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.
- 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
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
A lot of optimizations happend since the NNUE was introduced
and since then some parts of the code were left unused. This
got to the point where asserts were have to be made just to
let people know that modifying something will not have any
effects or may even break everything due to the assumptions
being made. Removing these parts removes those inexisting
"false dependencies". Additionally:
* append_changed_indices now takes the king pos and stateinfo
explicitly, no more misleading pos parameter
* IndexList is removed in favor of a generic ValueList.
Feature transformer just instantiates the type it needs.
* The update cost and refresh requirement is deferred to the
feature set once again, but now doesn't go through the whole
FeatureSet machinery and just calls HalfKP directly.
* accumulator no longer has a singular dimension.
* The PS constants and the PieceSquareIndex array are made local
to the HalfKP feature set because they are specific to it and
DO differ for other feature sets.
* A few names are changed to more descriptive
Passed STC non-regression:
https://tests.stockfishchess.org/tests/view/608421dd95e7f1852abd2790
LLR: 2.95 (-2.94,2.94) <-2.50,0.50>
Total: 180008 W: 16186 L: 16258 D: 147564
Ptnml(0-2): 587, 12593, 63725, 12503, 596
closes https://github.com/official-stockfish/Stockfish/pull/3441
No functional change
This patch changes the pop_lsb() signature from Square pop_lsb(Bitboard*) to
Square pop_lsb(Bitboard&). This is more idomatic for C++ style signatures.
Passed a non-regression STC test:
LLR: 2.93 (-2.94,2.94) {-1.25,0.25}
Total: 21280 W: 1928 L: 1847 D: 17505
Ptnml(0-2): 71, 1427, 7558, 1518, 66
https://tests.stockfishchess.org/tests/view/6053a1e22433018de7a38e2f
We have verified that the generated binary is identical on gcc-10.
Closes https://github.com/official-stockfish/Stockfish/pull/3404
No functional change.
Size of the weights in the last layer is less than 512 bits. It leads to wrong data access for AVX512. There is no error because in current implementation it is guaranteed that there is an array of zeros after weights so zero multiplied by something is returned and sum is correct. It is a mistake that can lead to unexpected bugs in the future. Used AVX2 instructions for smaller input size.
No measurable slowdown on avx512.
closes https://github.com/official-stockfish/Stockfish/pull/3298
No functional change.
Reordered weights in such a way that accumulated sum fits to output.
Weights are grouped in blocks of four elements because four
int8 (weight type) corresponds to one int32 (output type).
No horizontal additions.
Grouped AVX512, AVX2 and SSSE3 implementations.
Repeated code was removed.
An earlier version passed STC:
LLR: 2.97 (-2.94,2.94) {-0.25,1.25}
Total: 15336 W: 1495 L: 1355 D: 12486
Ptnml(0-2): 44, 1054, 5350, 1158, 62
https://tests.stockfishchess.org/tests/view/5ff60e106019e097de3eefd5
Speedup depends on the architecture, up to 4% measured on a NNUE only bench.
closes https://github.com/official-stockfish/Stockfish/pull/3287
No functional change
Improves throughput by summing 2 intermediate dot products using 16 bit addition before upconverting to 32 bit.
Potential saturation is detected and the code-path is avoided in this case.
The saturation can't happen with the current nets,
but nets can be constructed that trigger this check.
STC https://tests.stockfishchess.org/tests/view/5fd40a861ac1691201888479
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 25544 W: 2451 L: 2296 D: 20797
Ptnml(0-2): 92, 1761, 8925, 1888, 106
about 5% speedup
closes https://github.com/official-stockfish/Stockfish/pull/3261
No functional change
This appears to be slightly faster than using a comparison against zero
to compute the high bits, on both old (like Pentium III) and new (like
Zen 2) hardware.
closes https://github.com/official-stockfish/Stockfish/pull/3254
No functional change.
in affine transform for AVX512/AVX2/SSSE3
The idea is to initialize sum with the first element instead of zero.
Reduce one add_epi32 and one set_zero SIMD instructions for each output dimension.
sum = 0; for i = 1 to n sum += a[i] ->
sum = a[1]; for i = 2 to n sum += a[i]
STC:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 69048 W: 7024 L: 6799 D: 55225
Ptnml(0-2): 260, 5175, 23458, 5342, 289
https://tests.stockfishchess.org/tests/view/5faf2cf467cbf42301d6aa06
closes https://github.com/official-stockfish/Stockfish/pull/3227
No functional change.
For the feature transformer the code is analogical to AVX2 since there was room for easy adaptation of wider simd registers.
For the smaller affine transforms that have 32 byte stride we keep 2 columns in one zmm register. We also unroll more aggressively so that in the end we have to do 16 parallel horizontal additions on ymm slices each consisting of 4 32-bit integers. The slices are embedded in 8 zmm registers.
These changes provide about 1.5% speedup for AVX-512 builds.
Closes https://github.com/official-stockfish/Stockfish/pull/3218
No functional change.
A non-functional speedup. Unroll the loops going over
the output dimensions in the affine transform layers by
a factor of 4 and perform 4 horizontal additions at a time.
Instead of doing naive horizontal additions on each vector
separately use hadd and shuffling between vectors to reduce
the number of instructions by using all lanes for all stages
of the horizontal adds.
passed STC of the initial version:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 17808 W: 1914 L: 1756 D: 14138
Ptnml(0-2): 76, 1330, 5948, 1460, 90
https://tests.stockfishchess.org/tests/view/5f9d516f6a2c112b60691da3
passed STC of the final version after cleanup:
LLR: 2.95 (-2.94,2.94) {-0.25,1.25}
Total: 16296 W: 1750 L: 1595 D: 12951
Ptnml(0-2): 72, 1192, 5479, 1319, 86
https://tests.stockfishchess.org/tests/view/5f9df5776a2c112b60691de3
closes https://github.com/official-stockfish/Stockfish/pull/3203
No functional change
This patch was inspired by c065abd which updates the accumulator,
if possible, based on the accumulator of two plies back if
the accumulator of the preceding ply is not available.
With this patch we look back even further in the position history
in an attempt to reduce the number of complete recomputations.
When we find a usable accumulator for the position N plies back,
we also update the accumulator of the position N-1 plies back
because that accumulator is most likely to be helpful later
when evaluating positions in sibling branches.
By not updating all intermediate accumulators immediately,
we avoid doing too much work that is not certain to be useful.
Overall, roughly 2-3% speedup.
This patch makes the code more specific to the net architecture,
changing input features of the net will require additional changes
to the incremental update code as discussed in the PR #3193 and #3191.
Passed STC:
https://tests.stockfishchess.org/tests/view/5f9056712c92c7fe3a8c60d0
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 10040 W: 1116 L: 968 D: 7956
Ptnml(0-2): 42, 722, 3365, 828, 63
closes https://github.com/official-stockfish/Stockfish/pull/3193
No functional change.
- Clean signature of functions in namespace NNUE
- Add comment for countermove based pruning
- Remove bestMoveCount variable
- Add const qualifier to kpp_board_index array
- Fix spaces in get_best_thread()
- Fix indention in capture LMR code in search.cpp
- Rename TtmemDeleter to LargePageDeleter
Closes https://github.com/official-stockfish/Stockfish/pull/3063
No functional change
Use TT memory functions to allocate memory for the NNUE weights. This
should provide a small speed-up on systems where large pages are not
automatically used, including Windows and some Linux distributions.
Further, since we now have a wrapper for std::aligned_alloc(), we can
simplify the TT memory management a bit:
- We no longer need to store separate pointers to the hash table and
its underlying memory allocation.
- We also get to merge the Linux-specific and default implementations
of aligned_ttmem_alloc().
Finally, we'll enable the VirtualAlloc code path with large page
support also for Win32.
STC: https://tests.stockfishchess.org/tests/view/5f66595823a84a47b9036fba
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 14896 W: 1854 L: 1686 D: 11356
Ptnml(0-2): 65, 1224, 4742, 1312, 105
closes https://github.com/official-stockfish/Stockfish/pull/3081
No functional change.
This fixes#3108 and removes some NNUE code that is currently not used.
At the moment, do_null_move() copies the accumulator from the previous
state into the new state, which is correct. It then clears the "computed_score"
flag because the side to move has changed, and with the other side to move
NNUE will return a completely different evaluation (normally with changed
sign but also with different NNUE-internal tempo bonus).
The problem is that do_null_move() clears the wrong flag. It clears the
computed_score flag of the old state, not of the new state. It turns out
that this almost never affects the search. For example, fixing it does not
change the current bench (but it does change the previous bench). This is
because the search code usually avoids calling evaluate() after a null move.
This PR corrects do_null_move() by removing the computed_score flag altogether.
The flag is not needed because nnue_evaluate() is never called twice on a position.
This PR also removes some unnecessary {}s and inserts a few blank lines
in the modified NNUE files in line with SF coding style.
Resulf ot STC non-regression test:
LLR: 2.95 (-2.94,2.94) {-1.25,0.25}
Total: 26328 W: 3118 L: 3012 D: 20198
Ptnml(0-2): 126, 2208, 8397, 2300, 133
https://tests.stockfishchess.org/tests/view/5f553ccc2d02727c56b36db1
closes https://github.com/official-stockfish/Stockfish/pull/3109
bench: 4109324
