Add NNUE evaluation

This patch ports the efficiently updatable neural network (NNUE) evaluation to Stockfish.

Both the NNUE and the classical evaluations are available, and can be used to
assign a value to a position that is later used in alpha-beta (PVS) search to find the
best move. The classical evaluation computes this value as a function of various chess
concepts, handcrafted by experts, tested and tuned using fishtest. The NNUE evaluation
computes this value with a neural network based on basic inputs. The network is optimized
and trained on the evalutions of millions of positions at moderate search depth.

The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward.
It can be evaluated efficiently on CPUs, and exploits the fact that only parts
of the neural network need to be updated after a typical chess move.
[The nodchip repository](https://github.com/nodchip/Stockfish) provides additional
tools to train and develop the NNUE networks.

This patch is the result of contributions of various authors, from various communities,
including: nodchip, ynasu87, yaneurao (initial port and NNUE authors), domschl, FireFather,
rqs, xXH4CKST3RXx, tttak, zz4032, joergoster, mstembera, nguyenpham, erbsenzaehler,
dorzechowski, and vondele.

This new evaluation needed various changes to fishtest and the corresponding infrastructure,
for which tomtor, ppigazzini, noobpwnftw, daylen, and vondele are gratefully acknowledged.

The first networks have been provided by gekkehenker and sergiovieri, with the latter
net (nn-97f742aaefcd.nnue) being the current default.

The evaluation function can be selected at run time with the `Use NNUE` (true/false) UCI option,
provided the `EvalFile` option points the the network file (depending on the GUI, with full path).

The performance of the NNUE evaluation relative to the classical evaluation depends somewhat on
the hardware, and is expected to improve quickly, but is currently on > 80 Elo on fishtest:

60000 @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f28fe6ea5abc164f05e4c4c
ELO: 92.77 +-2.1 (95%) LOS: 100.0%
Total: 60000 W: 24193 L: 8543 D: 27264
Ptnml(0-2): 609, 3850, 9708, 10948, 4885

40000 @ 20+0.2 th 8
https://tests.stockfishchess.org/tests/view/5f290229a5abc164f05e4c58
ELO: 89.47 +-2.0 (95%) LOS: 100.0%
Total: 40000 W: 12756 L: 2677 D: 24567
Ptnml(0-2): 74, 1583, 8550, 7776, 2017

At the same time, the impact on the classical evaluation remains minimal, causing no significant
regression:

sprt @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f2906a2a5abc164f05e4c5b
LLR: 2.94 (-2.94,2.94) {-6.00,-4.00}
Total: 34936 W: 6502 L: 6825 D: 21609
Ptnml(0-2): 571, 4082, 8434, 3861, 520

sprt @ 60+0.6 th 1
https://tests.stockfishchess.org/tests/view/5f2906cfa5abc164f05e4c5d
LLR: 2.93 (-2.94,2.94) {-6.00,-4.00}
Total: 10088 W: 1232 L: 1265 D: 7591
Ptnml(0-2): 49, 914, 3170, 843, 68

The needed networks can be found at https://tests.stockfishchess.org/nns
It is recommended to use the default one as indicated by the `EvalFile` UCI option.

Guidelines for testing new nets can be found at
https://github.com/glinscott/fishtest/wiki/Creating-my-first-test#nnue-net-tests

Integration has been discussed in various issues:
https://github.com/official-stockfish/Stockfish/issues/2823
https://github.com/official-stockfish/Stockfish/issues/2728

The integration branch will be closed after the merge:
https://github.com/official-stockfish/Stockfish/pull/2825
https://github.com/official-stockfish/Stockfish/tree/nnue-player-wip

closes https://github.com/official-stockfish/Stockfish/pull/2912

This will be an exciting time for computer chess, looking forward to seeing the evolution of
this approach.

Bench: 4746616

src/nnue/architectures/halfkp_256x2-32-32.h

@@ -0,0 +1,54 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Definition of input features and network structure used in NNUE evaluation function
+
+#ifndef NNUE_HALFKP_256X2_32_32_H_INCLUDED
+#define NNUE_HALFKP_256X2_32_32_H_INCLUDED
+
+#include "../features/feature_set.h"
+#include "../features/half_kp.h"
+
+#include "../layers/input_slice.h"
+#include "../layers/affine_transform.h"
+#include "../layers/clipped_relu.h"
+
+namespace Eval::NNUE {
+
+// Input features used in evaluation function
+using RawFeatures = Features::FeatureSet<
+    Features::HalfKP<Features::Side::kFriend>>;
+
+// Number of input feature dimensions after conversion
+constexpr IndexType kTransformedFeatureDimensions = 256;
+
+namespace Layers {
+
+// Define network structure
+using InputLayer = InputSlice<kTransformedFeatureDimensions * 2>;
+using HiddenLayer1 = ClippedReLU<AffineTransform<InputLayer, 32>>;
+using HiddenLayer2 = ClippedReLU<AffineTransform<HiddenLayer1, 32>>;
+using OutputLayer = AffineTransform<HiddenLayer2, 1>;
+
+}  // namespace Layers
+
+using Network = Layers::OutputLayer;
+
+}  // namespace Eval::NNUE
+
+#endif // #ifndef NNUE_HALFKP_256X2_32_32_H_INCLUDED

