional change.

3 years agoRevert "Simplify En Passant"
Guy Vreuls [Fri, 4 Jun 2021 10:20:27 +0000 (12:20 +0200)]
Revert "Simplify En Passant"

This reverts commit 9f8058bd26df1c3ca37b85f811026f1eb82e6524.

Fixes the memory leak discussed in pull request #3523
https://github.com/official-stockfish/Stockfish/pull/3523

Passed non-regression STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 76184 W: 6330 L: 6282 D: 63572
Ptnml(0-2): 202, 5047, 27564, 5059, 220
https://tests.stockfishchess.org/tests/view/60ba146c457376eb8bcaa2e2

closes https://github.com/official-stockfish/Stockfish/pull/3527

Benched to verify there is no functional change.

Bench: 4364128

3 years agoClean SIMD code a bit
Stéphane Nicolet [Fri, 4 Jun 2021 11:56:40 +0000 (13:56 +0200)]
Clean SIMD code a bit

Cleaner vector code structure in feature transformer. This patch just
regroups the parts of the inner loop for each SIMD instruction set.

Tested for non-regression:
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 115760 W: 9835 L: 9831 D: 96094
Ptnml(0-2): 326, 7776, 41715, 7694, 369
https://tests.stockfishchess.org/tests/view/60b96b39457376eb8bcaa26e

It would be nice if a future patch could use some of the macros at
the top of the file to unify the code between the distincts SIMD
instruction sets (of course, unifying the Relu will be the challenge).

closes https://github.com/official-stockfish/Stockfish/pull/3506

No functional change

3 years agoMakefile: better "make clean" for Windows
Stéphane Nicolet [Wed, 2 Jun 2021 06:10:54 +0000 (08:10 +0200)]
Makefile: better "make clean" for Windows

Make clean should be really clean on Windows.

Fixes issue https://github.com/official-stockfish/Stockfish/issues/3291
Closes https://github.com/official-stockfish/Stockfish/pull/3517

No functional change

3 years agoIntroducing NodeType Root
bmc4 [Thu, 3 Jun 2021 16:52:39 +0000 (13:52 -0300)]
Introducing NodeType Root

We transform rootNode into constexpr by adding a new NodeType `Root`,
which causes a speed up.

Local test:
```
Build Tester: 1.4.7.0
Windows 10 (Version 10.0, Build 0, 64-bit Edition)
Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz
SafeMode: No
Running In VM: No
HyperThreading Enabled: Yes
CPU Warmup: Yes
Command Line: bench
Tests per Build: 25
ANOVA: n/a

                Engine# (NPS)                     Speedup     Sp     Conf. 95%    S.S.
patch  (920.179,4) ---> master  (906.329,2)  --->  1,528%  20.336,5     Yes        No
```

---------

STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 98216 W: 8348 L: 8102 D: 81766
Ptnml(0-2): 295, 6357, 35549, 6621, 286
https://tests.stockfishchess.org/tests/view/60b797e2457376eb8bcaa0ab

Yellow LTC:
LLR: -2.95 (-2.94,2.94) <0.50,3.50>
Total: 76936 W: 2651 L: 2626 D: 71659
Ptnml(0-2): 29, 2233, 33916, 2264, 26
https://tests.stockfishchess.org/tests/view/60b80d6d457376eb8bcaa145

closes https://github.com/official-stockfish/Stockfish/pull/3522

No functional change

3 years agoMake extra time for bestMoveInstability dependent on rootdepth.
xoto10 [Thu, 3 Jun 2021 18:46:55 +0000 (19:46 +0100)]
Make extra time for bestMoveInstability dependent on rootdepth.

This change allocates more base time to moves and makes the additional time added for best move instability dependent on rootdepth.

STC 10+0.1 :
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 19432 W: 1711 L: 1553 D: 16168
Ptnml(0-2): 47, 1250, 6989, 1358, 72
https://tests.stockfishchess.org/tests/view/60b8cd41457376eb8bcaa1ad

LTC 60+0.6 :
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 22480 W: 810 L: 693 D: 20977
Ptnml(0-2): 9, 603, 9902, 714, 12
https://tests.stockfishchess.org/tests/view/60b8e5bf457376eb8bcaa1e6

closes https://github.com/official-stockfish/Stockfish/pull/3526

Bench 4364128

3 years agoUpdate default net to nn-7e66505906a6.nnue
Joost VandeVondele [Thu, 3 Jun 2021 11:53:28 +0000 (13:53 +0200)]
Update default net to nn-7e66505906a6.nnue

Trained with pytorch using the master branch and recommended settings,
the data used is the previous 64B binpack enhanced with a 2B binpack
generated using an opening book of positions for with the static eval
is significantly different from d9 search.

book           : https://drive.google.com/file/d/1rHcKY5rv34kwku6g89OhnE8Bkfq3UWau/view?usp=sharing
book generation: https://github.com/vondele/Stockfish/commit/3ce43ab0c4ce09c1fc5bca5ca27a248e67fddd24
binpack        : https://drive.google.com/file/d/1rHcKY5rv34kwku6g89OhnE8Bkfq3UWau/view?usp=sharing

