The new network caused some issues initially due to the very narrow neuron set between the first two FC layers. Necessary changes were hacked together to make it work. This patch is a mature approach to make the affine transform code faster, more readable, and easier to maintain should the layer sizes change again.
The following changes were made:
* ClippedReLU always produces a multiple of 32 outputs. This is about as good of a solution for AffineTransform's SIMD requirements as it can get without a bigger rewrite.
* All self-contained simd helpers are moved to a separate file (simd.h). Inline asm is utilized to work around GCC's issues with code generation and register assignment. See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101693, https://godbolt.org/z/da76fY1n7
* AffineTransform has 2 specializations. While it's more lines of code due to the boilerplate, the logic in both is significantly reduced, as these two are impossible to nicely combine into one.
1) The first specialization is for cases when there's >=128 inputs. It uses a different approach to perform the affine transform and can make full use of AVX512 without any edge cases. Furthermore, it has higher theoretical throughput because less loads are needed in the hot path, requiring only a fixed amount of instructions for horizontal additions at the end, which are amortized by the large number of inputs.
2) The second specialization is made to handle smaller layers where performance is still necessary but edge cases need to be handled. AVX512 implementation for this was ommited by mistake, a remnant from the temporary implementation for the new... This could be easily reintroduced if needed. A slightly more detailed description of both implementations is in the code.
Overall it should be a minor speedup, as shown on fishtest:
passed STC:
LLR: 2.96 (-2.94,2.94) <-0.50,2.50>
Total: 51520 W: 4074 L: 3888 D: 43558
Ptnml(0-2): 111, 3136, 19097, 3288, 128
and various tests shown in the pull request
closes https://github.com/official-stockfish/Stockfish/pull/3663
No functional change
Introduces a new NNUE network architecture and associated network parameters
The summary of the changes:
* Position for each perspective mirrored such that the king is on e..h files. Cuts the feature transformer size in half, while preserving enough knowledge to be good. See https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.b40q4rb1w7on.
* The number of neurons after the feature transformer increased two-fold, to 1024x2. This is possibly mostly due to the now very optimized feature transformer update code.
* The number of neurons after the second layer is reduced from 16 to 8, to reduce the speed impact. This, perhaps surprisingly, doesn't harm the strength much. See https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.6qkocr97fezq
The AffineTransform code did not work out-of-the box with the smaller number of neurons after the second layer, so some temporary changes have been made to add a special case for InputDimensions == 8. Also additional 0 padding is added to the output for some archs that cannot process inputs by <=8 (SSE2, NEON). VNNI uses an implementation that can keep all outputs in the registers while reducing the number of loads by 3 for each 16 inputs, thanks to the reduced number of output neurons. However GCC is particularily bad at optimization here (and perhaps why the current way the affine transform is done even passed sprt) (see https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit# for details) and more work will be done on this in the following days. I expect the current VNNI implementation to be improved and extended to other architectures.
The network was trained with a slightly modified version of the pytorch trainer (https://github.com/glinscott/nnue-pytorch); the changes are in https://github.com/glinscott/nnue-pytorch/pull/143
The training utilized 2 datasets.
dataset A - https://drive.google.com/file/d/1VlhnHL8f-20AXhGkILujnNXHwy9T-MQw/view?usp=sharing
dataset B - as described in ba01f4b954
The training process was as following:
train on dataset A for 350 epochs, take the best net in terms of elo at 20k nodes per move (it's fine to take anything from later stages of training).
convert the .ckpt to .pt
--resume-from-model from the .pt file, train on dataset B for <600 epochs, take the best net. Lambda=0.8, applied before the loss function.
The first training command:
python3 train.py \
../nnue-pytorch-training/data/large_gensfen_multipvdiff_100_d9.binpack \
../nnue-pytorch-training/data/large_gensfen_multipvdiff_100_d9.binpack \
--gpus "$3," \
--threads 1 \
--num-workers 1 \
--batch-size 16384 \
--progress_bar_refresh_rate 20 \
--smart-fen-skipping \
--random-fen-skipping 3 \
--features=HalfKAv2_hm^ \
--lambda=1.0 \
--max_epochs=600 \
--default_root_dir ../nnue-pytorch-training/experiment_$1/run_$2
The second training command:
python3 serialize.py \
--features=HalfKAv2_hm^ \
../nnue-pytorch-training/experiment_131/run_6/default/version_0/checkpoints/epoch-499.ckpt \
../nnue-pytorch-training/experiment_$1/base/base.pt
python3 train.py \
../nnue-pytorch-training/data/michael_commit_b94a65.binpack \
../nnue-pytorch-training/data/michael_commit_b94a65.binpack \
--gpus "$3," \
--threads 1 \
--num-workers 1 \
--batch-size 16384 \
--progress_bar_refresh_rate 20 \
--smart-fen-skipping \
--random-fen-skipping 3 \
--features=HalfKAv2_hm^ \
--lambda=0.8 \
--max_epochs=600 \
--resume-from-model ../nnue-pytorch-training/experiment_$1/base/base.pt \
--default_root_dir ../nnue-pytorch-training/experiment_$1/run_$2
STC: https://tests.stockfishchess.org/tests/view/611120b32a8a49ac5be798c4
LLR: 2.97 (-2.94,2.94) <-0.50,2.50>
Total: 22480 W: 2434 L: 2251 D: 17795
Ptnml(0-2): 101, 1736, 7410, 1865, 128
LTC: https://tests.stockfishchess.org/tests/view/611152b32a8a49ac5be798ea
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 9776 W: 442 L: 333 D: 9001
Ptnml(0-2): 5, 295, 4180, 402, 6
closes https://github.com/official-stockfish/Stockfish/pull/3646
bench: 5189338
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
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
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
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
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/
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
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
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
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
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
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
bmc4