src/nnue/evaluate_nnue.cpp

@@ -0,0 +1,178 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Code for calculating NNUE evaluation function
+
+#include <fstream>
+#include <iostream>
+#include <set>
+
+#include "../evaluate.h"
+#include "../position.h"
+#include "../misc.h"
+#include "../uci.h"
+
+#include "evaluate_nnue.h"
+
+ExtPieceSquare kpp_board_index[PIECE_NB] = {
+ // convention: W - us, B - them
+ // viewed from other side, W and B are reversed
+    { PS_NONE,     PS_NONE     },
+    { PS_W_PAWN,   PS_B_PAWN   },
+    { PS_W_KNIGHT, PS_B_KNIGHT },
+    { PS_W_BISHOP, PS_B_BISHOP },
+    { PS_W_ROOK,   PS_B_ROOK   },
+    { PS_W_QUEEN,  PS_B_QUEEN  },
+    { PS_W_KING,   PS_B_KING   },
+    { PS_NONE,     PS_NONE     },
+    { PS_NONE,     PS_NONE     },
+    { PS_B_PAWN,   PS_W_PAWN   },
+    { PS_B_KNIGHT, PS_W_KNIGHT },
+    { PS_B_BISHOP, PS_W_BISHOP },
+    { PS_B_ROOK,   PS_W_ROOK   },
+    { PS_B_QUEEN,  PS_W_QUEEN  },
+    { PS_B_KING,   PS_W_KING   },
+    { PS_NONE,     PS_NONE     }
+};
+
+
+namespace Eval::NNUE {
+
+  // Input feature converter
+  AlignedPtr<FeatureTransformer> feature_transformer;
+
+  // Evaluation function
+  AlignedPtr<Network> network;
+
+  // Evaluation function file name
+  std::string fileName;
+
+  namespace Detail {
+
+  // Initialize the evaluation function parameters
+  template <typename T>
+  void Initialize(AlignedPtr<T>& pointer) {
+
+    pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
+    std::memset(pointer.get(), 0, sizeof(T));
+  }
+
+  // Read evaluation function parameters
+  template <typename T>
+  bool ReadParameters(std::istream& stream, const AlignedPtr<T>& pointer) {
+
+    std::uint32_t header;
+    stream.read(reinterpret_cast<char*>(&header), sizeof(header));
+    if (!stream || header != T::GetHashValue()) return false;
+    return pointer->ReadParameters(stream);
+  }
+
+  }  // namespace Detail
+
+  // Initialize the evaluation function parameters
+  void Initialize() {
+
+    Detail::Initialize(feature_transformer);
+    Detail::Initialize(network);
+  }
+
+  // Read network header
+  bool ReadHeader(std::istream& stream,
+    std::uint32_t* hash_value, std::string* architecture) {
+
+    std::uint32_t version, size;
+    stream.read(reinterpret_cast<char*>(&version), sizeof(version));
+    stream.read(reinterpret_cast<char*>(hash_value), sizeof(*hash_value));
+    stream.read(reinterpret_cast<char*>(&size), sizeof(size));
+    if (!stream || version != kVersion) return false;
+    architecture->resize(size);
+    stream.read(&(*architecture)[0], size);
+    return !stream.fail();
+  }
+
+  // Read network parameters
+  bool ReadParameters(std::istream& stream) {
+
+    std::uint32_t hash_value;
+    std::string architecture;
+    if (!ReadHeader(stream, &hash_value, &architecture)) return false;
+    if (hash_value != kHashValue) return false;
+    if (!Detail::ReadParameters(stream, feature_transformer)) return false;
+    if (!Detail::ReadParameters(stream, network)) return false;
+    return stream && stream.peek() == std::ios::traits_type::eof();
+  }
+
+  // Proceed with the difference calculation if possible
+  static void UpdateAccumulatorIfPossible(const Position& pos) {
+
+    feature_transformer->UpdateAccumulatorIfPossible(pos);
+  }
+
+  // Calculate the evaluation value
+  static Value ComputeScore(const Position& pos, bool refresh) {
+
+    auto& accumulator = pos.state()->accumulator;
+    if (!refresh && accumulator.computed_score) {
+      return accumulator.score;
+    }
+
+    alignas(kCacheLineSize) TransformedFeatureType
+        transformed_features[FeatureTransformer::kBufferSize];
+    feature_transformer->Transform(pos, transformed_features, refresh);
+    alignas(kCacheLineSize) char buffer[Network::kBufferSize];
+    const auto output = network->Propagate(transformed_features, buffer);
+
+    auto score = static_cast<Value>(output[0] / FV_SCALE);
+
+    accumulator.score = score;
+    accumulator.computed_score = true;
+    return accumulator.score;
+  }
+
+  // Load the evaluation function file
+  bool load_eval_file(const std::string& evalFile) {
+
+    Initialize();
+    fileName = evalFile;
+
+    std::ifstream stream(evalFile, std::ios::binary);
+
+    const bool result = ReadParameters(stream);
+
+    return result;
+  }
+
+  // Evaluation function. Perform differential calculation.
+  Value evaluate(const Position& pos) {
+    Value v = ComputeScore(pos, false);
+    v = Utility::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
+
+    return v;
+  }
+
+  // Evaluation function. Perform full calculation.
+  Value compute_eval(const Position& pos) {
+    return ComputeScore(pos, true);
+  }
+
+  // Proceed with the difference calculation if possible
+  void update_eval(const Position& pos) {
+    UpdateAccumulatorIfPossible(pos);
+  }
+
+} // namespace Eval::NNUE

src/nnue/evaluate_nnue.h

@@ -0,0 +1,48 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// header used in NNUE evaluation function
+
+#ifndef NNUE_EVALUATE_NNUE_H_INCLUDED
+#define NNUE_EVALUATE_NNUE_H_INCLUDED
+
+#include "nnue_feature_transformer.h"
+
+#include <memory>
+
+namespace Eval::NNUE {
+
+  // Hash value of evaluation function structure
+  constexpr std::uint32_t kHashValue =
+      FeatureTransformer::GetHashValue() ^ Network::GetHashValue();
+
+  // Deleter for automating release of memory area
+  template <typename T>
+  struct AlignedDeleter {
+    void operator()(T* ptr) const {
+      ptr->~T();
+      std_aligned_free(ptr);
+    }
+  };
+
+  template <typename T>
+  using AlignedPtr = std::unique_ptr<T, AlignedDeleter<T>>;
+
+}  // namespace Eval::NNUE
+
+#endif // #ifndef NNUE_EVALUATE_NNUE_H_INCLUDED