-------

Data generation command:

generate_training_data depth 9 count 31250000 random_multi_pv 2 random_multi_pv_diff 100 random_move_max_ply 8 random_move_count 3 set_recommended_uci_options eval_limit 32000 output_file_name output.binpack book wrongNNUE.epd seed ${RANDOM}${RANDOM}

Training command:

python train.py ../../all_d9_fishd9_d8_d10_wrong_shuffle.binpack ../../all_d9_fishd9_d8_d10_wrong_shuffle.binpack  --gpus 1 --threads 2 --num-workers 2 --batch-size 16384 --progress_bar_refresh_rate 300 --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^   --lambda=1.0  --max_epochs=400 --seed $RANDOM --default_root_dir exp/run_5

-------

passed STC:
https://tests.stockfishchess.org/tests/view/60b7c79a457376eb8bcaa104
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 64592 W: 6254 L: 6028 D: 52310
Ptnml(0-2): 255, 4785, 22020, 4951, 285

passed LTC:
https://tests.stockfishchess.org/tests/view/60b85307457376eb8bcaa182
LLR: 2.96 (-2.94,2.94) <0.50,3.50>
Total: 45560 W: 1998 L: 1826 D: 41736
Ptnml(0-2): 36, 1604, 19335, 1762, 43

closes https://github.com/official-stockfish/Stockfish/pull/3521

Bench: 4364128

3 years agoTypography change for bench
Stéphane Nicolet [Wed, 2 Jun 2021 06:37:00 +0000 (08:37 +0200)]
Typography change for bench

3 years agoRemove formerPV variable.
Stefan Geschwentner [Mon, 31 May 2021 18:26:05 +0000 (20:26 +0200)]
Remove formerPV variable.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 75672 W: 6546 L: 6496 D: 62630
Ptnml(0-2): 238, 5274, 26761, 5326, 237
https://tests.stockfishchess.org/tests/view/60b349c0ec0c03148cbed055

LTC:
LLR: 2.98 (-2.94,2.94) <-2.50,0.50>
Total: 137816 W: 4676 L: 4689 D: 128451
Ptnml(0-2): 52, 4237, 60354, 4202, 63
https://tests.stockfishchess.org/tests/view/60b38970ec0c03148cbed075

closes https://github.com/official-stockfish/Stockfish/pull/3515

Bench: 4892288

3 years agoPre-initialize ss->ply
J. Oster [Mon, 31 May 2021 15:46:40 +0000 (17:46 +0200)]
Pre-initialize ss->ply

We pre-initialize ss->ply over the whole stack. There is no need
to re-assign the same value(s) over and over again while searching.
Probably a tiny speedup on longer searches.

Tested for no regression:

STC
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 25784 W: 2205 L: 2101 D: 21478
Ptnml(0-2): 62, 1660, 9368, 1716, 86
https://tests.stockfishchess.org/tests/view/60b516c6457376eb8bca9dfa

LTC
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 26200 W: 944 L: 878 D: 24378
Ptnml(0-2): 12, 732, 11545, 800, 11
https://tests.stockfishchess.org/tests/view/60b53652457376eb8bca9e0e

closes https://github.com/official-stockfish/Stockfish/pull/3516

No functional change.

3 years agoCheck Extension with Static Evaluation
candirufish [Mon, 31 May 2021 04:47:35 +0000 (06:47 +0200)]
Check Extension with Static Evaluation

extension for checking moves, at higher depth and more decisive positions.

stc:
LLR: 2.97 (-2.94,2.94) <-0.50,2.50>
Total: 87008 W: 7337 L: 7100 D: 72571
Ptnml(0-2): 264, 5737, 31270, 5964, 269
https://tests.stockfishchess.org/tests/view/60b1034787a1a67ae56c47b6

ltc:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 79320 W: 2629 L: 2432 D: 74259
Ptnml(0-2): 29, 2205, 35000, 2392, 34
https://tests.stockfishchess.org/tests/view/60b1ae0b87a1a67ae56c487c

closes https://github.com/official-stockfish/Stockfish/pull/3514

Bench: 4447112

3 years agoFix export of the feature transformer.
Tomasz Sobczyk [Sat, 29 May 2021 09:40:40 +0000 (11:40 +0200)]
Fix export of the feature transformer.

