The sdot instruction computes (and accumulates) a signed dot product,
which is quite handy for Stockfish's NNUE code. The instruction is
optional for Armv8.2 and Armv8.3, and mandatory for Armv8.4 and above.
The commit adds a new 'arm-dotprod' architecture with enabled dot
product support. It also enables dot product support for the existing
'apple-silicon' architecture, which is at least Armv8.5.
The following local speed test was performed on an Apple M1 with
ARCH=apple-silicon. I had to remove CPU pinning from the benchmark
script. However, the results were still consistent: Checking both
binaries against themselves reported a speedup of +0.0000 and +0.0005,
respectively.
```
Result of 100 runs
==================
base (...ish.037ef3e1) = 1917997 +/- 7152
test (...fish.dotprod) = 2159682 +/- 9066
diff = +241684 +/- 2923
speedup = +0.1260
P(speedup > 0) = 1.0000
CPU: 10 x arm
Hyperthreading: off
```
Fixes#4193
closes https://github.com/official-stockfish/Stockfish/pull/4400
No functional change
The accumulator should be an earlyclobber because it is written before
all input operands are read. Otherwise, the asm code computes a wrong
result if the accumulator shares a register with one of the other input
operands (which happens if we pass in the same expression for the
accumulator and the operand).
Closes https://github.com/official-stockfish/Stockfish/pull/4339
No functional change
Removed sprintf() which generated a warning, because of security reasons.
Replace NULL with nullptr
Replace typedef with using
Do not inherit from std::vector. Use composition instead.
optimize mutex-unlocking
closes https://github.com/official-stockfish/Stockfish/pull/4327
No functional change
If a global function has no previous declaration, either the declaration
is missing in the corresponding header file or the function should be
declared static. Static functions are local to the translation unit,
which allows the compiler to apply some optimizations earlier (when
compiling the translation unit rather than during link-time
optimization).
The commit enables the warning for gcc, clang, and mingw. It also fixes
the reported warnings by declaring the functions static or by adding a
header file (benchmark.h).
closes https://github.com/official-stockfish/Stockfish/pull/4325
No functional change
Normalizes the internal value as reported by evaluate or search
to the UCI centipawn result used in output. This value is derived from
the win_rate_model() such that Stockfish outputs an advantage of
"100 centipawns" for a position if the engine has a 50% probability to win
from this position in selfplay at fishtest LTC time control.
The reason to introduce this normalization is that our evaluation is, since NNUE,
no longer related to the classical parameter PawnValueEg (=208). This leads to
the current evaluation changing quite a bit from release to release, for example,
the eval needed to have 50% win probability at fishtest LTC (in cp and internal Value):
June 2020 : 113cp (237)
June 2021 : 115cp (240)
April 2022 : 134cp (279)
July 2022 : 167cp (348)
With this patch, a 100cp advantage will have a fixed interpretation,
i.e. a 50% win chance. To keep this value steady, it will be needed to update the win_rate_model()
from time to time, based on fishtest data. This analysis can be performed with
a set of scripts currently available at https://github.com/vondele/WLD_model
fixes https://github.com/official-stockfish/Stockfish/issues/4155
closes https://github.com/official-stockfish/Stockfish/pull/4216
No functional change
This patch chooses the delta value (which skews the nnue evaluation between positional and materialistic)
depending on the material: If the material is low, delta will be higher and the evaluation is shifted
to the positional value. If the material is high, the evaluation will be shifted to the psqt value.
I don't think slightly negative values of delta should be a concern.
Passed STC:
https://tests.stockfishchess.org/tests/view/62418513b3b383e86185766f
LLR: 2.94 (-2.94,2.94) <0.00,2.50>
Total: 28808 W: 7832 L: 7564 D: 13412
Ptnml(0-2): 147, 3186, 7505, 3384, 182
Passed LTC:
https://tests.stockfishchess.org/tests/view/62419137b3b383e861857842
LLR: 2.96 (-2.94,2.94) <0.50,3.00>
Total: 58632 W: 15776 L: 15450 D: 27406
Ptnml(0-2): 42, 5889, 17149, 6173, 63
closes https://github.com/official-stockfish/Stockfish/pull/3971
Bench: 7588855
This commit generalizes the feature transform to use vec_t macros
that are architecture defined instead of using a seperate code path for each one.