src/nnue/features/feature_set.h

@@ -0,0 +1,135 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// A class template that represents the input feature set of the NNUE evaluation function
+
+#ifndef NNUE_FEATURE_SET_H_INCLUDED
+#define NNUE_FEATURE_SET_H_INCLUDED
+
+#include "features_common.h"
+#include <array>
+
+namespace Eval::NNUE::Features {
+
+  // Class template that represents a list of values
+  template <typename T, T... Values>
+  struct CompileTimeList;
+
+  template <typename T, T First, T... Remaining>
+  struct CompileTimeList<T, First, Remaining...> {
+    static constexpr bool Contains(T value) {
+      return value == First || CompileTimeList<T, Remaining...>::Contains(value);
+    }
+    static constexpr std::array<T, sizeof...(Remaining) + 1>
+        kValues = {{First, Remaining...}};
+  };
+
+  // Base class of feature set
+  template <typename Derived>
+  class FeatureSetBase {
+
+   public:
+    // Get a list of indices for active features
+    template <typename IndexListType>
+    static void AppendActiveIndices(
+        const Position& pos, TriggerEvent trigger, IndexListType active[2]) {
+
+      for (Color perspective : { WHITE, BLACK }) {
+        Derived::CollectActiveIndices(
+            pos, trigger, perspective, &active[perspective]);
+      }
+    }
+
+    // Get a list of indices for recently changed features
+    template <typename PositionType, typename IndexListType>
+    static void AppendChangedIndices(
+        const PositionType& pos, TriggerEvent trigger,
+        IndexListType removed[2], IndexListType added[2], bool reset[2]) {
+
+      const auto& dp = pos.state()->dirtyPiece;
+      if (dp.dirty_num == 0) return;
+
+      for (Color perspective : { WHITE, BLACK }) {
+        reset[perspective] = false;
+        switch (trigger) {
+          case TriggerEvent::kFriendKingMoved:
+            reset[perspective] =
+                dp.pieceId[0] == PIECE_ID_KING + perspective;
+            break;
+          default:
+            assert(false);
+            break;
+        }
+        if (reset[perspective]) {
+          Derived::CollectActiveIndices(
+              pos, trigger, perspective, &added[perspective]);
+        } else {
+          Derived::CollectChangedIndices(
+              pos, trigger, perspective,
+              &removed[perspective], &added[perspective]);
+        }
+      }
+    }
+  };
+
+  // Class template that represents the feature set
+  template <typename FeatureType>
+  class FeatureSet<FeatureType> : public FeatureSetBase<FeatureSet<FeatureType>> {
+
+   public:
+    // Hash value embedded in the evaluation file
+    static constexpr std::uint32_t kHashValue = FeatureType::kHashValue;
+    // Number of feature dimensions
+    static constexpr IndexType kDimensions = FeatureType::kDimensions;
+    // Maximum number of simultaneously active features
+    static constexpr IndexType kMaxActiveDimensions =
+        FeatureType::kMaxActiveDimensions;
+    // Trigger for full calculation instead of difference calculation
+    using SortedTriggerSet =
+        CompileTimeList<TriggerEvent, FeatureType::kRefreshTrigger>;
+    static constexpr auto kRefreshTriggers = SortedTriggerSet::kValues;
+
+   private:
+    // Get a list of indices for active features
+    static void CollectActiveIndices(
+        const Position& pos, const TriggerEvent trigger, const Color perspective,
+        IndexList* const active) {
+      if (FeatureType::kRefreshTrigger == trigger) {
+        FeatureType::AppendActiveIndices(pos, perspective, active);
+      }
+    }
+
+    // Get a list of indices for recently changed features
+    static void CollectChangedIndices(
+        const Position& pos, const TriggerEvent trigger, const Color perspective,
+        IndexList* const removed, IndexList* const added) {
+
+      if (FeatureType::kRefreshTrigger == trigger) {
+        FeatureType::AppendChangedIndices(pos, perspective, removed, added);
+      }
+    }
+
+    // Make the base class and the class template that recursively uses itself a friend
+    friend class FeatureSetBase<FeatureSet>;
+    template <typename... FeatureTypes>
+    friend class FeatureSet;
+  };
+
+}  // namespace Eval::NNUE::Features
+
+#endif // #ifndef NNUE_FEATURE_SET_H_INCLUDED

src/nnue/features/features_common.h

@@ -0,0 +1,45 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+//Common header of input features of NNUE evaluation function
+
+#ifndef NNUE_FEATURES_COMMON_H_INCLUDED
+#define NNUE_FEATURES_COMMON_H_INCLUDED
+
+#include "../../evaluate.h"
+#include "../nnue_common.h"
+
+namespace Eval::NNUE::Features {
+
+  class IndexList;
+
+  template <typename... FeatureTypes>
+  class FeatureSet;
+
+  // Trigger to perform full calculations instead of difference only
+  enum class TriggerEvent {
+    kFriendKingMoved // calculate full evaluation when own king moves
+  };
+
+  enum class Side {
+    kFriend // side to move
+  };
+
+}  // namespace Eval::NNUE::Features
+
+#endif // #ifndef NNUE_FEATURES_COMMON_H_INCLUDED

src/nnue/features/half_kp.cpp

@@ -0,0 +1,92 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+//Definition of input features HalfKP of NNUE evaluation function
+
+#include "half_kp.h"
+#include "index_list.h"
+
+namespace Eval::NNUE::Features {
+
+  // Find the index of the feature quantity from the king position and PieceSquare
+  template <Side AssociatedKing>
+  inline IndexType HalfKP<AssociatedKing>::MakeIndex(Square sq_k, PieceSquare p) {
+    return static_cast<IndexType>(PS_END) * static_cast<IndexType>(sq_k) + p;
+  }
+
+  // Get pieces information
+  template <Side AssociatedKing>
+  inline void HalfKP<AssociatedKing>::GetPieces(
+      const Position& pos, Color perspective,
+      PieceSquare** pieces, Square* sq_target_k) {
+
+    *pieces = (perspective == BLACK) ?
+        pos.eval_list()->piece_list_fb() :
+        pos.eval_list()->piece_list_fw();
+    const PieceId target = (AssociatedKing == Side::kFriend) ?
+        static_cast<PieceId>(PIECE_ID_KING + perspective) :
+        static_cast<PieceId>(PIECE_ID_KING + ~perspective);
+    *sq_target_k = static_cast<Square>(((*pieces)[target] - PS_W_KING) % SQUARE_NB);
+  }
+
+  // Get a list of indices for active features
+  template <Side AssociatedKing>
+  void HalfKP<AssociatedKing>::AppendActiveIndices(
+      const Position& pos, Color perspective, IndexList* active) {
+
+    // Do nothing if array size is small to avoid compiler warning
+    if (RawFeatures::kMaxActiveDimensions < kMaxActiveDimensions) return;
+
+    PieceSquare* pieces;
+    Square sq_target_k;
+    GetPieces(pos, perspective, &pieces, &sq_target_k);
+    for (PieceId i = PIECE_ID_ZERO; i < PIECE_ID_KING; ++i) {
+      if (pieces[i] != PS_NONE) {
+        active->push_back(MakeIndex(sq_target_k, pieces[i]));
+      }
+    }
+  }
+
+  // Get a list of indices for recently changed features
+  template <Side AssociatedKing>
+  void HalfKP<AssociatedKing>::AppendChangedIndices(
+      const Position& pos, Color perspective,
+      IndexList* removed, IndexList* added) {
+
+    PieceSquare* pieces;
+    Square sq_target_k;
+    GetPieces(pos, perspective, &pieces, &sq_target_k);
+    const auto& dp = pos.state()->dirtyPiece;
+    for (int i = 0; i < dp.dirty_num; ++i) {
+      if (dp.pieceId[i] >= PIECE_ID_KING) continue;
+      const auto old_p = static_cast<PieceSquare>(
+          dp.old_piece[i].from[perspective]);
+      if (old_p != PS_NONE) {
+        removed->push_back(MakeIndex(sq_target_k, old_p));
+      }
+      const auto new_p = static_cast<PieceSquare>(
+          dp.new_piece[i].from[perspective]);
+      if (new_p != PS_NONE) {
+        added->push_back(MakeIndex(sq_target_k, new_p));
+      }
+    }
+  }
+
+  template class HalfKP<Side::kFriend>;
+
+}  // namespace Eval::NNUE::Features