PSQT export was missing.

fixes #3507

closes https://github.com/official-stockfish/Stockfish/pull/3508

No functional change

3 years agoSimplify NNUE / classical evaluation selection
Joost VandeVondele [Sat, 29 May 2021 13:44:57 +0000 (15:44 +0200)]
Simplify NNUE / classical evaluation selection

for the new network architecture these rules can be simplified,
closer to the original PSQT difference based again.

passed STC
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 22656 W: 1979 L: 1868 D: 18809
Ptnml(0-2): 70, 1496, 8087, 1603, 72
https://tests.stockfishchess.org/tests/view/60b24579db3c4776cb89d122

passed LTC
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 30224 W: 1015 L: 953 D: 28256
Ptnml(0-2): 4, 860, 13330, 906, 12
https://tests.stockfishchess.org/tests/view/60b27613db3c4776cb89d145

closes https://github.com/official-stockfish/Stockfish/pull/3511

Bench: 3937626

3 years agoRemove Stat Reset at beta cutoff
VoyagerOne [Thu, 27 May 2021 20:03:56 +0000 (16:03 -0400)]
Remove Stat Reset at beta cutoff

STC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 63936 W: 5350 L: 5288 D: 53298
Ptnml(0-2): 184, 4295, 22954, 4345, 190
https://tests.stockfishchess.org/tests/view/60affb4c12066fd299795c64

LTC:
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 35856 W: 1201 L: 1142 D: 33513
Ptnml(0-2): 7, 1031, 15795, 1086, 9
https://tests.stockfishchess.org/tests/view/60b0537812066fd299795cc6

closes https://github.com/official-stockfish/Stockfish/pull/3505

bench: 3831936

3 years agoDo not use lazy evaluation inside NNUE
Stéphane Nicolet [Wed, 26 May 2021 23:10:00 +0000 (01:10 +0200)]
Do not use lazy evaluation inside NNUE

This simplification patch implements two changes:

1. it simplifies away the so-called "lazy" path in the NNUE evaluation internals,
   where we trusted the psqt head alone to avoid the costly "positional" head in
   some cases;
2. it raises a little bit the NNUEThreshold1 in evaluate.cpp (from 682 to 800),
   which increases the limit where we switched from NNUE eval to Classical eval.

Both effects increase the number of positional evaluations done by our new net
architecture, but the results of our tests below seem to indicate that the loss
of speed will be compensated by the gain of eval quality.

STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 26280 W: 2244 L: 2137 D: 21899
Ptnml(0-2): 72, 1755, 9405, 1810, 98
https://tests.stockfishchess.org/tests/view/60ae73f112066fd299795a51

LTC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 20592 W: 750 L: 677 D: 19165
Ptnml(0-2): 9, 614, 8980, 681, 12
https://tests.stockfishchess.org/tests/view/60ae88e812066fd299795a82

closes https://github.com/official-stockfish/Stockfish/pull/3503

Bench: 3817907

3 years agoLess reduction for capture/promotions.
Stefan Geschwentner [Wed, 26 May 2021 14:17:37 +0000 (16:17 +0200)]
Less reduction for capture/promotions.

Exclude captures/promotions at expected cut nodes (which also not a
former PV node) from LMR if a response to the first previous
opponent move.

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 288656 W: 24886 L: 24413 D: 239357
Ptnml(0-2): 900, 19738, 102578, 20213, 899
https://tests.stockfishchess.org/tests/view/60ad505112066fd29979595b

LTC:
LLR: 2.97 (-2.94,2.94) <0.50,3.50>
Total: 31344 W: 1107 L: 975 D: 29262
Ptnml(0-2): 12, 879, 13757, 1013, 11
https://tests.stockfishchess.org/tests/view/60adffce12066fd2997959d2

closes https://github.com/official-stockfish/Stockfish/pull/3500

Bench: 3827710

3 years agoSimplify the thread term for reduction formula
IIvec [Thu, 20 May 2021 18:02:32 +0000 (20:02 +0200)]
Simplify the thread term for reduction formula

Dependance on Threads.size() was removed Search::init() for the Reductions[] initialization.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 17376 W: 1024 L: 929 D: 15423
Ptnml(0-2): 24, 781, 6989, 864, 30
https://tests.stockfishchess.org/tests/view/60ac110812066fd2997957dc

LTC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 145552 W: 3656 L: 3673 D: 138223
Ptnml(0-2): 37, 3351, 66014, 3340, 34
https://tests.stockfishchess.org/tests/view/60ac267412066fd299795825

closes https://github.com/official-stockfish/Stockfish/pull/3502

Bench 3864295

3 years agoExpose the lazy threshold for the feature transformer PSQT as a parameter.
Tomasz Sobczyk [Tue, 25 May 2021 11:09:40 +0000 (13:09 +0200)]
Expose the lazy threshold for the feature transformer PSQT as a parameter.

Definition of the lazy threshold moved to evaluate.cpp where all others are.
Lazy threshold only used for real searches, not used for the "eval" call.
This preserves the purity of NNUE evaluation, which is useful to verify
consistency between the engine and the NNUE trainer.

closes https://github.com/official-stockfish/Stockfish/pull/3499

No functional change

3 years agoIncreased reduction for captures in LMR
bmc4 [Mon, 24 May 2021 11:15:04 +0000 (08:15 -0300)]
Increased reduction for captures in LMR

It now does, in LMR, an increased on reduction by 1 for captures in cut nodes.

STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 30656 W: 2565 L: 2397 D: 25694
Ptnml(0-2): 63, 2012, 11029, 2142, 82
https://tests.stockfishchess.org/tests/view/60a96733ce8ea25a3ef04178

LTC:
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 124840 W: 4139 L: 3878 D: 116823
Ptnml(0-2): 48, 3480, 55100, 3747, 45
https://tests.stockfishchess.org/tests/view/60a995f5ce8ea25a3ef041b7

closes https://github.com/official-stockfish/Stockfish/pull/3494

bench: 3864295

3 years agoSometimes change the (materialist, positional) balance
Stéphane Nicolet [Sat, 22 May 2021 17:44:15 +0000 (19:44 +0200)]
Sometimes change the (materialist, positional) balance

Our new nets output two values for the side to move in the last layer.
We can interpret the first value as a material evaluation of the
position, and the second one as the dynamic, positional value of the
location of pieces.

This patch changes the balance for the (materialist, positional) parts
of the score from (128, 128) to (121, 135) when the piece material is
equal between the two players, but keeps the standard (128, 128) balance
when one player is at least an exchange up.

Passed STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 15936 W: 1421 L: 1266 D: 13249
Ptnml(0-2): 37, 1037, 5694, 1134, 66
https://tests.stockfishchess.org/tests/view/60a82df9ce8ea25a3ef0408f

Passed LTC:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 13904 W: 516 L: 410 D: 12978
Ptnml(0-2): 4, 374, 6088, 484, 2
https://tests.stockfishchess.org/tests/view/60a8bbf9ce8ea25a3ef04101

closes https://github.com/official-stockfish/Stockfish/pull/3492

Bench: 3856635

3 years agoTuning Search
bmc4 [Sat, 22 May 2021 05:47:23 +0000 (02:47 -0300)]
Tuning Search

This patch tunes constant in search.cpp

STC:
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 30648 W: 2580 L: 2410 D: 25658
Ptnml(0-2): 80, 1969, 11093, 2065, 117
https://tests.stockfishchess.org/tests/view/60a71d3cce8ea25a3ef03fae

LTC:
LLR: 2.95 (-2.94,2.94) <0.50,3.50>
Total: 52896 W: 1776 L: 1617 D: 49503
Ptnml(0-2): 13, 1462, 23347, 1605, 21
https://tests.stockfishchess.org/tests/view/60a794ddce8ea25a3ef0400a

closes https://github.com/official-stockfish/Stockfish/pull/3491

Bench: 4004731

3 years agoSimplify reduction for consecutive fails
bmc4 [Sat, 22 May 2021 05:41:52 +0000 (02:41 -0300)]
Simplify reduction for consecutive fails

Revert the heuristic introduced in #3184, by which we reduced more
the late sons of the root position after consecutive fail highs.

---
Before new net architecture:

STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 226336 W: 20373 L: 20500 D: 185463
Ptnml(0-2): 755, 16087, 79595, 15992, 739
https://tests.stockfishchess.org/tests/view/609dec205085663412d08e9d

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 67432 W: 2411 L: 2375 D: 62646
Ptnml(0-2): 33, 1944, 29714, 2004, 21
https://tests.stockfishchess.org/tests/view/609ee30f5085663412d08fc3

---
After new net architecture:

STC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 141752 W: 11591 L: 11617 D: 118544
Ptnml(0-2): 387, 9231, 51674, 9189, 395
https://tests.stockfishchess.org/tests/view/60a4320ace8ea25a3ef03cfd

LTC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 294072 W: 9825 L: 9950 D: 274297
Ptnml(0-2): 121, 8610, 129681, 8521, 103
https://tests.stockfishchess.org/tests/view/60a51b5ece8ea25a3ef03dcd
---

closes https://github.com/official-stockfish/Stockfish/pull/3490

Bench: 3752892

3 years agoUpdate default net to nn-7756374aaed3.nnue
Joost VandeVondele [Sat, 22 May 2021 05:00:53 +0000 (07:00 +0200)]
Update default net to nn-7756374aaed3.nnue

trained with pytorch using the master branch and recommended settings,
same data set as previously used:

python train.py ../../all_d9_fishd9_d8_d10_shuffle.binpack ../../all_d9_fishd9_d8_d10_shuffle.binpack \
        --gpus 1 --threads 2 --num-workers 2 --batch-size 16384 --progress_bar_refresh_rate 300 \
        --smart-fen-skipping --random-fen-skipping 3 --features=HalfKAv2^   --lambda=1.0 \
        --max_epochs=400 --seed $RANDOM --default_root_dir exp/run_8

passed STC:
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 21424 W: 2078 L: 1907 D: 17439
Ptnml(0-2): 80, 1512, 7385, 1627, 108
https://tests.stockfishchess.org/tests/view/60a6c749ce8ea25a3ef03f4d

passed LTC:
LLR: 2.94 (-2.94,2.94) <0.50,3.50>
Total: 67912 W: 2851 L: 2648 D: 62413
Ptnml(0-2): 40, 2348, 28984, 2537, 47
https://tests.stockfishchess.org/tests/view/60a722ecce8ea25a3ef03fb9

closes https://github.com/official-stockfish/Stockfish/pull/3489

Bench: 3779522

3 years agoCompact position structures
Guy Vreuls [Fri, 21 May 2021 18:22:29 +0000 (20:22 +0200)]
Compact position structures

Reorder the structures data members in position.h to reduce padding.