It should make some old architectures (MMX, including improvements by Fanael) faster
and make further such improvements easier in the future.
Includes some corrections to CI for mingw.
closes https://github.com/official-stockfish/Stockfish/pull/3955
closes https://github.com/official-stockfish/Stockfish/pull/3928
No functional change
Architecture:
The diagram of the "SFNNv4" architecture:
https://user-images.githubusercontent.com/8037982/153455685-cbe3a038-e158-4481-844d-9d5fccf5c33a.png
The most important architectural changes are the following:
* 1024x2 [activated] neurons are pairwise, elementwise multiplied (not quite pairwise due to implementation details, see diagram), which introduces a non-linearity that exhibits similar benefits to previously tested sigmoid activation (quantmoid4), while being slightly faster.
* The following layer has therefore 2x less inputs, which we compensate by having 2 more outputs. It is possible that reducing the number of outputs might be beneficial (as we had it as low as 8 before). The layer is now 1024->16.
* The 16 outputs are split into 15 and 1. The 1-wide output is added to the network output (after some necessary scaling due to quantization differences). The 15-wide is activated and follows the usual path through a set of linear layers. The additional 1-wide output is at least neutral, but has shown a slightly positive trend in training compared to networks without it (all 16 outputs through the usual path), and allows possibly an additional stage of lazy evaluation to be introduced in the future.
Additionally, the inference code was rewritten and no longer uses a recursive implementation. This was necessitated by the splitting of the 16-wide intermediate result into two, which was impossible to do with the old implementation with ugly hacks. This is hopefully overall for the better.
First session:
The first session was training a network from scratch (random initialization). The exact trainer used was slightly different (older) from the one used in the second session, but it should not have a measurable effect. The purpose of this session is to establish a strong network base for the second session. Small deviations in strength do not harm the learnability in the second session.
The training was done using the following command:
python3 train.py \
/home/sopel/nnue/nnue-pytorch-training/data/nodes5000pv2_UHO.binpack \
/home/sopel/nnue/nnue-pytorch-training/data/nodes5000pv2_UHO.binpack \
--gpus "$3," \
--threads 4 \
--num-workers 4 \
--batch-size 16384 \
--progress_bar_refresh_rate 20 \
--random-fen-skipping 3 \
--features=HalfKAv2_hm^ \
--lambda=1.0 \
--gamma=0.992 \
--lr=8.75e-4 \
--max_epochs=400 \
--default_root_dir ../nnue-pytorch-training/experiment_$1/run_$2
Every 20th net was saved and its playing strength measured against some baseline at 25k nodes per move with pure NNUE evaluation (modified binary). The exact setup is not important as long as it's consistent. The purpose is to sift good candidates from bad ones.
The dataset can be found https://drive.google.com/file/d/1UQdZN_LWQ265spwTBwDKo0t1WjSJKvWY/view
Second session:
The second training session was done starting from the best network (as determined by strength testing) from the first session. It is important that it's resumed from a .pt model and NOT a .ckpt model. The conversion can be performed directly using serialize.py
The LR schedule was modified to use gamma=0.995 instead of gamma=0.992 and LR=4.375e-4 instead of LR=8.75e-4 to flatten the LR curve and allow for longer training. The training was then running for 800 epochs instead of 400 (though it's possibly mostly noise after around epoch 600).
The training was done using the following command:
The training was done using the following command:
python3 train.py \
/data/sopel/nnue/nnue-pytorch-training/data/T60T70wIsRightFarseerT60T74T75T76.binpack \
/data/sopel/nnue/nnue-pytorch-training/data/T60T70wIsRightFarseerT60T74T75T76.binpack \
--gpus "$3," \
--threads 4 \
--num-workers 4 \
--batch-size 16384 \
--progress_bar_refresh_rate 20 \
--random-fen-skipping 3 \
--features=HalfKAv2_hm^ \
--lambda=1.0 \
--gamma=0.995 \
--lr=4.375e-4 \
--max_epochs=800 \
--resume-from-model /data/sopel/nnue/nnue-pytorch-training/data/exp295/nn-epoch399.pt \
--default_root_dir ../nnue-pytorch-training/experiment_$1/run_$run_id
In particular note that we now use lambda=1.0 instead of lambda=0.8 (previous nets), because tests show that WDL-skipping introduced by vondele performs better with lambda=1.0. Nets were being saved every 20th epoch. In total 16 runs were made with these settings and the best nets chosen according to playing strength at 25k nodes per move with pure NNUE evaluation - these are the 4 nets that have been put on fishtest.
The dataset can be found either at ftp://ftp.chessdb.cn/pub/sopel/data_sf/T60T70wIsRightFarseerT60T74T75T76.binpack in its entirety (download might be painfully slow because hosted in China) or can be assembled in the following way:
Get the 5640ad48ae/script/interleave_binpacks.py script.
Download T60T70wIsRightFarseer.binpack https://drive.google.com/file/d/1_sQoWBl31WAxNXma2v45004CIVltytP8/view
Download farseerT74.binpack http://trainingdata.farseer.org/T74-May13-End.7z
Download farseerT75.binpack http://trainingdata.farseer.org/T75-June3rd-End.7z
Download farseerT76.binpack http://trainingdata.farseer.org/T76-Nov10th-End.7z
Run python3 interleave_binpacks.py T60T70wIsRightFarseer.binpack farseerT74.binpack farseerT75.binpack farseerT76.binpack T60T70wIsRightFarseerT60T74T75T76.binpack
Tests:
STC: https://tests.stockfishchess.org/tests/view/6203fb85d71106ed12a407b7
LLR: 2.94 (-2.94,2.94) <0.00,2.50>
Total: 16952 W: 4775 L: 4521 D: 7656
Ptnml(0-2): 133, 1818, 4318, 2076, 131
LTC: https://tests.stockfishchess.org/tests/view/62041e68d71106ed12a40e85
LLR: 2.94 (-2.94,2.94) <0.50,3.00>
Total: 14944 W: 4138 L: 3907 D: 6899
Ptnml(0-2): 21, 1499, 4202, 1728, 22
closes https://github.com/official-stockfish/Stockfish/pull/3927
Bench: 4919707
This patch optimizes the NEON implementation in two ways.
The activation layer after the feature transformer is rewritten to make it easier for the compiler to see through dependencies and unroll. This in itself is a minimal, but a positive improvement. Other architectures could benefit from this too in the future. This is not an algorithmic change.
The affine transform for large matrices (first layer after FT) on NEON now utilizes the same optimized code path as >=SSSE3, which makes the memory accesses more sequential and makes better use of the available registers, which allows for code that has longer dependency chains.
Benchmarks from Redshift#161, profile-build with apple clang
george@Georges-MacBook-Air nets % ./stockfish-b82d93 bench 2>&1 | tail -4 (current master)
===========================
Total time (ms) : 2167
Nodes searched : 4667742
Nodes/second : 2154011
george@Georges-MacBook-Air nets % ./stockfish-7377b8 bench 2>&1 | tail -4 (this patch)
===========================
Total time (ms) : 1842
Nodes searched : 4667742
Nodes/second : 2534061
This is a solid 18% improvement overall, larger in a bench with NNUE-only, not mixed.
Improvement is also observed on armv7-neon (Raspberry Pi, and older phones), around 5% speedup.
No changes for architectures other than NEON.
closes https://github.com/official-stockfish/Stockfish/pull/3837
No functional changes.
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
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
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
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
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
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
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
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.