src/nnue/features/half_kp.h

@@ -0,0 +1,67 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+//Definition of input features HalfKP of NNUE evaluation function
+
+#ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED
+#define NNUE_FEATURES_HALF_KP_H_INCLUDED
+
+#include "../../evaluate.h"
+#include "features_common.h"
+
+namespace Eval::NNUE::Features {
+
+  // Feature HalfKP: Combination of the position of own king
+  // and the position of pieces other than kings
+  template <Side AssociatedKing>
+  class HalfKP {
+
+   public:
+    // Feature name
+    static constexpr const char* kName = "HalfKP(Friend)";
+    // Hash value embedded in the evaluation file
+    static constexpr std::uint32_t kHashValue =
+        0x5D69D5B9u ^ (AssociatedKing == Side::kFriend);
+    // Number of feature dimensions
+    static constexpr IndexType kDimensions =
+        static_cast<IndexType>(SQUARE_NB) * static_cast<IndexType>(PS_END);
+    // Maximum number of simultaneously active features
+    static constexpr IndexType kMaxActiveDimensions = PIECE_ID_KING;
+    // Trigger for full calculation instead of difference calculation
+    static constexpr TriggerEvent kRefreshTrigger = TriggerEvent::kFriendKingMoved;
+
+    // Get a list of indices for active features
+    static void AppendActiveIndices(const Position& pos, Color perspective,
+                                    IndexList* active);
+
+    // Get a list of indices for recently changed features
+    static void AppendChangedIndices(const Position& pos, Color perspective,
+                                     IndexList* removed, IndexList* added);
+
+    // Index of a feature for a given king position and another piece on some square
+    static IndexType MakeIndex(Square sq_k, PieceSquare p);
+
+   private:
+    // Get pieces information
+    static void GetPieces(const Position& pos, Color perspective,
+                          PieceSquare** pieces, Square* sq_target_k);
+  };
+
+}  // namespace Eval::NNUE::Features
+
+#endif // #ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED

src/nnue/features/index_list.h

@@ -0,0 +1,64 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Definition of index list of input features
+
+#ifndef NNUE_FEATURES_INDEX_LIST_H_INCLUDED
+#define NNUE_FEATURES_INDEX_LIST_H_INCLUDED
+
+#include "../../position.h"
+#include "../nnue_architecture.h"
+
+namespace Eval::NNUE::Features {
+
+  // Class template used for feature index list
+  template <typename T, std::size_t MaxSize>
+  class ValueList {
+
+   public:
+    std::size_t size() const { return size_; }
+    void resize(std::size_t size) { size_ = size; }
+    void push_back(const T& value) { values_[size_++] = value; }
+    T& operator[](std::size_t index) { return values_[index]; }
+    T* begin() { return values_; }
+    T* end() { return values_ + size_; }
+    const T& operator[](std::size_t index) const { return values_[index]; }
+    const T* begin() const { return values_; }
+    const T* end() const { return values_ + size_; }
+
+    void swap(ValueList& other) {
+      const std::size_t max_size = std::max(size_, other.size_);
+      for (std::size_t i = 0; i < max_size; ++i) {
+        std::swap(values_[i], other.values_[i]);
+      }
+      std::swap(size_, other.size_);
+    }
+
+   private:
+    T values_[MaxSize];
+    std::size_t size_ = 0;
+  };
+
+  //Type of feature index list
+  class IndexList
+      : public ValueList<IndexType, RawFeatures::kMaxActiveDimensions> {
+  };
+
+}  // namespace Eval::NNUE::Features
+
+#endif // NNUE_FEATURES_INDEX_LIST_H_INCLUDED