Passed STC:
https://tests.stockfishchess.org/tests/view/60a8011fce8ea25a3ef04069
LLR: 2.94 (-2.94,2.94) <-0.50,2.50>
Total: 14120 W: 1214 L: 1067 D: 11839
Ptnml(0-2): 26, 857, 5161, 976, 40

---

Also tested for speed locally by Joost:

Result of  50 runs
==================
base (./stockfish.master       ) =    2254919  +/- 4439
test (./stockfish.patch        ) =    2274003  +/- 5278
diff                             =     +19084  +/- 6386
==================
speedup        = +0.0085
P(speedup > 0) =  1.0000

---

closes https://github.com/official-stockfish/Stockfish/pull/3488

No functional change.

3 years agoRemove Tempo
Stéphane Nicolet [Tue, 18 May 2021 23:24:51 +0000 (01:24 +0200)]
Remove Tempo

The Tempo variable was introduced 10 years ago in our search because the
classical evaluation function was antisymmetrical in White and Black by design
to gain speed:

    Eval(White to play) = -Eval(Black to play)

Nowadays our neural networks know which side is to play in a position when
they evaluate a position and are trained on real games, so the neural network
encodes the advantage of moving as an output of search. This patch shows that
the Tempo variable is not necessary anymore.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 33512 W: 2805 L: 2709 D: 27998
Ptnml(0-2): 80, 2209, 12095, 2279, 93
https://tests.stockfishchess.org/tests/view/60a44ceace8ea25a3ef03d30

LTC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 53920 W: 1807 L: 1760 D: 50353
Ptnml(0-2): 16, 1617, 23650, 1658, 19
https://tests.stockfishchess.org/tests/view/60a477f0ce8ea25a3ef03d49

We also tried a match (20000 games) at STC using purely classical, result was neutral:
https://tests.stockfishchess.org/tests/view/60a4eebcce8ea25a3ef03db5

Note: there are two locations left in search.cpp where we assume antisymmetry
of evaluation (in relation with a speed optimization for null moves in lines
770 and 1439), but as the values are just used for heuristic pruning this
approximation should not hurt too much because the order of magnitude is still
true most of the time.

closes https://github.com/official-stockfish/Stockfish/pull/3481

Bench: 4015864

3 years agoSimplify usage of LMR for captures
Vizvezdenec [Wed, 19 May 2021 17:57:04 +0000 (20:57 +0300)]
Simplify usage of LMR for captures

This patch simplifies a lot of "enablers" for LMR when move is a capture or promotion.
After it we will have only 2 conditions - if node is a cutNode
or if it's an allNode that was not in PV,
so all captures or promotions wouldn't go thru LMR at any PVnodes.

passed STC
https://tests.stockfishchess.org/tests/view/60a40117ce8ea25a3ef03ca7
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 58976 W: 4875 L: 4807 D: 49294
Ptnml(0-2): 176, 3897, 21270, 3973, 172

passed LTC
https://tests.stockfishchess.org/tests/view/60a43ff8ce8ea25a3ef03d18
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 65272 W: 2203 L: 2165 D: 60904
Ptnml(0-2): 28, 1936, 28668, 1978, 26

closes https://github.com/official-stockfish/Stockfish/pull/3480

bench 4110764

3 years agoUse if instead of goto
Prokop Randáček [Wed, 12 May 2021 18:15:21 +0000 (20:15 +0200)]
Use if instead of goto

This PR inverts the if and removes goto in the generate_all function.

closes https://github.com/official-stockfish/Stockfish/pull/3461

No functional change

3 years agoUse packed 32-bit MMX operations for updating the PSQT accumulator
Fanael Linithien [Tue, 18 May 2021 17:17:59 +0000 (19:17 +0200)]
Use packed 32-bit MMX operations for updating the PSQT accumulator

This improves the speed of NNUE by a bit on old hardware that code path
is intended for, like a Pentium III 1.13 GHz:

10 repeats of "./stockfish bench 16 1 13 default depth NNUE":

Before:
54 642 504 897 cycles (± 0.12%)
62 301 937 829 instructions (± 0.03%)

After:
54 320 821 928 cycles (± 0.13%)
62 084 742 699 instructions (± 0.02%)

Speed of go depth 20 from startpos:

Before: 53103 nps
After: 53856 nps

closes https://github.com/official-stockfish/Stockfish/pull/3476

No functional change.

3 years agoUse Markdown syntax in the readme
Yohaan Seth Nathan [Tue, 18 May 2021 18:52:59 +0000 (00:22 +0530)]
Use Markdown syntax in the readme

provide direct links to the mentioned files.

closes https://github.com/official-stockfish/Stockfish/pull/3477

No Functional Change

3 years agoDo more continuation history based pruning
Vizvezdenec [Tue, 18 May 2021 07:02:20 +0000 (10:02 +0300)]
Do more continuation history based pruning

This patch increases lmrDepth threshold for continuation history based pruning in search.
This part of code for a long time was known to be really TC sensitive - decreasing
this threshold easily passed lower time controls but failed badly at LTC,
on the other hand it increase was part of a tuning that resulted
in being negative at STC but was +12 elo at 180+1.8.

