Update nnue architecture to SFNNv16 and net nn-89cb98a217f7.nnue

Failed STC (https://tests.stockfishchess.org/tests/view/6a4f36355529b8472df7ff7e):
LLR: -2.94 (-2.94,2.94) <0.00,2.00>
Total: 144768 W: 38003 L: 38051 D: 68714
Ptnml(0-2): 521, 17154, 37050, 17170, 489

Passed LTC (https://tests.stockfishchess.org/tests/view/6a501dad5529b8472df80103):
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 189888 W: 49613 L: 48984 D: 91291
Ptnml(0-2): 122, 20406, 53256, 21041, 119

Passed VLTC (https://tests.stockfishchess.org/tests/view/6a52892c5529b8472df80481):
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 46124 W: 12156 L: 11865 D: 22103
Ptnml(0-2): 12, 4523, 13696, 4824, 7

Passed VVLTC (https://tests.stockfishchess.org/tests/view/6a566a285529b8472df80b84):
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 44698 W: 11805 L: 11507 D: 21386
Ptnml(0-2): 9, 3819, 14395, 4117, 9

This PR adds pawn pair features, a new set of features that is indexed by pairs of pawns occupying the same or adjacent files (hence "3-wide"). Because these are a superset of pawn–pawn interactions in threat inputs (including the recently added pawn-pusher inputs), those are removed. HalfKA features remain unchanged.

Pawn-pair features were invented by Jonathan for Pawnocchio, based on his observation that in a net trained on *all* pairs of pawns, the most important pawn pairs were those differing by at most one file.

sscg13 made the necessary changes to the trainer (https://github.com/official-stockfish/nnue-pytorch/pull/502) and initial changes to inference. The slowdown was originally far too large (some 10%); following major optimizations from Jonathan and a bit from me, the slowdown is small enough for a comfortable LTC pass; on my machine I get a 3.5% slowdown overall. I expect there are further "easy" speedups in the new code, so this gap should narrow a bit.

Nettest PR: https://github.com/vondele/nettest/pull/359

closes https://github.com/official-stockfish/Stockfish/pull/6982

Bench: 2718396

Co-Authored-By: anematode <timothy.herchen@gmail.com>
Co-Authored-By: Jonathan Hallström <lmj.hallstrom@gmail.com>
This commit is contained in:
sscg13
2026-07-20 11:06:41 +02:00
committed by Joost VandeVondele
co-authored by anematode Jonathan Hallström
parent 4d75fd4f70
commit f4bcd40409
16 changed files with 440 additions and 135 deletions
+2 -2
View File
@@ -80,14 +80,14 @@ SRCS = attacks.cpp benchmark.cpp bitboard.cpp evaluate.cpp main.cpp \
misc.cpp movegen.cpp movepick.cpp position.cpp \
search.cpp thread.cpp timeman.cpp tt.cpp uci.cpp ucioption.cpp tune.cpp syzygy/tbprobe.cpp \
nnue/nnue_accumulator.cpp nnue/nnue_misc.cpp nnue/network.cpp \
nnue/features/half_ka_v2_hm.cpp nnue/features/full_threats.cpp \
nnue/features/half_ka_v2_hm.cpp nnue/features/full_threats.cpp nnue/features/pp_3wide.cpp \
engine.cpp score.cpp memory.cpp
OTHER_SRCS = universal/entry_x86.cpp universal/entry_arm64.cpp universal/entry_riscv64.cpp universal/nnue_embed.cpp
HEADERS = attacks.h benchmark.h bitboard.h evaluate.h misc.h movegen.h movepick.h history.h \
nnue/nnue_misc.h nnue/features/half_ka_v2_hm.h nnue/features/full_threats.h \
nnue/layers/affine_transform.h nnue/layers/affine_transform_sparse_input.h \
nnue/features/pp_3wide.h nnue/layers/affine_transform.h nnue/layers/affine_transform_sparse_input.h \
nnue/layers/clipped_relu.h nnue/layers/sqr_clipped_relu.h nnue/nnue_accumulator.h \
nnue/nnue_architecture.h nnue/nnue_common.h nnue/nnue_feature_transformer.h nnue/simd.h \
nnue/nnz_helper.h position.h search.h syzygy/tbprobe.h thread.h thread_native.h timeman.h \