src/nnue/layers/affine_transform.h

@@ -0,0 +1,215 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Definition of layer AffineTransform of NNUE evaluation function
+
+#ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
+#define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
+
+#include <iostream>
+#include "../nnue_common.h"
+
+namespace Eval::NNUE::Layers {
+
+  // Affine transformation layer
+  template <typename PreviousLayer, IndexType OutputDimensions>
+  class AffineTransform {
+   public:
+    // Input/output type
+    using InputType = typename PreviousLayer::OutputType;
+    using OutputType = std::int32_t;
+    static_assert(std::is_same<InputType, std::uint8_t>::value, "");
+
+    // Number of input/output dimensions
+    static constexpr IndexType kInputDimensions =
+        PreviousLayer::kOutputDimensions;
+    static constexpr IndexType kOutputDimensions = OutputDimensions;
+    static constexpr IndexType kPaddedInputDimensions =
+        CeilToMultiple<IndexType>(kInputDimensions, kMaxSimdWidth);
+
+    // Size of forward propagation buffer used in this layer
+    static constexpr std::size_t kSelfBufferSize =
+        CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize);
+
+    // Size of the forward propagation buffer used from the input layer to this layer
+    static constexpr std::size_t kBufferSize =
+        PreviousLayer::kBufferSize + kSelfBufferSize;
+
+    // Hash value embedded in the evaluation file
+    static constexpr std::uint32_t GetHashValue() {
+      std::uint32_t hash_value = 0xCC03DAE4u;
+      hash_value += kOutputDimensions;
+      hash_value ^= PreviousLayer::GetHashValue() >> 1;
+      hash_value ^= PreviousLayer::GetHashValue() << 31;
+      return hash_value;
+    }
+
+   // Read network parameters
+    bool ReadParameters(std::istream& stream) {
+      if (!previous_layer_.ReadParameters(stream)) return false;
+      stream.read(reinterpret_cast<char*>(biases_),
+                  kOutputDimensions * sizeof(BiasType));
+      stream.read(reinterpret_cast<char*>(weights_),
+                  kOutputDimensions * kPaddedInputDimensions *
+                  sizeof(WeightType));
+      return !stream.fail();
+    }
+
+    // Forward propagation
+    const OutputType* Propagate(
+        const TransformedFeatureType* transformed_features, char* buffer) const {
+      const auto input = previous_layer_.Propagate(
+          transformed_features, buffer + kSelfBufferSize);
+      const auto output = reinterpret_cast<OutputType*>(buffer);
+
+  #if defined(USE_AVX512)
+      constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
+      const __m512i kOnes = _mm512_set1_epi16(1);
+      const auto input_vector = reinterpret_cast<const __m512i*>(input);
+
+  #elif defined(USE_AVX2)
+      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      const __m256i kOnes = _mm256_set1_epi16(1);
+      const auto input_vector = reinterpret_cast<const __m256i*>(input);
+
+  #elif defined(USE_SSSE3)
+      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      const __m128i kOnes = _mm_set1_epi16(1);
+      const auto input_vector = reinterpret_cast<const __m128i*>(input);
+
+  #elif defined(USE_NEON)
+      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
+  #endif
+
+      for (IndexType i = 0; i < kOutputDimensions; ++i) {
+        const IndexType offset = i * kPaddedInputDimensions;
+
+  #if defined(USE_AVX512)
+        __m512i sum = _mm512_setzero_si512();
+        const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            __m512i product = _mm512_maddubs_epi16(_mm512_loadu_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #else
+            __m512i product = _mm512_maddubs_epi16(_mm512_load_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #endif
+
+            product = _mm512_madd_epi16(product, kOnes);
+            sum = _mm512_add_epi32(sum, product);
+        }
+        output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
+
+        // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
+        // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit)
+        // and we have to do one more 256bit chunk.
+        if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2)
+        {
+            const auto iv_256  = reinterpret_cast<const __m256i*>(input);
+            const auto row_256 = reinterpret_cast<const __m256i*>(&weights_[offset]);
+            int j = kNumChunks * 2;
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)  // See HACK comment below in AVX2.
+            __m256i sum256 = _mm256_maddubs_epi16(_mm256_loadu_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+  #else
+            __m256i sum256 = _mm256_maddubs_epi16(_mm256_load_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+  #endif
+
+            sum256 = _mm256_madd_epi16(sum256, _mm256_set1_epi16(1));
+            sum256 = _mm256_hadd_epi32(sum256, sum256);
+            sum256 = _mm256_hadd_epi32(sum256, sum256);
+            const __m128i lo = _mm256_extracti128_si256(sum256, 0);
+            const __m128i hi = _mm256_extracti128_si256(sum256, 1);
+            output[i] += _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi);
+        }
+
+  #elif defined(USE_AVX2)
+        __m256i sum = _mm256_setzero_si256();
+        const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m256i product = _mm256_maddubs_epi16(
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
+            //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
+            //       even though alignas is specified.
+            _mm256_loadu_si256
+  #else
+            _mm256_load_si256
+  #endif
+
+            (&input_vector[j]), _mm256_load_si256(&row[j]));
+          product = _mm256_madd_epi16(product, kOnes);
+          sum = _mm256_add_epi32(sum, product);
+        }
+        sum = _mm256_hadd_epi32(sum, sum);
+        sum = _mm256_hadd_epi32(sum, sum);
+        const __m128i lo = _mm256_extracti128_si256(sum, 0);
+        const __m128i hi = _mm256_extracti128_si256(sum, 1);
+        output[i] = _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi) + biases_[i];
+
+  #elif defined(USE_SSSE3)
+        __m128i sum = _mm_cvtsi32_si128(biases_[i]);
+        const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m128i product = _mm_maddubs_epi16(
+              _mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
+          product = _mm_madd_epi16(product, kOnes);
+          sum = _mm_add_epi32(sum, product);
+        }
+        sum = _mm_hadd_epi32(sum, sum);
+        sum = _mm_hadd_epi32(sum, sum);
+        output[i] = _mm_cvtsi128_si32(sum);
+
+  #elif defined(USE_NEON)
+        int32x4_t sum = {biases_[i]};
+        const auto row = reinterpret_cast<const int8x8_t*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]);
+          product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]);
+          sum = vpadalq_s16(sum, product);
+        }
+        output[i] = sum[0] + sum[1] + sum[2] + sum[3];
+
+  #else
+        OutputType sum = biases_[i];
+        for (IndexType j = 0; j < kInputDimensions; ++j) {
+          sum += weights_[offset + j] * input[j];
+        }
+        output[i] = sum;
+  #endif
+
+      }
+      return output;
+    }
+
+   private:
+    using BiasType = OutputType;
+    using WeightType = std::int8_t;
+
+    PreviousLayer previous_layer_;
+
+    alignas(kCacheLineSize) BiasType biases_[kOutputDimensions];
+    alignas(kCacheLineSize)
+        WeightType weights_[kOutputDimensions * kPaddedInputDimensions];
+  };
+
+}  // namespace Eval::NNUE::Layers
+
+#endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED

src/nnue/layers/clipped_relu.h

@@ -0,0 +1,186 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Definition of layer ClippedReLU of NNUE evaluation function
+
+#ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
+#define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
+
+#include "../nnue_common.h"
+
+namespace Eval::NNUE::Layers {
+
+  // Clipped ReLU
+  template <typename PreviousLayer>
+  class ClippedReLU {
+   public:
+    // Input/output type
+    using InputType = typename PreviousLayer::OutputType;
+    using OutputType = std::uint8_t;
+    static_assert(std::is_same<InputType, std::int32_t>::value, "");
+
+    // Number of input/output dimensions
+    static constexpr IndexType kInputDimensions =
+        PreviousLayer::kOutputDimensions;
+    static constexpr IndexType kOutputDimensions = kInputDimensions;
+
+    // Size of forward propagation buffer used in this layer
+    static constexpr std::size_t kSelfBufferSize =
+        CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize);
+
+    // Size of the forward propagation buffer used from the input layer to this layer
+    static constexpr std::size_t kBufferSize =
+        PreviousLayer::kBufferSize + kSelfBufferSize;
+
+    // Hash value embedded in the evaluation file
+    static constexpr std::uint32_t GetHashValue() {
+      std::uint32_t hash_value = 0x538D24C7u;
+      hash_value += PreviousLayer::GetHashValue();
+      return hash_value;
+    }
+
+    // Read network parameters
+    bool ReadParameters(std::istream& stream) {
+      return previous_layer_.ReadParameters(stream);
+    }
+
+    // Forward propagation
+    const OutputType* Propagate(
+        const TransformedFeatureType* transformed_features, char* buffer) const {
+      const auto input = previous_layer_.Propagate(
+          transformed_features, buffer + kSelfBufferSize);
+      const auto output = reinterpret_cast<OutputType*>(buffer);
+
+  #if defined(USE_AVX2)
+      constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
+      const __m256i kZero = _mm256_setzero_si256();
+      const __m256i kOffsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+      const auto in = reinterpret_cast<const __m256i*>(input);
+      const auto out = reinterpret_cast<__m256i*>(output);
+      for (IndexType i = 0; i < kNumChunks; ++i) {
+        const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+          // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
+          //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
+          //       even though alignas is specified.
+          _mm256_loadu_si256
+  #else
+          _mm256_load_si256
+  #endif
+
+          (&in[i * 4 + 0]),
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+          _mm256_loadu_si256
+  #else
+          _mm256_load_si256
+  #endif
+
+          (&in[i * 4 + 1])), kWeightScaleBits);
+        const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+          _mm256_loadu_si256
+  #else
+          _mm256_load_si256
+  #endif
+
+          (&in[i * 4 + 2]),
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+          _mm256_loadu_si256
+  #else
+          _mm256_load_si256
+  #endif
+
+          (&in[i * 4 + 3])), kWeightScaleBits);
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+        _mm256_storeu_si256
+  #else
+        _mm256_store_si256
+  #endif
+
+          (&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
+            _mm256_packs_epi16(words0, words1), kZero), kOffsets));
+      }
+      constexpr IndexType kStart = kNumChunks * kSimdWidth;
+
+  #elif defined(USE_SSSE3)
+      constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
+
+  #ifdef USE_SSE41
+      const __m128i kZero = _mm_setzero_si128();
+  #else
+      const __m128i k0x80s = _mm_set1_epi8(-128);
+  #endif
+
+      const auto in = reinterpret_cast<const __m128i*>(input);
+      const auto out = reinterpret_cast<__m128i*>(output);
+      for (IndexType i = 0; i < kNumChunks; ++i) {
+        const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
+            _mm_load_si128(&in[i * 4 + 0]),
+            _mm_load_si128(&in[i * 4 + 1])), kWeightScaleBits);
+        const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
+            _mm_load_si128(&in[i * 4 + 2]),
+            _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits);
+        const __m128i packedbytes = _mm_packs_epi16(words0, words1);
+        _mm_store_si128(&out[i],
+
+  #ifdef USE_SSE41
+          _mm_max_epi8(packedbytes, kZero)
+  #else
+          _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
+  #endif
+
+        );
+      }
+      constexpr IndexType kStart = kNumChunks * kSimdWidth;
+
+  #elif defined(USE_NEON)
+      constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2);
+      const int8x8_t kZero = {0};
+      const auto in = reinterpret_cast<const int32x4_t*>(input);
+      const auto out = reinterpret_cast<int8x8_t*>(output);
+      for (IndexType i = 0; i < kNumChunks; ++i) {
+        int16x8_t shifted;
+        const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
+        pack[0] = vqshrn_n_s32(in[i * 2 + 0], kWeightScaleBits);
+        pack[1] = vqshrn_n_s32(in[i * 2 + 1], kWeightScaleBits);
+        out[i] = vmax_s8(vqmovn_s16(shifted), kZero);
+      }
+      constexpr IndexType kStart = kNumChunks * (kSimdWidth / 2);
+  #else
+      constexpr IndexType kStart = 0;
+  #endif
+
+      for (IndexType i = kStart; i < kInputDimensions; ++i) {
+        output[i] = static_cast<OutputType>(
+            std::max(0, std::min(127, input[i] >> kWeightScaleBits)));
+      }
+      return output;
+    }
+
+   private:
+    PreviousLayer previous_layer_;
+  };
+
+}  // namespace Eval::NNUE::Layers
+
+#endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED

src/nnue/layers/input_slice.h

@@ -0,0 +1,68 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// NNUE evaluation function layer InputSlice definition
+
+#ifndef NNUE_LAYERS_INPUT_SLICE_H_INCLUDED
+#define NNUE_LAYERS_INPUT_SLICE_H_INCLUDED
+
+#include "../nnue_common.h"
+
+namespace Eval::NNUE::Layers {
+
+// Input layer
+template <IndexType OutputDimensions, IndexType Offset = 0>
+class InputSlice {
+ public:
+  // Need to maintain alignment
+  static_assert(Offset % kMaxSimdWidth == 0, "");
+
+  // Output type
+  using OutputType = TransformedFeatureType;
+
+  // Output dimensionality
+  static constexpr IndexType kOutputDimensions = OutputDimensions;
+
+  // Size of forward propagation buffer used from the input layer to this layer
+  static constexpr std::size_t kBufferSize = 0;
+
+  // Hash value embedded in the evaluation file
+  static constexpr std::uint32_t GetHashValue() {
+    std::uint32_t hash_value = 0xEC42E90Du;
+    hash_value ^= kOutputDimensions ^ (Offset << 10);
+    return hash_value;
+  }
+
+  // Read network parameters
+  bool ReadParameters(std::istream& /*stream*/) {
+    return true;
+  }
+
+  // Forward propagation
+  const OutputType* Propagate(
+      const TransformedFeatureType* transformed_features,
+      char* /*buffer*/) const {
+    return transformed_features + Offset;
+  }
+
+ private:
+};
+
+}  // namespace Layers
+
+#endif // #ifndef NNUE_LAYERS_INPUT_SLICE_H_INCLUDED

src/nnue/nnue_accumulator.h

@@ -0,0 +1,39 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Class for difference calculation of NNUE evaluation function
+
+#ifndef NNUE_ACCUMULATOR_H_INCLUDED
+#define NNUE_ACCUMULATOR_H_INCLUDED
+
+#include "nnue_architecture.h"
+
+namespace Eval::NNUE {
+
+  // Class that holds the result of affine transformation of input features
+  struct alignas(32) Accumulator {
+    std::int16_t
+        accumulation[2][kRefreshTriggers.size()][kTransformedFeatureDimensions];
+    Value score;
+    bool computed_accumulation;
+    bool computed_score;
+  };
+
+}  // namespace Eval::NNUE
+
+#endif // NNUE_ACCUMULATOR_H_INCLUDED