After recent simplification of special conditions that sometimes
increase it from 4 to 5 it was logical to overall test at longer
time controls if 5 is better than 4 with deeper searches.

reduces strenght on STC
https://tests.stockfishchess.org/tests/view/60a3a8bbce8ea25a3ef03c74
ELO: -2.57 +-2.0 (95%) LOS: 0.6%
Total: 20000 W: 1820 L: 1968 D: 16212
Ptnml(0-2): 68, 1582, 6836, 1458, 56

Passed LTC with STC bounds
https://tests.stockfishchess.org/tests/view/60a027395085663412d090ce
LLR: 2.93 (-2.94,2.94) <-0.50,2.50>
Total: 175256 W: 6774 L: 6548 D: 161934
Ptnml(0-2): 91, 5808, 75604, 6034, 91

Passed VLTC with LTC bounds
https://tests.stockfishchess.org/tests/view/60a2bccce229097940a037a7
LLR: 2.96 (-2.94,2.94) <0.50,3.50>
Total: 65736 W: 1224 L: 1092 D: 63420
Ptnml(0-2): 5, 1012, 30706, 1136, 9

closes https://github.com/official-stockfish/Stockfish/pull/3473

bench 3689330

3 years agoNew NNUE architecture and net
Tomasz Sobczyk [Tue, 18 May 2021 15:36:26 +0000 (17:36 +0200)]
New NNUE architecture and net

Introduces a new NNUE network architecture and associated network parameters,
as obtained by a new pytorch trainer.

The network is already very strong at short TC, without regression at longer TC,
and has potential for further improvements.

https://tests.stockfishchess.org/tests/view/60a159c65085663412d0921d
TC: 10s+0.1s, 1 thread
ELO: 21.74 +-3.4 (95%) LOS: 100.0%
Total: 10000 W: 1559 L: 934 D: 7507
Ptnml(0-2): 38, 701, 2972, 1176, 113

https://tests.stockfishchess.org/tests/view/60a187005085663412d0925b
TC: 60s+0.6s, 1 thread
ELO: 5.85 +-1.7 (95%) LOS: 100.0%
Total: 20000 W: 1381 L: 1044 D: 17575
Ptnml(0-2): 27, 885, 7864, 1172, 52

https://tests.stockfishchess.org/tests/view/60a2beede229097940a03806
TC: 20s+0.2s, 8 threads
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 34272 W: 1610 L: 1452 D: 31210
Ptnml(0-2): 30, 1285, 14350, 1439, 32

https://tests.stockfishchess.org/tests/view/60a2d687e229097940a03c72
TC: 60s+0.6s, 8 threads
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 45544 W: 1262 L: 1214 D: 43068
Ptnml(0-2): 12, 1129, 20442, 1177, 12

The network has been trained (by vondele) using the https://github.com/glinscott/nnue-pytorch/ trainer (started by glinscott),
specifically the branch https://github.com/Sopel97/nnue-pytorch/tree/experiment_56.
The data used are in 64 billion positions (193GB total) generated and scored with the current master net
d8: https://drive.google.com/file/d/1hOOYSDKgOOp38ZmD0N4DV82TOLHzjUiF/view?usp=sharing
d9: https://drive.google.com/file/d/1VlhnHL8f-20AXhGkILujnNXHwy9T-MQw/view?usp=sharing
d10: https://drive.google.com/file/d/1ZC5upzBYMmMj1gMYCkt6rCxQG0GnO3Kk/view?usp=sharing
fishtest_d9: https://drive.google.com/file/d/1GQHt0oNgKaHazwJFTRbXhlCN3FbUedFq/view?usp=sharing

This network also contains a few architectural changes with respect to the current master:

    Size changed from 256x2-32-32-1 to 512x2-16-32-1
        ~15-20% slower
        ~2x larger
        adds a special path for 16 valued ClippedReLU
        fixes affine transform code for 16 inputs/outputs, buy using InputDimensions instead of PaddedInputDimensions
            this is safe now because the inputs are processed in groups of 4 in the current affine transform code
    The feature set changed from HalfKP to HalfKAv2
        Includes information about the kings like HalfKA
        Packs king features better, resulting in 8% size reduction compared to HalfKA
    The board is flipped for the black's perspective, instead of rotated like in the current master
    PSQT values for each feature
        the feature transformer now outputs a part that is fowarded directly to the output and allows learning piece values more directly than the previous network architecture. The effect is visible for high imbalance positions, where the current master network outputs evaluations skewed towards zero.
        8 PSQT values per feature, chosen based on (popcount(pos.pieces()) - 1) / 4
        initialized to classical material values on the start of the training
    8 subnetworks (512x2->16->32->1), chosen based on (popcount(pos.pieces()) - 1) / 4
        only one subnetwork is evaluated for any position, no or marginal speed loss

A diagram of the network is available: https://user-images.githubusercontent.com/8037982/118656988-553a1700-b7eb-11eb-82ef-56a11cbebbf2.png
A more complete description: https://github.com/glinscott/nnue-pytorch/blob/master/docs/nnue.md

closes https://github.com/official-stockfish/Stockfish/pull/3474

Bench: 3806488

3 years agoSmall clean-ups
Stéphane Nicolet [Mon, 17 May 2021 07:13:34 +0000 (09:13 +0200)]
Small clean-ups

- Comment for Countemove pruning -> Continuation history
- Fix comment in input_slice.h
- Shorter lines in Makefile
- Comment for scale factor
- Fix comment for pinners in see_ge()
- Change Thread.id() signature to size_t
- Trailing space in reprosearch.sh
- Add Douglas Matos Gomes to the AUTHORS file
- Introduce comment for undo_null_move()
- Use Stockfish coding style for export_net()
- Change date in AUTHORS file

closes https://github.com/official-stockfish/Stockfish/pull/3416

No functional change

3 years agoSimplification for countermoves based pruning
Vizvezdenec [Fri, 14 May 2021 21:55:45 +0000 (00:55 +0300)]
Simplification for countermoves based pruning

Simplify away two extra conditions in countermoves based pruning.
These conditions (both of them) were introduced quite a long time ago
via speculative LTCs and seem to no longer bring any benefit.

passed STC
https://tests.stockfishchess.org/tests/view/609e81f35085663412d08f31
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 28488 W: 2487 L: 2382 D: 23619
Ptnml(0-2): 87, 1919, 10123, 2032, 83

passed LTC
https://tests.stockfishchess.org/tests/view/609e9c085085663412d08f59
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 33176 W: 1219 L: 1155 D: 30802
Ptnml(0-2): 13, 1036, 14423, 1106, 10

closes https://github.com/official-stockfish/Stockfish/pull/3468

Bench: 4749514

3 years agoSimplify LMR rules for statScore
bmc4 [Fri, 14 May 2021 02:47:41 +0000 (23:47 -0300)]
Simplify LMR rules for statScore

We simplify two parts of LMR which seem not to bring strength anymore.

---

Individual Tests:
https://tests.stockfishchess.org/tests/view/609d1cc15085663412d0856a
https://tests.stockfishchess.org/tests/view/609cb0cc7746e3dc74ffae8d
https://tests.stockfishchess.org/tests/view/609d1c9f5085663412d08568

---

LTC:
LLR: 2.97 (-2.94,2.94) <-2.50,0.50>
Total: 84184 W: 3093 L: 3066 D: 78025
Ptnml(0-2): 47, 2755, 36458, 2788, 44
https://tests.stockfishchess.org/tests/view/609d84615085663412d08e2f

---

While at it, we also update the Elo estimate of the previous rule in LMR, see:
https://tests.stockfishchess.org/tests/view/609a933c3a33eb67a844f7ca
https://tests.stockfishchess.org/tests/view/609a959c3a33eb67a844f7d5
https://tests.stockfishchess.org/tests/view/609afff73a33eb67a844f870

---

closes https://github.com/official-stockfish/Stockfish/pull/3464

Bench: 4156523

3 years agoRemove early return in Probcut code
bmc4 [Thu, 13 May 2021 14:12:56 +0000 (11:12 -0300)]
Remove early return in Probcut code

We simplify away early return in ProbCut, as it seems not to bring any strength anymore.

STC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 42632 W: 3705 L: 3617 D: 35310
Ptnml(0-2): 123, 2947, 15110, 2991, 145
https://tests.stockfishchess.org/tests/view/609c49da7746e3dc74ffae02

LTC:
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 35384 W: 1314 L: 1251 D: 32819
Ptnml(0-2): 11, 1130, 15355, 1177, 19
https://tests.stockfishchess.org/tests/view/609c71467746e3dc74ffae47

---

While at it, we also update the Elo estimate of ProbCut
(see https://tests.stockfishchess.org/tests/view/609bfb597746e3dc74ffabe3).

closes https://github.com/official-stockfish/Stockfish/pull/3462

bench: 3764662

3 years agoRemove BoolConditions from tuning code
Unai Corzo [Fri, 14 May 2021 15:35:32 +0000 (17:35 +0200)]
Remove BoolConditions from tuning code

Remove BoolConditions from tuning code, as the feature does not work
and the code has not be touched in years.

No functional change

3 years agoSimplify LMR rule for non-checking captures
bmc4 [Wed, 12 May 2021 12:38:26 +0000 (09:38 -0300)]
Simplify LMR rule for non-checking captures

We simplify away the complicated rule in LMR for "non-checking captures
likely to be bad", as it seems not to bring any strength anymore.

STC:
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 55256 W: 4972 L: 4897 D: 45387
Ptnml(0-2): 177, 3976, 19234, 4077, 164
https://tests.stockfishchess.org/tests/view/609adf3b3a33eb67a844f842

LTC:
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 10344 W: 437 L: 353 D: 9554
Ptnml(0-2): 1, 322, 4449, 392, 8
https://tests.stockfishchess.org/tests/view/609b3dfa3a33eb67a844f88e

--

While at it, we also update the Elo estimate of the previous rule in LMR
(see https://tests.stockfishchess.org/tests/view/609af2a63a33eb67a844f867).

closes https://github.com/official-stockfish/Stockfish/pull/3460

Bench: 3840688

3 years agoE2K: added support for MCST Elbrus 2000 CPU architecture
EntityFX [Mon, 12 Apr 2021 10:42:35 +0000 (13:42 +0300)]
E2K: added support for MCST Elbrus 2000 CPU architecture

e2k (Elbrus 2000) - this is a VLIW/EPIC architecture,
the like Intel Itanium (IA-64) architecture.
The architecture has half native / half software support
for most Intel/AMD SIMD (e.g. MMX/SSE/SSE2/SSE3/SSSE3/SSE4.1/SSE4.2/AES/AVX/AVX2 & 3DNow!/SSE4a/XOP/FMA4) via intrinsics.

https://en.wikipedia.org/wiki/Elbrus_2000

closes https://github.com/official-stockfish/Stockfish/pull/3425

No functional change

3 years agoRemove coordination between searching threads
bmc4 [Sun, 9 May 2021 19:33:41 +0000 (16:33 -0300)]
Remove coordination between searching threads

In summary, this revert #2204, as it seems not to bring any strength anymore, so it's no long needed.

STC (5+0.05 @ 8 threads):
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 105728 W: 6406 L: 6393 D: 92929
Ptnml(0-2): 154, 5479, 41599, 5464, 168
https://tests.stockfishchess.org/tests/view/6096994095e7f1852abd3154

LTC (20+0.2 @ 8 threads):
LLR: 2.96 (-2.94,2.94) <-2.50,0.50>
Total: 26336 W: 774 L: 712 D: 24850
Ptnml(0-2): 9, 641, 11810, 695, 13
https://tests.stockfishchess.org/tests/view/6097c62995e7f1852abd31e8

closes https://github.com/official-stockfish/Stockfish/pull/3459

No functional change.

3 years agoSimplify LMR
bmc4 [Tue, 11 May 2021 14:04:02 +0000 (11:04 -0300)]
Simplify LMR

as it seems not to bring any strength and thus is no longer needed.

Tests for updating elo estimates:
https://tests.stockfishchess.org/tests/view/6099ff123a33eb67a844f789
https://tests.stockfishchess.org/tests/view/60953e6695e7f1852abd305b

Individual simplification tests:
https://tests.stockfishchess.org/tests/view/6098cfc73a33eb67a844f6a1
https://tests.stockfishchess.org/tests/view/6095539495e7f1852abd308b

LTC:
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 96984 W: 3624 L: 3608 D: 89752
Ptnml(0-2): 45, 3222, 41939, 3244, 42
https://tests.stockfishchess.org/tests/view/6099921a3a33eb67a844f74f

closes https://github.com/official-stockfish/Stockfish/pull/3458

bench: 3836428

3 years agoExporting the currently loaded network file
Tomasz Sobczyk [Fri, 7 May 2021 10:24:12 +0000 (12:24 +0200)]
Exporting the currently loaded network file

This PR adds an ability to export any currently loaded network.
The export_net command now takes an optional filename parameter.
If the loaded net is not the embedded net the filename parameter is required.

Two changes were required to support this:

* the "architecture" string, which is really just a some kind of description in the net, is now saved into netDescription on load and correctly saved on export.
* the AffineTransform scrambles weights for some architectures and sparsifies them, such that retrieving the index is hard. This is solved by having a temporary scrambled<->unscrambled index lookup table when loading the network, and the actual index is saved for each individual weight that makes it to canSaturate16. This increases the size of the canSaturate16 entries by 6 bytes.

closes https://github.com/official-stockfish/Stockfish/pull/3456

No functional change

3 years agoCleanup of likelyFailLow logic
Vizvezdenec [Wed, 5 May 2021 16:03:20 +0000 (19:03 +0300)]
Cleanup of likelyFailLow logic

This patch broadens and simplifies definition of PvNode that is likely to fail low.
New definition can be described as following "If node was already researched
at depth >= current depth and failed low there" which is more logical than the
previous version and takes less space + allows to not recompute it every time during move loop.

Passed simplification STC
https://tests.stockfishchess.org/tests/view/609148bf95e7f1852abd2e82
LLR: 2.93 (-2.94,2.94) <-2.50,0.50>
Total: 20128 W: 1865 L: 1751 D: 16512
Ptnml(0-2): 63, 1334, 7165, 1430, 72

Passed simplification LTC
https://tests.stockfishchess.org/tests/view/6091691295e7f1852abd2e8b
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 95128 W: 3498 L: 3481 D: 88149
Ptnml(0-2): 41, 2956, 41549, 2981, 37

closes https://github.com/official-stockfish/Stockfish/pull/3455

Bench: 3933037