src/nnue/nnue_architecture.h

@@ -0,0 +1,38 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Input features and network structure used in NNUE evaluation function
+
+#ifndef NNUE_ARCHITECTURE_H_INCLUDED
+#define NNUE_ARCHITECTURE_H_INCLUDED
+
+// Defines the network structure
+#include "architectures/halfkp_256x2-32-32.h"
+
+namespace Eval::NNUE {
+
+  static_assert(kTransformedFeatureDimensions % kMaxSimdWidth == 0, "");
+  static_assert(Network::kOutputDimensions == 1, "");
+  static_assert(std::is_same<Network::OutputType, std::int32_t>::value, "");
+
+  // Trigger for full calculation instead of difference calculation
+  constexpr auto kRefreshTriggers = RawFeatures::kRefreshTriggers;
+
+}  // namespace Eval::NNUE
+
+#endif // #ifndef NNUE_ARCHITECTURE_H_INCLUDED

src/nnue/nnue_common.h

@@ -0,0 +1,77 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// Constants used in NNUE evaluation function
+
+#ifndef NNUE_COMMON_H_INCLUDED
+#define NNUE_COMMON_H_INCLUDED
+
+#if defined(USE_AVX2)
+#include <immintrin.h>
+
+#elif defined(USE_SSE41)
+#include <smmintrin.h>
+
+#elif defined(USE_SSSE3)
+#include <tmmintrin.h>
+
+#elif defined(USE_SSE2)
+#include <emmintrin.h>
+
+#elif defined(USE_NEON)
+#include <arm_neon.h>
+#endif
+
+namespace Eval::NNUE {
+
+  // Version of the evaluation file
+  constexpr std::uint32_t kVersion = 0x7AF32F16u;
+
+  // Constant used in evaluation value calculation
+  constexpr int FV_SCALE = 16;
+  constexpr int kWeightScaleBits = 6;
+
+  // Size of cache line (in bytes)
+  constexpr std::size_t kCacheLineSize = 64;
+
+  // SIMD width (in bytes)
+  #if defined(USE_AVX2)
+  constexpr std::size_t kSimdWidth = 32;
+
+  #elif defined(USE_SSE2)
+  constexpr std::size_t kSimdWidth = 16;
+
+  #elif defined(USE_NEON)
+  constexpr std::size_t kSimdWidth = 16;
+  #endif
+
+  constexpr std::size_t kMaxSimdWidth = 32;
+
+  // Type of input feature after conversion
+  using TransformedFeatureType = std::uint8_t;
+  using IndexType = std::uint32_t;
+
+  // Round n up to be a multiple of base
+  template <typename IntType>
+  constexpr IntType CeilToMultiple(IntType n, IntType base) {
+    return (n + base - 1) / base * base;
+  }
+
+}  // namespace Eval::NNUE
+
+#endif // #ifndef NNUE_COMMON_H_INCLUDED

src/nnue/nnue_feature_transformer.h

@@ -0,0 +1,355 @@
+/*
+  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
+  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+
+  Stockfish is free software: you can redistribute it and/or modify
+  it under the terms of the GNU General Public License as published by
+  the Free Software Foundation, either version 3 of the License, or
+  (at your option) any later version.
+
+  Stockfish is distributed in the hope that it will be useful,
+  but WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  GNU General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program.  If not, see <http://www.gnu.org/licenses/>.
+*/
+
+// A class that converts the input features of the NNUE evaluation function
+
+#ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED
+#define NNUE_FEATURE_TRANSFORMER_H_INCLUDED
+
+#include "nnue_common.h"
+#include "nnue_architecture.h"
+#include "features/index_list.h"
+
+#include <cstring> // std::memset()
+
+namespace Eval::NNUE {
+
+  // Input feature converter
+  class FeatureTransformer {
+
+   private:
+    // Number of output dimensions for one side
+    static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
+
+   public:
+    // Output type
+    using OutputType = TransformedFeatureType;
+
+    // Number of input/output dimensions
+    static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
+    static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
+
+    // Size of forward propagation buffer
+    static constexpr std::size_t kBufferSize =
+        kOutputDimensions * sizeof(OutputType);
+
+    // Hash value embedded in the evaluation file
+    static constexpr std::uint32_t GetHashValue() {
+      return RawFeatures::kHashValue ^ kOutputDimensions;
+    }
+
+    // Read network parameters
+    bool ReadParameters(std::istream& stream) {
+      stream.read(reinterpret_cast<char*>(biases_),
+                  kHalfDimensions * sizeof(BiasType));
+      stream.read(reinterpret_cast<char*>(weights_),
+                  kHalfDimensions * kInputDimensions * sizeof(WeightType));
+      return !stream.fail();
+    }
+
+    // Proceed with the difference calculation if possible
+    bool UpdateAccumulatorIfPossible(const Position& pos) const {
+      const auto now = pos.state();
+      if (now->accumulator.computed_accumulation) {
+        return true;
+      }
+      const auto prev = now->previous;
+      if (prev && prev->accumulator.computed_accumulation) {
+        UpdateAccumulator(pos);
+        return true;
+      }
+      return false;
+    }
+
+    // Convert input features
+    void Transform(const Position& pos, OutputType* output, bool refresh) const {
+      if (refresh || !UpdateAccumulatorIfPossible(pos)) {
+        RefreshAccumulator(pos);
+      }
+      const auto& accumulation = pos.state()->accumulator.accumulation;
+
+  #if defined(USE_AVX2)
+      constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
+      constexpr int kControl = 0b11011000;
+      const __m256i kZero = _mm256_setzero_si256();
+
+  #elif defined(USE_SSSE3)
+      constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
+
+  #ifdef USE_SSE41
+      const __m128i kZero = _mm_setzero_si128();
+  #else
+      const __m128i k0x80s = _mm_set1_epi8(-128);
+  #endif
+
+  #elif defined(USE_NEON)
+      constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+      const int8x8_t kZero = {0};
+  #endif
+
+      const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
+      for (IndexType p = 0; p < 2; ++p) {
+        const IndexType offset = kHalfDimensions * p;
+
+  #if defined(USE_AVX2)
+        auto out = reinterpret_cast<__m256i*>(&output[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m256i sum0 =
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
+            //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
+            //       even though alignas is specified.
+            _mm256_loadu_si256
+  #else
+            _mm256_load_si256
+  #endif
+
+            (&reinterpret_cast<const __m256i*>(
+              accumulation[perspectives[p]][0])[j * 2 + 0]);
+          __m256i sum1 =
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            _mm256_loadu_si256
+  #else
+            _mm256_load_si256
+  #endif
+
+            (&reinterpret_cast<const __m256i*>(
+              accumulation[perspectives[p]][0])[j * 2 + 1]);
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+          _mm256_storeu_si256
+  #else
+          _mm256_store_si256
+  #endif
+
+          (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
+              _mm256_packs_epi16(sum0, sum1), kZero), kControl));
+        }
+
+  #elif defined(USE_SSSE3)
+        auto out = reinterpret_cast<__m128i*>(&output[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
+              accumulation[perspectives[p]][0])[j * 2 + 0]);
+          __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
+              accumulation[perspectives[p]][0])[j * 2 + 1]);
+      const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
+
+          _mm_store_si128(&out[j],
+
+  #ifdef USE_SSE41
+            _mm_max_epi8(packedbytes, kZero)
+  #else
+            _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
+  #endif
+
+          );
+        }
+
+  #elif defined(USE_NEON)
+        const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          int16x8_t sum = reinterpret_cast<const int16x8_t*>(
+              accumulation[perspectives[p]][0])[j];
+          out[j] = vmax_s8(vqmovn_s16(sum), kZero);
+        }
+
+  #else
+        for (IndexType j = 0; j < kHalfDimensions; ++j) {
+          BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
+          output[offset + j] = static_cast<OutputType>(
+              std::max<int>(0, std::min<int>(127, sum)));
+        }
+  #endif
+
+      }
+    }
+
+   private:
+    // Calculate cumulative value without using difference calculation
+    void RefreshAccumulator(const Position& pos) const {
+      auto& accumulator = pos.state()->accumulator;
+      IndexType i = 0;
+      Features::IndexList active_indices[2];
+      RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
+                                       active_indices);
+      for (Color perspective : { WHITE, BLACK }) {
+        std::memcpy(accumulator.accumulation[perspective][i], biases_,
+                   kHalfDimensions * sizeof(BiasType));
+        for (const auto index : active_indices[perspective]) {
+          const IndexType offset = kHalfDimensions * index;
+
+  #if defined(USE_AVX2)
+          auto accumulation = reinterpret_cast<__m256i*>(
+              &accumulator.accumulation[perspective][i][0]);
+          auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+          constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+          for (IndexType j = 0; j < kNumChunks; ++j) {
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            _mm256_storeu_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadu_si256(&accumulation[j]), column[j]));
+  #else
+            accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
+  #endif
+          }
+
+  #elif defined(USE_SSE2)
+          auto accumulation = reinterpret_cast<__m128i*>(
+              &accumulator.accumulation[perspective][i][0]);
+          auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+          constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+          for (IndexType j = 0; j < kNumChunks; ++j) {
+            accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
+          }
+
+  #elif defined(USE_NEON)
+          auto accumulation = reinterpret_cast<int16x8_t*>(
+              &accumulator.accumulation[perspective][i][0]);
+          auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
+          constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+          for (IndexType j = 0; j < kNumChunks; ++j) {
+            accumulation[j] = vaddq_s16(accumulation[j], column[j]);
+          }
+
+  #else
+          for (IndexType j = 0; j < kHalfDimensions; ++j) {
+            accumulator.accumulation[perspective][i][j] += weights_[offset + j];
+          }
+  #endif
+
+        }
+      }
+
+      accumulator.computed_accumulation = true;
+      accumulator.computed_score = false;
+    }
+
+    // Calculate cumulative value using difference calculation
+    void UpdateAccumulator(const Position& pos) const {
+      const auto prev_accumulator = pos.state()->previous->accumulator;
+      auto& accumulator = pos.state()->accumulator;
+      IndexType i = 0;
+      Features::IndexList removed_indices[2], added_indices[2];
+      bool reset[2];
+      RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
+                                        removed_indices, added_indices, reset);
+      for (Color perspective : { WHITE, BLACK }) {
+
+  #if defined(USE_AVX2)
+        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        auto accumulation = reinterpret_cast<__m256i*>(
+            &accumulator.accumulation[perspective][i][0]);
+
+  #elif defined(USE_SSE2)
+        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        auto accumulation = reinterpret_cast<__m128i*>(
+            &accumulator.accumulation[perspective][i][0]);
+
+  #elif defined(USE_NEON)
+        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        auto accumulation = reinterpret_cast<int16x8_t*>(
+            &accumulator.accumulation[perspective][i][0]);
+  #endif
+
+        if (reset[perspective]) {
+          std::memcpy(accumulator.accumulation[perspective][i], biases_,
+                      kHalfDimensions * sizeof(BiasType));
+        } else {
+          std::memcpy(accumulator.accumulation[perspective][i],
+                      prev_accumulator.accumulation[perspective][i],
+                      kHalfDimensions * sizeof(BiasType));
+          // Difference calculation for the deactivated features
+          for (const auto index : removed_indices[perspective]) {
+            const IndexType offset = kHalfDimensions * index;
+
+  #if defined(USE_AVX2)
+            auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
+            }
+
+  #elif defined(USE_SSE2)
+            auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
+            }
+
+  #elif defined(USE_NEON)
+            auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = vsubq_s16(accumulation[j], column[j]);
+            }
+
+  #else
+            for (IndexType j = 0; j < kHalfDimensions; ++j) {
+              accumulator.accumulation[perspective][i][j] -=
+                  weights_[offset + j];
+            }
+  #endif
+
+          }
+        }
+        { // Difference calculation for the activated features
+          for (const auto index : added_indices[perspective]) {
+            const IndexType offset = kHalfDimensions * index;
+
+  #if defined(USE_AVX2)
+            auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
+            }
+
+  #elif defined(USE_SSE2)
+            auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
+            }
+
+  #elif defined(USE_NEON)
+            auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
+            for (IndexType j = 0; j < kNumChunks; ++j) {
+              accumulation[j] = vaddq_s16(accumulation[j], column[j]);
+            }
+
+  #else
+            for (IndexType j = 0; j < kHalfDimensions; ++j) {
+              accumulator.accumulation[perspective][i][j] +=
+                  weights_[offset + j];
+            }
+  #endif
+
+          }
+        }
+      }
+
+      accumulator.computed_accumulation = true;
+      accumulator.computed_score = false;
+    }
+
+    using BiasType = std::int16_t;
+    using WeightType = std::int16_t;
+
+    alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
+    alignas(kCacheLineSize)
+        WeightType weights_[kHalfDimensions * kInputDimensions];
+  };
+
+}  // namespace Eval::NNUE
+
+#endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED