Add transform minimize_binpack for minimizing existing .binpack datasets. (#4447 )

Takes advantage of the sample skipping rules that are used during training (capture, check, or VALUE_NONE). Adds positions to keep continuity, which improves compression.
Tools transform option for filtering data for training nn-335a9b2d8a80.nnue (#4324 )
2026-05-20 12:07:43 +00:00 · 2023-04-25 19:21:29 +02:00 · 2023-02-05 14:21:48 +01:00 · 2022-12-12 22:55:39 +01:00 · 2022-06-03 06:36:46 +02:00 · 2022-05-30 21:03:44 +02:00
103 changed files with 23085 additions and 3506 deletions
@@ -0,0 +1,260 @@
 name: Stockfish
 on:
  push:
    branches:
      - master
      - tools
      - github_ci
      - github_ci_armv7
  pull_request:
    branches:
      - master
      - tools
 jobs:
  Stockfish:
    name: ${{ matrix.config.name }}
    runs-on: ${{ matrix.config.os }}
    env:
      COMPILER: ${{ matrix.config.compiler }}
      COMP: ${{ matrix.config.comp }}
    strategy:
      matrix:
        config:
          # set the variable for the required tests:
          # run_expensive_tests: true
          # run_32bit_tests: true
          # run_64bit_tests: true
          - {
              name: "Ubuntu 20.04 GCC",
              os: ubuntu-20.04,
              compiler: g++,
              comp: gcc,
              run_expensive_tests: true,
              run_64bit_tests: true,
              shell: 'bash {0}'
            }
          - {
              name: "Ubuntu 20.04 Clang",
              os: ubuntu-20.04,
              compiler: clang++,
              comp: clang,
              run_64bit_tests: true,
              shell: 'bash {0}'
            }
          - {
              name: "MacOS 10.15 Apple Clang",
              os: macos-10.15,
              compiler: clang++,
              comp: clang,
              run_64bit_tests: true,
              shell: 'bash {0}'
            }
          - {
              name: "MacOS 10.15 GCC 10",
              os: macos-10.15,
              compiler: g++-10,
              comp: gcc,
              run_64bit_tests: true,
              shell: 'bash {0}'
            }
          - {
              name: "Windows 2022 Mingw-w64 GCC x86_64",
              os: windows-2022,
              compiler: g++,
              comp: mingw,
              run_64bit_tests: true,
              msys_sys: 'mingw64',
              msys_env: 'x86_64-gcc',
              shell: 'msys2 {0}'
            }
    defaults:
      run:
        working-directory: src
        shell: ${{ matrix.config.shell }}
    steps:
      - uses: actions/checkout@v2
        with:
          fetch-depth: 0
      - name: Download required linux packages
        if: runner.os == 'Linux'
        run: |
          sudo apt update
          sudo apt install expect valgrind g++-multilib
      - name: Setup msys and install required packages
        if: runner.os == 'Windows'
        uses: msys2/setup-msys2@v2
        with:
          msystem: ${{matrix.config.msys_sys}}
          install: mingw-w64-${{matrix.config.msys_env}} make git expect
      - name: Download the used network from the fishtest framework
        run: |
          make net
      - name: Extract the bench number from the commit history
        run: |
          git log HEAD | grep "\b[Bb]ench[ :]\+[0-9]\{7\}" | head -n 1 | sed "s/[^0-9]*\([0-9]*\).*/\1/g" > git_sig
          [ -s git_sig ] && echo "benchref=$(cat git_sig)" >> $GITHUB_ENV && echo "Reference bench:" $(cat git_sig) || echo "No bench found"
      - name: Check compiler
        run: |
          $COMPILER -v
      - name: Test help target
        run: |
          make help
      # x86-32 tests
      - name: Test debug x86-32 build
        if: ${{ matrix.config.run_32bit_tests }}
        run: |
          export CXXFLAGS="-D_GLIBCXX_DEBUG"
          make clean
          make -j2 ARCH=x86-32 optimize=no debug=yes build
          ../tests/signature.sh $benchref
      - name: Test x86-32 build
        if: ${{ matrix.config.run_32bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-32 build
          ../tests/signature.sh $benchref
      - name: Test x86-32-sse41-popcnt build
        if: ${{ matrix.config.run_32bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-32-sse41-popcnt build
          ../tests/signature.sh $benchref
      - name: Test x86-32-sse2 build
        if: ${{ matrix.config.run_32bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-32-sse2 build
          ../tests/signature.sh $benchref
      - name: Test general-32 build
        if: ${{ matrix.config.run_32bit_tests }}
        run: |
          make clean
          make -j2 ARCH=general-32 build
          ../tests/signature.sh $benchref
      # x86-64 tests
      - name: Test debug x86-64-modern build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          export CXXFLAGS="-D_GLIBCXX_DEBUG"
          make clean
          make -j2 ARCH=x86-64-modern optimize=no debug=yes build
          ../tests/signature.sh $benchref
      - name: Test x86-64-modern build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-modern build
          ../tests/signature.sh $benchref
      - name: Test x86-64-ssse3 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-ssse3 build
          ../tests/signature.sh $benchref
      - name: Test x86-64-sse3-popcnt build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-sse3-popcnt build
          ../tests/signature.sh $benchref
      - name: Test x86-64 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64 build
          ../tests/signature.sh $benchref
      - name: Test general-64 build
        if: matrix.config.run_64bit_tests
        run: |
          make clean
          make -j2 ARCH=general-64 build
          ../tests/signature.sh $benchref
      # x86-64 with newer extensions tests
      - name: Compile x86-64-avx2 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-avx2 build
      - name: Compile x86-64-bmi2 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-bmi2 build
      - name: Compile x86-64-avx512 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-avx512 build
      - name: Compile x86-64-vnni512 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-vnni512 build
      - name: Compile x86-64-vnni256 build
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-vnni256 build
      # Other tests
      - name: Check perft and search reproducibility
        if: ${{ matrix.config.run_64bit_tests }}
        run: |
          make clean
          make -j2 ARCH=x86-64-modern build
          ../tests/perft.sh
          ../tests/reprosearch.sh
      # Sanitizers
      - name: Run under valgrind
        if: ${{ matrix.config.run_expensive_tests }}
        run: |
          export CXXFLAGS="-O1 -fno-inline"
          make clean
          make -j2 ARCH=x86-64-modern debug=yes optimize=no build > /dev/null
          ../tests/instrumented.sh --valgrind
          ../tests/instrumented.sh --valgrind-thread
      - name: Run with UB sanitizer
        if: ${{ matrix.config.run_expensive_tests }}
        run: |
          export CXXFLAGS="-O1 -fno-inline"
          make clean
          make -j2 ARCH=x86-64-modern sanitize=undefined optimize=no debug=yes build > /dev/null
          ../tests/instrumented.sh --sanitizer-undefined
      - name: Run with thread sanitizer
        if: ${{ matrix.config.run_expensive_tests }}
        run: |
          export CXXFLAGS="-O1 -fno-inline"
          make clean
          make -j2 ARCH=x86-64-modern sanitize=thread optimize=no debug=yes build > /dev/null
          ../tests/instrumented.sh --sanitizer-thread
@@ -0,0 +1,12 @@
 # Files from build
 **/*.o
 **/*.s
 src/.depend
 # Built binary
 src/stockfish*
 src/-lstdc++.res
 # Neural network for the NNUE evaluation
 **/*.nnue
@@ -1,80 +0,0 @@
 language: cpp
 dist: bionic
 matrix:
  include:
    - os: linux
      compiler: gcc
      addons:
        apt:
          packages: ['g++-8', 'g++-8-multilib', 'g++-multilib', 'valgrind', 'expect', 'curl']
      env:
        - COMPILER=g++-8
        - COMP=gcc
    - os: linux
      compiler: clang
      addons:
        apt:
          packages: ['clang-10', 'llvm-10-dev', 'g++-multilib', 'valgrind', 'expect', 'curl']
      env:
        - COMPILER=clang++-10
        - COMP=clang
    - os: osx
      osx_image: xcode12
      compiler: gcc
      env:
        - COMPILER=g++
        - COMP=gcc
    - os: osx
      osx_image: xcode12
      compiler: clang
      env:
        - COMPILER=clang++
        - COMP=clang
 branches:
  only:
   - master
 before_script:
  - cd src
 script:
  # Obtain bench reference from git log
  - git log HEAD | grep "\b[Bb]ench[ :]\+[0-9]\{7\}" | head -n 1 | sed "s/[^0-9]*\([0-9]*\).*/\1/g" > git_sig
  - export benchref=$(cat git_sig)
  - echo "Reference bench:" $benchref
  #
  # Compiler version string
  - $COMPILER -v
  #
  # Verify bench number against various builds
  - export CXXFLAGS="-Werror -D_GLIBCXX_DEBUG"
  - make clean && make -j2 ARCH=x86-64 optimize=no debug=yes build && ../tests/signature.sh $benchref
  - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32 optimize=no debug=yes build && ../tests/signature.sh $benchref; fi
  - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32 build && ../tests/signature.sh $benchref; fi
  #
  # Check perft and reproducible search
  - export CXXFLAGS="-Werror"
  - make clean && make -j2 ARCH=x86-64 build
  - ../tests/perft.sh
  - ../tests/reprosearch.sh
  #
  # Valgrind
  #
  - export CXXFLAGS="-O1 -fno-inline"
  - if [ -x "$(command -v valgrind )" ]; then make clean && make -j2 ARCH=x86-64 debug=yes optimize=no build > /dev/null && ../tests/instrumented.sh --valgrind; fi
  - if [ -x "$(command -v valgrind )" ]; then ../tests/instrumented.sh --valgrind-thread; fi
  #
  # Sanitizer
  #
  - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-64 sanitize=undefined optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-undefined; fi
  - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-64 sanitize=thread    optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-thread; fi
@@ -1,4 +1,4 @@
-# List of authors for Stockfish, as of August 4, 2020
+# List of authors for Stockfish
 # Founders of the Stockfish project and fishtest infrastructure
 Tord Romstad (romstad)
@@ -19,32 +19,43 @@ Alain Savard (Rocky640)
 Alayan Feh (Alayan-stk-2)
 Alexander Kure
 Alexander Pagel (Lolligerhans)
 Alfredo Menezes (lonfom169)
 Ali AlZhrani (Cooffe)
 Andrei Vetrov (proukornew)
 Andrew Grant (AndyGrant)
 Andrey Neporada (nepal)
 Andy Duplain
 Antoine Champion (antoinechampion)
 Aram Tumanian (atumanian)
 Arjun Temurnikar
 Artem Solopiy (EntityFX)
 Auguste Pop
 Balint Pfliegel
 Ben Chaney (Chaneybenjamini)
 Ben Koshy (BKSpurgeon)
 Bill Henry (VoyagerOne)
 Bojun Guo (noobpwnftw, Nooby)
 braich
 Brian Sheppard (SapphireBrand, briansheppard-toast)
 Bruno de Melo Costa (BM123499)
 Bryan Cross (crossbr)
 candirufish
 Chess13234
 Chris Cain (ceebo)
 Dale Weiler (graphitemaster)
 Dan Schmidt (dfannius)
 Daniel Axtens (daxtens)
 Daniel Dugovic (ddugovic)
-Dariusz Orzechowski
+Dariusz Orzechowski (dorzechowski)
 David Zar
 Daylen Yang (daylen)
 Deshawn Mohan-Smith (GoldenRare)
 Dieter Dobbelaere (ddobbelaere)
 DiscanX
 Dominik Schlösser (domschl)
 double-beep
 Douglas Matos Gomes (dsmsgms)
 Dubslow
 Eduardo Cáceres (eduherminio)
 Eelco de Groot (KingDefender)
 Elvin Liu (solarlight2)
@@ -53,12 +64,15 @@ Ernesto Gatti
 Linmiao Xu (linrock)
 Fabian Beuke (madnight)
 Fabian Fichter (ianfab)
 Fanael Linithien (Fanael)
 fanon
 Fauzi Akram Dabat (FauziAkram)
 Felix Wittmann
 gamander
 Gary Heckman (gheckman)
 George Sobala (gsobala)
 gguliash
 Giacomo Lorenzetti (G-Lorenz)
 Gian-Carlo Pascutto (gcp)
 Gontran Lemaire (gonlem)
 Goodkov Vasiliy Aleksandrovich (goodkov)
@@ -79,21 +93,26 @@ Jean Gauthier (OuaisBla)
 Jean-Francois Romang (jromang)
 Jekaa
 Jerry Donald Watson (jerrydonaldwatson)
 jjoshua2
 Jonathan Calovski (Mysseno)
-Jonathan Dumale (SFisGOD)
+Jonathan Buladas Dumale (SFisGOD)
 Joost VandeVondele (vondele)
 Jörg Oster (joergoster)
 Joseph Ellis (jhellis3)
 Joseph R. Prostko
 Julian Willemer (NightlyKing)
 jundery
 Justin Blanchard (UncombedCoconut)
 Kelly Wilson
 Ken Takusagawa
 Kian E (KJE-98)
 kinderchocolate
 Kiran Panditrao (Krgp)
 Kojirion
 Krystian Kuzniarek (kuzkry)
 Leonardo Ljubičić (ICCF World Champion)
 Leonid Pechenik (lp--)
 Liam Keegan (lkeegan)
 Linus Arver (listx)
 loco-loco
 Lub van den Berg (ElbertoOne)
@@ -104,8 +123,11 @@ Maciej Żenczykowski (zenczykowski)
 Malcolm Campbell (xoto10)
 Mark Tenzer (31m059)
 marotear
 Matt Ginsberg (mattginsberg)
 Matthew Lai (matthewlai)
 Matthew Sullivan (Matt14916)
 Max A. (Disservin)
 Maxim Molchanov (Maxim)
 Michael An (man)
 Michael Byrne (MichaelB7)
 Michael Chaly (Vizvezdenec)
@@ -114,6 +136,7 @@ Michael Whiteley (protonspring)
 Michel Van den Bergh (vdbergh)
 Miguel Lahoz (miguel-l)
 Mikael Bäckman (mbootsector)
 Mike Babigian (Farseer)
 Mira
 Miroslav Fontán (Hexik)
 Moez Jellouli (MJZ1977)
@@ -125,6 +148,8 @@ Niklas Fiekas (niklasf)
 Nikolay Kostov (NikolayIT)
 Nguyen Pham (nguyenpham)
 Norman Schmidt (FireFather)
 notruck
 Ofek Shochat (OfekShochat, ghostway)
 Ondrej Mosnáček (WOnder93)
 Oskar Werkelin Ahlin
 Pablo Vazquez
@@ -133,6 +158,7 @@ Pascal Romaret
 Pasquale Pigazzini (ppigazzini)
 Patrick Jansen (mibere)
 pellanda
 Peter Schneider (pschneider1968)
 Peter Zsifkovits (CoffeeOne)
 Praveen Kumar Tummala (praveentml)
 Rahul Dsilva (silversolver1)
@@ -145,19 +171,23 @@ Rodrigo Exterckötter Tjäder
 Ron Britvich (Britvich)
 Ronald de Man (syzygy1, syzygy)
 rqs
 Rui Coelho (ruicoelhopedro)
 Ryan Schmitt
 Ryan Takker
 Sami Kiminki (skiminki)
 Sebastian Buchwald (UniQP)
 Sergei Antonov (saproj)
 Sergei Ivanov (svivanov72)
 Sergio Vieri (sergiovieri)
 sf-x
 Shane Booth (shane31)
 Shawn Varghese (xXH4CKST3RXx)
 Siad Daboul (Topologist)
 Stefan Geschwentner (locutus2)
 Stefano Cardanobile (Stefano80)
 Steinar Gunderson (sesse)
 Stéphane Nicolet (snicolet)
 Prokop Randáček (ProkopRandacek)
 Thanar2
 thaspel
 theo77186
@@ -165,11 +195,13 @@ Tom Truscott
 Tom Vijlbrief (tomtor)
 Tomasz Sobczyk (Sopel97)
 Torsten Franz (torfranz, tfranzer)
 Torsten Hellwig (Torom)
 Tracey Emery (basepr1me)
 tttak
 Unai Corzo (unaiic)
 Uri Blass (uriblass)
 Vince Negri (cuddlestmonkey)
 xefoci7612
 zz4032
@@ -1,46 +1,51 @@
 ## Overview
-[![Build Status](https://travis-ci.org/official-stockfish/Stockfish.svg?branch=master)](https://travis-ci.org/official-stockfish/Stockfish)
+[![Build Status](https://github.com/official-stockfish/Stockfish/actions/workflows/stockfish.yml/badge.svg)](https://github.com/official-stockfish/Stockfish/actions)
 [![Build Status](https://ci.appveyor.com/api/projects/status/github/official-stockfish/Stockfish?branch=master&svg=true)](https://ci.appveyor.com/project/mcostalba/stockfish/branch/master)
 [Stockfish](https://stockfishchess.org) is a free, powerful UCI chess engine
-derived from Glaurung 2.1. It features two evaluation functions, the classical
+derived from Glaurung 2.1. Stockfish is not a complete chess program and requires a
-evaluation based on handcrafted terms, and the NNUE evaluation based on
+UCI-compatible graphical user interface (GUI) (e.g. XBoard with PolyGlot, Scid,
-efficiently updateable neural networks. The classical evaluation runs efficiently
+Cute Chess, eboard, Arena, Sigma Chess, Shredder, Chess Partner or Fritz) in order
-on most 64bit CPU architectures, while the NNUE evaluation benefits strongly from the
+to be used comfortably. Read the documentation for your GUI of choice for information
-vector intrinsics available on modern CPUs (avx2 or similar).
+about how to use Stockfish with it.
 Stockfish is not a complete chess program and requires a
 UCI-compatible GUI (e.g. XBoard with PolyGlot, Scid, Cute Chess, eboard, Arena,
 Sigma Chess, Shredder, Chess Partner or Fritz) in order to be used comfortably.
 Read the documentation for your GUI of choice for information about how to use
 Stockfish with it.
 The Stockfish engine features two evaluation functions for chess. The efficiently 
 updatable neural network (NNUE) based evaluation is the default and by far the strongest.
 The classical evaluation based on handcrafted terms remains available. The strongest 
 network is integrated in the binary and downloaded automatically during the build process.
 The NNUE evaluation benefits from the vector intrinsics available on most CPUs (sse2, 
 avx2, neon, or similar).
 ## Files
 This distribution of Stockfish consists of the following files:
-  * Readme.md, the file you are currently reading.
+  * [README.md](https://github.com/official-stockfish/Stockfish/blob/master/README.md),
    the file you are currently reading.
-  * Copying.txt, a text file containing the GNU General Public License version 3.
+  * [Copying.txt](https://github.com/official-stockfish/Stockfish/blob/master/Copying.txt),
    a text file containing the GNU General Public License version 3.
-  * src, a subdirectory containing the full source code, including a Makefile
+  * [AUTHORS](https://github.com/official-stockfish/Stockfish/blob/master/AUTHORS), 
    a text file with the list of authors for the project
  * [src](https://github.com/official-stockfish/Stockfish/tree/master/src),
    a subdirectory containing the full source code, including a Makefile
    that can be used to compile Stockfish on Unix-like systems.
-To use the NNUE evaluation an additional data file with neural network parameters
+  * a file with the .nnue extension, storing the neural network for the NNUE
-needs to be downloaded. The filename for the default set can be found as the default
+    evaluation. Binary distributions will have this file embedded.
 value of the `EvalFile` UCI option, with the format
 `nn-[SHA256 first 12 digits].nnue` (e.g. nn-c157e0a5755b.nnue). This file can be downloaded from
 ```
 https://tests.stockfishchess.org/api/nn/[filename]
 ```
 replacing `[filename]` as needed.
 ## The UCI protocol and available options
-## UCI options
+The Universal Chess Interface (UCI) is a standard protocol used to communicate with
 a chess engine, and is the recommended way to do so for typical graphical user interfaces
 (GUI) or chess tools. Stockfish implements the majority of its options as described
 in [the UCI protocol](https://www.shredderchess.com/download/div/uci.zip).
-Currently, Stockfish has the following UCI options:
+Developers can see the default values for UCI options available in Stockfish by typing
 `./stockfish uci` in a terminal, but the majority of users will typically see them and
 change them via a chess GUI. This is a list of available UCI options in Stockfish:
  * #### Threads
    The number of CPU threads used for searching a position. For best performance, set
@@ -49,6 +54,9 @@ Currently, Stockfish has the following UCI options:
  * #### Hash
    The size of the hash table in MB. It is recommended to set Hash after setting Threads.
  * #### Clear Hash
    Clear the hash table.
  * #### Ponder
    Let Stockfish ponder its next move while the opponent is thinking.
@@ -58,19 +66,14 @@ Currently, Stockfish has the following UCI options:
  * #### Use NNUE
    Toggle between the NNUE and classical evaluation functions. If set to "true",
-    the network parameters must be availabe to load from file (see also EvalFile).
+    the network parameters must be available to load from file (see also EvalFile),
    if they are not embedded in the binary.
  * #### EvalFile
    The name of the file of the NNUE evaluation parameters. Depending on the GUI the
-    filename should include the full path to the folder/directory that contains the file.
+    filename might have to include the full path to the folder/directory that contains
-
+    the file. Other locations, such as the directory that contains the binary and the
-  * #### Contempt
+    working directory, are also searched.
    A positive value for contempt favors middle game positions and avoids draws,
    effective for the classical evaluation only.
  * #### Analysis Contempt
    By default, contempt is set to prefer the side to move. Set this option to "White"
    or "Black" to analyse with contempt for that side, or "Off" to disable contempt.
  * #### UCI_AnalyseMode
    An option handled by your GUI.
@@ -103,14 +106,14 @@ Currently, Stockfish has the following UCI options:
    Example: `C:\tablebases\wdl345;C:\tablebases\wdl6;D:\tablebases\dtz345;D:\tablebases\dtz6`
    It is recommended to store .rtbw files on an SSD. There is no loss in storing
-    the .rtbz files on a regular HD. It is recommended to verify all md5 checksums
+    the .rtbz files on a regular HDD. It is recommended to verify all md5 checksums
    of the downloaded tablebase files (`md5sum -c checksum.md5`) as corruption will
    lead to engine crashes.
  * #### SyzygyProbeDepth
    Minimum remaining search depth for which a position is probed. Set this option
-    to a higher value to probe less agressively if you experience too much slowdown
+    to a higher value to probe less aggressively if you experience too much slowdown
-    (in terms of nps) due to TB probing.
+    (in terms of nps) due to tablebase probing.
  * #### Syzygy50MoveRule
    Disable to let fifty-move rule draws detected by Syzygy tablebase probes count
@@ -132,42 +135,119 @@ Currently, Stockfish has the following UCI options:
    Tells the engine to use nodes searched instead of wall time to account for
    elapsed time. Useful for engine testing.
  * #### Clear Hash
    Clear the hash table.
  * #### Debug Log File
    Write all communication to and from the engine into a text file.
-## classical and NNUE evaluation
+For developers the following non-standard commands might be of interest, mainly useful for debugging:
  * #### bench *ttSize threads limit fenFile limitType evalType*
    Performs a standard benchmark using various options. The signature of a version 
    (standard node count) is obtained using all defaults. `bench` is currently 
    `bench 16 1 13 default depth mixed`.
  * #### compiler
    Give information about the compiler and environment used for building a binary.
  * #### d
    Display the current position, with ascii art and fen.
  * #### eval
    Return the evaluation of the current position.
  * #### export_net [filename]
    Exports the currently loaded network to a file.
    If the currently loaded network is the embedded network and the filename
    is not specified then the network is saved to the file matching the name
    of the embedded network, as defined in evaluate.h.
    If the currently loaded network is not the embedded network (some net set
    through the UCI setoption) then the filename parameter is required and the
    network is saved into that file.
  * #### flip
    Flips the side to move.
 ### Generating Training Data
 To generate training data from the classic eval, use the generate_training_data command with the setting "Use NNUE" set to "false". The given example is generation in its simplest form. There are more commands.
 ```
 uci
 setoption name PruneAtShallowDepth value false
 setoption name Use NNUE value false
 setoption name Threads value X
 setoption name Hash value Y
 setoption name SyzygyPath value path
 isready
 generate_training_data depth A count B keep_draws 1 eval_limit 32000
 ```
 - `A` is the searched depth per move, or how far the engine looks forward. This value is an integer.
 - `B` is the amount of positions generated. This value is also an integer.
 Specify how many threads and how much memory you would like to use with the `x` and `y` values. The option SyzygyPath is not necessary, but if you would like to use it, you must first have Syzygy endgame tablebases on your computer, which you can find [here](http://oics.olympuschess.com/tracker/index.php). You will need to have a torrent client to download these tablebases, as that is probably the fastest way to obtain them. The `path` is the path to the folder containing those tablebases. It does not have to be surrounded in quotes.
 This will create a file named "training_data.binpack" in the same folder as the binary containing the generated training data. Once generation is done, you can rename the file to something like "1billiondepth12.binpack" to remember the depth and quantity of the positions and move it to a folder named "trainingdata" in the same directory as the binaries.
 You will also need validation data that is used for loss calculation and accuracy computation. Validation data is generated in the same way as training data, but generally at most 1 million positions should be used as there's no need for more and it would just slow the learning process down. It may also be better to slightly increase the depth for validation data. After generation you can rename the validation data file to "val.binpack" and drop it in a folder named "validationdata" in the same directory to make it easier.
 ## Training data formats.
 Currently there are 3 training data formats. Two of them are supported directly.
 - `.bin` - the original training data format. Uses 40 bytes per entry. Is supported directly by the `generate_training_data` command.
 - `.plain` - a human readable training data format. This one is not supported directly by the `generate_training_data` command. It should not be used for data exchange because it's less compact than other formats. It is mostly useful for inspection of the data.
 - `.binpack` - a compact binary training data format that exploits positions chains to further reduce size. It uses on average between 2 to 3 bytes per entry when generating data with `generate_training_data`. It is supported directly by `generate_training_data` command. It is currently the default for the `generate_training_data` command. A more in depth description can be found [here](docs/binpack.md)
 ### Conversion between formats.
 There is a builting converted that support all 3 formats described above. Any of them can be converted to any other. For more information and usage guide see [here](docs/convert.md).
 ## A note on classical evaluation versus NNUE evaluation
 Both approaches assign a value to a position that is 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 (e.g. piece positions only). The network is optimized and trained
-on the evalutions of millions of positions at moderate search depth.
+on the evaluations 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
+[The nodchip repository](https://github.com/nodchip/Stockfish) provided the first 
-tools to train and develop the NNUE networks.
+version of the needed tools to train and develop the NNUE networks. Today, more 
 advanced training tools are available in 
 [the nnue-pytorch repository](https://github.com/glinscott/nnue-pytorch/), 
 while data generation tools are available in 
 [a dedicated branch](https://github.com/official-stockfish/Stockfish/tree/tools).
 On CPUs supporting modern vector instructions (avx2 and similar), the NNUE evaluation
-results in stronger playing strength, even if the nodes per second computed by the engine
+results in much stronger playing strength, even if the nodes per second computed by
-is somewhat lower (roughly 60% of nps is typical).
+the engine is somewhat lower (roughly 80% of nps is typical).
-Note that the NNUE evaluation depends on the Stockfish binary and the network parameter
+Notes:
 file (see EvalFile). Not every parameter file is compatible with a given Stockfish binary.
 The default value of the EvalFile UCI option is the name of a network that is guaranteed
 to be compatible with that binary.
-## What to expect from Syzygybases?
+1) the NNUE evaluation depends on the Stockfish binary and the network parameter file 
 (see the EvalFile UCI option). Not every parameter file is compatible with a given
 Stockfish binary, but the default value of the EvalFile UCI option is the name of a 
 network that is guaranteed to be compatible with that binary.
 2) to use the NNUE evaluation, the additional data file with neural network parameters
 needs to be available. Normally, this file is already embedded in the binary or it can
 be downloaded. The filename for the default (recommended) net can be found as the default
 value of the `EvalFile` UCI option, with the format `nn-[SHA256 first 12 digits].nnue`
 (for instance, `nn-c157e0a5755b.nnue`). This file can be downloaded from
 ```
 https://tests.stockfishchess.org/api/nn/[filename]
 ```
 replacing `[filename]` as needed.
 ## What to expect from the Syzygy tablebases?
 If the engine is searching a position that is not in the tablebases (e.g.
 a position with 8 pieces), it will access the tablebases during the search.
 If the engine reports a very large score (typically 153.xx), this means
-that it has found a winning line into a tablebase position.
+it has found a winning line into a tablebase position.
 If the engine is given a position to search that is in the tablebases, it
 will use the tablebases at the beginning of the search to preselect all
@@ -175,14 +255,14 @@ good moves, i.e. all moves that preserve the win or preserve the draw while
 taking into account the 50-move rule.
 It will then perform a search only on those moves. **The engine will not move
 immediately**, unless there is only a single good move. **The engine likely
-will not report a mate score even if the position is known to be won.**
+will not report a mate score, even if the position is known to be won.**
 It is therefore clear that this behaviour is not identical to what one might
 be used to with Nalimov tablebases. There are technical reasons for this
 difference, the main technical reason being that Nalimov tablebases use the
-DTM metric (distance-to-mate), while Syzygybases use a variation of the
+DTM metric (distance-to-mate), while the Syzygy tablebases use a variation of the
 DTZ metric (distance-to-zero, zero meaning any move that resets the 50-move
-counter). This special metric is one of the reasons that Syzygybases are
+counter). This special metric is one of the reasons that the Syzygy tablebases are
 more compact than Nalimov tablebases, while still storing all information
 needed for optimal play and in addition being able to take into account
 the 50-move rule.
@@ -191,8 +271,8 @@ the 50-move rule.
 Stockfish supports large pages on Linux and Windows. Large pages make
 the hash access more efficient, improving the engine speed, especially
-on large hash sizes. Typical increases are 5..10% in terms of nps, but
+on large hash sizes. Typical increases are 5..10% in terms of nodes per
-speed increases up to 30% have been measured. The support is
+second, but speed increases up to 30% have been measured. The support is
 automatic. Stockfish attempts to use large pages when available and
 will fall back to regular memory allocation when this is not the case.
@@ -200,17 +280,17 @@ will fall back to regular memory allocation when this is not the case.
 Large page support on Linux is obtained by the Linux kernel
 transparent huge pages functionality. Typically, transparent huge pages
-are already enabled and no configuration is needed.
+are already enabled, and no configuration is needed.
 ### Support on Windows
 The use of large pages requires "Lock Pages in Memory" privilege. See
 [Enable the Lock Pages in Memory Option (Windows)](https://docs.microsoft.com/en-us/sql/database-engine/configure-windows/enable-the-lock-pages-in-memory-option-windows)
-on how to enable this privilege. Logout/login may be needed
+on how to enable this privilege, then run [RAMMap](https://docs.microsoft.com/en-us/sysinternals/downloads/rammap)
-afterwards. Due to memory fragmentation, it may not always be
+to double-check that large pages are used. We suggest that you reboot
-possible to allocate large pages even when enabled. A reboot
+your computer after you have enabled large pages, because long Windows
-might alleviate this problem. To determine whether large pages
+sessions suffer from memory fragmentation, which may prevent Stockfish
-are in use, see the engine log.
+from getting large pages: a fresh session is better in this regard.
 ## Compiling Stockfish yourself from the sources
@@ -225,26 +305,26 @@ targets with corresponding descriptions.
 ```
    cd src
    make help
    make net
    make build ARCH=x86-64-modern
 ```
-When not using the Makefile to compile (for instance with Microsoft MSVC) you
+When not using the Makefile to compile (for instance, with Microsoft MSVC) you
 need to manually set/unset some switches in the compiler command line; see
 file *types.h* for a quick reference.
-When reporting an issue or a bug, please tell us which version and
+When reporting an issue or a bug, please tell us which Stockfish version
-compiler you used to create your executable. These informations can
+and which compiler you used to create your executable. This information
-be found by typing the following commands in a console:
+can be found by typing the following command in a console:
 ```
-    ./stockfish
+    ./stockfish compiler
    compiler
 ```
 ## Understanding the code base and participating in the project
-Stockfish's improvement over the last couple of years has been a great
+Stockfish's improvement over the last decade has been a great community
-community effort. There are a few ways to help contribute to its growth.
+effort. There are a few ways to help contribute to its growth.
 ### Donating hardware
@@ -265,8 +345,9 @@ generic rather than being focused on Stockfish's precise implementation.
 Nevertheless, a helpful resource.
 * The latest source can always be found on [GitHub](https://github.com/official-stockfish/Stockfish).
-Discussions about Stockfish take place in the [FishCooking](https://groups.google.com/forum/#!forum/fishcooking)
+Discussions about Stockfish take place these days mainly in the [FishCooking](https://groups.google.com/forum/#!forum/fishcooking)
-group and engine testing is done on [Fishtest](https://tests.stockfishchess.org/tests).
+group and on the [Stockfish Discord channel](https://discord.gg/nv8gDtt).
 The engine testing is done on [Fishtest](https://tests.stockfishchess.org/tests).
 If you want to help improve Stockfish, please read this [guideline](https://github.com/glinscott/fishtest/wiki/Creating-my-first-test)
 first, where the basics of Stockfish development are explained.
@@ -274,16 +355,17 @@ first, where the basics of Stockfish development are explained.
 ## Terms of use
 Stockfish is free, and distributed under the **GNU General Public License version 3**
-(GPL v3). Essentially, this means that you are free to do almost exactly
+(GPL v3). Essentially, this means you are free to do almost exactly
 what you want with the program, including distributing it among your
-friends, making it available for download from your web site, selling
+friends, making it available for download from your website, selling
 it (either by itself or as part of some bigger software package), or
 using it as the starting point for a software project of your own.
 The only real limitation is that whenever you distribute Stockfish in
-some way, you must always include the full source code, or a pointer
+some way, you MUST always include the license and the full source code
-to where the source code can be found. If you make any changes to the
+(or a pointer to where the source code can be found) to generate the 
-source code, these changes must also be made available under the GPL.
+exact binary you are distributing. If you make any changes to the
 source code, these changes must also be made available under the GPL v3.
 For full details, read the copy of the GPL v3 found in the file named
-*Copying.txt*.
+[*Copying.txt*](https://github.com/official-stockfish/Stockfish/blob/master/Copying.txt).
@@ -1,154 +1,235 @@
-Contributors with >10,000 CPU hours as of January 7, 2020
+Contributors to Fishtest with >10,000 CPU hours, as of 2022-04-14.
 Thank you!
-Username                  CPU Hours   Games played
+Username                                CPU Hours     Games played
--------------------------------------------------
+------------------------------------------------------------------
-noobpwnftw                  9305707      695548021
+noobpwnftw                               31714850       2267266129
-mlang                        780050       61648867
+mlang                                     2954099        198421098
-dew                          621626       43921547
+technologov                               2324150        102449398
-mibere                       524702       42238645
+dew                                       1670874         99276012
-crunchy                      354587       27344275
+grandphish2                               1134273         68070459
-cw                           354495       27274181
+okrout                                     901194         77738874
-fastgm                       332801       22804359
+TueRens                                    821388         50207666
-JojoM                        295750       20437451
+tvijlbrief                                 795993         51894442
-CSU_Dynasty                  262015       21828122
+pemo                                       744463         32486677
-Fisherman                    232181       18939229
+JojoM                                      724378         43660674
-ctoks                        218866       17622052
+mibere                                     703840         46867607
-glinscott                    201989       13780820
+linrock                                    626939         17408017
-tvijlbrief                   201204       15337115
+gvreuls                                    534079         34352532
-velislav                     188630       14348485
+cw                                         507221         34006775
-gvreuls                      187164       15149976
+fastgm                                     489749         29344518
-bking_US                     180289       11876016
+crunchy                                    427035         27344275
-nordlandia                   172076       13467830
+CSU_Dynasty                                424643         28525220
-leszek                       157152       11443978
+ctoks                                      415771         27364603
-Thanar                       148021       12365359
+oz                                         369200         27017658
-spams                        141975       10319326
+bcross                                     342642         23671289
-drabel                       138073       11121749
+Fisherman                                  327231         21829379
-vdv                          137850        9394330
+velislav                                   325670         20911076
-mgrabiak                     133578       10454324
+leszek                                     321295         19874113
-TueRens                      132485       10878471
+Dantist                                    274747         16910258
-bcross                       129683       11557084
+mgrabiak                                   237604         15418700
-marrco                       126078        9356740
+robal                                      217959         13840386
-sqrt2                        125830        9724586
+glinscott                                  217799         13780820
-robal                        122873        9593418
+nordlandia                                 211692         13484886
-vdbergh                      120766        8926915
+drabel                                     201967         13798360
-malala                       115926        8002293
+bking_US                                   198894         11876016
-CoffeeOne                    114241        5004100
+mhoram                                     194862         12261809
-dsmith                       113189        7570238
+Thanar                                     179852         12365359
-BrunoBanani                  104644        7436849
+vdv                                        175544          9904472
-Data                          92328        8220352
+spams                                      157128         10319326
-mhoram                        89333        6695109
+rpngn                                      154081          9652139
-davar                         87924        7009424
+marrco                                     150300          9402229
-xoto                          81094        6869316
+sqrt2                                      147963          9724586
-ElbertoOne                    80899        7023771
+vdbergh                                    137430          8955097
-grandphish2                   78067        6160199
+CoffeeOne                                  137100          5024116
-brabos                        77212        6186135
+malala                                     136182          8002293
-psk                           75733        5984901
+xoto                                       133759          9159372
-BRAVONE                       73875        5054681
+davar                                      125240          8117121
-sunu                          70771        5597972
+dsmith                                     122059          7570238
-sterni1971                    70605        5590573
+amicic                                     119659          7937885
-MaZePallas                    66886        5188978
+Data                                       113305          8220352
-Vizvezdenec                   63708        4967313
+BrunoBanani                                112960          7436849
-nssy                          63462        5259388
+CypressChess                               108321          7759588
-jromang                       61634        4940891
+DesolatedDodo                              106811          6776980
-teddybaer                     61231        5407666
+MaZePallas                                 102823          6633619
-Pking_cda                     60099        5293873
+sterni1971                                 100532          5880772
-solarlight                    57469        5028306
+sunu                                       100167          7040199
-dv8silencer                   56913        3883992
+ElbertoOne                                  99028          7023771
-tinker                        54936        4086118
+skiminki                                    98123          6478402
-renouve                       49732        3501516
+brabos                                      92118          6186135
-Freja                         49543        3733019
+cuistot                                     90358          5351004
-robnjr                        46972        4053117
+psk                                         89957          5984901
-rap                           46563        3219146
+racerschmacer                               85712          6119648
-Bobo1239                      46036        3817196
+Vizvezdenec                                 83761          5344740
-ttruscott                     45304        3649765
+zeryl                                       83680          5250995
-racerschmacer                 44881        3975413
+sschnee                                     83003          4840890
-finfish                       44764        3370515
+0x3C33                                      82614          5271253
-eva42                         41783        3599691
+BRAVONE                                     81239          5054681
-biffhero                      40263        3111352
+nssy                                        76497          5259388
-bigpen0r                      39817        3291647
+teddybaer                                   75125          5407666
-mhunt                         38871        2691355
+jromang                                     74796          5175825
-ronaldjerum                   38820        3240695
+Pking_cda                                   73776          5293873
-Antihistamine                 38785        2761312
+Calis007                                    72477          4088576
-pb00067                       38038        3086320
+solarlight                                  70517          5028306
-speedycpu                     37591        3003273
+dv8silencer                                 70287          3883992
-rkl                           37207        3289580
+Bobo1239                                    68515          4652287
-VoyagerOne                    37050        3441673
+manap                                       66273          4121774
-jbwiebe                       35320        2805433
+yurikvelo                                   65716          4457300
-cuistot                       34191        2146279
+tinker                                      64333          4268790
-homyur                        33927        2850481
+Wolfgang                                    62644          3817410
-manap                         32873        2327384
+qurashee                                    61208          3429862
-gri                           32538        2515779
+robnjr                                      57262          4053117
-oryx                          31267        2899051
+Freja                                       56938          3733019
-EthanOConnor                  30959        2090311
+ttruscott                                   56010          3680085
-SC                            30832        2730764
+rkl                                         55132          4164467
-csnodgrass                    29505        2688994
+renouve                                     53811          3501516
-jmdana                        29458        2205261
+megaman7de                                  52434          3243016
-strelock                      28219        2067805
+MaxKlaxxMiner                               51977          3153032
-jkiiski                       27832        1904470
+finfish                                     51360          3370515
-Pyafue                        27533        1902349
+eva42                                       51272          3599691
-Garf                          27515        2747562
+eastorwest                                  51058          3451555
-eastorwest                    27421        2317535
+rap                                         49985          3219146
-slakovv                       26903        2021889
+pb00067                                     49727          3298270
-Prcuvu                        24835        2170122
+Spprtr                                      48920          3161711
-anst                          24714        2190091
+bigpen0r                                    47667          3336927
-hyperbolic.tom                24319        2017394
+ronaldjerum                                 47654          3240695
-Patrick_G                     23687        1801617
+biffhero                                    46564          3111352
-Sharaf_DG                     22896        1786697
+Fifis                                       45843          3088497
-nabildanial                   22195        1519409
+VoyagerOne                                  45476          3452465
-chriswk                       21931        1868317
+speedycpu                                   43842          3003273
-achambord                     21665        1767323
+jbwiebe                                     43305          2805433
-Zirie                         20887        1472937
+Antihistamine                               41788          2761312
-team-oh                       20217        1636708
+mhunt                                       41735          2691355
-Isidor                        20096        1680691
+homyur                                      39893          2850481
-ncfish1                       19931        1520927
+gri                                         39871          2515779
-nesoneg                       19875        1463031
+armo9494                                    39064          2832326
-Spprtr                        19853        1548165
+oryx                                        38867          2976992
-JanErik                       19849        1703875
+SC                                          37299          2731694
-agg177                        19478        1395014
+Garf                                        37213          2986270
-SFTUser                       19231        1567999
+tolkki963                                   37059          2154330
-xor12                         19017        1680165
+csnodgrass                                  36207          2688994
-sg4032                        18431        1641865
+jmdana                                      36157          2210661
-rstoesser                     18118        1293588
+strelock                                    34716          2074055
-MazeOfGalious                 17917        1629593
+DMBK                                        34010          2482916
-j3corre                       17743         941444
+EthanOConnor                                33370          2090311
-cisco2015                     17725        1690126
+slakovv                                     32915          2021889
-ianh2105                      17706        1632562
+gopeto                                      30993          2028106
-dex                           17678        1467203
+manapbk                                     30987          1810399
-jundery                       17194        1115855
+Prcuvu                                      30377          2170122
-iisiraider                    17019        1101015
+anst                                        30301          2190091
-horst.prack                   17012        1465656
+jkiiski                                     30136          1904470
-Adrian.Schmidt123             16563        1281436
+hyperbolic.tom                              29840          2017394
-purplefishies                 16342        1092533
+chuckstablers                               29659          2093438
-wei                           16274        1745989
+Pyafue                                      29650          1902349
-ville                         16144        1384026
+ncfish1                                     29105          1704011
-eudhan                        15712        1283717
+belzedar94                                  27935          1789106
-OuaisBla                      15581         972000
+OuaisBla                                    27636          1578800
-DragonLord                    15559        1162790
+chriswk                                     26902          1868317
-dju                           14716         875569
+achambord                                   26582          1767323
-chris                         14479        1487385
+Patrick_G                                   26276          1801617
-0xB00B1ES                     14079        1001120
+yorkman                                     26193          1992080
-OssumOpossum                  13776        1007129
+SFTUser                                     25182          1675689
-enedene                       13460         905279
+nabildanial                                 24942          1519409
-bpfliegel                     13346         884523
+Sharaf_DG                                   24765          1786697
-Ente                          13198        1156722
+rodneyc                                     24275          1410450
-IgorLeMasson                  13087        1147232
+agg177                                      23890          1395014
-jpulman                       13000         870599
+JanErik                                     23408          1703875
-ako027ako                     12775        1173203
+Isidor                                      23388          1680691
-Nikolay.IT                    12352        1068349
+Norabor                                     23339          1602636
-Andrew Grant                  12327         895539
+Ente                                        23270          1651432
-joster                        12008         950160
+cisco2015                                   22897          1762669
-AdrianSA                      11996         804972
+MarcusTullius                               22688          1274821
-Nesa92                        11455        1111993
+Zirie                                       22542          1472937
-fatmurphy                     11345         853210
+team-oh                                     22272          1636708
-Dark_wizzie                   11108        1007152
+MazeOfGalious                               21978          1629593
-modolief                      10869         896470
+sg4032                                      21947          1643265
-mschmidt                      10757         803401
+ianh2105                                    21725          1632562
-infinity                      10594         727027
+xor12                                       21628          1680365
-mabichito                     10524         749391
+dex                                         21612          1467203
-Thomas A. Anderson            10474         732094
+nesoneg                                     21494          1463031
-thijsk                        10431         719357
+Roady                                       21323          1433822
-Flopzee                       10339         894821
+sphinx                                      21211          1384728
-crocogoat                     10104        1013854
+user213718                                  21196          1397710
-SapphireBrand                 10104         969604
+spcc                                        21065          1311338
-stocky                        10017         699440
+jjoshua2                                    21001          1423089
 horst.prack                                 20878          1465656
 0xB00B1ES                                   20590          1208666
 j3corre                                     20405           941444
 kdave                                       20364          1389254
 Adrian.Schmidt123                           20316          1281436
 Ulysses                                     20217          1351500
 markkulix                                   19976          1115258
 wei                                         19973          1745989
 rstoesser                                   19569          1293588
 eudhan                                      19274          1283717
 fishtester                                  18995          1238686
 vulcan                                      18871          1729392
 jundery                                     18445          1115855
 iisiraider                                  18247          1101015
 ville                                       17883          1384026
 chris                                       17698          1487385
 purplefishies                               17595          1092533
 dju                                         17353           978595
 Wencey                                      17125           805964
 DragonLord                                  17014          1162790
 thirdlife                                   16996           447356
 IgorLeMasson                                16064          1147232
 ako027ako                                   15671          1173203
 AndreasKrug                                 15550          1194497
 Nikolay.IT                                  15154          1068349
 Andrew Grant                                15114           895539
 scuzzi                                      14928           953313
 OssumOpossum                                14857          1007129
 Karby                                       14808           867120
 jsys14                                      14652           855642
 enedene                                     14476           905279
 bpfliegel                                   14298           884523
 mpx86                                       14019           759568
 jpulman                                     13982           870599
 crocogoat                                   13803          1117422
 joster                                      13794           950160
 Nesa92                                      13786          1114691
 mbeier                                      13650          1044928
 Hjax                                        13535           915487
 Dark_wizzie                                 13422          1007152
 Jopo12321                                   13367           678852
 Rudolphous                                  13244           883140
 Machariel                                   13010           863104
 mabichito                                   12903           749391
 thijsk                                      12886           722107
 AdrianSA                                    12860           804972
 infinigon                                   12807           937332
 Flopzee                                     12698           894821
 fatmurphy                                   12547           853210
 SapphireBrand                               12416           969604
 modolief                                    12386           896470
 Farseer                                     12249           694108
 pgontarz                                    12151           848794
 pirt                                        12008           923149
 stocky                                      11954           699440
 mschmidt                                    11941           803401
 dbernier                                    11609           818636
 Maxim                                       11543           836024
 infinity                                    11470           727027
 aga                                         11409           695071
 torbjo                                      11395           729145
 Thomas A. Anderson                          11372           732094
 savage84                                    11358           670860
 FormazChar                                  11349           850327
 d64                                         11263           789184
 MooTheCow                                   11237           720174
 snicolet                                    11106           869170
 ali-al-zhrani                               11098           768494
 whelanh                                     11067           235676
 Jackfish                                    10978           720078
 deflectooor                                 10886           520116
 basepi                                      10637           744851
 Cubox                                       10621           826448
 michaelrpg                                  10509           739239
 OIVAS7572                                   10420           995586
 dzjp                                        10343           732529
 Garruk                                      10334           704065
 ols                                         10259           570669
 lbraesch                                    10252           647825
 qoo_charly_cai                              10212           620407
 Naven94                                     10069           503192
@@ -1,75 +0,0 @@
 version: 1.0.{build}
 clone_depth: 50
 branches:
  only:
    - master
    - nnue-player-wip
 # Operating system (build VM template)
 os: Visual Studio 2019
 # Build platform, i.e. x86, x64, AnyCPU. This setting is optional.
 platform:
  - x86
  - x64
 # build Configuration, i.e. Debug, Release, etc.
 configuration:
  - Debug
  - Release
 matrix:
  # The build fail immediately once one of the job fails
  fast_finish: true
 # Scripts that are called at very beginning, before repo cloning
 init:
  - cmake --version
  - msbuild /version
 before_build:
  - ps: |
      # Get sources
      $src = get-childitem -Path *.cpp -Recurse | select -ExpandProperty FullName
      $src = $src -join ' '
      $src = $src.Replace("\", "/")
      # Build CMakeLists.txt
      $t = 'cmake_minimum_required(VERSION 3.17)',
           'project(Stockfish)',
           'set(CMAKE_CXX_STANDARD 17)',
           'set(CMAKE_CXX_STANDARD_REQUIRED ON)',
           'set (CMAKE_CXX_EXTENSIONS OFF)',
           'set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_SOURCE_DIR}/src)',
           'set(source_files', $src, ')',
           'add_executable(stockfish ${source_files})'
      # Write CMakeLists.txt withouth BOM
      $MyPath = (Get-Item -Path "." -Verbose).FullName + '\CMakeLists.txt'
      $Utf8NoBomEncoding = New-Object System.Text.UTF8Encoding $False
      [System.IO.File]::WriteAllLines($MyPath, $t, $Utf8NoBomEncoding)
      # Obtain bench reference from git log
      $b = git log HEAD | sls "\b[Bb]ench[ :]+[0-9]{7}" | select -first 1
      $bench = $b -match '\D+(\d+)' | % { $matches[1] }
      Write-Host "Reference bench:" $bench
      $g = "Visual Studio 16 2019"
      If (${env:PLATFORM} -eq 'x64') { $a = "x64" }
      If (${env:PLATFORM} -eq 'x86') { $a = "Win32" }
      cmake -G "${g}" -A ${a} .
      Write-Host "Generated files for: " $g $a
 build_script:
  - cmake --build . --config %CONFIGURATION% -- /verbosity:minimal
 before_test:
  - cd src/%CONFIGURATION%
  - stockfish bench 2> out.txt >NUL
  - ps: |
      # Verify bench number
      $s = (gc "./out.txt" | out-string)
      $r = ($s -match 'Nodes searched \D+(\d+)' | % { $matches[1] })
      Write-Host "Engine bench:" $r
      Write-Host "Reference bench:" $bench
      If ($r -ne $bench) { exit 1 }
@@ -0,0 +1,42 @@
 # Binpack
 Binpack is a binary training data storage format designed to take advantage of position chains differing by a single move. Therefore it is very good at compactly storing data generated from real games (as opposed to random positions for example sourced from an opening book).
 It is currently implemented through a single header library in `extra/nnue_data_binpack_format.h`.
 Below follows a rough description of the format in a BNF-like notation.
 ```
 [[nodiscard]] std::uint16_t signedToUnsigned(std::int16_t a) {
    std::uint16_t r;
    std::memcpy(&r, &a, sizeof(std::uint16_t));
    if (r & 0x8000) r ^= 0x7FFF; // flip value bits if negative
    r = (r << 1) | (r >> 15); // store sign bit at bit 0
    return r;
 }
 file := <block>*
 block := BINP<chain>*
 chain := <stem><movetext>
 stem := <pos><move><score><ply_and_result><rule50> (32 bytes)
 pos := https://github.com/Sopel97/nnue_data_compress/blob/master/src/chess/Position.h#L1166 (24 bytes)
 move := https://github.com/Sopel97/nnue_data_compress/blob/master/src/chess/Chess.h#L1044 (2 bytes)
 score := signedToUnsigned(score) (2 bytes, big endian)
 ply_and_result := ply bitwise_or (signedToUnsigned(result) << 14) (2 bytes, big endian)
 rule50 := rule_50_counter (2 bytes, big endian)
    // this is a small defect from old version,
    I didn't want to break backwards compatibility. Effectively means that there's
    one byte left for something else in the future because rule50 always fits in one byte.
 movetext := <count><move_and_score>*
 count := number of plies in the movetext (2 bytes, big endian). Can be 0.
 move_and_score := <encoded_move><encoded_score> (~2 bytes)
 encoded_move := oof this one is complicated to explain.
    https://github.com/Sopel97/nnue_data_compress/blob/master/src/compress_file.cpp#L827.
    https://github.com/Sopel97/chess_pos_db/blob/master/docs/bcgn/variable_length.md
 encoded_score := https://en.wikipedia.org/wiki/Variable-width_encoding
    with block size of 4 bits + 1 bit for extension bit.
    Encoded value is signedToUnsigned(-prev_score - current_score)
    (scores are always seen from the perspective of side to move in <pos>, that's why the '-' before prev_score)
 ```
@@ -0,0 +1,18 @@
 # Convert
 `convert` allows conversion of training data between any of `.plain`, `.bin`, and `.binpack`.
 As all commands in stockfish `convert` can be invoked either from command line (as `stockfish.exe convert ...`) or in the interactive prompt.
 The syntax of this command is as follows:
 ```
 convert from_path to_path [append] [validate]
 ```
 `from_path` is the path to the file to convert from. The type of the data is deduced based on its extension (one of `.plain`, `.bin`, `.binpack`).
 `to_path` is the path to an output file. The type of the data is deduced from its extension. If the file does not exist it is created.
 `append` and `validate` can come in any order and are optional.
 If `append` not specified then the output file will be truncated prior to any writes. If `append` is specified then the converted training data will be appended to the end of the output file.
 If `validate` is specified then the conversion will stop on the first illegal move found and a diagnostic will be shown.
@@ -0,0 +1,67 @@
 # generate_training_data
 `generate_training_data` command allows generation of training data from self-play in a manner that suits training better than traditional games. It introduces random moves to diversify openings, and fixed depth evaluation.
 As all commands in stockfish `generate_training_data` can be invoked either from command line (as `stockfish.exe generate_training_data ...`, but this is not recommended because it's not possible to specify UCI options before `generate_training_data` executes) or in the interactive prompt.
 It is recommended to set the `PruneAtShallowDepth` UCI option to `false` as it will increase the quality of fixed depth searches.
 It is recommended to keep the `EnableTranspositionTable` UCI option at the default `true` value as it will make the generation process faster without noticably harming the uniformity of the data.
 `generate_training_data` takes named parameters in the form of `generate_training_data param_1_name param_1_value param_2_name param_2_value ...`.
 Currently the following options are available:
 `set_recommended_uci_options` - this is a modifier not a parameter, no value follows it. If specified then some UCI options are set to recommended values.
 `depth` - sets minimum and maximum depth of evaluation of each position. Default: 3.
 `mindepth` - minimum depth of evaluation of each position. If not specified then the same as `depth`.
 `maxdepth` - minimum depth of evaluation of each position. If not specified then the same as `depth`.
 `nodes` - the number of nodes to use for evaluation of each position. This number is multiplied by the number of PVs of the current search. This does NOT override the `depth` and `depth2` options. If specified then whichever of depth or nodes limit is reached first applies.
 `count` - the number of training data entries to generate. 1 entry == 1 position. If both `count` and `max_time_*` are specified the data generation process will end when any of conditions is fullfilled. Default: 8000000000 (8B).
 `max_time_seconds`, `max_time_minutes`, `max_time_hours` - specifies the maximum runtime for the data generation. The data generation will NOT be interrupted while a self-play game is in progress. If both `count` and `max_time_*` are specified the data generation process will end when any of conditions is fullfilled. Default: \~250 years.
 `output_file_name` - the name of the file to output to. If the extension is not present or doesn't match the selected training data format the right extension will be appened. Default: generated_kifu
 `eval_limit` - evaluations with higher absolute value than this will not be written and will terminate a self-play game. Should not exceed 10000 which is VALUE_KNOWN_WIN, but is only hardcapped at mate in 2 (\~30000). Default: 3000
 `random_move_min_ply` - the minimal ply at which a random move may be executed instead of a move chosen by search. Default: 1.
 `random_move_max_ply` - the maximal ply at which a random move may be executed instead of a move chosen by search. Default: 24.
 `random_move_count` - maximum number of random moves in a single self-play game. Default: 5.
 `random_move_like_apery` - either 0 or 1. If 1 then random king moves will be followed by a random king move from the opponent whenever possible with 50% probability. Default: 0.
 `random_multi_pv` - the number of PVs used for determining the random move. If not specified then a truly random move will be chosen. If specified then a multiPV search will be performed the random move will be one of the moves chosen by the search.
 `random_multi_pv_diff` - Makes the multiPV random move selection consider only moves that are at most `random_multi_pv_diff` worse than the next best move. Default: 30000 (all multiPV moves).
 `random_multi_pv_depth` - the depth to use for multiPV search for random move. Default: `depth2`.
 `random_multi_pv_nodes` - the maximum number of nodes for a multiPV search for random move. Default: `nodes`.
 `write_min_ply` - minimum ply for which the training data entry will be emitted. Default: 16.
 `write_max_ply` - maximum ply for which the training data entry will be emitted. Default: 400.
 `book` - a path to an opening book to use for the starting positions. Currently only .epd format is supported. If not specified then the starting position is always the standard chess starting position.
 `save_every` - the number of training data entries per file. If not specified then there will be always one file. If specified there may be more than one file generated (each having at most `save_every` training data entries) and each file will have a unique number attached.
 `random_file_name` - if specified then the output filename will be chosen randomly. Overrides `output_file_name`.
 `keep_draws` - either 0 or 1. If 1 then training data from drawn games will be emitted too. Default: 1.
 `adjudicate_draws_by_score` - either 0 or 1. If 1 then drawn games will be adjudicated when the score remains 0 for at least 8 plies after ply 80. Default: 1.
 `adjudicate_draws_by_insufficient_mating_material` - either 0 or 1. If 1 then position with insufficient material will be adjudicated as draws. Default: 1.
 `data_format` - format of the training data to use. Either `bin` or `binpack`. Default: `binpack`.
 `seed` - seed for the PRNG. Can be either a number or a string. If it's a string then its hash will be used. If not specified then the current time will be used.
@@ -0,0 +1,41 @@
 # generate_training_data_nonpv
 `generate_training_data_nonpv` command allows generation of training data from self-play in a manner that suits training better than traditional games. It plays fixed nodes self play games for exploration and records [some of] the evaluated positions. Then rescores them with fixed depth search.
 As all commands in stockfish `generate_training_data_nonpv` can be invoked either from command line (as `stockfish.exe generate_training_data_nonpv ...`, but this is not recommended because it's not possible to specify UCI options before `generate_training_data_nonpv` executes) or in the interactive prompt.
 It is recommended to set the `PruneAtShallowDepth` UCI option to `false` as it will increase the quality of fixed depth searches.
 It is recommended to keep the `EnableTranspositionTable` UCI option at the default `true` value as it will make the generation process faster without noticably harming the uniformity of the data.
 `generate_training_data_nonpv` takes named parameters in the form of `generate_training_data_nonpv param_1_name param_1_value param_2_name param_2_value ...`.
 Currently the following options are available:
 `depth` - the search depth to use for rescoring. Default: 3.
 `count` - the number of training data entries to generate. 1 entry == 1 position. Default: 1000000 (1M).
 `exploration_min_nodes` - the min number of nodes to use for exploraton during selfplay. Default: 5000.
 `exploration_max_nodes` - the max number of nodes to use for exploraton during selfplay. The number of nodes is chosen from a uniform distribution between min and max. Default: 15000.
 `exploration_save_rate` - the ratio of positions seen during exploration self play games that are saved for later rescoring. Default: 0.01 (meaning 1 in 100 positions seen during search get saved for rescoring).
 `output_file` - the name of the file to output to. If the extension is not present or doesn't match the selected training data format the right extension will be appened. Default: generated_gensfen_nonpv
 `eval_limit` - evaluations with higher absolute value than this will not be written and will terminate a self-play game. Should not exceed 10000 which is VALUE_KNOWN_WIN, but is only hardcapped at mate in 2 (\~30000). Default: 4000
 `exploration_eval_limit` - same as `eval_limit` but used during exploration with a value from fixed depth search.
 `exploration_min_pieces` - the min number of pieces in the self play games to start the fixed depth search. Note that even if there's N pieces on the board the fixed nodes search usually reaches positions with less pieces and they are saved too. Default: 8.
 `exploration_max_ply` the max ply for the exploration self play. Default: 200.
 `smart_fen_skipping` - this is a flag option. When specified some position that are not good candidates for teaching are removed from the output. This includes positions where the best move is a capture or promotion, and position where a king is in check.
 `book` - a path to an opening book to use for the starting positions. Currently only .epd format is supported. If not specified then the starting position is always the standard chess starting position.
 `data_format` - format of the training data to use. Either `bin` or `binpack`. Default: `binpack`.
 `seed` - seed for the PRNG. Can be either a number or a string. If it's a string then its hash will be used. If not specified then the current time will be used.
@@ -0,0 +1,41 @@
 # Stats
 `gather_statistics` command allows gathering various statistics from a .bin or a .binpack file. The syntax is `gather_statistics (GROUP)* input_file FILENAME`. There can be many groups specified. Any statistic gatherer that belongs to at least one of the specified groups will be used.
 Simplest usage: `stockfish.exe gather_statistics all input_file a.binpack`
 Any name that doesn't designate an argument name or is not an argument will be interpreted as a group name.
 ## Parameters
 `input_file` - the path to the .bin or .binpack input file to read
 `output_file` - optional path to the output file to write the results too. Results are always written on the console, so if this is specified the results will be written in both places.
 `max_count` - the maximum number of positions to process. Default: no limit.
 ## Groups
 `all` - a special group designating all statistics gatherers available.
 `position_count` - the total number of positions in the file.
 `king`, `king_square_count` - the number of times a king was on each square. Output is layed out as a chessboard, with the 8th rank being the topmost. Separate values for white and black kings.
 `move`, `move_from_count` - same as `king_square_count` but for from_sq(move)
 `move`, `move_to_count` - same as `king_square_count` but for to_sq(move)
 `move`, `move_type` - the number of moves with each type. Includes normal, captures, castling, promotions, enpassant. The groups are not disjoint.
 `move`, `moved_piece_type` - the number of times a piece of each type was moved
 `piece_count` - the histogram of the number of pieces on the board
 `ply_discontinuities` - the number of times the ply jumped by a value different than 1 between two consecutive positions. Usually the number of games.
 `material_imbalance` - the histogram of imbalances, with values computed using "simple eval", that is pawn=1, bishop=knight=3, rook=5, queen=9
 `results` - the distribution of game results
 `endgames_6man` - distribution of endgame configurations for <=6 pieces (including kings)
@@ -0,0 +1,42 @@
 # Transform
 `transform` command exposes subcommands that perform some specific transformation over data. The call syntax is `transform <subcommand>`. Currently implemented subcommands are listed and described below.
 ## `nudged_static`
 `transform nudged_static` takes named parameters in the form of `transform nudged_static param_1_name param_1_value param_2_name param_2_value ...` and flag parameters which don't require values.
 This command goes through positions in the input files and replaces the scores with new ones - generated from static eval - but slightly adjusted based on the scores in the original input file.
 Currently the following options are available:
 `input_file` - path to the input file. Supports bin and binpack formats. Default: in.binpack.
 `output_file` - path to the output file. Supports bin and binpack formats. Default: out.binpack.
 `absolute` - states that the adjustment should be bounded by an absolute value. After this token follows the maximum absolute adjustment. Values are always adjusted towards scores in the input file. This is the default mode. Default maximum adjustement: 5.
 `relative` - states that the adjustment should be bounded by a value relative in magnitude to the static eval value. After this token follows the maximum relative change - a floating point value greater than 0. For example a value of 0.1 only allows changing the static eval by at most 10% towards the score from the input file.
 `interpolate` states that the output score should be a value interpolated between static eval and the score from the input file. After this token follows the interpolation constant `t`. `t` of 0 means that only static eval is used. `t` of 1 means that only score from the input file is used. `t` of 0.5 means that the static eval and input score are averaged. It accepts values outside of range `<0, 1>`, but the usefulness is questionable.
 ## `rescore`
 `transform rescore` takes named parameters in the form of `transform rescore param_1_name param_1_value param_2_name param_2_value ...` and flag parameters which don't require values.
 This tool respects the UCI option `Threads` and uses all available threads.
 This command takes a path to the input file that is either a .epd file which contains one FEN per line or a .bin or .binpack file and outputs a .bin or .binpack file with these positions rescored with specified depth search.
 Currently the following options are available:
 `input_file` - path to the input file. Default: in.binpack.
 `output_file` - path to the output .bin or .binpack file. The file is opened in append mode. Default: out.binpack.
 `depth` - the search depth to use for rescoring. Default: 3.
 `keep_moves` - whether to keep moves from the input file if available. Allows to keep compression in .binpack. Default: 1.
 `research_count` - number of additional searches of depth N done on the same position before using the eval. Default: 0.
@@ -0,0 +1,12 @@
 # validate_training_data
 `validate_training_data` allows validation of training data of types `.plain`, `.bin`, and `.binpack`.
 As all commands in stockfish `validate_training_data` can be invoked either from command line (as `stockfish.exe validate_training_data ...`) or in the interactive prompt.
 The syntax of this command is as follows:
 ```
 validate_training_data in_path
 ```
 `in_path` is the path to the file to validate. The type of the data is deduced based on its extension (one of `.plain`, `.bin`, `.binpack`).
@@ -0,0 +1,52 @@
 # `pgn_to_plain`
 This script converts pgn files into text file to apply `learn convert_bin` command. You need to import [python-chess](https://pypi.org/project/python-chess/) to use this script.
    pip install python-chess
 # Example of Qhapaq's finetune using `pgn_to_plain`
 ## Download data
 You can download data from [here](http://rebel13.nl/index.html)
 ## Convert pgn files
 **Important : convert text will be superheavy (approx 200 byte / position)**
    python pgn_to_plain.py --pgn "pgn/*.pgn" --start_ply 1 --output converted_pgn.txt
 `--pgn` option supports wildcard. When you use pgn files with elo >= 3300, You will get 1.7 GB text file.
 ## Convert into training data
 ### Example build command
    make nnue-learn ARCH=x86-64
 See `src/Makefile` for detail.
 ### Convert
    ./stockfish
    learn convert_bin converted_pgn.txt output_file_name pgn_bin.bin
 	learn shuffle pgn_bin.bin
 You also need to prepare validation data for training like following.
 	python pgn_to_plain.py --pgn "pgn/ccrl-40-15-3400.pgn" --start_ply 1 --output ccrl-40-15-3400.txt
 	./stockfish
    learn convert_bin ccrl-40-15-3400.txt ccrl-40-15-3400_plain.bin
 ### Learn
    ./stockfish
 	setoption name Threads value 8
    learn shuffled_sfen.bin newbob_decay 0.5  validation_set_file_name ccrl-40-15-3400_plain.bin  nn_batch_size 50000 batchsize 1000000 eval_save_interval 8000000 eta 0.05 lambda 0.0 eval_limit 3000 mirror_percentage 0 use_draw_in_training 1
@@ -0,0 +1,42 @@
 import sys
 ENTRY_SIZE = 40
 NUM_ENTRIES_IN_CHUNK = 1024*1024
 def copy(infile, outfile, count, times):
    if times > 1:
        outfile.write(infile.read(count*ENTRY_SIZE)*times)
    else:
        offset = 0
        while offset < count:
            to_read = NUM_ENTRIES_IN_CHUNK if offset + NUM_ENTRIES_IN_CHUNK <= count else count - offset
            outfile.write(infile.read(to_read*ENTRY_SIZE))
            offset += NUM_ENTRIES_IN_CHUNK
 def work():
    filename = sys.argv[1]
    offset = int(sys.argv[2])
    count = int(sys.argv[3])
    times = int(sys.argv[4]) if len(sys.argv) >= 5 else 1
    with open(filename, 'rb') as infile:
        infile.seek(offset * ENTRY_SIZE)
        filename_parts = filename.split('.')
        out_path = '.'.join(filename_parts[:-1]) + '_' + str(offset) + '_' + str(count) + '_' + str(times) + '.' + filename_parts[-1]
        with open(out_path, 'wb') as outfile:
            copy(infile, outfile, count, times)
 def show_help():
    print('Usage: python extract_bin.py filename offset count [times]')
    print('filename - the path to the .bin file to process')
    print('offset - the number of sfens to skip')
    print('count - the number of sfens to extract')
    print('times - the number of times to repeat the extracted sfens. Default = 1')
    print('The result is saved in a new file named `filename.stem`_`offset`_`count`_`times`.bin')
 if len(sys.argv) < 4:
    show_help()
 else:
    work()
@@ -0,0 +1,86 @@
 import struct
 import sys
 import os
 import random
 from pathlib import Path
 def copy_next_chunk(in_file, out_file):
    chunk_header = in_file.read(8)
    assert chunk_header[0:4] == b"BINP"
    size = struct.unpack("<I", chunk_header[4:])[0]
    out_file.write(chunk_header)
    data = in_file.read(size)
    out_file.write(data)
    return size + 8
 def main():
    if len(sys.argv) < 4:
        print("Usage: python interleave_binpacks.py infile1 ... infileN outfile")
        print("       The output binpack, will contain all data from the input files.")
        print("       Data is read sequentially from the input, randomly alternating between files.")
        return
    # open last arg as output file name
    out_filename = sys.argv[-1]
    print("outfile: ", out_filename)
    if Path(out_filename).exists():
        print(
            "Output path {} already exists. Please specify a path to a file that does not exist.".format(
                out_filename
            )
        )
        return
    out_file = open(out_filename, "wb")
    # open other args as input file names, and get their sizes
    in_filenames = []
    for i in range(1, len(sys.argv) - 1):
        in_filenames.append(sys.argv[i])
    print("infiles: ", in_filenames)
    in_files = []
    in_files_remaining = []
    for in_filename in in_filenames:
        in_file = open(in_filename, "rb")
        in_files.append(in_file)
        file_size = os.path.getsize(in_filename)
        in_files_remaining.append(file_size)
    # randomly pick a file, with a probability related to their sizes.
    # copy from the front and keep track of remaining sizes
    total_remaining = sum(in_files_remaining)
    print("Merging {} bytes ".format(total_remaining))
    total_size = 0
    report_every = 100
    prev_mib = -report_every
    while total_remaining > 0:
        where = random.randrange(total_remaining)
        i = 0
        while where >= in_files_remaining[i]:
            where -= in_files_remaining[i]
            i += 1
        size = copy_next_chunk(in_files[i], out_file)
        in_files_remaining[i] -= size
        total_remaining -= size
        total_size += size
        mib = total_size // 1024 // 1024
        if mib // 100 != prev_mib // 100:
            print("Copied {} MiB".format(mib))
            prev_mib = mib
    out_file.close()
    for in_file in in_files:
        in_file.close()
    print("Merged  {} bytes".format(total_size))
 main()
@@ -0,0 +1,110 @@
 import chess.pgn
 import argparse
 import glob
 import re
 from typing import List
 # todo close in c++ tools using pgn-extract
 # https://www.cs.kent.ac.uk/people/staff/djb/pgn-extract/help.html#-w
 commentRe = re.compile("([+-]*M*[0-9.]*)/([0-9]*)")
 mateRe = re.compile("([+-])M([0-9]*)")
 flip_black = False
 def parse_result(result_str:str, board:chess.Board) -> int:
    if result_str == "1/2-1/2":
        return 0
    if result_str == "0-1":
        if board.turn == chess.WHITE:
            return -1
        else:
            return 1
    elif result_str == "1-0":
        if board.turn == chess.WHITE:
            return 1
        else:
            return -1
    else:
        print("illegal result", result_str)
        raise ValueError
 def game_sanity_check(game: chess.pgn.Game) -> bool:
    if not game.headers["Result"] in ["1/2-1/2", "0-1", "1-0"]:
        print("invalid result", game.headers["Result"])
        return False
    return True
 def parse_comment_for_score(comment_str: str, board: chess.Board) -> int:
    global commentRe
    global mateRe
    global flip_black
    try:
      m = commentRe.search(comment_str)
      if m:
         score = m.group(1)
         # depth = int(m.group(2))
         m = mateRe.search(score)
         if m:
            if m.group(1) == "+":
               score =  32000 - int(m.group(2))
            else:
               score = -32000 + int(m.group(2))
         else:
            score = int(float(score) * 208) # pawn to SF PawnValueEg
         if flip_black and board.turn == chess.BLACK:
            score = -score
      else:
         score = 0
    except:
      score = 0
    return score
 def parse_game(game: chess.pgn.Game, writer, start_play: int=1)->None:
    board: chess.Board = game.board()
    if not game_sanity_check(game):
        return
    result: str = game.headers["Result"]
    ply = 0
    for node in game.mainline():
        move = node.move
        if ply >= start_play:
            comment: str = node.comment
            writer.write("fen " + board.fen() + "\n")
            writer.write("move " + str(move) + "\n")
            writer.write("score " + str(parse_comment_for_score(comment, board)) + "\n")
            writer.write("ply " + str(ply)+"\n")
            writer.write("result " + str(parse_result(result, board)) +"\n")
            writer.write("e\n")
        ply += 1
        board.push(move)
 def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--pgn", type=str, required=True)
    parser.add_argument("--start_ply", type=int, default=1)
    parser.add_argument("--output", type=str, default="plain.txt")
    parser.add_argument("--flip_black_score", action='store_true', dest='flip_black_score', help="Flip black score. Default: False")
    args = parser.parse_args()
    global flip_black
    flip_black = args.flip_black_score
    pgn_files: List[str] = glob.glob(args.pgn)
    pgn_files = sorted(pgn_files, key=lambda x:float(re.findall("-(\d+).pgn",x)[0] if re.findall("-(\d+).pgn",x) else 0.0))
    f = open(args.output, 'w')
    for pgn_file in pgn_files:
        print("parse", pgn_file)
        pgn_loader = open(pgn_file)
        while True:
            game = chess.pgn.read_game(pgn_loader)
            if game is None:
                break
            parse_game(game, f, args.start_ply)
    f.close()
 if __name__=="__main__":
    main()
@@ -0,0 +1,89 @@
 import struct
 import sys
 import os
 import random
 from pathlib import Path
 def index_binpack(file):
    print('Indexing...')
    index = []
    offset = 0
    report_every = 100
    prev_mib = -report_every
    while file.peek():
        chunk_header = file.read(8)
        assert chunk_header[0:4] == b'BINP'
        size = struct.unpack('<I', chunk_header[4:])[0]
        file.seek(size, os.SEEK_CUR)
        index.append((offset, size + 8))
        offset += size + 8
        mib = offset // 1024 // 1024
        if mib // 100 != prev_mib // 100:
            print('Indexed {} MiB'.format(mib))
            prev_mib = mib
    return index
 def copy_binpack_indexed(in_file, index, out_files):
    print('Copying...')
    total_size = 0
    report_every = 100
    prev_mib = -report_every
    nextfile = 0
    for offset, size in index:
        in_file.seek(offset, os.SEEK_SET)
        data = in_file.read(size)
        assert len(data) == size
        out_files[nextfile].write(data)
        nextfile = (nextfile + 1) % len(out_files)
        total_size += size
        mib = total_size // 1024 // 1024
        if mib // 100 != prev_mib // 100:
            print('Copied {} MiB'.format(mib))
            prev_mib = mib
 def main():
    if len(sys.argv) < 3:
        print('Usage: python shuffle_binpack.py infile outfile [split_count]')
        return
    in_filename = sys.argv[1]
    if len(sys.argv) > 3:
       # split the infile in split_count pieces, creating new outfile names based on the provided name
       basefile = sys.argv[2]
       split_count = int(sys.argv[3])
       base=os.path.splitext(basefile)[0]
       ext=os.path.splitext(basefile)[1]
       out_filenames = []
       for i in range(split_count):
           out_filenames.append(base+"_{}".format(i)+ext)
    else:
       out_filenames = [sys.argv[2]]
    for out_filename in out_filenames:
      if (Path(out_filename).exists()):
          print('Output path {} already exists. Please specify a path to a file that does not exist.'.format(out_filename))
          return
    print(out_filenames)
    in_file = open(in_filename, 'rb')
    index = index_binpack(in_file)
    print('Shuffling...')
    random.shuffle(index)
    out_files = []
    for out_filename in out_filenames:
        out_files.append(open(out_filename, 'wb'))
    copy_binpack_indexed(in_file, index, out_files)
    in_file.close()
    for out_file in out_files:
        out_file.close()
 main()
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -87,16 +87,20 @@ const vector<string> Defaults = {
  // Chess 960
  "setoption name UCI_Chess960 value true",
  "bbqnnrkr/pppppppp/8/8/8/8/PPPPPPPP/BBQNNRKR w HFhf - 0 1 moves g2g3 d7d5 d2d4 c8h3 c1g5 e8d6 g5e7 f7f6",
  "nqbnrkrb/pppppppp/8/8/8/8/PPPPPPPP/NQBNRKRB w KQkq - 0 1",
  "setoption name UCI_Chess960 value false"
 };
 } // namespace
 namespace Stockfish {
 /// setup_bench() builds a list of UCI commands to be run by bench. There
 /// are five parameters: TT size in MB, number of search threads that
 /// should be used, the limit value spent for each position, a file name
-/// where to look for positions in FEN format and the type of the limit:
+/// where to look for positions in FEN format, the type of the limit:
-/// depth, perft, nodes and movetime (in millisecs).
+/// depth, perft, nodes and movetime (in millisecs), and evaluation type
 /// mixed (default), classical, NNUE.
 ///
 /// bench -> search default positions up to depth 13
 /// bench 64 1 15 -> search default positions up to depth 15 (TT = 64MB)
@@ -115,6 +119,7 @@ vector<string> setup_bench(const Position& current, istream& is) {
  string limit     = (is >> token) ? token : "13";
  string fenFile   = (is >> token) ? token : "default";
  string limitType = (is >> token) ? token : "depth";
  string evalType  = (is >> token) ? token : "mixed";
  go = limitType == "eval" ? "eval" : "go " + limitType + " " + limit;
@@ -146,14 +151,25 @@ vector<string> setup_bench(const Position& current, istream& is) {
  list.emplace_back("setoption name Hash value " + ttSize);
  list.emplace_back("ucinewgame");
  size_t posCounter = 0;
  for (const string& fen : fens)
      if (fen.find("setoption") != string::npos)
          list.emplace_back(fen);
      else
      {
          if (evalType == "classical" || (evalType == "mixed" && posCounter % 2 == 0))
              list.emplace_back("setoption name Use NNUE value false");
          else if (evalType == "NNUE" || (evalType == "mixed" && posCounter % 2 != 0))
              list.emplace_back("setoption name Use NNUE value true");
          list.emplace_back("position fen " + fen);
          list.emplace_back(go);
          ++posCounter;
      }
  list.emplace_back("setoption name Use NNUE value true");
  return list;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,6 +23,8 @@
 #include "bitboard.h"
 #include "types.h"
 namespace Stockfish {
 namespace {
  // There are 24 possible pawn squares: files A to D and ranks from 2 to 7.
@@ -66,7 +68,6 @@ namespace {
 } // namespace
 bool Bitbases::probe(Square wksq, Square wpsq, Square bksq, Color stm) {
  assert(file_of(wpsq) <= FILE_D);
@@ -96,7 +97,6 @@ void Bitbases::init() {
          KPKBitbase.set(idx);
 }
 namespace {
  KPKPosition::KPKPosition(unsigned idx) {
@@ -150,8 +150,8 @@ namespace {
    Bitboard b = attacks_bb<KING>(ksq[stm]);
    while (b)
-        r |= stm == WHITE ? db[index(BLACK, ksq[BLACK] , pop_lsb(&b), psq)]
+        r |= stm == WHITE ? db[index(BLACK, ksq[BLACK], pop_lsb(b), psq)]
-                          : db[index(WHITE, pop_lsb(&b),  ksq[WHITE], psq)];
+                          : db[index(WHITE, pop_lsb(b), ksq[WHITE], psq)];
    if (stm == WHITE)
    {
@@ -168,3 +168,5 @@ namespace {
  }
 } // namespace
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -22,11 +22,14 @@
 #include "bitboard.h"
 #include "misc.h"
 namespace Stockfish {
 uint8_t PopCnt16[1 << 16];
 uint8_t SquareDistance[SQUARE_NB][SQUARE_NB];
 Bitboard SquareBB[SQUARE_NB];
 Bitboard LineBB[SQUARE_NB][SQUARE_NB];
 Bitboard BetweenBB[SQUARE_NB][SQUARE_NB];
 Bitboard PseudoAttacks[PIECE_TYPE_NB][SQUARE_NB];
 Bitboard PawnAttacks[COLOR_NB][SQUARE_NB];
@@ -39,13 +42,22 @@ namespace {
  Bitboard BishopTable[0x1480]; // To store bishop attacks
  void init_magics(PieceType pt, Bitboard table[], Magic magics[]);
 }
 /// safe_destination() returns the bitboard of target square for the given step
 /// from the given square. If the step is off the board, returns empty bitboard.
 inline Bitboard safe_destination(Square s, int step) {
    Square to = Square(s + step);
    return is_ok(to) && distance(s, to) <= 2 ? square_bb(to) : Bitboard(0);
 }
 /// Bitboards::pretty() returns an ASCII representation of a bitboard suitable
 /// to be printed to standard output. Useful for debugging.
-const std::string Bitboards::pretty(Bitboard b) {
+std::string Bitboards::pretty(Bitboard b) {
  std::string s = "+---+---+---+---+---+---+---+---+\n";
@@ -96,12 +108,17 @@ void Bitboards::init() {
      for (PieceType pt : { BISHOP, ROOK })
          for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2)
          {
              if (PseudoAttacks[pt][s1] & s2)
-                  LineBB[s1][s2] = (attacks_bb(pt, s1, 0) & attacks_bb(pt, s2, 0)) | s1 | s2;
+              {
                  LineBB[s1][s2]    = (attacks_bb(pt, s1, 0) & attacks_bb(pt, s2, 0)) | s1 | s2;
                  BetweenBB[s1][s2] = (attacks_bb(pt, s1, square_bb(s2)) & attacks_bb(pt, s2, square_bb(s1)));
              }
              BetweenBB[s1][s2] |= s2;
          }
  }
 }
 namespace {
  Bitboard sliding_attack(PieceType pt, Square sq, Bitboard occupied) {
@@ -110,10 +127,10 @@ namespace {
    Direction   RookDirections[4] = {NORTH, SOUTH, EAST, WEST};
    Direction BishopDirections[4] = {NORTH_EAST, SOUTH_EAST, SOUTH_WEST, NORTH_WEST};
-    for(Direction d : (pt == ROOK ? RookDirections : BishopDirections))
+    for (Direction d : (pt == ROOK ? RookDirections : BishopDirections))
    {
        Square s = sq;
-        while(safe_destination(s, d) && !(occupied & s))
+        while (safe_destination(s, d) && !(occupied & s))
            attacks |= (s += d);
    }
@@ -201,3 +218,5 @@ namespace {
    }
  }
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,19 +23,21 @@
 #include "types.h"
 namespace Stockfish {
 namespace Bitbases {
 void init();
 bool probe(Square wksq, Square wpsq, Square bksq, Color us);
-}
+} // namespace Stockfish::Bitbases
 namespace Bitboards {
 void init();
-const std::string pretty(Bitboard b);
+std::string pretty(Bitboard b);
-}
+} // namespace Stockfish::Bitboards
 constexpr Bitboard AllSquares = ~Bitboard(0);
 constexpr Bitboard DarkSquares = 0xAA55AA55AA55AA55ULL;
@@ -73,6 +75,7 @@ extern uint8_t PopCnt16[1 << 16];
 extern uint8_t SquareDistance[SQUARE_NB][SQUARE_NB];
 extern Bitboard SquareBB[SQUARE_NB];
 extern Bitboard BetweenBB[SQUARE_NB][SQUARE_NB];
 extern Bitboard LineBB[SQUARE_NB][SQUARE_NB];
 extern Bitboard PseudoAttacks[PIECE_TYPE_NB][SQUARE_NB];
 extern Bitboard PawnAttacks[COLOR_NB][SQUARE_NB];
@@ -209,23 +212,29 @@ constexpr Bitboard adjacent_files_bb(Square s) {
 inline Bitboard line_bb(Square s1, Square s2) {
  assert(is_ok(s1) && is_ok(s2));
  return LineBB[s1][s2];
 }
-/// between_bb() returns a bitboard representing squares that are linearly
+/// between_bb(s1, s2) returns a bitboard representing the squares in the semi-open
-/// between the two given squares (excluding the given squares). If the given
+/// segment between the squares s1 and s2 (excluding s1 but including s2). If the
-/// squares are not on a same file/rank/diagonal, we return 0. For instance,
+/// given squares are not on a same file/rank/diagonal, it returns s2. For instance,
-/// between_bb(SQ_C4, SQ_F7) will return a bitboard with squares D5 and E6.
+/// between_bb(SQ_C4, SQ_F7) will return a bitboard with squares D5, E6 and F7, but
 /// between_bb(SQ_E6, SQ_F8) will return a bitboard with the square F8. This trick
 /// allows to generate non-king evasion moves faster: the defending piece must either
 /// interpose itself to cover the check or capture the checking piece.
 inline Bitboard between_bb(Square s1, Square s2) {
-  Bitboard b = line_bb(s1, s2) & ((AllSquares << s1) ^ (AllSquares << s2));
+
-  return b & (b - 1); //exclude lsb
+  assert(is_ok(s1) && is_ok(s2));
  return BetweenBB[s1][s2];
 }
-/// forward_ranks_bb() returns a bitboard representing the squares on the ranks
+/// forward_ranks_bb() returns a bitboard representing the squares on the ranks in
-/// in front of the given one, from the point of view of the given color. For instance,
+/// front of the given one, from the point of view of the given color. For instance,
 /// forward_ranks_bb(BLACK, SQ_D3) will return the 16 squares on ranks 1 and 2.
 constexpr Bitboard forward_ranks_bb(Color c, Square s) {
@@ -279,16 +288,6 @@ inline int edge_distance(File f) { return std::min(f, File(FILE_H - f)); }
 inline int edge_distance(Rank r) { return std::min(r, Rank(RANK_8 - r)); }
 /// safe_destination() returns the bitboard of target square for the given step
 /// from the given square. If the step is off the board, returns empty bitboard.
 inline Bitboard safe_destination(Square s, int step)
 {
    Square to = Square(s + step);
    return is_ok(to) && distance(s, to) <= 2 ? square_bb(to) : Bitboard(0);
 }
 /// attacks_bb(Square) returns the pseudo attacks of the give piece type
 /// assuming an empty board.
@@ -422,13 +421,20 @@ inline Square msb(Bitboard b) {
 #endif
 /// least_significant_square_bb() returns the bitboard of the least significant
 /// square of a non-zero bitboard. It is equivalent to square_bb(lsb(bb)).
 inline Bitboard least_significant_square_bb(Bitboard b) {
  assert(b);
  return b & -b;
 }
 /// pop_lsb() finds and clears the least significant bit in a non-zero bitboard
-inline Square pop_lsb(Bitboard* b) {
+inline Square pop_lsb(Bitboard& b) {
-  assert(*b);
+  assert(b);
-  const Square s = lsb(*b);
+  const Square s = lsb(b);
-  *b &= *b - 1;
+  b &= b - 1;
  return s;
 }
@@ -440,4 +446,6 @@ inline Square frontmost_sq(Color c, Bitboard b) {
  return c == WHITE ? msb(b) : lsb(b);
 }
 } // namespace Stockfish
 #endif // #ifndef BITBOARD_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -22,6 +22,8 @@
 #include "endgame.h"
 #include "movegen.h"
 namespace Stockfish {
 namespace {
  // Used to drive the king towards the edge of the board
@@ -553,8 +555,8 @@ ScaleFactor Endgame<KRPPKRP>::operator()(const Position& pos) const {
  assert(verify_material(pos, strongSide, RookValueMg, 2));
  assert(verify_material(pos, weakSide,   RookValueMg, 1));
-  Square strongPawn1 = pos.squares<PAWN>(strongSide)[0];
+  Square strongPawn1 = lsb(pos.pieces(strongSide, PAWN));
-  Square strongPawn2 = pos.squares<PAWN>(strongSide)[1];
+  Square strongPawn2 = msb(pos.pieces(strongSide, PAWN));
  Square weakKing = pos.square<KING>(weakSide);
  // Does the stronger side have a passed pawn?
@@ -638,8 +640,8 @@ ScaleFactor Endgame<KBPPKB>::operator()(const Position& pos) const {
      return SCALE_FACTOR_NONE;
  Square weakKing = pos.square<KING>(weakSide);
-  Square strongPawn1 = pos.squares<PAWN>(strongSide)[0];
+  Square strongPawn1 = lsb(pos.pieces(strongSide, PAWN));
-  Square strongPawn2 = pos.squares<PAWN>(strongSide)[1];
+  Square strongPawn2 = msb(pos.pieces(strongSide, PAWN));
  Square blockSq1, blockSq2;
  if (relative_rank(strongSide, strongPawn1) > relative_rank(strongSide, strongPawn2))
@@ -741,3 +743,5 @@ ScaleFactor Endgame<KPKP>::operator()(const Position& pos) const {
  // it's probably at least a draw even with the pawn.
  return Bitbases::probe(strongKing, strongPawn, weakKing, us) ? SCALE_FACTOR_NONE : SCALE_FACTOR_DRAW;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -28,6 +28,7 @@
 #include "position.h"
 #include "types.h"
 namespace Stockfish {
 /// EndgameCode lists all supported endgame functions by corresponding codes
@@ -120,4 +121,6 @@ namespace Endgames {
  }
 }
 } // namespace Stockfish
 #endif // #ifndef ENDGAME_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -20,46 +20,146 @@
 #include <cassert>
 #include <cstdlib>
 #include <cstring>   // For std::memset
 #include <fstream>
 #include <iomanip>
 #include <sstream>
 #include <iostream>
 #include <streambuf>
 #include <vector>
 #include "nnue/evaluate_nnue.h"
 #include "bitboard.h"
 #include "evaluate.h"
 #include "material.h"
 #include "misc.h"
 #include "pawns.h"
 #include "thread.h"
 #include "timeman.h"
 #include "uci.h"
 #include "incbin/incbin.h"
 // Macro to embed the default efficiently updatable neural network (NNUE) file
 // data in the engine binary (using incbin.h, by Dale Weiler).
 // This macro invocation will declare the following three variables
 //     const unsigned char        gEmbeddedNNUEData[];  // a pointer to the embedded data
 //     const unsigned char *const gEmbeddedNNUEEnd;     // a marker to the end
 //     const unsigned int         gEmbeddedNNUESize;    // the size of the embedded file
 // Note that this does not work in Microsoft Visual Studio.
 #if !defined(_MSC_VER) && !defined(NNUE_EMBEDDING_OFF)
  INCBIN(EmbeddedNNUE, EvalFileDefaultName);
 #else
  const unsigned char        gEmbeddedNNUEData[1] = {0x0};
  const unsigned char *const gEmbeddedNNUEEnd = &gEmbeddedNNUEData[1];
  const unsigned int         gEmbeddedNNUESize = 1;
 #endif
 using namespace std;
 namespace Stockfish {
 namespace Eval {
-  bool useNNUE;
+  namespace NNUE {
-  std::string eval_file_loaded="None";
+    string currentEvalFileName = "None";
    UseNNUEMode useNNUE;
-  void init_NNUE() {
+    static UseNNUEMode nnue_mode_from_option(const UCI::Option& mode)
    {
      if (mode == "false")
        return UseNNUEMode::False;
      else if (mode == "true")
         return UseNNUEMode::True;
      else if (mode == "pure")
        return UseNNUEMode::Pure;
-    useNNUE = Options["Use NNUE"];
+      return UseNNUEMode::False;
-    std::string eval_file = std::string(Options["EvalFile"]);
+    }
    if (useNNUE && eval_file_loaded != eval_file)
        if (Eval::NNUE::load_eval_file(eval_file))
            eval_file_loaded = eval_file;
  }
-  void verify_NNUE() {
+  /// NNUE::init() tries to load a NNUE network at startup time, or when the engine
  /// receives a UCI command "setoption name EvalFile value nn-[a-z0-9]{12}.nnue"
  /// The name of the NNUE network is always retrieved from the EvalFile option.
  /// We search the given network in three locations: internally (the default
  /// network may be embedded in the binary), in the active working directory and
  /// in the engine directory. Distro packagers may define the DEFAULT_NNUE_DIRECTORY
  /// variable to have the engine search in a special directory in their distro.
-    std::string eval_file = std::string(Options["EvalFile"]);
+  void NNUE::init() {
-    if (useNNUE && eval_file_loaded != eval_file)
+
    useNNUE = nnue_mode_from_option(Options["Use NNUE"]);
    if (useNNUE == UseNNUEMode::False)
        return;
    string eval_file = string(Options["EvalFile"]);
    if (eval_file.empty())
        eval_file = EvalFileDefaultName;
    #if defined(DEFAULT_NNUE_DIRECTORY)
    #define stringify2(x) #x
    #define stringify(x) stringify2(x)
    vector<string> dirs = { "<internal>" , "" , CommandLine::binaryDirectory , stringify(DEFAULT_NNUE_DIRECTORY) };
    #else
    vector<string> dirs = { "<internal>" , "" , CommandLine::binaryDirectory };
    #endif
    for (string directory : dirs)
        if (currentEvalFileName != eval_file)
        {
            if (directory != "<internal>")
            {
                ifstream stream(directory + eval_file, ios::binary);
                if (load_eval(eval_file, stream))
                    currentEvalFileName = eval_file;
            }
            if (directory == "<internal>" && eval_file == EvalFileDefaultName)
            {
                // C++ way to prepare a buffer for a memory stream
                class MemoryBuffer : public basic_streambuf<char> {
                    public: MemoryBuffer(char* p, size_t n) { setg(p, p, p + n); setp(p, p + n); }
                };
                MemoryBuffer buffer(const_cast<char*>(reinterpret_cast<const char*>(gEmbeddedNNUEData)),
                                    size_t(gEmbeddedNNUESize));
                (void) gEmbeddedNNUEEnd; // Silence warning on unused variable
                istream stream(&buffer);
                if (load_eval(eval_file, stream))
                    currentEvalFileName = eval_file;
            }
        }
  }
  /// NNUE::verify() verifies that the last net used was loaded successfully
  void NNUE::verify() {
    string eval_file = string(Options["EvalFile"]);
    if (eval_file.empty())
        eval_file = EvalFileDefaultName;
    if (useNNUE != UseNNUEMode::False && currentEvalFileName != eval_file)
    {
-        std::cerr << "Use of NNUE evaluation, but the file " << eval_file << " was not loaded successfully. "
+
-                  << "These network evaluation parameters must be available, compatible with this version of the code. "
+        string msg1 = "If the UCI option \"Use NNUE\" is set to true, network evaluation parameters compatible with the engine must be available.";
-                  << "The UCI option EvalFile might need to specify the full path, including the directory/folder name, to the file." << std::endl;
+        string msg2 = "The option is set to true, but the network file " + eval_file + " was not loaded successfully.";
-        std::exit(EXIT_FAILURE);
+        string msg3 = "The UCI option EvalFile might need to specify the full path, including the directory name, to the network file.";
        string msg4 = "The default net can be downloaded from: https://tests.stockfishchess.org/api/nn/" + std::string(EvalFileDefaultName);
        string msg5 = "The engine will be terminated now.";
        sync_cout << "info string ERROR: " << msg1 << sync_endl;
        sync_cout << "info string ERROR: " << msg2 << sync_endl;
        sync_cout << "info string ERROR: " << msg3 << sync_endl;
        sync_cout << "info string ERROR: " << msg4 << sync_endl;
        sync_cout << "info string ERROR: " << msg5 << sync_endl;
        exit(EXIT_FAILURE);
    }
-    if (useNNUE)
+    if (useNNUE != UseNNUEMode::False)
-        sync_cout << "info string NNUE evaluation using " << eval_file << " enabled." << sync_endl;
+        sync_cout << "info string NNUE evaluation using " << eval_file << " enabled" << sync_endl;
    else
-        sync_cout << "info string classical evaluation enabled." << sync_endl;
+        sync_cout << "info string classical evaluation enabled" << sync_endl;
  }
 }
@@ -97,7 +197,7 @@ namespace Trace {
    else
        os << scores[t][WHITE] << " | " << scores[t][BLACK];
-    os << " | " << scores[t][WHITE] - scores[t][BLACK] << "\n";
+    os << " | " << scores[t][WHITE] - scores[t][BLACK] << " |\n";
    return os;
  }
 }
@@ -107,17 +207,17 @@ using namespace Trace;
 namespace {
  // Threshold for lazy and space evaluation
-  constexpr Value LazyThreshold1  = Value(1400);
+  constexpr Value LazyThreshold1    =  Value(3631);
-  constexpr Value LazyThreshold2  = Value(1300);
+  constexpr Value LazyThreshold2    =  Value(2084);
-  constexpr Value SpaceThreshold = Value(12222);
+  constexpr Value SpaceThreshold    =  Value(11551);
  // KingAttackWeights[PieceType] contains king attack weights by piece type
-  constexpr int KingAttackWeights[PIECE_TYPE_NB] = { 0, 0, 81, 52, 44, 10 };
+  constexpr int KingAttackWeights[PIECE_TYPE_NB] = { 0, 0, 76, 46, 45, 14 };
  // SafeCheck[PieceType][single/multiple] contains safe check bonus by piece type,
  // higher if multiple safe checks are possible for that piece type.
  constexpr int SafeCheck[][2] = {
-      {}, {}, {792, 1283}, {645, 967}, {1084, 1897}, {772, 1119}
+      {}, {}, {805, 1292}, {650, 984}, {1071, 1886}, {730, 1128}
  };
 #define S(mg, eg) make_score(mg, eg)
@@ -125,73 +225,76 @@ namespace {
  // MobilityBonus[PieceType-2][attacked] contains bonuses for middle and end game,
  // indexed by piece type and number of attacked squares in the mobility area.
  constexpr Score MobilityBonus[][32] = {
-    { S(-62,-81), S(-53,-56), S(-12,-31), S( -4,-16), S(  3,  5), S( 13, 11), // Knight
+    { S(-62,-79), S(-53,-57), S(-12,-31), S( -3,-17), S(  3,  7), S( 12, 13), // Knight
-      S( 22, 17), S( 28, 20), S( 33, 25) },
+      S( 21, 16), S( 28, 21), S( 37, 26) },
-    { S(-48,-59), S(-20,-23), S( 16, -3), S( 26, 13), S( 38, 24), S( 51, 42), // Bishop
+    { S(-47,-59), S(-20,-25), S( 14, -8), S( 29, 12), S( 39, 21), S( 53, 40), // Bishop
-      S( 55, 54), S( 63, 57), S( 63, 65), S( 68, 73), S( 81, 78), S( 81, 86),
+      S( 53, 56), S( 60, 58), S( 62, 65), S( 69, 72), S( 78, 78), S( 83, 87),
-      S( 91, 88), S( 98, 97) },
+      S( 91, 88), S( 96, 98) },
-    { S(-60,-78), S(-20,-17), S(  2, 23), S(  3, 39), S(  3, 70), S( 11, 99), // Rook
+    { S(-60,-82), S(-24,-15), S(  0, 17) ,S(  3, 43), S(  4, 72), S( 14,100), // Rook
-      S( 22,103), S( 31,121), S( 40,134), S( 40,139), S( 41,158), S( 48,164),
+      S( 20,102), S( 30,122), S( 41,133), S(41 ,139), S( 41,153), S( 45,160),
-      S( 57,168), S( 57,169), S( 62,172) },
+      S( 57,165), S( 58,170), S( 67,175) },
-    { S(-30,-48), S(-12,-30), S( -8, -7), S( -9, 19), S( 20, 40), S( 23, 55), // Queen
+    { S(-29,-49), S(-16,-29), S( -8, -8), S( -8, 17), S( 18, 39), S( 25, 54), // Queen
-      S( 23, 59), S( 35, 75), S( 38, 78), S( 53, 96), S( 64, 96), S( 65,100),
+      S( 23, 59), S( 37, 73), S( 41, 76), S( 54, 95), S( 65, 95) ,S( 68,101),
-      S( 65,121), S( 66,127), S( 67,131), S( 67,133), S( 72,136), S( 72,141),
+      S( 69,124), S( 70,128), S( 70,132), S( 70,133) ,S( 71,136), S( 72,140),
-      S( 77,147), S( 79,150), S( 93,151), S(108,168), S(108,168), S(108,171),
+      S( 74,147), S( 76,149), S( 90,153), S(104,169), S(105,171), S(106,171),
-      S(110,182), S(114,182), S(114,192), S(116,219) }
+      S(112,178), S(114,185), S(114,187), S(119,221) }
  };
  // BishopPawns[distance from edge] contains a file-dependent penalty for pawns on
  // squares of the same color as our bishop.
  constexpr Score BishopPawns[int(FILE_NB) / 2] = {
    S(3, 8), S(3, 9), S(2, 7), S(3, 7)
  };
  // KingProtector[knight/bishop] contains penalty for each distance unit to own king
-  constexpr Score KingProtector[] = { S(8, 9), S(6, 9) };
+  constexpr Score KingProtector[] = { S(9, 9), S(7, 9) };
  // Outpost[knight/bishop] contains bonuses for each knight or bishop occupying a
  // pawn protected square on rank 4 to 6 which is also safe from a pawn attack.
-  constexpr Score Outpost[] = { S(56, 36), S(30, 23) };
+  constexpr Score Outpost[] = { S(54, 34), S(31, 25) };
  // PassedRank[Rank] contains a bonus according to the rank of a passed pawn
  constexpr Score PassedRank[RANK_NB] = {
-    S(0, 0), S(10, 28), S(17, 33), S(15, 41), S(62, 72), S(168, 177), S(276, 260)
+    S(0, 0), S(2, 38), S(15, 36), S(22, 50), S(64, 81), S(166, 184), S(284, 269)
  };
-  // RookOnFile[semiopen/open] contains bonuses for each rook when there is
+  constexpr Score RookOnClosedFile = S(10, 5);
-  // no (friendly) pawn on the rook file.
+  constexpr Score RookOnOpenFile[] = { S(18, 8), S(49, 26) };
  constexpr Score RookOnFile[] = { S(19, 7), S(48, 29) };
  // ThreatByMinor/ByRook[attacked PieceType] contains bonuses according to
  // which piece type attacks which one. Attacks on lesser pieces which are
  // pawn-defended are not considered.
  constexpr Score ThreatByMinor[PIECE_TYPE_NB] = {
-    S(0, 0), S(5, 32), S(57, 41), S(77, 56), S(88, 119), S(79, 161)
+    S(0, 0), S(6, 37), S(64, 50), S(82, 57), S(103, 130), S(81, 163)
  };
  constexpr Score ThreatByRook[PIECE_TYPE_NB] = {
-    S(0, 0), S(3, 46), S(37, 68), S(42, 60), S(0, 38), S(58, 41)
+    S(0, 0), S(3, 44), S(36, 71), S(44, 59), S(0, 39), S(60, 39)
  };
  constexpr Value CorneredBishop = Value(50);
  // Assorted bonuses and penalties
-  constexpr Score BadOutpost          = S( -7, 36);
+  constexpr Score UncontestedOutpost  = S(  0, 10);
  constexpr Score BishopOnKingRing    = S( 24,  0);
  constexpr Score BishopPawns         = S(  3,  7);
  constexpr Score BishopXRayPawns     = S(  4,  5);
  constexpr Score CorneredBishop      = S( 50, 50);
  constexpr Score FlankAttacks        = S(  8,  0);
-  constexpr Score Hanging             = S( 69, 36);
+  constexpr Score Hanging             = S( 72, 40);
  constexpr Score KnightOnQueen       = S( 16, 11);
  constexpr Score LongDiagonalBishop  = S( 45,  0);
  constexpr Score MinorBehindPawn     = S( 18,  3);
-  constexpr Score PassedFile          = S( 11,  8);
+  constexpr Score PassedFile          = S( 13,  8);
-  constexpr Score PawnlessFlank       = S( 17, 95);
+  constexpr Score PawnlessFlank       = S( 19, 97);
-  constexpr Score QueenInfiltration   = S( -2, 14);
+  constexpr Score ReachableOutpost    = S( 33, 19);
-  constexpr Score ReachableOutpost    = S( 31, 22);
+  constexpr Score RestrictedPiece     = S(  6,  7);
  constexpr Score RestrictedPiece     = S(  7,  7);
  constexpr Score RookOnKingRing      = S( 16,  0);
-  constexpr Score RookOnQueenFile     = S(  6, 11);
+  constexpr Score SliderOnQueen       = S( 62, 21);
-  constexpr Score SliderOnQueen       = S( 60, 18);
+  constexpr Score ThreatByKing        = S( 24, 87);
  constexpr Score ThreatByKing        = S( 24, 89);
  constexpr Score ThreatByPawnPush    = S( 48, 39);
-  constexpr Score ThreatBySafePawn    = S(173, 94);
+  constexpr Score ThreatBySafePawn    = S(167, 99);
  constexpr Score TrappedRook         = S( 55, 13);
  constexpr Score WeakQueenProtection = S( 14,  0);
-  constexpr Score WeakQueen           = S( 56, 15);
+  constexpr Score WeakQueen           = S( 57, 19);
 #undef S
@@ -282,8 +385,8 @@ namespace {
    attackedBy2[Us] = dblAttackByPawn | (attackedBy[Us][KING] & attackedBy[Us][PAWN]);
    // Init our king safety tables
-    Square s = make_square(Utility::clamp(file_of(ksq), FILE_B, FILE_G),
+    Square s = make_square(std::clamp(file_of(ksq), FILE_B, FILE_G),
-                           Utility::clamp(rank_of(ksq), RANK_2, RANK_7));
+                           std::clamp(rank_of(ksq), RANK_2, RANK_7));
    kingRing[Us] = attacks_bb<KING>(s) | s;
    kingAttackersCount[Them] = popcount(kingRing[Us] & pe->pawn_attacks(Them));
@@ -303,15 +406,16 @@ namespace {
    constexpr Direction Down = -pawn_push(Us);
    constexpr Bitboard OutpostRanks = (Us == WHITE ? Rank4BB | Rank5BB | Rank6BB
                                                   : Rank5BB | Rank4BB | Rank3BB);
-    const Square* pl = pos.squares<Pt>(Us);
+    Bitboard b1 = pos.pieces(Us, Pt);
    Bitboard b, bb;
    Score score = SCORE_ZERO;
    attackedBy[Us][Pt] = 0;
-    for (Square s = *pl; s != SQ_NONE; s = *++pl)
+    while (b1)
    {
        Square s = pop_lsb(b1);
        // Find attacked squares, including x-ray attacks for bishops and rooks
        b = Pt == BISHOP ? attacks_bb<BISHOP>(s, pos.pieces() ^ pos.pieces(QUEEN))
          : Pt ==   ROOK ? attacks_bb<  ROOK>(s, pos.pieces() ^ pos.pieces(QUEEN) ^ pos.pieces(Us, ROOK))
@@ -338,21 +442,21 @@ namespace {
            score += BishopOnKingRing;
        int mob = popcount(b & mobilityArea[Us]);
        mobility[Us] += MobilityBonus[Pt - 2][mob];
        if (Pt == BISHOP || Pt == KNIGHT)
        {
            // Bonus if the piece is on an outpost square or can reach one
-            // Reduced bonus for knights (BadOutpost) if few relevant targets
+            // Bonus for knights (UncontestedOutpost) if few relevant targets
-            bb = OutpostRanks & attackedBy[Us][PAWN] & ~pe->pawn_attacks_span(Them);
+            bb = OutpostRanks & (attackedBy[Us][PAWN] | shift<Down>(pos.pieces(PAWN)))
                              & ~pe->pawn_attacks_span(Them);
            Bitboard targets = pos.pieces(Them) & ~pos.pieces(PAWN);
            if (   Pt == KNIGHT
                && bb & s & ~CenterFiles // on a side outpost
                && !(b & targets)        // no relevant attacks
                && (!more_than_one(targets & (s & QueenSide ? QueenSide : KingSide))))
-                score += BadOutpost;
+                score += UncontestedOutpost * popcount(pos.pieces(PAWN) & (s & QueenSide ? QueenSide : KingSide));
            else if (bb & s)
                score += Outpost[Pt == BISHOP];
            else if (Pt == KNIGHT && bb & b & ~pos.pieces(Us))
@@ -365,14 +469,14 @@ namespace {
            // Penalty if the piece is far from the king
            score -= KingProtector[Pt == BISHOP] * distance(pos.square<KING>(Us), s);
-            if (Pt == BISHOP)
+            if constexpr (Pt == BISHOP)
            {
                // Penalty according to the number of our pawns on the same color square as the
                // bishop, bigger when the center files are blocked with pawns and smaller
                // when the bishop is outside the pawn chain.
                Bitboard blocked = pos.pieces(Us, PAWN) & shift<Down>(pos.pieces());
-                score -= BishopPawns * pos.pawns_on_same_color_squares(Us, s)
+                score -= BishopPawns[edge_distance(file_of(s))] * pos.pawns_on_same_color_squares(Us, s)
                                     * (!(attackedBy[Us][PAWN] & s) + popcount(blocked & CenterFiles));
                // Penalty for all enemy pawns x-rayed
@@ -390,45 +494,48 @@ namespace {
                {
                    Direction d = pawn_push(Us) + (file_of(s) == FILE_A ? EAST : WEST);
                    if (pos.piece_on(s + d) == make_piece(Us, PAWN))
-                        score -= !pos.empty(s + d + pawn_push(Us))                ? CorneredBishop * 4
+                        score -= !pos.empty(s + d + pawn_push(Us)) ? 4 * make_score(CorneredBishop, CorneredBishop)
-                                : pos.piece_on(s + d + d) == make_piece(Us, PAWN) ? CorneredBishop * 2
+                                                                   : 3 * make_score(CorneredBishop, CorneredBishop);
                                                                                  : CorneredBishop;
                }
            }
        }
-        if (Pt == ROOK)
+        if constexpr (Pt == ROOK)
        {
-            // Bonus for rook on the same file as a queen
+            // Bonuses for rook on a (semi-)open or closed file
            if (file_bb(s) & pos.pieces(QUEEN))
                score += RookOnQueenFile;
            // Bonus for rook on an open or semi-open file
            if (pos.is_on_semiopen_file(Us, s))
                score += RookOnFile[pos.is_on_semiopen_file(Them, s)];
            // Penalty when trapped by the king, even more if the king cannot castle
            else if (mob <= 3)
            {
-                File kf = file_of(pos.square<KING>(Us));
+                score += RookOnOpenFile[pos.is_on_semiopen_file(Them, s)];
-                if ((kf < FILE_E) == (file_of(s) < kf))
+            }
-                    score -= TrappedRook * (1 + !pos.castling_rights(Us));
+            else
            {
                // If our pawn on this file is blocked, increase penalty
                if ( pos.pieces(Us, PAWN)
                   & shift<Down>(pos.pieces())
                   & file_bb(s))
                {
                    score -= RookOnClosedFile;
                }
                // Penalty when trapped by the king, even more if the king cannot castle
                if (mob <= 3)
                {
                    File kf = file_of(pos.square<KING>(Us));
                    if ((kf < FILE_E) == (file_of(s) < kf))
                        score -= TrappedRook * (1 + !pos.castling_rights(Us));
                }
            }
        }
-        if (Pt == QUEEN)
+        if constexpr (Pt == QUEEN)
        {
            // Penalty if any relative pin or discovered attack against the queen
            Bitboard queenPinners;
            if (pos.slider_blockers(pos.pieces(Them, ROOK, BISHOP), s, queenPinners))
                score -= WeakQueen;
            // Bonus for queen on weak square in enemy camp
            if (relative_rank(Us, s) > RANK_4 && (~pe->pawn_attacks_span(Them) & s))
                score += QueenInfiltration;
        }
    }
-    if (T)
+    if constexpr (T)
        Trace::add(Pt, Us, score);
    return score;
@@ -504,18 +611,18 @@ namespace {
    int kingFlankAttack  = popcount(b1) + popcount(b2);
    int kingFlankDefense = popcount(b3);
-    kingDanger +=        kingAttackersCount[Them] * kingAttackersWeight[Them]
+    kingDanger +=        kingAttackersCount[Them] * kingAttackersWeight[Them] // (~10 Elo)
-                 + 185 * popcount(kingRing[Us] & weak)
+                 + 183 * popcount(kingRing[Us] & weak)                        // (~15 Elo)
-                 + 148 * popcount(unsafeChecks)
+                 + 148 * popcount(unsafeChecks)                               // (~4 Elo)
-                 +  98 * popcount(pos.blockers_for_king(Us))
+                 +  98 * popcount(pos.blockers_for_king(Us))                  // (~2 Elo)
-                 +  69 * kingAttacksCount[Them]
+                 +  69 * kingAttacksCount[Them]                               // (~0.5 Elo)
-                 +   3 * kingFlankAttack * kingFlankAttack / 8
+                 +   3 * kingFlankAttack * kingFlankAttack / 8                // (~0.5 Elo)
-                 +       mg_value(mobility[Them] - mobility[Us])
+                 +       mg_value(mobility[Them] - mobility[Us])              // (~0.5 Elo)
-                 - 873 * !pos.count<QUEEN>(Them)
+                 - 873 * !pos.count<QUEEN>(Them)                              // (~24 Elo)
-                 - 100 * bool(attackedBy[Us][KNIGHT] & attackedBy[Us][KING])
+                 - 100 * bool(attackedBy[Us][KNIGHT] & attackedBy[Us][KING])  // (~5 Elo)
-                 -   6 * mg_value(score) / 8
+                 -   6 * mg_value(score) / 8                                  // (~8 Elo)
-                 -   4 * kingFlankDefense
+                 -   4 * kingFlankDefense                                     // (~5 Elo)
-                 +  37;
+                 +  37;                                                       // (~0.5 Elo)
    // Transform the kingDanger units into a Score, and subtract it from the evaluation
    if (kingDanger > 100)
@@ -528,7 +635,7 @@ namespace {
    // Penalty if king flank is under attack, potentially moving toward the king
    score -= FlankAttacks * kingFlankAttack;
-    if (T)
+    if constexpr (T)
        Trace::add(KING, Us, score);
    return score;
@@ -567,11 +674,11 @@ namespace {
    {
        b = (defended | weak) & (attackedBy[Us][KNIGHT] | attackedBy[Us][BISHOP]);
        while (b)
-            score += ThreatByMinor[type_of(pos.piece_on(pop_lsb(&b)))];
+            score += ThreatByMinor[type_of(pos.piece_on(pop_lsb(b)))];
        b = weak & attackedBy[Us][ROOK];
        while (b)
-            score += ThreatByRook[type_of(pos.piece_on(pop_lsb(&b)))];
+            score += ThreatByRook[type_of(pos.piece_on(pop_lsb(b)))];
        if (weak & attackedBy[Us][KING])
            score += ThreatByKing;
@@ -629,7 +736,7 @@ namespace {
        score += SliderOnQueen * popcount(b & safe & attackedBy2[Us]) * (1 + queenImbalance);
    }
-    if (T)
+    if constexpr (T)
        Trace::add(THREAT, Us, score);
    return score;
@@ -669,7 +776,7 @@ namespace {
    while (b)
    {
-        Square s = pop_lsb(&b);
+        Square s = pop_lsb(b);
        assert(!(pos.pieces(Them, PAWN) & forward_file_bb(Us, s + Up)));
@@ -683,8 +790,8 @@ namespace {
            Square blockSq = s + Up;
            // Adjust bonus based on the king's proximity
-            bonus += make_score(0, (  (king_proximity(Them, blockSq) * 19) / 4
+            bonus += make_score(0, (  king_proximity(Them, blockSq) * 19 / 4
-                                     - king_proximity(Us,   blockSq) *  2) * w);
+                                    - king_proximity(Us,   blockSq) *  2) * w);
            // If blockSq is not the queening square then consider also a second push
            if (r != RANK_7)
@@ -699,14 +806,16 @@ namespace {
                bb = forward_file_bb(Them, s) & pos.pieces(ROOK, QUEEN);
                if (!(pos.pieces(Them) & bb))
-                    unsafeSquares &= attackedBy[Them][ALL_PIECES];
+                    unsafeSquares &= attackedBy[Them][ALL_PIECES] | pos.pieces(Them);
-                // If there are no enemy attacks on passed pawn span, assign a big bonus.
+                // If there are no enemy pieces or attacks on passed pawn span, assign a big bonus.
                // Or if there is some, but they are all attacked by our pawns, assign a bit smaller bonus.
                // Otherwise assign a smaller bonus if the path to queen is not attacked
                // and even smaller bonus if it is attacked but block square is not.
-                int k = !unsafeSquares                    ? 35 :
+                int k = !unsafeSquares                    ? 36 :
-                        !(unsafeSquares & squaresToQueen) ? 20 :
+                !(unsafeSquares & ~attackedBy[Us][PAWN])  ? 30 :
-                        !(unsafeSquares & blockSq)        ?  9 :
+                        !(unsafeSquares & squaresToQueen) ? 17 :
                        !(unsafeSquares & blockSq)        ?  7 :
                                                             0 ;
                // Assign a larger bonus if the block square is defended
@@ -720,7 +829,7 @@ namespace {
        score += bonus - PassedFile * edge_distance(file_of(s));
    }
-    if (T)
+    if constexpr (T)
        Trace::add(PASSED, Us, score);
    return score;
@@ -728,7 +837,7 @@ namespace {
  // Evaluation::space() computes a space evaluation for a given side, aiming to improve game
-  // play in the opening. It is based on the number of safe squares on the 4 central files
+  // play in the opening. It is based on the number of safe squares on the four central files
  // on ranks 2 to 4. Completely safe squares behind a friendly pawn are counted twice.
  // Finally, the space bonus is multiplied by a weight which decreases according to occupancy.
@@ -755,11 +864,13 @@ namespace {
    behind |= shift<Down>(behind);
    behind |= shift<Down+Down>(behind);
    // Compute space score based on the number of safe squares and number of our pieces
    // increased with number of total blocked pawns in position.
    int bonus = popcount(safe) + popcount(behind & safe & ~attackedBy[Them][ALL_PIECES]);
    int weight = pos.count<ALL_PIECES>(Us) - 3 + std::min(pe->blocked_count(), 9);
    Score score = make_score(bonus * weight * weight / 16, 0);
-    if (T)
+    if constexpr (T)
        Trace::add(SPACE, Us, score);
    return score;
@@ -774,7 +885,7 @@ namespace {
  Value Evaluation<T>::winnable(Score score) const {
    int outflanking =  distance<File>(pos.square<KING>(WHITE), pos.square<KING>(BLACK))
-                     - distance<Rank>(pos.square<KING>(WHITE), pos.square<KING>(BLACK));
+                    + int(rank_of(pos.square<KING>(WHITE)) - rank_of(pos.square<KING>(BLACK)));
    bool pawnsOnBothFlanks =   (pos.pieces(PAWN) & QueenSide)
                            && (pos.pieces(PAWN) & KingSide);
@@ -801,7 +912,7 @@ namespace {
    // Now apply the bonus: note that we find the attacking side by extracting the
    // sign of the midgame or endgame values, and that we carefully cap the bonus
    // so that the midgame and endgame scores do not change sign after the bonus.
-    int u = ((mg > 0) - (mg < 0)) * Utility::clamp(complexity + 50, -abs(mg), 0);
+    int u = ((mg > 0) - (mg < 0)) * std::clamp(complexity + 50, -abs(mg), 0);
    int v = ((eg > 0) - (eg < 0)) * std::max(complexity, -abs(eg));
    mg += u;
@@ -811,28 +922,42 @@ namespace {
    Color strongSide = eg > VALUE_DRAW ? WHITE : BLACK;
    int sf = me->scale_factor(pos, strongSide);
-    // If scale factor is not already specific, scale down via general heuristics
+    // If scale factor is not already specific, scale up/down via general heuristics
    if (sf == SCALE_FACTOR_NORMAL)
    {
        if (pos.opposite_bishops())
        {
            // For pure opposite colored bishops endgames use scale factor
            // based on the number of passed pawns of the strong side.
            if (   pos.non_pawn_material(WHITE) == BishopValueMg
                && pos.non_pawn_material(BLACK) == BishopValueMg)
                sf = 18 + 4 * popcount(pe->passed_pawns(strongSide));
            // For every other opposite colored bishops endgames use scale factor
            // based on the number of all pieces of the strong side.
            else
                sf = 22 + 3 * pos.count<ALL_PIECES>(strongSide);
        }
        // For rook endgames with strong side not having overwhelming pawn number advantage
        // and its pawns being on one flank and weak side protecting its pieces with a king
        // use lower scale factor.
        else if (  pos.non_pawn_material(WHITE) == RookValueMg
                && pos.non_pawn_material(BLACK) == RookValueMg
                && pos.count<PAWN>(strongSide) - pos.count<PAWN>(~strongSide) <= 1
                && bool(KingSide & pos.pieces(strongSide, PAWN)) != bool(QueenSide & pos.pieces(strongSide, PAWN))
                && (attacks_bb<KING>(pos.square<KING>(~strongSide)) & pos.pieces(~strongSide, PAWN)))
            sf = 36;
        // For queen vs no queen endgames use scale factor
        // based on number of minors of side that doesn't have queen.
        else if (pos.count<QUEEN>() == 1)
            sf = 37 + 3 * (pos.count<QUEEN>(WHITE) == 1 ? pos.count<BISHOP>(BLACK) + pos.count<KNIGHT>(BLACK)
                                                        : pos.count<BISHOP>(WHITE) + pos.count<KNIGHT>(WHITE));
        // In every other case use scale factor based on
        // the number of pawns of the strong side reduced if pawns are on a single flank.
        else
-            sf = std::min(sf, 36 + 7 * pos.count<PAWN>(strongSide));
+            sf = std::min(sf, 36 + 7 * pos.count<PAWN>(strongSide)) - 4 * !pawnsOnBothFlanks;
        // Reduce scale factor in case of pawns being on a single flank
        sf -= 4 * !pawnsOnBothFlanks;
    }
    // Interpolate between the middlegame and (scaled by 'sf') endgame score
@@ -840,7 +965,7 @@ namespace {
       + eg * int(PHASE_MIDGAME - me->game_phase()) * ScaleFactor(sf) / SCALE_FACTOR_NORMAL;
    v /= PHASE_MIDGAME;
-    if (T)
+    if constexpr (T)
    {
        Trace::add(WINNABLE, make_score(u, eg * ScaleFactor(sf) / SCALE_FACTOR_NORMAL - eg_value(score)));
        Trace::add(TOTAL, make_score(mg, eg * ScaleFactor(sf) / SCALE_FACTOR_NORMAL));
@@ -870,7 +995,7 @@ namespace {
    // Initialize score by reading the incrementally updated scores included in
    // the position object (material + piece square tables) and the material
    // imbalance. Score is computed internally from the white point of view.
-    Score score = pos.psq_score() + me->imbalance() + pos.this_thread()->contempt;
+    Score score = pos.psq_score() + me->imbalance() + pos.this_thread()->trend;
    // Probe the pawn hash table
    pe = Pawns::probe(pos);
@@ -878,7 +1003,9 @@ namespace {
    // Early exit if score is high
    auto lazy_skip = [&](Value lazyThreshold) {
-        return abs(mg_value(score) + eg_value(score)) / 2 > lazyThreshold + pos.non_pawn_material() / 64;
+        return abs(mg_value(score) + eg_value(score)) >   lazyThreshold
                                                        + std::abs(pos.this_thread()->bestValue) * 5 / 4
                                                        + pos.non_pawn_material() / 32;
    };
    if (lazy_skip(LazyThreshold1))
@@ -912,7 +1039,7 @@ make_v:
    Value v = winnable(score);
    // In case of tracing add all remaining individual evaluation terms
-    if (T)
+    if constexpr (T)
    {
        Trace::add(MATERIAL, pos.psq_score());
        Trace::add(IMBALANCE, me->imbalance());
@@ -924,15 +1051,44 @@ make_v:
    v = (v / 16) * 16;
    // Side to move point of view
-    v = (pos.side_to_move() == WHITE ? v : -v) + Tempo;
+    v = (pos.side_to_move() == WHITE ? v : -v);
    // Damp down the evaluation linearly when shuffling
    v = v * (100 - pos.rule50_count()) / 100;
    return v;
  }
-} // namespace
+
  /// Fisher Random Chess: correction for cornered bishops, to fix chess960 play with NNUE
  Value fix_FRC(const Position& pos) {
    constexpr Bitboard Corners =  1ULL << SQ_A1 | 1ULL << SQ_H1 | 1ULL << SQ_A8 | 1ULL << SQ_H8;
    if (!(pos.pieces(BISHOP) & Corners))
        return VALUE_ZERO;
    int correction = 0;
    if (   pos.piece_on(SQ_A1) == W_BISHOP
        && pos.piece_on(SQ_B2) == W_PAWN)
        correction -= CorneredBishop;
    if (   pos.piece_on(SQ_H1) == W_BISHOP
        && pos.piece_on(SQ_G2) == W_PAWN)
        correction -= CorneredBishop;
    if (   pos.piece_on(SQ_A8) == B_BISHOP
        && pos.piece_on(SQ_B7) == B_PAWN)
        correction += CorneredBishop;
    if (   pos.piece_on(SQ_H8) == B_BISHOP
        && pos.piece_on(SQ_G7) == B_PAWN)
        correction += CorneredBishop;
    return pos.side_to_move() == WHITE ?  Value(3 * correction)
                                       : -Value(3 * correction);
  }
 } // namespace Eval
 /// evaluate() is the evaluator for the outer world. It returns a static
@@ -940,10 +1096,56 @@ make_v:
 Value Eval::evaluate(const Position& pos) {
-  if (Eval::useNNUE)
+  pos.this_thread()->on_eval();
-      return NNUE::evaluate(pos);
+
-  else
+  Value v;
-      return Evaluation<NO_TRACE>(pos).value();
+
  if (NNUE::useNNUE == NNUE::UseNNUEMode::Pure) {
      v = NNUE::evaluate(pos);
      // Guarantee evaluation does not hit the tablebase range
      v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
      return v;
  }
  // Deciding between classical and NNUE eval (~10 Elo): for high PSQ imbalance we use classical,
  // but we switch to NNUE during long shuffling or with high material on the board.
  bool useClassical = (pos.this_thread()->depth > 9 || pos.count<ALL_PIECES>() > 7) &&
          abs(eg_value(pos.psq_score())) * 5 > (856 + pos.non_pawn_material() / 64) * (10 + pos.rule50_count());
  // Deciding between classical and NNUE eval (~10 Elo): for high PSQ imbalance we use classical,
  // but we switch to NNUE during long shuffling or with high material on the board.
  if (NNUE::useNNUE == NNUE::UseNNUEMode::False || useClassical)
  {
      v = Evaluation<NO_TRACE>(pos).value();          // classical
      useClassical = abs(v) >= 297;
  }
  // If result of a classical evaluation is much lower than threshold fall back to NNUE
  if (NNUE::useNNUE != NNUE::UseNNUEMode::False && !useClassical)
  {
       Value nnue     = NNUE::evaluate(pos, true);     // NNUE
       int scale      = 1080 + 110 * pos.non_pawn_material() / 5120;
       Color stm      = pos.side_to_move();
       Value optimism = pos.this_thread()->optimism[stm];
       Value psq      = (stm == WHITE ? 1 : -1) * eg_value(pos.psq_score());
       int complexity = (278 * abs(nnue - psq)) / 256;
       optimism = optimism * (251 + complexity) / 256;
       v = (nnue * scale + optimism * (scale - 852)) / 1024;
       if (pos.is_chess960())
           v += fix_FRC(pos);
  }
  // Damp down the evaluation linearly when shuffling
  v = v * (195 - pos.rule50_count()) / 211;
  // Guarantee evaluation does not hit the tablebase range
  v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
  return v;
 }
 /// trace() is like evaluate(), but instead of returning a value, it returns
@@ -951,7 +1153,7 @@ Value Eval::evaluate(const Position& pos) {
 /// descriptions and values of each evaluation term. Useful for debugging.
 /// Trace scores are from white's point of view
-std::string Eval::trace(const Position& pos) {
+std::string Eval::trace(Position& pos) {
  if (pos.checkers())
      return "Final evaluation: none (in check)";
@@ -961,42 +1163,62 @@ std::string Eval::trace(const Position& pos) {
  Value v;
-  if (Eval::useNNUE)
+  std::memset(scores, 0, sizeof(scores));
  {
      v = NNUE::evaluate(pos);
  }
  else
  {
      std::memset(scores, 0, sizeof(scores));
-      pos.this_thread()->contempt = SCORE_ZERO; // Reset any dynamic contempt
+  // Reset any global variable used in eval
  pos.this_thread()->depth           = 0;
  pos.this_thread()->trend           = SCORE_ZERO;
  pos.this_thread()->bestValue       = VALUE_ZERO;
  pos.this_thread()->optimism[WHITE] = VALUE_ZERO;
  pos.this_thread()->optimism[BLACK] = VALUE_ZERO;
-      v = Evaluation<TRACE>(pos).value();
+  v = Evaluation<TRACE>(pos).value();
-      ss << std::showpoint << std::noshowpos << std::fixed << std::setprecision(2)
+  ss << std::showpoint << std::noshowpos << std::fixed << std::setprecision(2)
-         << "     Term    |    White    |    Black    |    Total   \n"
+     << " Contributing terms for the classical eval:\n"
-         << "             |   MG    EG  |   MG    EG  |   MG    EG \n"
+     << "+------------+-------------+-------------+-------------+\n"
-         << " ------------+-------------+-------------+------------\n"
+     << "|    Term    |    White    |    Black    |    Total    |\n"
-         << "    Material | " << Term(MATERIAL)
+     << "|            |   MG    EG  |   MG    EG  |   MG    EG  |\n"
-         << "   Imbalance | " << Term(IMBALANCE)
+     << "+------------+-------------+-------------+-------------+\n"
-         << "       Pawns | " << Term(PAWN)
+     << "|   Material | " << Term(MATERIAL)
-         << "     Knights | " << Term(KNIGHT)
+     << "|  Imbalance | " << Term(IMBALANCE)
-         << "     Bishops | " << Term(BISHOP)
+     << "|      Pawns | " << Term(PAWN)
-         << "       Rooks | " << Term(ROOK)
+     << "|    Knights | " << Term(KNIGHT)
-         << "      Queens | " << Term(QUEEN)
+     << "|    Bishops | " << Term(BISHOP)
-         << "    Mobility | " << Term(MOBILITY)
+     << "|      Rooks | " << Term(ROOK)
-         << " King safety | " << Term(KING)
+     << "|     Queens | " << Term(QUEEN)
-         << "     Threats | " << Term(THREAT)
+     << "|   Mobility | " << Term(MOBILITY)
-         << "      Passed | " << Term(PASSED)
+     << "|King safety | " << Term(KING)
-         << "       Space | " << Term(SPACE)
+     << "|    Threats | " << Term(THREAT)
-         << "    Winnable | " << Term(WINNABLE)
+     << "|     Passed | " << Term(PASSED)
-         << " ------------+-------------+-------------+------------\n"
+     << "|      Space | " << Term(SPACE)
-         << "       Total | " << Term(TOTAL);
+     << "|   Winnable | " << Term(WINNABLE)
-  }
+     << "+------------+-------------+-------------+-------------+\n"
     << "|      Total | " << Term(TOTAL)
     << "+------------+-------------+-------------+-------------+\n";
  if (NNUE::useNNUE != NNUE::UseNNUEMode::False)
      ss << '\n' << NNUE::trace(pos) << '\n';
  ss << std::showpoint << std::showpos << std::fixed << std::setprecision(2) << std::setw(15);
  v = pos.side_to_move() == WHITE ? v : -v;
  ss << "\nClassical evaluation   " << to_cp(v) << " (white side)\n";
  if (NNUE::useNNUE != NNUE::UseNNUEMode::False)
  {
      v = NNUE::evaluate(pos, false);
      v = pos.side_to_move() == WHITE ? v : -v;
      ss << "NNUE evaluation        " << to_cp(v) << " (white side)\n";
  }
-  ss << "\nFinal evaluation: " << to_cp(v) << " (white side)\n";
+  v = evaluate(pos);
  v = pos.side_to_move() == WHITE ? v : -v;
  ss << "Final evaluation       " << to_cp(v) << " (white side)";
  if (NNUE::useNNUE != NNUE::UseNNUEMode::False)
     ss << " [with scaled NNUE, hybrid, ...]";
  ss << "\n";
  return ss.str();
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -20,30 +20,49 @@
 #define EVALUATE_H_INCLUDED
 #include <string>
 #include <optional>
 #include "types.h"
 namespace Stockfish {
 class Position;
 namespace Eval {
-  std::string trace(const Position& pos);
+  std::string trace(Position& pos);
  Value evaluate(const Position& pos);
-  extern bool useNNUE;
+  // The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
-  extern std::string eval_file_loaded;
+  // for the build process (profile-build and fishtest) to work. Do not change the
-  void init_NNUE();
+  // name of the macro, as it is used in the Makefile.
-  void verify_NNUE();
+  #define EvalFileDefaultName   "nn-3c0aa92af1da.nnue"
  namespace NNUE {
    enum struct UseNNUEMode
    {
      False,
      True,
      Pure
    };
-    Value evaluate(const Position& pos);
+    extern UseNNUEMode useNNUE;
-    Value compute_eval(const Position& pos);
+    extern std::string currentEvalFileName;
-    void  update_eval(const Position& pos);
+
-    bool  load_eval_file(const std::string& evalFile);
+    std::string trace(Position& pos);
    Value evaluate(const Position& pos, bool adjusted = false);
    void init();
    void verify();
    bool load_eval(std::string name, std::istream& stream);
    bool save_eval(std::ostream& stream);
    bool save_eval(const std::optional<std::string>& filename);
  } // namespace NNUE
 } // namespace Eval
 } // namespace Stockfish
 #endif // #ifndef EVALUATE_H_INCLUDED
@@ -0,0 +1,26 @@
 The file "incbin.h" is free and unencumbered software released into
 the public domain by Dale Weiler, see:
   <https://github.com/graphitemaster/incbin>
 Anyone is free to copy, modify, publish, use, compile, sell, or
 distribute this software, either in source code form or as a compiled
 binary, for any purpose, commercial or non-commercial, and by any
 means.
 In jurisdictions that recognize copyright laws, the author or authors
 of this software dedicate any and all copyright interest in the
 software to the public domain. We make this dedication for the benefit
 of the public at large and to the detriment of our heirs and
 successors. We intend this dedication to be an overt act of
 relinquishment in perpetuity of all present and future rights to this
 software under copyright law.
 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
 IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
 OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 OTHER DEALINGS IN THE SOFTWARE.
 For more information, please refer to <http://unlicense.org/>
@@ -0,0 +1,368 @@
 /**
 * @file incbin.h
 * @author Dale Weiler
 * @brief Utility for including binary files
 *
 * Facilities for including binary files into the current translation unit and
 * making use from them externally in other translation units.
 */
 #ifndef INCBIN_HDR
 #define INCBIN_HDR
 #include <limits.h>
 #if   defined(__AVX512BW__) || \
      defined(__AVX512CD__) || \
      defined(__AVX512DQ__) || \
      defined(__AVX512ER__) || \
      defined(__AVX512PF__) || \
      defined(__AVX512VL__) || \
      defined(__AVX512F__)
 # define INCBIN_ALIGNMENT_INDEX 6
 #elif defined(__AVX__)      || \
      defined(__AVX2__)
 # define INCBIN_ALIGNMENT_INDEX 5
 #elif defined(__SSE__)      || \
      defined(__SSE2__)     || \
      defined(__SSE3__)     || \
      defined(__SSSE3__)    || \
      defined(__SSE4_1__)   || \
      defined(__SSE4_2__)   || \
      defined(__neon__)
 # define INCBIN_ALIGNMENT_INDEX 4
 #elif ULONG_MAX != 0xffffffffu
 # define INCBIN_ALIGNMENT_INDEX 3
 # else
 # define INCBIN_ALIGNMENT_INDEX 2
 #endif
 /* Lookup table of (1 << n) where `n' is `INCBIN_ALIGNMENT_INDEX' */
 #define INCBIN_ALIGN_SHIFT_0 1
 #define INCBIN_ALIGN_SHIFT_1 2
 #define INCBIN_ALIGN_SHIFT_2 4
 #define INCBIN_ALIGN_SHIFT_3 8
 #define INCBIN_ALIGN_SHIFT_4 16
 #define INCBIN_ALIGN_SHIFT_5 32
 #define INCBIN_ALIGN_SHIFT_6 64
 /* Actual alignment value */
 #define INCBIN_ALIGNMENT \
    INCBIN_CONCATENATE( \
        INCBIN_CONCATENATE(INCBIN_ALIGN_SHIFT, _), \
        INCBIN_ALIGNMENT_INDEX)
 /* Stringize */
 #define INCBIN_STR(X) \
    #X
 #define INCBIN_STRINGIZE(X) \
    INCBIN_STR(X)
 /* Concatenate */
 #define INCBIN_CAT(X, Y) \
    X ## Y
 #define INCBIN_CONCATENATE(X, Y) \
    INCBIN_CAT(X, Y)
 /* Deferred macro expansion */
 #define INCBIN_EVAL(X) \
    X
 #define INCBIN_INVOKE(N, ...) \
    INCBIN_EVAL(N(__VA_ARGS__))
 /* Green Hills uses a different directive for including binary data */
 #if defined(__ghs__)
 #  if (__ghs_asm == 2)
 #    define INCBIN_MACRO ".file"
 /* Or consider the ".myrawdata" entry in the ld file */
 #  else
 #    define INCBIN_MACRO "\tINCBIN"
 #  endif
 #else
 #  define INCBIN_MACRO ".incbin"
 #endif
 #ifndef _MSC_VER
 #  define INCBIN_ALIGN \
    __attribute__((aligned(INCBIN_ALIGNMENT)))
 #else
 #  define INCBIN_ALIGN __declspec(align(INCBIN_ALIGNMENT))
 #endif
 #if defined(__arm__) || /* GNU C and RealView */ \
    defined(__arm) || /* Diab */ \
    defined(_ARM) /* ImageCraft */
 #  define INCBIN_ARM
 #endif
 #ifdef __GNUC__
 /* Utilize .balign where supported */
 #  define INCBIN_ALIGN_HOST ".balign " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n"
 #  define INCBIN_ALIGN_BYTE ".balign 1\n"
 #elif defined(INCBIN_ARM)
 /*
 * On arm assemblers, the alignment value is calculated as (1 << n) where `n' is
 * the shift count. This is the value passed to `.align'
 */
 #  define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT_INDEX) "\n"
 #  define INCBIN_ALIGN_BYTE ".align 0\n"
 #else
 /* We assume other inline assembler's treat `.align' as `.balign' */
 #  define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n"
 #  define INCBIN_ALIGN_BYTE ".align 1\n"
 #endif
 /* INCBIN_CONST is used by incbin.c generated files */
 #if defined(__cplusplus)
 #  define INCBIN_EXTERNAL extern "C"
 #  define INCBIN_CONST    extern const
 #else
 #  define INCBIN_EXTERNAL extern
 #  define INCBIN_CONST    const
 #endif
 /**
 * @brief Optionally override the linker section into which data is emitted.
 *
 * @warning If you use this facility, you'll have to deal with platform-specific linker output
 * section naming on your own
 *
 * Overriding the default linker output section, e.g for esp8266/Arduino:
 * @code
 * #define INCBIN_OUTPUT_SECTION ".irom.text"
 * #include "incbin.h"
 * INCBIN(Foo, "foo.txt");
 * // Data is emitted into program memory that never gets copied to RAM
 * @endcode
 */
 #if !defined(INCBIN_OUTPUT_SECTION)
 #  if defined(__APPLE__)
 #    define INCBIN_OUTPUT_SECTION         ".const_data"
 #  else
 #    define INCBIN_OUTPUT_SECTION         ".rodata"
 #  endif
 #endif
 #if defined(__APPLE__)
 /* The directives are different for Apple branded compilers */
 #  define INCBIN_SECTION         INCBIN_OUTPUT_SECTION "\n"
 #  define INCBIN_GLOBAL(NAME)    ".globl " INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n"
 #  define INCBIN_INT             ".long "
 #  define INCBIN_MANGLE          "_"
 #  define INCBIN_BYTE            ".byte "
 #  define INCBIN_TYPE(...)
 #else
 #  define INCBIN_SECTION         ".section " INCBIN_OUTPUT_SECTION "\n"
 #  define INCBIN_GLOBAL(NAME)    ".global " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n"
 #  if defined(__ghs__)
 #    define INCBIN_INT           ".word "
 #  else
 #    define INCBIN_INT           ".int "
 #  endif
 #  if defined(__USER_LABEL_PREFIX__)
 #    define INCBIN_MANGLE        INCBIN_STRINGIZE(__USER_LABEL_PREFIX__)
 #  else
 #    define INCBIN_MANGLE        ""
 #  endif
 #  if defined(INCBIN_ARM)
 /* On arm assemblers, `@' is used as a line comment token */
 #    define INCBIN_TYPE(NAME)    ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", %object\n"
 #  elif defined(__MINGW32__) || defined(__MINGW64__)
 /* Mingw doesn't support this directive either */
 #    define INCBIN_TYPE(NAME)
 #  else
 /* It's safe to use `@' on other architectures */
 #    define INCBIN_TYPE(NAME)    ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", @object\n"
 #  endif
 #  define INCBIN_BYTE            ".byte "
 #endif
 /* List of style types used for symbol names */
 #define INCBIN_STYLE_CAMEL 0
 #define INCBIN_STYLE_SNAKE 1
 /**
 * @brief Specify the prefix to use for symbol names.
 *
 * By default this is `g', producing symbols of the form:
 * @code
 * #include "incbin.h"
 * INCBIN(Foo, "foo.txt");
 *
 * // Now you have the following symbols:
 * // const unsigned char gFooData[];
 * // const unsigned char *const gFooEnd;
 * // const unsigned int gFooSize;
 * @endcode
 *
 * If however you specify a prefix before including: e.g:
 * @code
 * #define INCBIN_PREFIX incbin
 * #include "incbin.h"
 * INCBIN(Foo, "foo.txt");
 *
 * // Now you have the following symbols instead:
 * // const unsigned char incbinFooData[];
 * // const unsigned char *const incbinFooEnd;
 * // const unsigned int incbinFooSize;
 * @endcode
 */
 #if !defined(INCBIN_PREFIX)
 #  define INCBIN_PREFIX g
 #endif
 /**
 * @brief Specify the style used for symbol names.
 *
 * Possible options are
 * - INCBIN_STYLE_CAMEL "CamelCase"
 * - INCBIN_STYLE_SNAKE "snake_case"
 *
 * Default option is *INCBIN_STYLE_CAMEL* producing symbols of the form:
 * @code
 * #include "incbin.h"
 * INCBIN(Foo, "foo.txt");
 *
 * // Now you have the following symbols:
 * // const unsigned char <prefix>FooData[];
 * // const unsigned char *const <prefix>FooEnd;
 * // const unsigned int <prefix>FooSize;
 * @endcode
 *
 * If however you specify a style before including: e.g:
 * @code
 * #define INCBIN_STYLE INCBIN_STYLE_SNAKE
 * #include "incbin.h"
 * INCBIN(foo, "foo.txt");
 *
 * // Now you have the following symbols:
 * // const unsigned char <prefix>foo_data[];
 * // const unsigned char *const <prefix>foo_end;
 * // const unsigned int <prefix>foo_size;
 * @endcode
 */
 #if !defined(INCBIN_STYLE)
 #  define INCBIN_STYLE INCBIN_STYLE_CAMEL
 #endif
 /* Style lookup tables */
 #define INCBIN_STYLE_0_DATA Data
 #define INCBIN_STYLE_0_END End
 #define INCBIN_STYLE_0_SIZE Size
 #define INCBIN_STYLE_1_DATA _data
 #define INCBIN_STYLE_1_END _end
 #define INCBIN_STYLE_1_SIZE _size
 /* Style lookup: returning identifier */
 #define INCBIN_STYLE_IDENT(TYPE) \
    INCBIN_CONCATENATE( \
        INCBIN_STYLE_, \
        INCBIN_CONCATENATE( \
            INCBIN_EVAL(INCBIN_STYLE), \
            INCBIN_CONCATENATE(_, TYPE)))
 /* Style lookup: returning string literal */
 #define INCBIN_STYLE_STRING(TYPE) \
    INCBIN_STRINGIZE( \
        INCBIN_STYLE_IDENT(TYPE)) \
 /* Generate the global labels by indirectly invoking the macro with our style
 * type and concatenating the name against them. */
 #define INCBIN_GLOBAL_LABELS(NAME, TYPE) \
    INCBIN_INVOKE( \
        INCBIN_GLOBAL, \
        INCBIN_CONCATENATE( \
            NAME, \
            INCBIN_INVOKE( \
                INCBIN_STYLE_IDENT, \
                TYPE))) \
    INCBIN_INVOKE( \
        INCBIN_TYPE, \
        INCBIN_CONCATENATE( \
            NAME, \
            INCBIN_INVOKE( \
                INCBIN_STYLE_IDENT, \
                TYPE)))
 /**
 * @brief Externally reference binary data included in another translation unit.
 *
 * Produces three external symbols that reference the binary data included in
 * another translation unit.
 *
 * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
 * "Data", as well as "End" and "Size" after. An example is provided below.
 *
 * @param NAME The name given for the binary data
 *
 * @code
 * INCBIN_EXTERN(Foo);
 *
 * // Now you have the following symbols:
 * // extern const unsigned char <prefix>FooData[];
 * // extern const unsigned char *const <prefix>FooEnd;
 * // extern const unsigned int <prefix>FooSize;
 * @endcode
 */
 #define INCBIN_EXTERN(NAME) \
    INCBIN_EXTERNAL const INCBIN_ALIGN unsigned char \
        INCBIN_CONCATENATE( \
            INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
            INCBIN_STYLE_IDENT(DATA))[]; \
    INCBIN_EXTERNAL const INCBIN_ALIGN unsigned char *const \
    INCBIN_CONCATENATE( \
        INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
        INCBIN_STYLE_IDENT(END)); \
    INCBIN_EXTERNAL const unsigned int \
        INCBIN_CONCATENATE( \
            INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
            INCBIN_STYLE_IDENT(SIZE))
 /**
 * @brief Include a binary file into the current translation unit.
 *
 * Includes a binary file into the current translation unit, producing three symbols
 * for objects that encode the data and size respectively.
 *
 * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
 * "Data", as well as "End" and "Size" after. An example is provided below.
 *
 * @param NAME The name to associate with this binary data (as an identifier.)
 * @param FILENAME The file to include (as a string literal.)
 *
 * @code
 * INCBIN(Icon, "icon.png");
 *
 * // Now you have the following symbols:
 * // const unsigned char <prefix>IconData[];
 * // const unsigned char *const <prefix>IconEnd;
 * // const unsigned int <prefix>IconSize;
 * @endcode
 *
 * @warning This must be used in global scope
 * @warning The identifiers may be different if INCBIN_STYLE is not default
 *
 * To externally reference the data included by this in another translation unit
 * please @see INCBIN_EXTERN.
 */
 #ifdef _MSC_VER
 #define INCBIN(NAME, FILENAME) \
    INCBIN_EXTERN(NAME)
 #else
 #define INCBIN(NAME, FILENAME) \
    __asm__(INCBIN_SECTION \
            INCBIN_GLOBAL_LABELS(NAME, DATA) \
            INCBIN_ALIGN_HOST \
            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) ":\n" \
            INCBIN_MACRO " \"" FILENAME "\"\n" \
            INCBIN_GLOBAL_LABELS(NAME, END) \
            INCBIN_ALIGN_BYTE \
            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) ":\n" \
                INCBIN_BYTE "1\n" \
            INCBIN_GLOBAL_LABELS(NAME, SIZE) \
            INCBIN_ALIGN_HOST \
            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(SIZE) ":\n" \
                INCBIN_INT INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) " - " \
                           INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) "\n" \
            INCBIN_ALIGN_HOST \
            ".text\n" \
    ); \
    INCBIN_EXTERN(NAME)
 #endif
 #endif
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -18,23 +18,25 @@
 #include <iostream>
 #include "nnue/evaluate_nnue.h"
 #include "bitboard.h"
 #include "endgame.h"
 #include "position.h"
 #include "psqt.h"
 #include "search.h"
 #include "syzygy/tbprobe.h"
 #include "thread.h"
 #include "tt.h"
 #include "uci.h"
 #include "syzygy/tbprobe.h"
-namespace PSQT {
+using namespace Stockfish;
  void init();
 }
 int main(int argc, char* argv[]) {
  std::cout << engine_info() << std::endl;
  CommandLine::init(argc, argv);
  UCI::init(Options);
  Tune::init();
  PSQT::init();
@@ -44,7 +46,7 @@ int main(int argc, char* argv[]) {
  Endgames::init();
  Threads.set(size_t(Options["Threads"]));
  Search::clear(); // After threads are up
-  Eval::init_NNUE();
+  Eval::NNUE::init();
  UCI::loop(argc, argv);
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,32 +24,39 @@
 using namespace std;
 namespace Stockfish {
 namespace {
  #define S(mg, eg) make_score(mg, eg)
  // Polynomial material imbalance parameters
-  constexpr int QuadraticOurs[][PIECE_TYPE_NB] = {
+  // One Score parameter for each pair (our piece, another of our pieces)
-    //            OUR PIECES
+  constexpr Score QuadraticOurs[][PIECE_TYPE_NB] = {
-    // pair pawn knight bishop rook queen
+    // OUR PIECE 2
-    {1438                               }, // Bishop pair
+    // bishop pair    pawn         knight       bishop       rook           queen
-    {  40,   38                         }, // Pawn
+    {S(1419, 1455)                                                                  }, // Bishop pair
-    {  32,  255, -62                    }, // Knight      OUR PIECES
+    {S( 101,   28), S( 37,  39)                                                     }, // Pawn
-    {   0,  104,   4,    0              }, // Bishop
+    {S(  57,   64), S(249, 187), S(-49, -62)                                        }, // Knight      OUR PIECE 1
-    { -26,   -2,  47,   105,  -208      }, // Rook
+    {S(   0,    0), S(118, 137), S( 10,  27), S(  0,   0)                           }, // Bishop
-    {-189,   24, 117,   133,  -134, -6  }  // Queen
+    {S( -63,  -68), S( -5,   3), S(100,  81), S(132, 118), S(-246, -244)            }, // Rook
    {S(-210, -211), S( 37,  14), S(147, 141), S(161, 105), S(-158, -174), S(-9,-31) }  // Queen
  };
-  constexpr int QuadraticTheirs[][PIECE_TYPE_NB] = {
+  // One Score parameter for each pair (our piece, their piece)
-    //           THEIR PIECES
+  constexpr Score QuadraticTheirs[][PIECE_TYPE_NB] = {
-    // pair pawn knight bishop rook queen
+    // THEIR PIECE
-    {                                   }, // Bishop pair
+    // bishop pair   pawn         knight       bishop       rook         queen
-    {  36,                              }, // Pawn
+    {                                                                               }, // Bishop pair
-    {   9,   63,                        }, // Knight      OUR PIECES
+    {S(  33,  30)                                                                   }, // Pawn
-    {  59,   65,  42,                   }, // Bishop
+    {S(  46,  18), S(106,  84)                                                      }, // Knight      OUR PIECE
-    {  46,   39,  24,   -24,            }, // Rook
+    {S(  75,  35), S( 59,  44), S( 60,  15)                                         }, // Bishop
-    {  97,  100, -42,   137,  268,      }  // Queen
+    {S(  26,  35), S(  6,  22), S( 38,  39), S(-12,  -2)                            }, // Rook
    {S(  97,  93), S(100, 163), S(-58, -91), S(112, 192), S(276, 225)               }  // Queen
  };
  #undef S
  // Endgame evaluation and scaling functions are accessed directly and not through
  // the function maps because they correspond to more than one material hash key.
  Endgame<KXK>    EvaluateKXK[] = { Endgame<KXK>(WHITE),    Endgame<KXK>(BLACK) };
@@ -67,7 +74,7 @@ namespace {
  bool is_KBPsK(const Position& pos, Color us) {
    return   pos.non_pawn_material(us) == BishopValueMg
-          && pos.count<PAWN  >(us) >= 1;
+          && pos.count<PAWN>(us) >= 1;
  }
  bool is_KQKRPs(const Position& pos, Color us) {
@@ -82,11 +89,11 @@ namespace {
  /// piece type for both colors.
  template<Color Us>
-  int imbalance(const int pieceCount[][PIECE_TYPE_NB]) {
+  Score imbalance(const int pieceCount[][PIECE_TYPE_NB]) {
    constexpr Color Them = ~Us;
-    int bonus = 0;
+    Score bonus = SCORE_ZERO;
    // Second-degree polynomial material imbalance, by Tord Romstad
    for (int pt1 = NO_PIECE_TYPE; pt1 <= QUEEN; ++pt1)
@@ -130,7 +137,7 @@ Entry* probe(const Position& pos) {
  Value npm_w = pos.non_pawn_material(WHITE);
  Value npm_b = pos.non_pawn_material(BLACK);
-  Value npm   = Utility::clamp(npm_w + npm_b, EndgameLimit, MidgameLimit);
+  Value npm   = std::clamp(npm_w + npm_b, EndgameLimit, MidgameLimit);
  // Map total non-pawn material into [PHASE_ENDGAME, PHASE_MIDGAME]
  e->gamePhase = Phase(((npm - EndgameLimit) * PHASE_MIDGAME) / (MidgameLimit - EndgameLimit));
@@ -213,8 +220,10 @@ Entry* probe(const Position& pos) {
  { pos.count<BISHOP>(BLACK) > 1, pos.count<PAWN>(BLACK), pos.count<KNIGHT>(BLACK),
    pos.count<BISHOP>(BLACK)    , pos.count<ROOK>(BLACK), pos.count<QUEEN >(BLACK) } };
-  e->value = int16_t((imbalance<WHITE>(pieceCount) - imbalance<BLACK>(pieceCount)) / 16);
+  e->score = (imbalance<WHITE>(pieceCount) - imbalance<BLACK>(pieceCount)) / 16;
  return e;
 }
 } // namespace Material
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,7 +24,7 @@
 #include "position.h"
 #include "types.h"
-namespace Material {
+namespace Stockfish::Material {
 /// Material::Entry contains various information about a material configuration.
 /// It contains a material imbalance evaluation, a function pointer to a special
@@ -37,8 +37,8 @@ namespace Material {
 struct Entry {
-  Score imbalance() const { return make_score(value, value); }
+  Score imbalance() const { return score; }
-  Phase game_phase() const { return gamePhase; }
+  Phase game_phase() const { return (Phase)gamePhase; }
  bool specialized_eval_exists() const { return evaluationFunction != nullptr; }
  Value evaluate(const Position& pos) const { return (*evaluationFunction)(pos); }
@@ -57,15 +57,15 @@ struct Entry {
  const EndgameBase<Value>* evaluationFunction;
  const EndgameBase<ScaleFactor>* scalingFunction[COLOR_NB]; // Could be one for each
                                                             // side (e.g. KPKP, KBPsK)
-  int16_t value;
+  Score score;
  int16_t gamePhase;
  uint8_t factor[COLOR_NB];
  Phase gamePhase;
 };
 typedef HashTable<Entry, 8192> Table;
 Entry* probe(const Position& pos);
-} // namespace Material
+} // namespace Stockfish::Material
 #endif // #ifndef MATERIAL_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -36,6 +36,8 @@ typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP,
                      PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
 typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY);
 typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
 typedef bool(*fun4_t)(USHORT, PGROUP_AFFINITY, USHORT, PUSHORT);
 typedef WORD(*fun5_t)();
 }
 #endif
@@ -51,11 +53,20 @@ typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
 #include <sys/mman.h>
 #endif
 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) || defined(__e2k__)
 #define POSIXALIGNEDALLOC
 #include <stdlib.h>
 #endif
 #include "misc.h"
 #include "thread.h"
 using namespace std;
 namespace Stockfish {
 SynchronizedRegionLogger sync_region_cout(std::cout);
 namespace {
 /// Version number. If Version is left empty, then compile date in the format
@@ -103,7 +114,14 @@ public:
    static Logger l;
-    if (!fname.empty() && !l.file.is_open())
+    if (l.file.is_open())
    {
        cout.rdbuf(l.out.buf);
        cin.rdbuf(l.in.buf);
        l.file.close();
    }
    if (!fname.empty())
    {
        l.file.open(fname, ifstream::out);
@@ -116,23 +134,18 @@ public:
        cin.rdbuf(&l.in);
        cout.rdbuf(&l.out);
    }
    else if (fname.empty() && l.file.is_open())
    {
        cout.rdbuf(l.out.buf);
        cin.rdbuf(l.in.buf);
        l.file.close();
    }
  }
 };
 } // namespace
 /// engine_info() returns the full name of the current Stockfish version. This
 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
 /// the program was compiled) or "Stockfish <Version>", depending on whether
 /// Version is empty.
-const string engine_info(bool to_uci) {
+string engine_info(bool to_uci) {
  const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
  string month, day, year;
@@ -155,7 +168,7 @@ const string engine_info(bool to_uci) {
 /// compiler_info() returns a string trying to describe the compiler we use
-const std::string compiler_info() {
+std::string compiler_info() {
  #define stringify2(x) #x
  #define stringify(x) stringify2(x)
@@ -184,6 +197,18 @@ const std::string compiler_info() {
     compiler += "(version ";
     compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
     compiler += ")";
  #elif defined(__e2k__) && defined(__LCC__)
    #define dot_ver2(n) \
      compiler += (char)'.'; \
      compiler += (char)('0' + (n) / 10); \
      compiler += (char)('0' + (n) % 10);
     compiler += "MCST LCC ";
     compiler += "(version ";
     compiler += std::to_string(__LCC__ / 100);
     dot_ver2(__LCC__ % 100)
     dot_ver2(__LCC_MINOR__)
     compiler += ")";
  #elif __GNUC__
     compiler += "g++ (GNUC) ";
     compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
@@ -214,26 +239,33 @@ const std::string compiler_info() {
  compiler += "\nCompilation settings include: ";
  compiler += (Is64Bit ? " 64bit" : " 32bit");
  #if defined(USE_VNNI)
    compiler += " VNNI";
  #endif
  #if defined(USE_AVX512)
    compiler += " AVX512";
  #endif
  compiler += (HasPext ? " BMI2" : "");
  #if defined(USE_AVX2)
    compiler += " AVX2";
  #endif
  #if defined(USE_SSE42)
    compiler += " SSE42";
  #endif
  #if defined(USE_SSE41)
    compiler += " SSE41";
  #endif
  #if defined(USE_SSSE3)
    compiler += " SSSE3";
  #endif
-  #if defined(USE_SSE3)
+  #if defined(USE_SSE2)
-    compiler += " SSE3";
+    compiler += " SSE2";
  #endif
-    compiler += (HasPext ? " BMI2" : "");
+  compiler += (HasPopCnt ? " POPCNT" : "");
-    compiler += (HasPopCnt ? " POPCNT" : "");
+  #if defined(USE_MMX)
    compiler += " MMX";
  #endif
  #if defined(USE_NEON)
    compiler += " NEON";
  #endif
  #if !defined(NDEBUG)
    compiler += " DEBUG";
  #endif
@@ -316,13 +348,16 @@ void prefetch(void* addr) {
 #endif
-/// Wrappers for systems where the c++17 implementation doesn't guarantee the availability of aligned_alloc.
+
-/// Memory allocated with std_aligned_alloc must be freed with std_aligned_free.
+/// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
-///
+/// does not guarantee the availability of aligned_alloc(). Memory allocated with
 /// std_aligned_alloc() must be freed with std_aligned_free().
 void* std_aligned_alloc(size_t alignment, size_t size) {
-#if defined(__APPLE__)
+
-  return aligned_alloc(alignment, size);
+#if defined(POSIXALIGNEDALLOC)
  void *mem;
  return posix_memalign(&mem, alignment, size) ? nullptr : mem;
 #elif defined(_WIN32)
  return _mm_malloc(size, alignment);
 #else
@@ -331,7 +366,8 @@ void* std_aligned_alloc(size_t alignment, size_t size) {
 }
 void std_aligned_free(void* ptr) {
-#if defined(__APPLE__)
+
 #if defined(POSIXALIGNEDALLOC)
  free(ptr);
 #elif defined(_WIN32)
  _mm_free(ptr);
@@ -340,25 +376,16 @@ void std_aligned_free(void* ptr) {
 #endif
 }
-/// aligned_ttmem_alloc() will return suitably aligned memory, and if possible use large pages.
+/// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
 /// The returned pointer is the aligned one, while the mem argument is the one that needs
 /// to be passed to free. With c++17 some of this functionality could be simplified.
-#if defined(__linux__) && !defined(__ANDROID__)
+#if defined(_WIN32)
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
+static void* aligned_large_pages_alloc_windows(size_t allocSize) {
-  constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page sizes
+  #if !defined(_WIN64)
-  size_t size = ((allocSize + alignment - 1) / alignment) * alignment; // multiple of alignment
+    (void)allocSize; // suppress unused-parameter compiler warning
-  if (posix_memalign(&mem, alignment, size))
+    return nullptr;
-     mem = nullptr;
+  #else
  madvise(mem, allocSize, MADV_HUGEPAGE);
  return mem;
 }
 #elif defined(_WIN64)
 static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
  HANDLE hProcessToken { };
  LUID luid { };
@@ -401,25 +428,14 @@ static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
  CloseHandle(hProcessToken);
  return mem;
  #endif
 }
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
+void* aligned_large_pages_alloc(size_t allocSize) {
  static bool firstCall = true;
  // Try to allocate large pages
-  mem = aligned_ttmem_alloc_large_pages(allocSize);
+  void* mem = aligned_large_pages_alloc_windows(allocSize);
  // Suppress info strings on the first call. The first call occurs before 'uci'
  // is received and in that case this output confuses some GUIs.
  if (!firstCall)
  {
      if (mem)
          sync_cout << "info string Hash table allocation: Windows large pages used." << sync_endl;
      else
          sync_cout << "info string Hash table allocation: Windows large pages not used." << sync_endl;
  }
  firstCall = false;
  // Fall back to regular, page aligned, allocation if necessary
  if (!mem)
@@ -430,37 +446,46 @@ void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
 #else
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
+void* aligned_large_pages_alloc(size_t allocSize) {
-  constexpr size_t alignment = 64; // assumed cache line size
+#if defined(__linux__)
-  size_t size = allocSize + alignment - 1; // allocate some extra space
+  constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
-  mem = malloc(size);
+#else
-  void* ret = reinterpret_cast<void*>((uintptr_t(mem) + alignment - 1) & ~uintptr_t(alignment - 1));
+  constexpr size_t alignment = 4096; // assumed small page size
-  return ret;
+#endif
  // round up to multiples of alignment
  size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
  void *mem = std_aligned_alloc(alignment, size);
 #if defined(MADV_HUGEPAGE)
  madvise(mem, size, MADV_HUGEPAGE);
 #endif
  return mem;
 }
 #endif
-/// aligned_ttmem_free() will free the previously allocated ttmem
+/// aligned_large_pages_free() will free the previously allocated ttmem
-#if defined(_WIN64)
+#if defined(_WIN32)
-void aligned_ttmem_free(void* mem) {
+void aligned_large_pages_free(void* mem) {
  if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
  {
      DWORD err = GetLastError();
-      std::cerr << "Failed to free transposition table. Error code: 0x" <<
+      std::cerr << "Failed to free large page memory. Error code: 0x"
-          std::hex << err << std::dec << std::endl;
+                << std::hex << err
                << std::dec << std::endl;
      exit(EXIT_FAILURE);
  }
 }
 #else
-void aligned_ttmem_free(void *mem) {
+void aligned_large_pages_free(void *mem) {
-  free(mem);
+  std_aligned_free(mem);
 }
 #endif
@@ -474,11 +499,11 @@ void bindThisThread(size_t) {}
 #else
-/// best_group() retrieves logical processor information using Windows specific
+/// best_node() retrieves logical processor information using Windows specific
-/// API and returns the best group id for the thread with index idx. Original
+/// API and returns the best node id for the thread with index idx. Original
 /// code from Texel by Peter Österlund.
-int best_group(size_t idx) {
+int best_node(size_t idx) {
  int threads = 0;
  int nodes = 0;
@@ -492,7 +517,8 @@ int best_group(size_t idx) {
  if (!fun1)
      return -1;
-  // First call to get returnLength. We expect it to fail due to null buffer
+  // First call to GetLogicalProcessorInformationEx() to get returnLength.
  // We expect the call to fail due to null buffer.
  if (fun1(RelationAll, nullptr, &returnLength))
      return -1;
@@ -500,7 +526,7 @@ int best_group(size_t idx) {
  SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
  ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
-  // Second call, now we expect to succeed
+  // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
  if (!fun1(RelationAll, buffer, &returnLength))
  {
      free(buffer);
@@ -550,24 +576,125 @@ int best_group(size_t idx) {
 void bindThisThread(size_t idx) {
  // Use only local variables to be thread-safe
-  int group = best_group(idx);
+  int node = best_node(idx);
-  if (group == -1)
+  if (node == -1)
      return;
  // Early exit if the needed API are not available at runtime
  HMODULE k32 = GetModuleHandle("Kernel32.dll");
  auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
  auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
  auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
  auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
  if (!fun2 || !fun3)
      return;
-  GROUP_AFFINITY affinity;
+  if (!fun4 || !fun5)
-  if (fun2(group, &affinity))
+  {
-      fun3(GetCurrentThread(), &affinity, nullptr);
+      GROUP_AFFINITY affinity;
      if (fun2(node, &affinity))                                                 // GetNumaNodeProcessorMaskEx
          fun3(GetCurrentThread(), &affinity, nullptr);                          // SetThreadGroupAffinity
  }
  else
  {
      // If a numa node has more than one processor group, we assume they are
      // sized equal and we spread threads evenly across the groups.
      USHORT elements, returnedElements;
      elements = fun5();                                                         // GetMaximumProcessorGroupCount
      GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
      if (fun4(node, affinity, elements, &returnedElements))                     // GetNumaNodeProcessorMask2
          fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr);  // SetThreadGroupAffinity
      free(affinity);
  }
 }
 #endif
 } // namespace WinProcGroup
 #ifdef _WIN32
 #include <direct.h>
 #define GETCWD _getcwd
 #else
 #include <unistd.h>
 #define GETCWD getcwd
 #endif
 namespace CommandLine {
 string argv0;            // path+name of the executable binary, as given by argv[0]
 string binaryDirectory;  // path of the executable directory
 string workingDirectory; // path of the working directory
 void init(int argc, char* argv[]) {
    (void)argc;
    string pathSeparator;
    // extract the path+name of the executable binary
    argv0 = argv[0];
 #ifdef _WIN32
    pathSeparator = "\\";
  #ifdef _MSC_VER
    // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
    // issues in some windows 10 versions, so check returned values carefully.
    char* pgmptr = nullptr;
    if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
        argv0 = pgmptr;
  #endif
 #else
    pathSeparator = "/";
 #endif
    // extract the working directory
    workingDirectory = "";
    char buff[40000];
    char* cwd = GETCWD(buff, 40000);
    if (cwd)
        workingDirectory = cwd;
    // extract the binary directory path from argv0
    binaryDirectory = argv0;
    size_t pos = binaryDirectory.find_last_of("\\/");
    if (pos == std::string::npos)
        binaryDirectory = "." + pathSeparator;
    else
        binaryDirectory.resize(pos + 1);
    // pattern replacement: "./" at the start of path is replaced by the working directory
    if (binaryDirectory.find("." + pathSeparator) == 0)
        binaryDirectory.replace(0, 1, workingDirectory);
 }
 } // namespace CommandLine
 // Returns a string that represents the current time. (Used when learning evaluation functions)
 std::string now_string()
 {
    // Using std::ctime(), localtime() gives a warning that MSVC is not secure.
    // This shouldn't happen in the C++ standard, but...
 #if defined(_MSC_VER)
  // C4996 : 'ctime' : This function or variable may be unsafe.Consider using ctime_s instead.
 #pragma warning(disable : 4996)
 #endif
    auto now = std::chrono::system_clock::now();
    auto tp = std::chrono::system_clock::to_time_t(now);
    auto result = string(std::ctime(&tp));
    // remove line endings if they are included at the end
    while (*result.rbegin() == '\n' || (*result.rbegin() == '\r'))
        result.pop_back();
    return result;
 }
 void sleep(int ms)
 {
    std::this_thread::sleep_for(std::chrono::milliseconds(ms));
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -19,32 +19,56 @@
 #ifndef MISC_H_INCLUDED
 #define MISC_H_INCLUDED
 #include <algorithm>
 #include <cassert>
 #include <chrono>
 #include <functional>
 #include <mutex>
 #include <ostream>
 #include <string>
 #include <vector>
 #include <iostream>
 #include <cstdint>
 #include <cmath>
 #include <cctype>
 #include <sstream>
 #include <deque>
 #include "types.h"
-const std::string engine_info(bool to_uci = false);
+namespace Stockfish {
-const std::string compiler_info();
+
 std::string engine_info(bool to_uci = false);
 std::string compiler_info();
 void prefetch(void* addr);
 void start_logger(const std::string& fname);
 void* std_aligned_alloc(size_t alignment, size_t size);
 void std_aligned_free(void* ptr);
-void* aligned_ttmem_alloc(size_t size, void*& mem);
+void* aligned_large_pages_alloc(size_t size); // memory aligned by page size, min alignment: 4096 bytes
-void aligned_ttmem_free(void* mem); // nop if mem == nullptr
+void aligned_large_pages_free(void* mem); // nop if mem == nullptr
 void dbg_hit_on(bool b);
 void dbg_hit_on(bool c, bool b);
 void dbg_mean_of(int v);
 void dbg_print();
 /// Debug macro to write to std::err if NDEBUG flag is set, and do nothing otherwise
 #if defined(NDEBUG)
 #define debug 1 && std::cerr
 #else
 #define debug 0 && std::cerr
 #endif
 inline void hit_any_key() {
 #ifndef NDEBUG
  debug << "Hit any key to continue..." << std::endl << std::flush;
  system("read");   // on Windows, should be system("pause");
 #endif
 }
 typedef std::chrono::milliseconds::rep TimePoint; // A value in milliseconds
 static_assert(sizeof(TimePoint) == sizeof(int64_t), "TimePoint should be 64 bits");
 inline TimePoint now() {
  return std::chrono::duration_cast<std::chrono::milliseconds>
        (std::chrono::steady_clock::now().time_since_epoch()).count();
@@ -65,15 +89,319 @@ std::ostream& operator<<(std::ostream&, SyncCout);
 #define sync_cout std::cout << IO_LOCK
 #define sync_endl std::endl << IO_UNLOCK
 namespace Utility {
-/// Clamp a value between lo and hi. Available in c++17.
+// align_ptr_up() : get the first aligned element of an array.
-template<class T> constexpr const T& clamp(const T& v, const T& lo, const T& hi) {
+// ptr must point to an array of size at least `sizeof(T) * N + alignment` bytes,
-  return v < lo ? lo : v > hi ? hi : v;
+// where N is the number of elements in the array.
 template <uintptr_t Alignment, typename T>
 T* align_ptr_up(T* ptr)
 {
  static_assert(alignof(T) < Alignment);
  const uintptr_t ptrint = reinterpret_cast<uintptr_t>(reinterpret_cast<char*>(ptr));
  return reinterpret_cast<T*>(reinterpret_cast<char*>((ptrint + (Alignment - 1)) / Alignment * Alignment));
 }
 // IsLittleEndian : true if and only if the binary is compiled on a little endian machine
 static inline const union { uint32_t i; char c[4]; } Le = { 0x01020304 };
 static inline const bool IsLittleEndian = (Le.c[0] == 4);
 // RunningAverage : a class to calculate a running average of a series of values.
 // For efficiency, all computations are done with integers.
 class RunningAverage {
  public:
      // Reset the running average to rational value p / q
      void set(int64_t p, int64_t q)
        { average = p * PERIOD * RESOLUTION / q; }
      // Update average with value v
      void update(int64_t v)
        { average = RESOLUTION * v + (PERIOD - 1) * average / PERIOD; }
      // Test if average is strictly greater than rational a / b
      bool is_greater(int64_t a, int64_t b) const
        { return b * average > a * (PERIOD * RESOLUTION); }
      int64_t value() const
        { return average / (PERIOD * RESOLUTION); }
  private :
      static constexpr int64_t PERIOD     = 4096;
      static constexpr int64_t RESOLUTION = 1024;
      int64_t average;
 };
 template <typename T, std::size_t MaxSize>
 class ValueList {
 public:
  std::size_t size() const { return size_; }
  void resize(std::size_t newSize) { size_ = newSize; }
  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 maxSize = std::max(size_, other.size_);
    for (std::size_t i = 0; i < maxSize; ++i) {
      std::swap(values_[i], other.values_[i]);
    }
    std::swap(size_, other.size_);
  }
 private:
  T values_[MaxSize];
  std::size_t size_ = 0;
 };
 // This logger allows printing many parts in a region atomically
 // but doesn't block the threads trying to append to other regions.
 // Instead if some region tries to pring while other region holds
 // the lock the messages are queued to be printed as soon as the
 // current region releases the lock.
 struct SynchronizedRegionLogger
 {
  using RegionId = std::uint64_t;
  struct Region
  {
    friend struct SynchronizedRegionLogger;
    Region() :
      logger(nullptr), region_id(0), is_held(false)
    {
    }
    Region(const Region&) = delete;
    Region& operator=(const Region&) = delete;
    Region(Region&& other) :
      logger(other.logger), region_id(other.region_id), is_held(other.is_held)
    {
      other.logger = nullptr;
      other.is_held = false;
    }
    Region& operator=(Region&& other) {
      if (is_held && logger != nullptr)
      {
        logger->release_region(region_id);
      }
      logger = other.logger;
      region_id = other.region_id;
      is_held = other.is_held;
      other.is_held = false;
      return *this;
    }
    ~Region() { unlock(); }
    void unlock() {
      if (is_held) {
        is_held = false;
        if (logger != nullptr)
          logger->release_region(region_id);
      }
    }
    Region& operator << (std::ostream&(*pManip)(std::ostream&)) {
      if (logger != nullptr)
        logger->write(region_id, pManip);
      return *this;
    }
    template <typename T>
    Region& operator << (const T& value) {
      if (logger != nullptr)
        logger->write(region_id, value);
      return *this;
    }
  private:
    SynchronizedRegionLogger* logger;
    RegionId region_id;
    bool is_held;
    Region(SynchronizedRegionLogger& log, RegionId id) :
      logger(&log), region_id(id), is_held(true)
    {
    }
  };
 private:
  struct RegionBookkeeping
  {
    RegionBookkeeping(RegionId rid) : id(rid), is_held(true) {}
    std::vector<std::string> pending_parts;
    RegionId id;
    bool is_held;
  };
  RegionId init_next_region()
  {
    static RegionId next_id = 0;
    std::lock_guard lock(mutex);
    const auto id = next_id++;
    regions.emplace_back(id);
    return id;
  }
  void write(RegionId id, std::ostream&(*pManip)(std::ostream&)) {
    std::lock_guard lock(mutex);
    if (regions.empty())
      return;
    if (id == regions.front().id) {
      // We can just directly print to the output because
      // we are at the front of the region queue.
      out << *pManip;
    } else {
      // We have to schedule the print until previous regions are
      // processed
      auto* region = find_region_nolock(id);
      if (region == nullptr)
        return;
      std::stringstream ss;
      ss << *pManip;
      region->pending_parts.emplace_back(std::move(ss).str());
    }
  }
  template <typename T>
  void write(RegionId id, const T& value) {
    std::lock_guard lock(mutex);
    if (regions.empty())
      return;
    if (id == regions.front().id) {
      // We can just directly print to the output because
      // we are at the front of the region queue.
      out << value;
    } else {
      // We have to schedule the print until previous regions are
      // processed
      auto* region = find_region_nolock(id);
      if (region == nullptr)
        return;
      std::stringstream ss;
      ss << value;
      region->pending_parts.emplace_back(std::move(ss).str());
    }
  }
  std::ostream& out;
  std::deque<RegionBookkeeping> regions;
  std::mutex mutex;
  RegionBookkeeping* find_region_nolock(RegionId id) {
    // Linear search because the amount of concurrent regions should be small.
    auto it = std::find_if(
      regions.begin(),
      regions.end(),
      [id](const RegionBookkeeping& r) { return r.id == id; });
    if (it == regions.end())
      return nullptr;
    else
      return &*it;
  }
  void release_region(RegionId id) {
    std::lock_guard lock(mutex);
    auto* region = find_region_nolock(id);
    if (region == nullptr)
      return;
    region->is_held = false;
    process_backlog_nolock();
  }
  void process_backlog_nolock()
  {
    while(!regions.empty()) {
      auto& region = regions.front();
      for(auto& part : region.pending_parts) {
        out << part;
      }
      // If the region is still held then we don't
      // want to start printing stuff from the next region.
      if (region.is_held)
        break;
      regions.pop_front();
    }
  }
 public:
  SynchronizedRegionLogger(std::ostream& s) :
    out(s)
  {
  }
  [[nodiscard]] Region new_region() {
    const auto id = init_next_region();
    return Region(*this, id);
  }
 };
 extern SynchronizedRegionLogger sync_region_cout;
 /// sigmoid(t, x0, y0, C, P, Q) implements a sigmoid-like function using only integers,
 /// with the following properties:
 ///
 ///  -  sigmoid is centered in (x0, y0)
 ///  -  sigmoid has amplitude [-P/Q , P/Q] instead of [-1 , +1]
 ///  -  limit is (y0 - P/Q) when t tends to -infinity
 ///  -  limit is (y0 + P/Q) when t tends to +infinity
 ///  -  the slope can be adjusted using C > 0, smaller C giving a steeper sigmoid
 ///  -  the slope of the sigmoid when t = x0 is P/(Q*C)
 ///  -  sigmoid is increasing with t when P > 0 and Q > 0
 ///  -  to get a decreasing sigmoid, change sign of P
 ///  -  mean value of the sigmoid is y0
 ///
 /// Use <https://www.desmos.com/calculator/jhh83sqq92> to draw the sigmoid
 inline int64_t sigmoid(int64_t t, int64_t x0,
                                  int64_t y0,
                                  int64_t  C,
                                  int64_t  P,
                                  int64_t  Q)
 {
   assert(C > 0);
   assert(Q != 0);
   return y0 + P * (t-x0) / (Q * (std::abs(t-x0) + C)) ;
 }
 /// xorshift64star Pseudo-Random Number Generator
 /// This class is based on original code written and dedicated
 /// to the public domain by Sebastiano Vigna (2014).
@@ -89,6 +417,19 @@ template<class T> constexpr const T& clamp(const T& v, const T& lo, const T& hi)
 /// For further analysis see
 ///   <http://vigna.di.unimi.it/ftp/papers/xorshift.pdf>
 static uint64_t string_hash(const std::string& str)
 {
  uint64_t h = 525201411107845655ull;
  for (auto c : str) {
    h ^= static_cast<uint64_t>(c);
    h *= 0x5bd1e9955bd1e995ull;
    h ^= h >> 47;
  }
  return h;
 }
 class PRNG {
  uint64_t s;
@@ -100,7 +441,9 @@ class PRNG {
  }
 public:
  PRNG() { set_seed_from_time(); }
  PRNG(uint64_t seed) : s(seed) { assert(seed); }
  PRNG(const std::string& seed) { set_seed(seed); }
  template<typename T> T rand() { return T(rand64()); }
@@ -108,8 +451,54 @@ public:
  /// Output values only have 1/8th of their bits set on average.
  template<typename T> T sparse_rand()
  { return T(rand64() & rand64() & rand64()); }
  // Returns a random number from 0 to n-1. (Not uniform distribution, but this is enough in reality)
  uint64_t rand(uint64_t n) { return rand<uint64_t>() % n; }
  // Return the random seed used internally.
  uint64_t get_seed() const { return s; }
  void set_seed(uint64_t seed) { s = seed; }
  uint64_t next_random_seed()
  {
    uint64_t seed = 0;
    for(int i = 0; i < 64; ++i)
    {
      const auto off = rand64() % 64;
      seed |= (rand64() & (uint64_t(1) << off)) >> off;
      seed <<= 1;
    }
    return seed;
  }
  void set_seed_from_time()
  {
      set_seed(std::chrono::system_clock::now().time_since_epoch().count());
  }
  void set_seed(const std::string& str)
  {
    if (str.empty())
    {
      set_seed_from_time();
    }
    else if (std::all_of(str.begin(), str.end(), [](char c) { return std::isdigit(c);} )) {
      set_seed(std::stoull(str));
    }
    else
    {
      set_seed(string_hash(str));
    }
  }
 };
 // Display a random seed. (For debugging)
 inline std::ostream& operator<<(std::ostream& os, PRNG& prng)
 {
  os << "PRNG::seed = " << std::hex << prng.get_seed() << std::dec;
  return os;
 }
 inline uint64_t mul_hi64(uint64_t a, uint64_t b) {
 #if defined(__GNUC__) && defined(IS_64BIT)
    __extension__ typedef unsigned __int128 uint128;
@@ -124,6 +513,74 @@ inline uint64_t mul_hi64(uint64_t a, uint64_t b) {
 #endif
 }
 // This bitset can be accessed concurrently, provided
 // the concurrent accesses are performed on distinct
 // instances of underlying type. That means the cuncurrent
 // accesses need to be spaced by at least
 // bits_per_bucket bits.
 // But at least best_concurrent_access_stride bits
 // is recommended to prevent false sharing.
 template <uint64_t N>
 struct LargeBitset
 {
 private:
    constexpr static uint64_t cache_line_size = 64;
 public:
    using UnderlyingType = uint64_t;
    constexpr static uint64_t num_bits = N;
    constexpr static uint64_t bits_per_bucket = 8 * sizeof(uint64_t);
    constexpr static uint64_t num_buckets = (num_bits + bits_per_bucket - 1) / bits_per_bucket;
    constexpr static uint64_t best_concurrent_access_stride = 8 * cache_line_size;
    LargeBitset()
    {
        std::fill(std::begin(bits), std::end(bits), 0);
    }
    void set(uint64_t idx)
    {
        const uint64_t bucket = idx / bits_per_bucket;
        const uint64_t bit = uint64_t(1) << (idx % bits_per_bucket);
        bits[bucket] |= bit;
    }
    bool test(uint64_t idx) const
    {
        const uint64_t bucket = idx / bits_per_bucket;
        const uint64_t bit = uint64_t(1) << (idx % bits_per_bucket);
        return bits[bucket] & bit;
    }
    uint64_t count() const
    {
        uint64_t c = 0;
        uint64_t i = 0;
        for (; i < num_buckets - 3; i += 4)
        {
            uint64_t c0 = popcount(bits[i+0]);
            uint64_t c1 = popcount(bits[i+1]);
            uint64_t c2 = popcount(bits[i+2]);
            uint64_t c3 = popcount(bits[i+3]);
            c0 += c1;
            c2 += c3;
            c += c0 + c2;
        }
        for (; i < num_buckets; ++i)
        {
            c += popcount(bits[i]);
        }
        return c;
    }
 private:
    alignas(cache_line_size) UnderlyingType bits[num_buckets];
 };
 /// Under Windows it is not possible for a process to run on more than one
 /// logical processor group. This usually means to be limited to use max 64
 /// cores. To overcome this, some special platform specific API should be
@@ -134,4 +591,118 @@ namespace WinProcGroup {
  void bindThisThread(size_t idx);
 }
 // Returns a string that represents the current time. (Used for log output when learning evaluation function)
 std::string now_string();
 void sleep(int ms);
 namespace Algo {
    // Fisher-Yates
    template <typename Rng, typename T>
    void shuffle(std::vector<T>& buf, Rng&& prng)
    {
        const auto size = buf.size();
        for (uint64_t i = 0; i < size; ++i)
            std::swap(buf[i], buf[prng.rand(size - i) + i]);
    }
    // split the string
    inline std::vector<std::string> split(const std::string& input, char delimiter) {
        std::istringstream stream(input);
        std::string field;
        std::vector<std::string> fields;
        while (std::getline(stream, field, delimiter)) {
            fields.push_back(field);
        }
        return fields;
    }
 }
 // --------------------
 //       Path
 // --------------------
 // Something like Path class in C#. File name manipulation.
 // Match with the C# method name.
 struct Path
 {
 	// Combine the path name and file name and return it.
 	// If the folder name is not an empty string, append it if there is no'/' or'\\' at the end.
 	static std::string combine(const std::string& folder, const std::string& filename)
 	{
 		if (folder.length() >= 1 && *folder.rbegin() != '/' && *folder.rbegin() != '\\')
 			return folder + "/" + filename;
 		return folder + filename;
 	}
 	// Get the file name part (excluding the folder name) from the full path expression.
 	static std::string get_file_name(const std::string& path)
 	{
 		// I don't know which "\" or "/" is used.
 		auto path_index1 = path.find_last_of("\\") + 1;
 		auto path_index2 = path.find_last_of("/") + 1;
 		auto path_index = std::max(path_index1, path_index2);
 		return path.substr(path_index);
 	}
 };
 // It is ignored when new even though alignas is specified & because it is ignored when the STL container allocates memory,
 // A custom allocator used for that.
 template <typename T>
 class AlignedAllocator {
 public:
  using value_type = T;
  AlignedAllocator() {}
  AlignedAllocator(const AlignedAllocator&) {}
  AlignedAllocator(AlignedAllocator&&) {}
  template <typename U> AlignedAllocator(const AlignedAllocator<U>&) {}
  T* allocate(std::size_t n) { return (T*)std_aligned_alloc(alignof(T), n * sizeof(T)); }
  void deallocate(T* p, std::size_t ) { std_aligned_free(p); }
 };
 template <typename T>
 class CacheLineAlignedAllocator {
 public:
    using value_type = T;
    constexpr static uint64_t cache_line_size = 64;
    CacheLineAlignedAllocator() {}
    CacheLineAlignedAllocator(const CacheLineAlignedAllocator&) {}
    CacheLineAlignedAllocator(CacheLineAlignedAllocator&&) {}
    template <typename U> CacheLineAlignedAllocator(const CacheLineAlignedAllocator<U>&) {}
    T* allocate(std::size_t n) { return (T*)std_aligned_alloc(cache_line_size, n * sizeof(T)); }
    void deallocate(T* p, std::size_t) { std_aligned_free(p); }
 };
 // --------------------
 //  Dependency Wrapper
 // --------------------
 namespace Dependency
 {
  // In the Linux environment, if you getline() the text file is'\r\n'
  // Since'\r' remains at the end, write a wrapper to remove this'\r'.
  // So when calling getline() on fstream,
  // just write getline() instead of std::getline() and use this function.
  extern bool getline(std::ifstream& fs, std::string& s);
 }
 namespace CommandLine {
  void init(int argc, char* argv[]);
  extern std::string binaryDirectory;  // path of the executable directory
  extern std::string workingDirectory; // path of the working directory
 }
 } // namespace Stockfish
 #endif // #ifndef MISC_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -21,24 +21,21 @@
 #include "movegen.h"
 #include "position.h"
 namespace Stockfish {
 namespace {
  template<GenType Type, Direction D>
-  ExtMove* make_promotions(ExtMove* moveList, Square to, Square ksq) {
+  ExtMove* make_promotions(ExtMove* moveList, Square to) {
    if (Type == CAPTURES || Type == EVASIONS || Type == NON_EVASIONS)
    {
        *moveList++ = make<PROMOTION>(to - D, to, QUEEN);
        if (attacks_bb<KNIGHT>(to) & ksq)
            *moveList++ = make<PROMOTION>(to - D, to, KNIGHT);
    }
    if (Type == QUIETS || Type == EVASIONS || Type == NON_EVASIONS)
    {
        *moveList++ = make<PROMOTION>(to - D, to, ROOK);
        *moveList++ = make<PROMOTION>(to - D, to, BISHOP);
-        if (!(attacks_bb<KNIGHT>(to) & ksq))
+        *moveList++ = make<PROMOTION>(to - D, to, KNIGHT);
            *moveList++ = make<PROMOTION>(to - D, to, KNIGHT);
    }
    return moveList;
@@ -55,20 +52,16 @@ namespace {
    constexpr Direction UpRight  = (Us == WHITE ? NORTH_EAST : SOUTH_WEST);
    constexpr Direction UpLeft   = (Us == WHITE ? NORTH_WEST : SOUTH_EAST);
-    const Square ksq = pos.square<KING>(Them);
+    const Bitboard emptySquares = ~pos.pieces();
-    Bitboard emptySquares;
+    const Bitboard enemies      =  Type == EVASIONS ? pos.checkers()
                                                    : pos.pieces(Them);
    Bitboard pawnsOn7    = pos.pieces(Us, PAWN) &  TRank7BB;
    Bitboard pawnsNotOn7 = pos.pieces(Us, PAWN) & ~TRank7BB;
    Bitboard enemies = (Type == EVASIONS ? pos.pieces(Them) & target:
                        Type == CAPTURES ? target : pos.pieces(Them));
    // Single and double pawn pushes, no promotions
    if (Type != CAPTURES)
    {
        emptySquares = (Type == QUIETS || Type == QUIET_CHECKS ? target : ~pos.pieces());
        Bitboard b1 = shift<Up>(pawnsNotOn7)   & emptySquares;
        Bitboard b2 = shift<Up>(b1 & TRank3BB) & emptySquares;
@@ -80,33 +73,24 @@ namespace {
        if (Type == QUIET_CHECKS)
        {
-            b1 &= pawn_attacks_bb(Them, ksq);
+            // To make a quiet check, you either make a direct check by pushing a pawn
-            b2 &= pawn_attacks_bb(Them, ksq);
+            // or push a blocker pawn that is not on the same file as the enemy king.
-
+            // Discovered check promotion has been already generated amongst the captures.
-            // Add pawn pushes which give discovered check. This is possible only
+            Square ksq = pos.square<KING>(Them);
-            // if the pawn is not on the same file as the enemy king, because we
+            Bitboard dcCandidatePawns = pos.blockers_for_king(Them) & ~file_bb(ksq);
-            // don't generate captures. Note that a possible discovery check
+            b1 &= pawn_attacks_bb(Them, ksq) | shift<   Up>(dcCandidatePawns);
-            // promotion has been already generated amongst the captures.
+            b2 &= pawn_attacks_bb(Them, ksq) | shift<Up+Up>(dcCandidatePawns);
            Bitboard dcCandidateQuiets = pos.blockers_for_king(Them) & pawnsNotOn7;
            if (dcCandidateQuiets)
            {
                Bitboard dc1 = shift<Up>(dcCandidateQuiets) & emptySquares & ~file_bb(ksq);
                Bitboard dc2 = shift<Up>(dc1 & TRank3BB) & emptySquares;
                b1 |= dc1;
                b2 |= dc2;
            }
        }
        while (b1)
        {
-            Square to = pop_lsb(&b1);
+            Square to = pop_lsb(b1);
            *moveList++ = make_move(to - Up, to);
        }
        while (b2)
        {
-            Square to = pop_lsb(&b2);
+            Square to = pop_lsb(b2);
            *moveList++ = make_move(to - Up - Up, to);
        }
    }
@@ -114,27 +98,24 @@ namespace {
    // Promotions and underpromotions
    if (pawnsOn7)
    {
        if (Type == CAPTURES)
            emptySquares = ~pos.pieces();
        if (Type == EVASIONS)
            emptySquares &= target;
        Bitboard b1 = shift<UpRight>(pawnsOn7) & enemies;
        Bitboard b2 = shift<UpLeft >(pawnsOn7) & enemies;
        Bitboard b3 = shift<Up     >(pawnsOn7) & emptySquares;
        if (Type == EVASIONS)
            b3 &= target;
        while (b1)
-            moveList = make_promotions<Type, UpRight>(moveList, pop_lsb(&b1), ksq);
+            moveList = make_promotions<Type, UpRight>(moveList, pop_lsb(b1));
        while (b2)
-            moveList = make_promotions<Type, UpLeft >(moveList, pop_lsb(&b2), ksq);
+            moveList = make_promotions<Type, UpLeft >(moveList, pop_lsb(b2));
        while (b3)
-            moveList = make_promotions<Type, Up     >(moveList, pop_lsb(&b3), ksq);
+            moveList = make_promotions<Type, Up     >(moveList, pop_lsb(b3));
    }
-    // Standard and en-passant captures
+    // Standard and en passant captures
    if (Type == CAPTURES || Type == EVASIONS || Type == NON_EVASIONS)
    {
        Bitboard b1 = shift<UpRight>(pawnsNotOn7) & enemies;
@@ -142,13 +123,13 @@ namespace {
        while (b1)
        {
-            Square to = pop_lsb(&b1);
+            Square to = pop_lsb(b1);
            *moveList++ = make_move(to - UpRight, to);
        }
        while (b2)
        {
-            Square to = pop_lsb(&b2);
+            Square to = pop_lsb(b2);
            *moveList++ = make_move(to - UpLeft, to);
        }
@@ -156,10 +137,8 @@ namespace {
        {
            assert(rank_of(pos.ep_square()) == relative_rank(Us, RANK_6));
-            // An en passant capture can be an evasion only if the checking piece
+            // An en passant capture cannot resolve a discovered check
-            // is the double pushed pawn and so is in the target. Otherwise this
+            if (Type == EVASIONS && (target & (pos.ep_square() + Up)))
            // is a discovery check and we are forced to do otherwise.
            if (Type == EVASIONS && !(target & (pos.ep_square() - Up)))
                return moveList;
            b1 = pawnsNotOn7 & pawn_attacks_bb(Them, pos.ep_square());
@@ -167,7 +146,7 @@ namespace {
            assert(b1);
            while (b1)
-                *moveList++ = make<ENPASSANT>(pop_lsb(&b1), pos.ep_square());
+                *moveList++ = make<EN_PASSANT>(pop_lsb(b1), pos.ep_square());
        }
    }
@@ -180,27 +159,19 @@ namespace {
    static_assert(Pt != KING && Pt != PAWN, "Unsupported piece type in generate_moves()");
-    const Square* pl = pos.squares<Pt>(Us);
+    Bitboard bb = pos.pieces(Us, Pt);
-    for (Square from = *pl; from != SQ_NONE; from = *++pl)
+    while (bb)
    {
-        if (Checks)
+        Square from = pop_lsb(bb);
        {
            if (    (Pt == BISHOP || Pt == ROOK || Pt == QUEEN)
                && !(attacks_bb<Pt>(from) & target & pos.check_squares(Pt)))
                continue;
            if (pos.blockers_for_king(~Us) & from)
                continue;
        }
        Bitboard b = attacks_bb<Pt>(from, pos.pieces()) & target;
-        if (Checks)
+        // To check, you either move freely a blocker or make a direct check.
        if (Checks && (Pt == QUEEN || !(pos.blockers_for_king(~Us) & from)))
            b &= pos.check_squares(Pt);
        while (b)
-            *moveList++ = make_move(from, pop_lsb(&b));
+            *moveList++ = make_move(from, pop_lsb(b));
    }
    return moveList;
@@ -209,46 +180,39 @@ namespace {
  template<Color Us, GenType Type>
  ExtMove* generate_all(const Position& pos, ExtMove* moveList) {
-    constexpr bool Checks = Type == QUIET_CHECKS; // Reduce template instantations
+
    static_assert(Type != LEGAL, "Unsupported type in generate_all()");
    constexpr bool Checks = Type == QUIET_CHECKS; // Reduce template instantiations
    const Square ksq = pos.square<KING>(Us);
    Bitboard target;
-    switch (Type)
+    // Skip generating non-king moves when in double check
    if (Type != EVASIONS || !more_than_one(pos.checkers()))
    {
-        case CAPTURES:
+        target = Type == EVASIONS     ?  between_bb(ksq, lsb(pos.checkers()))
-            target =  pos.pieces(~Us);
+               : Type == NON_EVASIONS ? ~pos.pieces( Us)
-            break;
+               : Type == CAPTURES     ?  pos.pieces(~Us)
-        case QUIETS:
+                                      : ~pos.pieces(   ); // QUIETS || QUIET_CHECKS
-        case QUIET_CHECKS:
+
-            target = ~pos.pieces();
+        moveList = generate_pawn_moves<Us, Type>(pos, moveList, target);
-            break;
+        moveList = generate_moves<Us, KNIGHT, Checks>(pos, moveList, target);
-        case EVASIONS:
+        moveList = generate_moves<Us, BISHOP, Checks>(pos, moveList, target);
-        {
+        moveList = generate_moves<Us,   ROOK, Checks>(pos, moveList, target);
-            Square checksq = lsb(pos.checkers());
+        moveList = generate_moves<Us,  QUEEN, Checks>(pos, moveList, target);
            target = between_bb(pos.square<KING>(Us), checksq) | checksq;
            break;
        }
        case NON_EVASIONS:
            target = ~pos.pieces(Us);
            break;
        default:
            static_assert(true, "Unsupported type in generate_all()");
    }
-    moveList = generate_pawn_moves<Us, Type>(pos, moveList, target);
+    if (!Checks || pos.blockers_for_king(~Us) & ksq)
    moveList = generate_moves<Us, KNIGHT, Checks>(pos, moveList, target);
    moveList = generate_moves<Us, BISHOP, Checks>(pos, moveList, target);
    moveList = generate_moves<Us,   ROOK, Checks>(pos, moveList, target);
    moveList = generate_moves<Us,  QUEEN, Checks>(pos, moveList, target);
    if (Type != QUIET_CHECKS && Type != EVASIONS)
    {
-        Square ksq = pos.square<KING>(Us);
+        Bitboard b = attacks_bb<KING>(ksq) & (Type == EVASIONS ? ~pos.pieces(Us) : target);
-        Bitboard b = attacks_bb<KING>(ksq) & target;
+        if (Checks)
-        while (b)
+            b &= ~attacks_bb<QUEEN>(pos.square<KING>(~Us));
            *moveList++ = make_move(ksq, pop_lsb(&b));
-        if ((Type != CAPTURES) && pos.can_castle(Us & ANY_CASTLING))
+        while (b)
-            for(CastlingRights cr : { Us & KING_SIDE, Us & QUEEN_SIDE } )
+            *moveList++ = make_move(ksq, pop_lsb(b));
        if ((Type == QUIETS || Type == NON_EVASIONS) && pos.can_castle(Us & ANY_CASTLING))
            for (CastlingRights cr : { Us & KING_SIDE, Us & QUEEN_SIDE } )
                if (!pos.castling_impeded(cr) && pos.can_castle(cr))
                    *moveList++ = make<CASTLING>(ksq, pos.castling_rook_square(cr));
    }
@@ -259,8 +223,10 @@ namespace {
 } // namespace
-/// <CAPTURES>     Generates all pseudo-legal captures plus queen and checking knight promotions
+/// <CAPTURES>     Generates all pseudo-legal captures plus queen promotions
-/// <QUIETS>       Generates all pseudo-legal non-captures and underpromotions(except checking knight)
+/// <QUIETS>       Generates all pseudo-legal non-captures and underpromotions
 /// <EVASIONS>     Generates all pseudo-legal check evasions when the side to move is in check
 /// <QUIET_CHECKS> Generates all pseudo-legal non-captures giving check, except castling and promotions
 /// <NON_EVASIONS> Generates all pseudo-legal captures and non-captures
 ///
 /// Returns a pointer to the end of the move list.
@@ -268,8 +234,8 @@ namespace {
 template<GenType Type>
 ExtMove* generate(const Position& pos, ExtMove* moveList) {
-  static_assert(Type == CAPTURES || Type == QUIETS || Type == NON_EVASIONS, "Unsupported type in generate()");
+  static_assert(Type != LEGAL, "Unsupported type in generate()");
-  assert(!pos.checkers());
+  assert((Type == EVASIONS) == (bool)pos.checkers());
  Color us = pos.side_to_move();
@@ -280,70 +246,11 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) {
 // Explicit template instantiations
 template ExtMove* generate<CAPTURES>(const Position&, ExtMove*);
 template ExtMove* generate<QUIETS>(const Position&, ExtMove*);
 template ExtMove* generate<EVASIONS>(const Position&, ExtMove*);
 template ExtMove* generate<QUIET_CHECKS>(const Position&, ExtMove*);
 template ExtMove* generate<NON_EVASIONS>(const Position&, ExtMove*);
 /// generate<QUIET_CHECKS> generates all pseudo-legal non-captures.
 /// Returns a pointer to the end of the move list.
 template<>
 ExtMove* generate<QUIET_CHECKS>(const Position& pos, ExtMove* moveList) {
  assert(!pos.checkers());
  Color us = pos.side_to_move();
  Bitboard dc = pos.blockers_for_king(~us) & pos.pieces(us) & ~pos.pieces(PAWN);
  while (dc)
  {
     Square from = pop_lsb(&dc);
     PieceType pt = type_of(pos.piece_on(from));
     Bitboard b = attacks_bb(pt, from, pos.pieces()) & ~pos.pieces();
     if (pt == KING)
         b &= ~attacks_bb<QUEEN>(pos.square<KING>(~us));
     while (b)
         *moveList++ = make_move(from, pop_lsb(&b));
  }
  return us == WHITE ? generate_all<WHITE, QUIET_CHECKS>(pos, moveList)
                     : generate_all<BLACK, QUIET_CHECKS>(pos, moveList);
 }
 /// generate<EVASIONS> generates all pseudo-legal check evasions when the side
 /// to move is in check. Returns a pointer to the end of the move list.
 template<>
 ExtMove* generate<EVASIONS>(const Position& pos, ExtMove* moveList) {
  assert(pos.checkers());
  Color us = pos.side_to_move();
  Square ksq = pos.square<KING>(us);
  Bitboard sliderAttacks = 0;
  Bitboard sliders = pos.checkers() & ~pos.pieces(KNIGHT, PAWN);
  // Find all the squares attacked by slider checkers. We will remove them from
  // the king evasions in order to skip known illegal moves, which avoids any
  // useless legality checks later on.
  while (sliders)
      sliderAttacks |= line_bb(ksq, pop_lsb(&sliders)) & ~pos.checkers();
  // Generate evasions for king, capture and non capture moves
  Bitboard b = attacks_bb<KING>(ksq) & ~pos.pieces(us) & ~sliderAttacks;
  while (b)
      *moveList++ = make_move(ksq, pop_lsb(&b));
  if (more_than_one(pos.checkers()))
      return moveList; // Double check, only a king move can save the day
  // Generate blocking evasions or captures of the checking piece
  return us == WHITE ? generate_all<WHITE, EVASIONS>(pos, moveList)
                     : generate_all<BLACK, EVASIONS>(pos, moveList);
 }
 /// generate<LEGAL> generates all the legal moves in the given position
 template<>
@@ -357,7 +264,7 @@ ExtMove* generate<LEGAL>(const Position& pos, ExtMove* moveList) {
  moveList = pos.checkers() ? generate<EVASIONS    >(pos, moveList)
                            : generate<NON_EVASIONS>(pos, moveList);
  while (cur != moveList)
-      if (   (pinned || from_sq(*cur) == ksq || type_of(*cur) == ENPASSANT)
+      if (  ((pinned && pinned & from_sq(*cur)) || from_sq(*cur) == ksq || type_of(*cur) == EN_PASSANT)
          && !pos.legal(*cur))
          *cur = (--moveList)->move;
      else
@@ -365,3 +272,5 @@ ExtMove* generate<LEGAL>(const Position& pos, ExtMove* moveList) {
  return moveList;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,6 +23,8 @@
 #include "types.h"
 namespace Stockfish {
 class Position;
 enum GenType {
@@ -66,8 +68,13 @@ struct MoveList {
    return std::find(begin(), end(), move) != end();
  }
  // returns the i th element
  const ExtMove at(size_t i) const { assert(0 <= i && i < size()); return begin()[i]; }
 private:
  ExtMove moveList[MAX_MOVES], *last;
 };
 } // namespace Stockfish
 #endif // #ifndef MOVEGEN_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -18,8 +18,11 @@
 #include <cassert>
 #include "bitboard.h"
 #include "movepick.h"
 namespace Stockfish {
 namespace {
  enum Stages {
@@ -54,11 +57,14 @@ namespace {
 /// ordering is at the current node.
 /// MovePicker constructor for the main search
-MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh, const LowPlyHistory* lp,
+MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh,
-                       const CapturePieceToHistory* cph, const PieceToHistory** ch, Move cm, const Move* killers, int pl)
+                                                             const CapturePieceToHistory* cph,
-           : pos(p), mainHistory(mh), lowPlyHistory(lp), captureHistory(cph), continuationHistory(ch),
+                                                             const PieceToHistory** ch,
-             ttMove(ttm), refutations{{killers[0], 0}, {killers[1], 0}, {cm, 0}}, depth(d), ply(pl) {
+                                                             Move cm,
-
+                                                             const Move* killers)
           : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch),
             ttMove(ttm), refutations{{killers[0], 0}, {killers[1], 0}, {cm, 0}}, depth(d)
 {
  assert(d > 0);
  stage = (pos.checkers() ? EVASION_TT : MAIN_TT) +
@@ -67,21 +73,24 @@ MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHist
 /// MovePicker constructor for quiescence search
 MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh,
-                       const CapturePieceToHistory* cph, const PieceToHistory** ch, Square rs)
+                                                             const CapturePieceToHistory* cph,
-           : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch), ttMove(ttm), recaptureSquare(rs), depth(d) {
+                                                             const PieceToHistory** ch,
-
+                                                             Square rs)
           : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch), ttMove(ttm), recaptureSquare(rs), depth(d)
 {
  assert(d <= 0);
  stage = (pos.checkers() ? EVASION_TT : QSEARCH_TT) +
-           !(ttm && (depth > DEPTH_QS_RECAPTURES || to_sq(ttm) == recaptureSquare)
+          !(   ttm
-                 && pos.pseudo_legal(ttm));
+            && (pos.checkers() || depth > DEPTH_QS_RECAPTURES || to_sq(ttm) == recaptureSquare)
            && pos.pseudo_legal(ttm));
 }
 /// MovePicker constructor for ProbCut: we generate captures with SEE greater
 /// than or equal to the given threshold.
-MovePicker::MovePicker(const Position& p, Move ttm, Value th, const CapturePieceToHistory* cph)
+MovePicker::MovePicker(const Position& p, Move ttm, Value th, Depth d, const CapturePieceToHistory* cph)
-           : pos(p), captureHistory(cph), ttMove(ttm), threshold(th) {
+           : pos(p), captureHistory(cph), ttMove(ttm), threshold(th), depth(d)
-
+{
  assert(!pos.checkers());
  stage = PROBCUT_TT + !(ttm && pos.capture(ttm)
@@ -97,18 +106,48 @@ void MovePicker::score() {
  static_assert(Type == CAPTURES || Type == QUIETS || Type == EVASIONS, "Wrong type");
-  for (auto& m : *this)
+  Bitboard threatened, threatenedByPawn, threatenedByMinor, threatenedByRook;
-      if (Type == CAPTURES)
+  if constexpr (Type == QUIETS)
-          m.value =  int(PieceValue[MG][pos.piece_on(to_sq(m))]) * 6
+  {
-                   + (*captureHistory)[pos.moved_piece(m)][to_sq(m)][type_of(pos.piece_on(to_sq(m)))];
+      Color us = pos.side_to_move();
      // squares threatened by pawns
      threatenedByPawn  = pos.attacks_by<PAWN>(~us);
      // squares threatened by minors or pawns
      threatenedByMinor = pos.attacks_by<KNIGHT>(~us) | pos.attacks_by<BISHOP>(~us) | threatenedByPawn;
      // squares threatened by rooks, minors or pawns
      threatenedByRook  = pos.attacks_by<ROOK>(~us) | threatenedByMinor;
-      else if (Type == QUIETS)
+      // pieces threatened by pieces of lesser material value
      threatened =  (pos.pieces(us, QUEEN) & threatenedByRook)
                  | (pos.pieces(us, ROOK)  & threatenedByMinor)
                  | (pos.pieces(us, KNIGHT, BISHOP) & threatenedByPawn);
  }
  else
  {
      // Silence unused variable warnings
      (void) threatened;
      (void) threatenedByPawn;
      (void) threatenedByMinor;
      (void) threatenedByRook;
  }
  for (auto& m : *this)
      if constexpr (Type == CAPTURES)
          m.value =  6 * int(PieceValue[MG][pos.piece_on(to_sq(m))])
                   +     (*captureHistory)[pos.moved_piece(m)][to_sq(m)][type_of(pos.piece_on(to_sq(m)))];
      else if constexpr (Type == QUIETS)
          m.value =      (*mainHistory)[pos.side_to_move()][from_to(m)]
                   + 2 * (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)]
-                   + 2 * (*continuationHistory[1])[pos.moved_piece(m)][to_sq(m)]
+                   +     (*continuationHistory[1])[pos.moved_piece(m)][to_sq(m)]
-                   + 2 * (*continuationHistory[3])[pos.moved_piece(m)][to_sq(m)]
+                   +     (*continuationHistory[3])[pos.moved_piece(m)][to_sq(m)]
                   +     (*continuationHistory[5])[pos.moved_piece(m)][to_sq(m)]
-                   + (ply < MAX_LPH ? std::min(4, depth / 3) * (*lowPlyHistory)[ply][from_to(m)] : 0);
+                   +     (threatened & from_sq(m) ?
                           (type_of(pos.moved_piece(m)) == QUEEN && !(to_sq(m) & threatenedByRook)  ? 50000
                          : type_of(pos.moved_piece(m)) == ROOK  && !(to_sq(m) & threatenedByMinor) ? 25000
                          :                                         !(to_sq(m) & threatenedByPawn)  ? 15000
                          :                                                                           0)
                          :                                                                           0);
      else // Type == EVASIONS
      {
@@ -116,8 +155,8 @@ void MovePicker::score() {
              m.value =  PieceValue[MG][pos.piece_on(to_sq(m))]
                       - Value(type_of(pos.moved_piece(m)));
          else
-              m.value =  (*mainHistory)[pos.side_to_move()][from_to(m)]
+              m.value =      (*mainHistory)[pos.side_to_move()][from_to(m)]
-                       + (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)]
+                       + 2 * (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)]
                       - (1 << 28);
      }
 }
@@ -141,7 +180,7 @@ Move MovePicker::select(Pred filter) {
 }
 /// MovePicker::next_move() is the most important method of the MovePicker class. It
-/// returns a new pseudo legal move every time it is called until there are no more
+/// returns a new pseudo-legal move every time it is called until there are no more
 /// moves left, picking the move with the highest score from a list of generated moves.
 Move MovePicker::next_move(bool skipQuiets) {
@@ -162,11 +201,12 @@ top:
      endMoves = generate<CAPTURES>(pos, cur);
      score<CAPTURES>();
      partial_insertion_sort(cur, endMoves, -3000 * depth);
      ++stage;
      goto top;
  case GOOD_CAPTURE:
-      if (select<Best>([&](){
+      if (select<Next>([&](){
                       return pos.see_ge(*cur, Value(-69 * cur->value / 1024)) ?
                              // Move losing capture to endBadCaptures to be tried later
                              true : (*endBadCaptures++ = *cur, false); }))
@@ -182,7 +222,7 @@ top:
          --endMoves;
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case REFUTATION:
      if (select<Next>([&](){ return    *cur != MOVE_NONE
@@ -190,7 +230,7 @@ top:
                                    &&  pos.pseudo_legal(*cur); }))
          return *(cur - 1);
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case QUIET_INIT:
      if (!skipQuiets)
@@ -203,7 +243,7 @@ top:
      }
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case QUIET:
      if (   !skipQuiets
@@ -217,7 +257,7 @@ top:
      endMoves = endBadCaptures;
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case BAD_CAPTURE:
      return select<Next>([](){ return true; });
@@ -228,16 +268,16 @@ top:
      score<EVASIONS>();
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case EVASION:
      return select<Best>([](){ return true; });
  case PROBCUT:
-      return select<Best>([&](){ return pos.see_ge(*cur, threshold); });
+      return select<Next>([&](){ return pos.see_ge(*cur, threshold); });
  case QCAPTURE:
-      if (select<Best>([&](){ return   depth > DEPTH_QS_RECAPTURES
+      if (select<Next>([&](){ return   depth > DEPTH_QS_RECAPTURES
                                    || to_sq(*cur) == recaptureSquare; }))
          return *(cur - 1);
@@ -246,14 +286,14 @@ top:
          return MOVE_NONE;
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case QCHECK_INIT:
      cur = moves;
      endMoves = generate<QUIET_CHECKS>(pos, cur);
      ++stage;
-      /* fallthrough */
+      [[fallthrough]];
  case QCHECK:
      return select<Next>([](){ return true; });
@@ -262,3 +302,5 @@ top:
  assert(false);
  return MOVE_NONE; // Silence warning
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -27,6 +27,8 @@
 #include "position.h"
 #include "types.h"
 namespace Stockfish {
 /// StatsEntry stores the stat table value. It is usually a number but could
 /// be a move or even a nested history. We use a class instead of naked value
 /// to directly call history update operator<<() on the entry so to use stats
@@ -84,13 +86,7 @@ enum StatsType { NoCaptures, Captures };
 /// unsuccessful during the current search, and is used for reduction and move
 /// ordering decisions. It uses 2 tables (one for each color) indexed by
 /// the move's from and to squares, see www.chessprogramming.org/Butterfly_Boards
-typedef Stats<int16_t, 10692, COLOR_NB, int(SQUARE_NB) * int(SQUARE_NB)> ButterflyHistory;
+typedef Stats<int16_t, 14365, COLOR_NB, int(SQUARE_NB) * int(SQUARE_NB)> ButterflyHistory;
 /// At higher depths LowPlyHistory records successful quiet moves near the root and quiet
 /// moves which are/were in the PV (ttPv)
 /// It is cleared with each new search and filled during iterative deepening
 constexpr int MAX_LPH = 4;
 typedef Stats<int16_t, 10692, MAX_LPH, int(SQUARE_NB) * int(SQUARE_NB)> LowPlyHistory;
 /// CounterMoveHistory stores counter moves indexed by [piece][to] of the previous
 /// move, see www.chessprogramming.org/Countermove_Heuristic
@@ -108,12 +104,12 @@ typedef Stats<int16_t, 29952, PIECE_NB, SQUARE_NB> PieceToHistory;
 typedef Stats<PieceToHistory, NOT_USED, PIECE_NB, SQUARE_NB> ContinuationHistory;
-/// MovePicker class is used to pick one pseudo legal move at a time from the
+/// MovePicker class is used to pick one pseudo-legal move at a time from the
 /// current position. The most important method is next_move(), which returns a
-/// new pseudo legal move each time it is called, until there are no moves left,
+/// new pseudo-legal move each time it is called, until there are no moves left,
-/// when MOVE_NONE is returned. In order to improve the efficiency of the alpha
+/// when MOVE_NONE is returned. In order to improve the efficiency of the
-/// beta algorithm, MovePicker attempts to return the moves which are most likely
+/// alpha-beta algorithm, MovePicker attempts to return the moves which are most
-/// to get a cut-off first.
+/// likely to get a cut-off first.
 class MovePicker {
  enum PickType { Next, Best };
@@ -121,18 +117,16 @@ class MovePicker {
 public:
  MovePicker(const MovePicker&) = delete;
  MovePicker& operator=(const MovePicker&) = delete;
-  MovePicker(const Position&, Move, Value, const CapturePieceToHistory*);
+  MovePicker(const Position&, Move, Depth, const ButterflyHistory*,
                                           const CapturePieceToHistory*,
                                           const PieceToHistory**,
                                           Move,
                                           const Move*);
  MovePicker(const Position&, Move, Depth, const ButterflyHistory*,
                                           const CapturePieceToHistory*,
                                           const PieceToHistory**,
                                           Square);
-  MovePicker(const Position&, Move, Depth, const ButterflyHistory*,
+  MovePicker(const Position&, Move, Value, Depth, const CapturePieceToHistory*);
                                           const LowPlyHistory*,
                                           const CapturePieceToHistory*,
                                           const PieceToHistory**,
                                           Move,
                                           const Move*,
                                           int);
  Move next_move(bool skipQuiets = false);
 private:
@@ -143,7 +137,6 @@ private:
  const Position& pos;
  const ButterflyHistory* mainHistory;
  const LowPlyHistory* lowPlyHistory;
  const CapturePieceToHistory* captureHistory;
  const PieceToHistory** continuationHistory;
  Move ttMove;
@@ -152,8 +145,9 @@ private:
  Square recaptureSquare;
  Value threshold;
  Depth depth;
  int ply;
  ExtMove moves[MAX_MOVES];
 };
 } // namespace Stockfish
 #endif // #ifndef MOVEPICK_H_INCLUDED
@@ -1,54 +0,0 @@
 /*
  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
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -18,161 +18,386 @@
 // Code for calculating NNUE evaluation function
 #include <fstream>
 #include <iostream>
 #include <set>
 #include <sstream>
 #include <iomanip>
 #include <fstream>
 #include "../evaluate.h"
 #include "../position.h"
 #include "../misc.h"
 #include "../uci.h"
 #include "../types.h"
 #include "evaluate_nnue.h"
-ExtPieceSquare kpp_board_index[PIECE_NB] = {
+namespace Stockfish::Eval::NNUE {
 // 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;
+  LargePagePtr<FeatureTransformer> featureTransformer;
  // Evaluation function
-  AlignedPtr<Network> network;
+  AlignedPtr<Network> network[LayerStacks];
  // Evaluation function file name
  std::string fileName;
  std::string netDescription;
  namespace Detail {
  // Initialize the evaluation function parameters
  template <typename T>
-  void Initialize(AlignedPtr<T>& pointer) {
+  void initialize(AlignedPtr<T>& pointer) {
    pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
    std::memset(pointer.get(), 0, sizeof(T));
  }
  template <typename T>
  void initialize(LargePagePtr<T>& pointer) {
    static_assert(alignof(T) <= 4096, "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
    pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(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) {
+  bool read_parameters(std::istream& stream, T& reference) {
    std::uint32_t header;
-    stream.read(reinterpret_cast<char*>(&header), sizeof(header));
+    header = read_little_endian<std::uint32_t>(stream);
-    if (!stream || header != T::GetHashValue()) return false;
+    if (!stream || header != T::get_hash_value()) return false;
-    return pointer->ReadParameters(stream);
+    return reference.read_parameters(stream);
  }
  // Write evaluation function parameters
  template <typename T>
  bool write_parameters(std::ostream& stream, const T& reference) {
    write_little_endian<std::uint32_t>(stream, T::get_hash_value());
    return reference.write_parameters(stream);
  }
  }  // namespace Detail
  // Initialize the evaluation function parameters
-  void Initialize() {
+  void initialize() {
-    Detail::Initialize(feature_transformer);
+    Detail::initialize(featureTransformer);
-    Detail::Initialize(network);
+    for (std::size_t i = 0; i < LayerStacks; ++i)
      Detail::initialize(network[i]);
  }
  // Read network header
-  bool ReadHeader(std::istream& stream,
+  bool read_header(std::istream& stream, std::uint32_t* hashValue, std::string* desc)
-    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));
+    version     = read_little_endian<std::uint32_t>(stream);
-    stream.read(reinterpret_cast<char*>(&size), sizeof(size));
+    *hashValue  = read_little_endian<std::uint32_t>(stream);
-    if (!stream || version != kVersion) return false;
+    size        = read_little_endian<std::uint32_t>(stream);
-    architecture->resize(size);
+    if (!stream || version != Version) return false;
-    stream.read(&(*architecture)[0], size);
+    desc->resize(size);
    stream.read(&(*desc)[0], size);
    return !stream.fail();
  }
  // Write network header
  bool write_header(std::ostream& stream, std::uint32_t hashValue, const std::string& desc)
  {
    write_little_endian<std::uint32_t>(stream, Version);
    write_little_endian<std::uint32_t>(stream, hashValue);
    write_little_endian<std::uint32_t>(stream, (std::uint32_t)desc.size());
    stream.write(&desc[0], desc.size());
    return !stream.fail();
  }
  // Read network parameters
-  bool ReadParameters(std::istream& stream) {
+  bool read_parameters(std::istream& stream) {
-    std::uint32_t hash_value;
+    std::uint32_t hashValue;
-    std::string architecture;
+    if (!read_header(stream, &hashValue, &netDescription)) return false;
-    if (!ReadHeader(stream, &hash_value, &architecture)) return false;
+    if (hashValue != HashValue) return false;
-    if (hash_value != kHashValue) return false;
+    if (!Detail::read_parameters(stream, *featureTransformer)) return false;
-    if (!Detail::ReadParameters(stream, feature_transformer)) return false;
+    for (std::size_t i = 0; i < LayerStacks; ++i)
-    if (!Detail::ReadParameters(stream, network)) return false;
+      if (!Detail::read_parameters(stream, *(network[i]))) return false;
    return stream && stream.peek() == std::ios::traits_type::eof();
  }
-  // Proceed with the difference calculation if possible
+  // Write network parameters
-  static void UpdateAccumulatorIfPossible(const Position& pos) {
+  bool write_parameters(std::ostream& stream) {
-    feature_transformer->UpdateAccumulatorIfPossible(pos);
+    if (!write_header(stream, HashValue, netDescription)) return false;
-  }
+    if (!Detail::write_parameters(stream, *featureTransformer)) return false;
-
+    for (std::size_t i = 0; i < LayerStacks; ++i)
-  // Calculate the evaluation value
+      if (!Detail::write_parameters(stream, *(network[i]))) return false;
-  static Value ComputeScore(const Position& pos, bool refresh) {
+    return (bool)stream;
    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 evaluate(const Position& pos, bool adjusted) {
    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;
+    // We manually align the arrays on the stack because with gcc < 9.3
    // overaligning stack variables with alignas() doesn't work correctly.
    constexpr uint64_t alignment = CacheLineSize;
    int delta = 10 - pos.non_pawn_material() / 1515;
 #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
    TransformedFeatureType transformedFeaturesUnaligned[
      FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
    auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
 #else
    alignas(alignment)
      TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
 #endif
    ASSERT_ALIGNED(transformedFeatures, alignment);
    const int bucket = (pos.count<ALL_PIECES>() - 1) / 4;
    const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
    const auto positional = network[bucket]->propagate(transformedFeatures);
    // Give more value to positional evaluation when adjusted flag is set
    if (adjusted)
        return static_cast<Value>(((128 - delta) * psqt + (128 + delta) * positional) / 128 / OutputScale);
    else
        return static_cast<Value>((psqt + positional) / OutputScale);
  }
-  // Evaluation function. Perform full calculation.
+  struct NnueEvalTrace {
-  Value compute_eval(const Position& pos) {
+    static_assert(LayerStacks == PSQTBuckets);
-    return ComputeScore(pos, true);
+
    Value psqt[LayerStacks];
    Value positional[LayerStacks];
    std::size_t correctBucket;
  };
  static NnueEvalTrace trace_evaluate(const Position& pos) {
    // We manually align the arrays on the stack because with gcc < 9.3
    // overaligning stack variables with alignas() doesn't work correctly.
    constexpr uint64_t alignment = CacheLineSize;
 #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
    TransformedFeatureType transformedFeaturesUnaligned[
      FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
    auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
 #else
    alignas(alignment)
      TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
 #endif
    ASSERT_ALIGNED(transformedFeatures, alignment);
    NnueEvalTrace t{};
    t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
    for (IndexType bucket = 0; bucket < LayerStacks; ++bucket) {
      const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket);
      const auto positional = network[bucket]->propagate(transformedFeatures);
      t.psqt[bucket] = static_cast<Value>( materialist / OutputScale );
      t.positional[bucket] = static_cast<Value>( positional / OutputScale );
    }
    return t;
  }
-  // Proceed with the difference calculation if possible
+  static const std::string PieceToChar(" PNBRQK  pnbrqk");
-  void update_eval(const Position& pos) {
+
-    UpdateAccumulatorIfPossible(pos);
+
  // format_cp_compact() converts a Value into (centi)pawns and writes it in a buffer.
  // The buffer must have capacity for at least 5 chars.
  static void format_cp_compact(Value v, char* buffer) {
    buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' ');
    int cp = std::abs(100 * v / PawnValueEg);
    if (cp >= 10000)
    {
        buffer[1] = '0' + cp / 10000; cp %= 10000;
        buffer[2] = '0' + cp / 1000; cp %= 1000;
        buffer[3] = '0' + cp / 100;
        buffer[4] = ' ';
    }
    else if (cp >= 1000)
    {
        buffer[1] = '0' + cp / 1000; cp %= 1000;
        buffer[2] = '0' + cp / 100; cp %= 100;
        buffer[3] = '.';
        buffer[4] = '0' + cp / 10;
    }
    else
    {
        buffer[1] = '0' + cp / 100; cp %= 100;
        buffer[2] = '.';
        buffer[3] = '0' + cp / 10; cp %= 10;
        buffer[4] = '0' + cp / 1;
    }
  }
-} // namespace Eval::NNUE
+
  // format_cp_aligned_dot() converts a Value into (centi)pawns and writes it in a buffer,
  // always keeping two decimals. The buffer must have capacity for at least 7 chars.
  static void format_cp_aligned_dot(Value v, char* buffer) {
    buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' ');
    double cp = 1.0 * std::abs(int(v)) / PawnValueEg;
    sprintf(&buffer[1], "%6.2f", cp);
  }
  // trace() returns a string with the value of each piece on a board,
  // and a table for (PSQT, Layers) values bucket by bucket.
  std::string trace(Position& pos) {
    std::stringstream ss;
    char board[3*8+1][8*8+2];
    std::memset(board, ' ', sizeof(board));
    for (int row = 0; row < 3*8+1; ++row)
      board[row][8*8+1] = '\0';
    // A lambda to output one box of the board
    auto writeSquare = [&board](File file, Rank rank, Piece pc, Value value) {
      const int x = ((int)file) * 8;
      const int y = (7 - (int)rank) * 3;
      for (int i = 1; i < 8; ++i)
         board[y][x+i] = board[y+3][x+i] = '-';
      for (int i = 1; i < 3; ++i)
         board[y+i][x] = board[y+i][x+8] = '|';
      board[y][x] = board[y][x+8] = board[y+3][x+8] = board[y+3][x] = '+';
      if (pc != NO_PIECE)
        board[y+1][x+4] = PieceToChar[pc];
      if (value != VALUE_NONE)
        format_cp_compact(value, &board[y+2][x+2]);
    };
    // We estimate the value of each piece by doing a differential evaluation from
    // the current base eval, simulating the removal of the piece from its square.
    Value base = evaluate(pos);
    base = pos.side_to_move() == WHITE ? base : -base;
    for (File f = FILE_A; f <= FILE_H; ++f)
      for (Rank r = RANK_1; r <= RANK_8; ++r)
      {
        Square sq = make_square(f, r);
        Piece pc = pos.piece_on(sq);
        Value v = VALUE_NONE;
        if (pc != NO_PIECE && type_of(pc) != KING)
        {
          auto st = pos.state();
          pos.remove_piece(sq);
          st->accumulator.computed[WHITE] = false;
          st->accumulator.computed[BLACK] = false;
          Value eval = evaluate(pos);
          eval = pos.side_to_move() == WHITE ? eval : -eval;
          v = base - eval;
          pos.put_piece(pc, sq);
          st->accumulator.computed[WHITE] = false;
          st->accumulator.computed[BLACK] = false;
        }
        writeSquare(f, r, pc, v);
      }
    ss << " NNUE derived piece values:\n";
    for (int row = 0; row < 3*8+1; ++row)
        ss << board[row] << '\n';
    ss << '\n';
    auto t = trace_evaluate(pos);
    ss << " NNUE network contributions "
       << (pos.side_to_move() == WHITE ? "(White to move)" : "(Black to move)") << std::endl
       << "+------------+------------+------------+------------+\n"
       << "|   Bucket   |  Material  | Positional |   Total    |\n"
       << "|            |   (PSQT)   |  (Layers)  |            |\n"
       << "+------------+------------+------------+------------+\n";
    for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket)
    {
      char buffer[3][8];
      std::memset(buffer, '\0', sizeof(buffer));
      format_cp_aligned_dot(t.psqt[bucket], buffer[0]);
      format_cp_aligned_dot(t.positional[bucket], buffer[1]);
      format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], buffer[2]);
      ss <<  "|  " << bucket    << "        "
         << " |  " << buffer[0] << "  "
         << " |  " << buffer[1] << "  "
         << " |  " << buffer[2] << "  "
         << " |";
      if (bucket == t.correctBucket)
          ss << " <-- this bucket is used";
      ss << '\n';
    }
    ss << "+------------+------------+------------+------------+\n";
    return ss.str();
  }
  // Load eval, from a file stream or a memory stream
  bool load_eval(std::string name, std::istream& stream) {
    initialize();
    fileName = name;
    return read_parameters(stream);
  }
  // Save eval, to a file stream or a memory stream
  bool save_eval(std::ostream& stream) {
    if (fileName.empty())
      return false;
    return write_parameters(stream);
  }
  /// Save eval, to a file given by its name
  bool save_eval(const std::optional<std::string>& filename) {
    std::string actualFilename;
    std::string msg;
    if (filename.has_value())
        actualFilename = filename.value();
    else
    {
        if (currentEvalFileName != EvalFileDefaultName)
        {
             msg = "Failed to export a net. A non-embedded net can only be saved if the filename is specified";
             sync_cout << msg << sync_endl;
             return false;
        }
        actualFilename = EvalFileDefaultName;
    }
    std::ofstream stream(actualFilename, std::ios_base::binary);
    bool saved = save_eval(stream);
    msg = saved ? "Network saved successfully to " + actualFilename
                : "Failed to export a net";
    sync_cout << msg << sync_endl;
    return saved;
  }
 } // namespace Stockfish::Eval::NNUE
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -25,11 +25,11 @@
 #include <memory>
-namespace Eval::NNUE {
+namespace Stockfish::Eval::NNUE {
  // Hash value of evaluation function structure
-  constexpr std::uint32_t kHashValue =
+  constexpr std::uint32_t HashValue =
-      FeatureTransformer::GetHashValue() ^ Network::GetHashValue();
+      FeatureTransformer::get_hash_value() ^ Network::get_hash_value();
  // Deleter for automating release of memory area
  template <typename T>
@@ -40,9 +40,20 @@ namespace Eval::NNUE {
    }
  };
  template <typename T>
  struct LargePageDeleter {
    void operator()(T* ptr) const {
      ptr->~T();
      aligned_large_pages_free(ptr);
    }
  };
  template <typename T>
  using AlignedPtr = std::unique_ptr<T, AlignedDeleter<T>>;
-}  // namespace Eval::NNUE
+  template <typename T>
  using LargePagePtr = std::unique_ptr<T, LargePageDeleter<T>>;
 }  // namespace Stockfish::Eval::NNUE
 #endif // #ifndef NNUE_EVALUATE_NNUE_H_INCLUDED
@@ -1,135 +0,0 @@
 /*
  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
@@ -0,0 +1,83 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
  Copyright (C) 2004-2022 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 HalfKAv2_hm of NNUE evaluation function
 #include "half_ka_v2_hm.h"
 #include "../../position.h"
 namespace Stockfish::Eval::NNUE::Features {
  // Orient a square according to perspective (rotates by 180 for black)
  inline Square HalfKAv2_hm::orient(Color perspective, Square s, Square ksq) {
    return Square(int(s) ^ (bool(perspective) * SQ_A8) ^ ((file_of(ksq) < FILE_E) * SQ_H1));
  }
  // Index of a feature for a given king position and another piece on some square
  inline IndexType HalfKAv2_hm::make_index(Color perspective, Square s, Piece pc, Square ksq) {
    Square o_ksq = orient(perspective, ksq, ksq);
    return IndexType(orient(perspective, s, ksq) + PieceSquareIndex[perspective][pc] + PS_NB * KingBuckets[o_ksq]);
  }
  // Get a list of indices for active features
  void HalfKAv2_hm::append_active_indices(
    const Position& pos,
    Color perspective,
    IndexList& active
  ) {
    Square ksq = pos.square<KING>(perspective);
    Bitboard bb = pos.pieces();
    while (bb)
    {
      Square s = pop_lsb(bb);
      active.push_back(make_index(perspective, s, pos.piece_on(s), ksq));
    }
  }
  // append_changed_indices() : get a list of indices for recently changed features
  void HalfKAv2_hm::append_changed_indices(
    Square ksq,
    const DirtyPiece& dp,
    Color perspective,
    IndexList& removed,
    IndexList& added
  ) {
    for (int i = 0; i < dp.dirty_num; ++i) {
      if (dp.from[i] != SQ_NONE)
        removed.push_back(make_index(perspective, dp.from[i], dp.piece[i], ksq));
      if (dp.to[i] != SQ_NONE)
        added.push_back(make_index(perspective, dp.to[i], dp.piece[i], ksq));
    }
  }
  int HalfKAv2_hm::update_cost(const StateInfo* st) {
    return st->dirtyPiece.dirty_num;
  }
  int HalfKAv2_hm::refresh_cost(const Position& pos) {
    return pos.count<ALL_PIECES>();
  }
  bool HalfKAv2_hm::requires_refresh(const StateInfo* st, Color perspective) {
    return st->dirtyPiece.piece[0] == make_piece(perspective, KING);
  }
 }  // namespace Stockfish::Eval::NNUE::Features
@@ -0,0 +1,124 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
  Copyright (C) 2004-2022 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_KA_V2_HM_H_INCLUDED
 #define NNUE_FEATURES_HALF_KA_V2_HM_H_INCLUDED
 #include "../nnue_common.h"
 #include "../../evaluate.h"
 #include "../../misc.h"
 namespace Stockfish {
  struct StateInfo;
 }
 namespace Stockfish::Eval::NNUE::Features {
  // Feature HalfKAv2_hm: Combination of the position of own king
  // and the position of pieces. Position mirrored such that king always on e..h files.
  class HalfKAv2_hm {
    // unique number for each piece type on each square
    enum {
      PS_NONE     =  0,
      PS_W_PAWN   =  0,
      PS_B_PAWN   =  1 * SQUARE_NB,
      PS_W_KNIGHT =  2 * SQUARE_NB,
      PS_B_KNIGHT =  3 * SQUARE_NB,
      PS_W_BISHOP =  4 * SQUARE_NB,
      PS_B_BISHOP =  5 * SQUARE_NB,
      PS_W_ROOK   =  6 * SQUARE_NB,
      PS_B_ROOK   =  7 * SQUARE_NB,
      PS_W_QUEEN  =  8 * SQUARE_NB,
      PS_B_QUEEN  =  9 * SQUARE_NB,
      PS_KING     =  10 * SQUARE_NB,
      PS_NB       =  11 * SQUARE_NB
    };
    static constexpr IndexType PieceSquareIndex[COLOR_NB][PIECE_NB] = {
      // convention: W - us, B - them
      // viewed from other side, W and B are reversed
      { PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_KING, PS_NONE,
        PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_KING, PS_NONE },
      { PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_KING, PS_NONE,
        PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_KING, PS_NONE }
    };
    // Orient a square according to perspective (rotates by 180 for black)
    static Square orient(Color perspective, Square s, Square ksq);
    // Index of a feature for a given king position and another piece on some square
    static IndexType make_index(Color perspective, Square s, Piece pc, Square ksq);
   public:
    // Feature name
    static constexpr const char* Name = "HalfKAv2_hm(Friend)";
    // Hash value embedded in the evaluation file
    static constexpr std::uint32_t HashValue = 0x7f234cb8u;
    // Number of feature dimensions
    static constexpr IndexType Dimensions =
        static_cast<IndexType>(SQUARE_NB) * static_cast<IndexType>(PS_NB) / 2;
    static constexpr int KingBuckets[64] = {
      -1, -1, -1, -1, 31, 30, 29, 28,
      -1, -1, -1, -1, 27, 26, 25, 24,
      -1, -1, -1, -1, 23, 22, 21, 20,
      -1, -1, -1, -1, 19, 18, 17, 16,
      -1, -1, -1, -1, 15, 14, 13, 12,
      -1, -1, -1, -1, 11, 10,  9,  8,
      -1, -1, -1, -1,  7,  6,  5,  4,
      -1, -1, -1, -1,  3,  2,  1,  0
    };
    // Maximum number of simultaneously active features.
    static constexpr IndexType MaxActiveDimensions = 32;
    using IndexList = ValueList<IndexType, MaxActiveDimensions>;
    // Get a list of indices for active features
    static void append_active_indices(
      const Position& pos,
      Color perspective,
      IndexList& active);
    // Get a list of indices for recently changed features
    static void append_changed_indices(
      Square ksq,
      const DirtyPiece& dp,
      Color perspective,
      IndexList& removed,
      IndexList& added
    );
    // Returns the cost of updating one perspective, the most costly one.
    // Assumes no refresh needed.
    static int update_cost(const StateInfo* st);
    static int refresh_cost(const Position& pos);
    // Returns whether the change stored in this StateInfo means that
    // a full accumulator refresh is required.
    static bool requires_refresh(const StateInfo* st, Color perspective);
  };
 }  // namespace Stockfish::Eval::NNUE::Features
 #endif // #ifndef NNUE_FEATURES_HALF_KA_V2_HM_H_INCLUDED
@@ -1,92 +0,0 @@
 /*
  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
@@ -1,67 +0,0 @@
 /*
  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
@@ -1,64 +0,0 @@
 /*
  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
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -22,180 +22,341 @@
 #define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
 #include <iostream>
 #include <algorithm>
 #include <type_traits>
 #include "../nnue_common.h"
 #include "../../simd.h"
-namespace Eval::NNUE::Layers {
+/*
  This file contains the definition for a fully connected layer (aka affine transform).
  Two approaches are employed, depending on the sizes of the transform.
-  // Affine transformation layer
+  Approach 1:
-  template <typename PreviousLayer, IndexType OutputDimensions>
+    - used when the PaddedInputDimensions >= 128
-  class AffineTransform {
+    - uses AVX512 if possible
-   public:
+    - processes inputs in batches of 2*InputSimdWidth
-    // Input/output type
+      - so in batches of 128 for AVX512
-    using InputType = typename PreviousLayer::OutputType;
+    - the weight blocks of size InputSimdWidth are transposed such that
-    using OutputType = std::int32_t;
+      access is sequential
-    static_assert(std::is_same<InputType, std::uint8_t>::value, "");
+    - N columns of the weight matrix are processed a time, where N
      depends on the architecture (the amount of registers)
    - accumulate + hadd is used
-    // Number of input/output dimensions
+  Approach 2:
-    static constexpr IndexType kInputDimensions =
+    - used when the PaddedInputDimensions < 128
-        PreviousLayer::kOutputDimensions;
+    - does not use AVX512
-    static constexpr IndexType kOutputDimensions = OutputDimensions;
+    - expected use-case is for when PaddedInputDimensions == 32 and InputDimensions <= 32.
-    static constexpr IndexType kPaddedInputDimensions =
+      - that's why AVX512 is hard to implement
-        CeilToMultiple<IndexType>(kInputDimensions, kMaxSimdWidth);
+    - expected use-case is small layers
      - not optimized as well as the approach 1
    - inputs are processed in chunks of 4, weights are respectively transposed
    - accumulation happens directly to int32s
 */
-    // Size of forward propagation buffer used in this layer
+namespace Stockfish::Eval::NNUE::Layers {
    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
+// Fallback implementation for older/other architectures.
-    static constexpr std::size_t kBufferSize =
+// Identical for both approaches. Requires the input to be padded to at least 16 values.
-        PreviousLayer::kBufferSize + kSelfBufferSize;
+#if !defined(USE_SSSE3)
  template <IndexType InputDimensions, IndexType PaddedInputDimensions, IndexType OutputDimensions>
  static void affine_transform_non_ssse3(std::int32_t* output, const std::int8_t* weights, const std::int32_t* biases, const std::uint8_t* input)
  {
 # if defined(USE_SSE2)
    // At least a multiple of 16, with SSE2.
    constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
    const __m128i Zeros = _mm_setzero_si128();
    const auto inputVector = reinterpret_cast<const __m128i*>(input);
-    // Hash value embedded in the evaluation file
+# elif defined(USE_MMX)
-    static constexpr std::uint32_t GetHashValue() {
+    constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 8;
-      std::uint32_t hash_value = 0xCC03DAE4u;
+    const __m64 Zeros = _mm_setzero_si64();
-      hash_value += kOutputDimensions;
+    const auto inputVector = reinterpret_cast<const __m64*>(input);
-      hash_value ^= PreviousLayer::GetHashValue() >> 1;
+
-      hash_value ^= PreviousLayer::GetHashValue() << 31;
+# elif defined(USE_NEON)
-      return hash_value;
+    constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
    const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
 # endif
    for (IndexType i = 0; i < OutputDimensions; ++i) {
      const IndexType offset = i * PaddedInputDimensions;
 # if defined(USE_SSE2)
      __m128i sumLo = _mm_cvtsi32_si128(biases[i]);
      __m128i sumHi = Zeros;
      const auto row = reinterpret_cast<const __m128i*>(&weights[offset]);
      for (IndexType j = 0; j < NumChunks; ++j) {
        __m128i row_j = _mm_load_si128(&row[j]);
        __m128i input_j = _mm_load_si128(&inputVector[j]);
        __m128i extendedRowLo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8);
        __m128i extendedRowHi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8);
        __m128i extendedInputLo = _mm_unpacklo_epi8(input_j, Zeros);
        __m128i extendedInputHi = _mm_unpackhi_epi8(input_j, Zeros);
        __m128i productLo = _mm_madd_epi16(extendedRowLo, extendedInputLo);
        __m128i productHi = _mm_madd_epi16(extendedRowHi, extendedInputHi);
        sumLo = _mm_add_epi32(sumLo, productLo);
        sumHi = _mm_add_epi32(sumHi, productHi);
      }
      __m128i sum = _mm_add_epi32(sumLo, sumHi);
      __m128i sumHigh_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
      sum = _mm_add_epi32(sum, sumHigh_64);
      __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
      sum = _mm_add_epi32(sum, sum_second_32);
      output[i] = _mm_cvtsi128_si32(sum);
 # elif defined(USE_MMX)
      __m64 sumLo = _mm_cvtsi32_si64(biases[i]);
      __m64 sumHi = Zeros;
      const auto row = reinterpret_cast<const __m64*>(&weights[offset]);
      for (IndexType j = 0; j < NumChunks; ++j) {
        __m64 row_j = row[j];
        __m64 input_j = inputVector[j];
        __m64 extendedRowLo = _mm_srai_pi16(_mm_unpacklo_pi8(row_j, row_j), 8);
        __m64 extendedRowHi = _mm_srai_pi16(_mm_unpackhi_pi8(row_j, row_j), 8);
        __m64 extendedInputLo = _mm_unpacklo_pi8(input_j, Zeros);
        __m64 extendedInputHi = _mm_unpackhi_pi8(input_j, Zeros);
        __m64 productLo = _mm_madd_pi16(extendedRowLo, extendedInputLo);
        __m64 productHi = _mm_madd_pi16(extendedRowHi, extendedInputHi);
        sumLo = _mm_add_pi32(sumLo, productLo);
        sumHi = _mm_add_pi32(sumHi, productHi);
      }
      __m64 sum = _mm_add_pi32(sumLo, sumHi);
      sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
      output[i] = _mm_cvtsi64_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 < NumChunks; ++j) {
        int16x8_t product = vmull_s8(inputVector[j * 2], row[j * 2]);
        product = vmlal_s8(product, inputVector[j * 2 + 1], row[j * 2 + 1]);
        sum = vpadalq_s16(sum, product);
      }
      output[i] = sum[0] + sum[1] + sum[2] + sum[3];
 # else
      std::int32_t sum = biases[i];
      for (IndexType j = 0; j < InputDimensions; ++j) {
        sum += weights[offset + j] * input[j];
      }
      output[i] = sum;
 # endif
    }
-   // Read network parameters
+# if defined(USE_MMX)
-    bool ReadParameters(std::istream& stream) {
+    _mm_empty();
-      if (!previous_layer_.ReadParameters(stream)) return false;
+# endif
-      stream.read(reinterpret_cast<char*>(biases_),
+  }
-                  kOutputDimensions * sizeof(BiasType));
+#endif
-      stream.read(reinterpret_cast<char*>(weights_),
+
-                  kOutputDimensions * kPaddedInputDimensions *
+  template <IndexType InDims, IndexType OutDims, typename Enabled = void>
-                  sizeof(WeightType));
+  class AffineTransform;
  // A specialization for large inputs.
  template <IndexType InDims, IndexType OutDims>
  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= 2*64)>> {
   public:
    // Input/output type
    using InputType = std::uint8_t;
    using OutputType = std::int32_t;
    // Number of input/output dimensions
    static constexpr IndexType InputDimensions = InDims;
    static constexpr IndexType OutputDimensions = OutDims;
    static constexpr IndexType PaddedInputDimensions =
      ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
    static constexpr IndexType PaddedOutputDimensions =
      ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
    using OutputBuffer = OutputType[PaddedOutputDimensions];
    static_assert(PaddedInputDimensions >= 128, "Something went wrong. This specialization should not have been chosen.");
 #if defined (USE_AVX512)
    static constexpr const IndexType InputSimdWidth = 64;
    static constexpr const IndexType MaxNumOutputRegs = 16;
 #elif defined (USE_AVX2)
    static constexpr const IndexType InputSimdWidth = 32;
    static constexpr const IndexType MaxNumOutputRegs = 8;
 #elif defined (USE_SSSE3)
    static constexpr const IndexType InputSimdWidth = 16;
    static constexpr const IndexType MaxNumOutputRegs = 8;
 #elif defined (USE_NEON)
    static constexpr const IndexType InputSimdWidth = 8;
    static constexpr const IndexType MaxNumOutputRegs = 8;
 #else
    // The fallback implementation will not have permuted weights.
    // We define these to avoid a lot of ifdefs later.
    static constexpr const IndexType InputSimdWidth = 1;
    static constexpr const IndexType MaxNumOutputRegs = 1;
 #endif
    // A big block is a region in the weight matrix of the size [PaddedInputDimensions, NumOutputRegs].
    // A small block is a region of size [InputSimdWidth, 1]
    static constexpr const IndexType NumOutputRegs = std::min(MaxNumOutputRegs, OutputDimensions);
    static constexpr const IndexType SmallBlockSize = InputSimdWidth;
    static constexpr const IndexType BigBlockSize = NumOutputRegs * PaddedInputDimensions;
    static constexpr const IndexType NumSmallBlocksInBigBlock = BigBlockSize / SmallBlockSize;
    static constexpr const IndexType NumSmallBlocksPerOutput = PaddedInputDimensions / SmallBlockSize;
    static constexpr const IndexType NumBigBlocks = OutputDimensions / NumOutputRegs;
    static_assert(OutputDimensions % NumOutputRegs == 0);
    // Hash value embedded in the evaluation file
    static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
      std::uint32_t hashValue = 0xCC03DAE4u;
      hashValue += OutputDimensions;
      hashValue ^= prevHash >> 1;
      hashValue ^= prevHash << 31;
      return hashValue;
    }
    /*
      Transposes the small blocks within a block.
      Effectively means that weights can be traversed sequentially during inference.
    */
    static IndexType get_weight_index(IndexType i)
    {
      const IndexType smallBlock = (i / SmallBlockSize) % NumSmallBlocksInBigBlock;
      const IndexType smallBlockCol = smallBlock / NumSmallBlocksPerOutput;
      const IndexType smallBlockRow = smallBlock % NumSmallBlocksPerOutput;
      const IndexType bigBlock   = i / BigBlockSize;
      const IndexType rest       = i % SmallBlockSize;
      const IndexType idx =
          bigBlock * BigBlockSize
        + smallBlockRow * SmallBlockSize * NumOutputRegs
        + smallBlockCol * SmallBlockSize
        + rest;
      return idx;
    }
    // Read network parameters
    bool read_parameters(std::istream& stream) {
      for (IndexType i = 0; i < OutputDimensions; ++i)
        biases[i] = read_little_endian<BiasType>(stream);
      for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
        weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
      return !stream.fail();
    }
    // Write network parameters
    bool write_parameters(std::ostream& stream) const {
      for (IndexType i = 0; i < OutputDimensions; ++i)
          write_little_endian<BiasType>(stream, biases[i]);
      for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
        write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
      return !stream.fail();
    }
    // Forward propagation
-    const OutputType* Propagate(
+    const OutputType* propagate(
-        const TransformedFeatureType* transformed_features, char* buffer) const {
+        const InputType* input, OutputType* output) const {
      const auto input = previous_layer_.Propagate(
          transformed_features, buffer + kSelfBufferSize);
      const auto output = reinterpret_cast<OutputType*>(buffer);
-  #if defined(USE_AVX512)
+#if defined (USE_AVX512)
-      constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
+      using acc_vec_t = __m512i;
-      const __m512i kOnes = _mm512_set1_epi16(1);
+      using bias_vec_t = __m128i;
-      const auto input_vector = reinterpret_cast<const __m512i*>(input);
+      using weight_vec_t = __m512i;
      using in_vec_t = __m512i;
      #define vec_zero _mm512_setzero_si512()
      #define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
      #define vec_hadd Simd::m512_hadd
      #define vec_haddx4 Simd::m512_haddx4
 #elif defined (USE_AVX2)
      using acc_vec_t = __m256i;
      using bias_vec_t = __m128i;
      using weight_vec_t = __m256i;
      using in_vec_t = __m256i;
      #define vec_zero _mm256_setzero_si256()
      #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
      #define vec_hadd Simd::m256_hadd
      #define vec_haddx4 Simd::m256_haddx4
 #elif defined (USE_SSSE3)
      using acc_vec_t = __m128i;
      using bias_vec_t = __m128i;
      using weight_vec_t = __m128i;
      using in_vec_t = __m128i;
      #define vec_zero _mm_setzero_si128()
      #define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
      #define vec_hadd Simd::m128_hadd
      #define vec_haddx4 Simd::m128_haddx4
 #elif defined (USE_NEON)
      using acc_vec_t = int32x4_t;
      using bias_vec_t = int32x4_t;
      using weight_vec_t = int8x8_t;
      using in_vec_t = int8x8_t;
      #define vec_zero {0}
      #define vec_add_dpbusd_32x2 Simd::neon_m128_add_dpbusd_epi32x2
      #define vec_hadd Simd::neon_m128_hadd
      #define vec_haddx4 Simd::neon_m128_haddx4
 #endif
-  #elif defined(USE_AVX2)
+#if defined (USE_SSSE3) || defined (USE_NEON)
-      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      const in_vec_t* invec = reinterpret_cast<const in_vec_t*>(input);
      const __m256i kOnes = _mm256_set1_epi16(1);
      const auto input_vector = reinterpret_cast<const __m256i*>(input);
-  #elif defined(USE_SSSE3)
+      // Perform accumulation to registers for each big block
-      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      for (IndexType bigBlock = 0; bigBlock < NumBigBlocks; ++bigBlock)
-      const __m128i kOnes = _mm_set1_epi16(1);
+      {
-      const auto input_vector = reinterpret_cast<const __m128i*>(input);
+        acc_vec_t acc[NumOutputRegs] = { vec_zero };
-  #elif defined(USE_NEON)
+        // Each big block has NumOutputRegs small blocks in each "row", one per register.
-      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+        // We process two small blocks at a time to save on one addition without VNNI.
-      const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
+        for (IndexType smallBlock = 0; smallBlock < NumSmallBlocksPerOutput; smallBlock += 2)
  #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 weight_vec_t* weightvec =
-            const auto row_256 = reinterpret_cast<const __m256i*>(&weights_[offset]);
+            reinterpret_cast<const weight_vec_t*>(
-            int j = kNumChunks * 2;
+                weights
              + bigBlock * BigBlockSize
              + smallBlock * SmallBlockSize * NumOutputRegs);
-  #if defined(__MINGW32__) || defined(__MINGW64__)  // See HACK comment below in AVX2.
+          const in_vec_t in0 = invec[smallBlock + 0];
-            __m256i sum256 = _mm256_maddubs_epi16(_mm256_loadu_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+          const in_vec_t in1 = invec[smallBlock + 1];
  #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));
+          for (IndexType k = 0; k < NumOutputRegs; ++k)
-            sum256 = _mm256_hadd_epi32(sum256, sum256);
+            vec_add_dpbusd_32x2(acc[k], in0, weightvec[k], in1, weightvec[k + NumOutputRegs]);
            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)
+        // Horizontally add all accumulators.
-        __m256i sum = _mm256_setzero_si256();
+        if constexpr (NumOutputRegs % 4 == 0)
-        const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
+        {
-        for (IndexType j = 0; j < kNumChunks; ++j) {
+          bias_vec_t* outputvec = reinterpret_cast<bias_vec_t*>(output);
-          __m256i product = _mm256_maddubs_epi16(
+          const bias_vec_t* biasvec = reinterpret_cast<const bias_vec_t*>(biases);
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+          for (IndexType k = 0; k < NumOutputRegs; k += 4)
-            // 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
+            const IndexType idx = (bigBlock * NumOutputRegs + k) / 4;
-            //       even though alignas is specified.
+            outputvec[idx] = vec_haddx4(acc[k+0], acc[k+1], acc[k+2], acc[k+3], biasvec[idx]);
-            _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);
+        else
-        sum = _mm256_hadd_epi32(sum, sum);
+        {
-        const __m128i lo = _mm256_extracti128_si256(sum, 0);
+          for (IndexType k = 0; k < NumOutputRegs; ++k)
-        const __m128i hi = _mm256_extracti128_si256(sum, 1);
+          {
-        output[i] = _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi) + biases_[i];
+            const IndexType idx = (bigBlock * NumOutputRegs + k);
-
+            output[idx] = vec_hadd(acc[k], biases[idx]);
-  #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
      }
 # undef vec_zero
 # undef vec_add_dpbusd_32x2
 # undef vec_hadd
 # undef vec_haddx4
 #else
      // Use old implementation for the other architectures.
      affine_transform_non_ssse3<
        InputDimensions,
        PaddedInputDimensions,
        OutputDimensions>(output, weights, biases, input);
 #endif
      return output;
    }
@@ -203,13 +364,176 @@ namespace Eval::NNUE::Layers {
    using BiasType = OutputType;
    using WeightType = std::int8_t;
-    PreviousLayer previous_layer_;
+    alignas(CacheLineSize) BiasType biases[OutputDimensions];
-
+    alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
    alignas(kCacheLineSize) BiasType biases_[kOutputDimensions];
    alignas(kCacheLineSize)
        WeightType weights_[kOutputDimensions * kPaddedInputDimensions];
  };
-}  // namespace Eval::NNUE::Layers
+  template <IndexType InDims, IndexType OutDims>
  class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < 2*64)>> {
   public:
    // Input/output type
    // Input/output type
    using InputType = std::uint8_t;
    using OutputType = std::int32_t;
    // Number of input/output dimensions
    static constexpr IndexType InputDimensions = InDims;
    static constexpr IndexType OutputDimensions = OutDims;
    static constexpr IndexType PaddedInputDimensions =
      ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
    static constexpr IndexType PaddedOutputDimensions =
      ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
    using OutputBuffer = OutputType[PaddedOutputDimensions];
    static_assert(PaddedInputDimensions < 128, "Something went wrong. This specialization should not have been chosen.");
 #if defined (USE_SSSE3)
    static constexpr const IndexType OutputSimdWidth = SimdWidth / 4;
    static constexpr const IndexType InputSimdWidth = SimdWidth;
 #endif
    // Hash value embedded in the evaluation file
    static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
      std::uint32_t hashValue = 0xCC03DAE4u;
      hashValue += OutputDimensions;
      hashValue ^= prevHash >> 1;
      hashValue ^= prevHash << 31;
      return hashValue;
    }
    static IndexType get_weight_index_scrambled(IndexType i)
    {
      return
        (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
        i / PaddedInputDimensions * 4 +
        i % 4;
    }
    static IndexType get_weight_index(IndexType i)
    {
 #if defined (USE_SSSE3)
      return get_weight_index_scrambled(i);
 #else
      return i;
 #endif
    }
    // Read network parameters
    bool read_parameters(std::istream& stream) {
      for (IndexType i = 0; i < OutputDimensions; ++i)
        biases[i] = read_little_endian<BiasType>(stream);
      for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
        weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
      return !stream.fail();
    }
    // Write network parameters
    bool write_parameters(std::ostream& stream) const {
      for (IndexType i = 0; i < OutputDimensions; ++i)
        write_little_endian<BiasType>(stream, biases[i]);
      for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
        write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
      return !stream.fail();
    }
    // Forward propagation
    const OutputType* propagate(
        const InputType* input, OutputType* output) const {
 #if defined (USE_AVX2)
      using vec_t = __m256i;
      #define vec_setzero _mm256_setzero_si256
      #define vec_set_32 _mm256_set1_epi32
      #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
      #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
      #define vec_add_dpbusd_32x4 Simd::m256_add_dpbusd_epi32x4
      #define vec_hadd Simd::m256_hadd
      #define vec_haddx4 Simd::m256_haddx4
 #elif defined (USE_SSSE3)
      using vec_t = __m128i;
      #define vec_setzero _mm_setzero_si128
      #define vec_set_32 _mm_set1_epi32
      #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
      #define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
      #define vec_add_dpbusd_32x4 Simd::m128_add_dpbusd_epi32x4
      #define vec_hadd Simd::m128_hadd
      #define vec_haddx4 Simd::m128_haddx4
 #endif
 #if defined (USE_SSSE3)
      const auto inputVector = reinterpret_cast<const vec_t*>(input);
      static_assert(OutputDimensions % OutputSimdWidth == 0 || OutputDimensions == 1);
      if constexpr (OutputDimensions % OutputSimdWidth == 0)
      {
        constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
        constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
        const auto input32 = reinterpret_cast<const std::int32_t*>(input);
        const vec_t* biasvec = reinterpret_cast<const vec_t*>(biases);
        vec_t acc[NumRegs];
        for (IndexType k = 0; k < NumRegs; ++k)
          acc[k] = biasvec[k];
        for (IndexType i = 0; i < NumChunks; i += 2)
        {
          const vec_t in0 = vec_set_32(input32[i + 0]);
          const vec_t in1 = vec_set_32(input32[i + 1]);
          const auto col0 = reinterpret_cast<const vec_t*>(&weights[(i + 0) * OutputDimensions * 4]);
          const auto col1 = reinterpret_cast<const vec_t*>(&weights[(i + 1) * OutputDimensions * 4]);
          for (IndexType k = 0; k < NumRegs; ++k)
            vec_add_dpbusd_32x2(acc[k], in0, col0[k], in1, col1[k]);
        }
        vec_t* outptr = reinterpret_cast<vec_t*>(output);
        for (IndexType k = 0; k < NumRegs; ++k)
          outptr[k] = acc[k];
      }
      else if constexpr (OutputDimensions == 1)
      {
        constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
        vec_t sum0 = vec_setzero();
        const auto row0 = reinterpret_cast<const vec_t*>(&weights[0]);
        for (int j = 0; j < (int)NumChunks; ++j)
        {
          const vec_t in = inputVector[j];
          vec_add_dpbusd_32(sum0, in, row0[j]);
        }
        output[0] = vec_hadd(sum0, biases[0]);
      }
 # undef vec_setzero
 # undef vec_set_32
 # undef vec_add_dpbusd_32
 # undef vec_add_dpbusd_32x2
 # undef vec_add_dpbusd_32x4
 # undef vec_hadd
 # undef vec_haddx4
 #else
      // Use old implementation for the other architectures.
      affine_transform_non_ssse3<
        InputDimensions,
        PaddedInputDimensions,
        OutputDimensions>(output, weights, biases, input);
 #endif
      return output;
    }
   private:
    using BiasType = OutputType;
    using WeightType = std::int8_t;
    alignas(CacheLineSize) BiasType biases[OutputDimensions];
    alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
  };
 }  // namespace Stockfish::Eval::NNUE::Layers
 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,164 +23,158 @@
 #include "../nnue_common.h"
-namespace Eval::NNUE::Layers {
+namespace Stockfish::Eval::NNUE::Layers {
  // Clipped ReLU
-  template <typename PreviousLayer>
+  template <IndexType InDims>
  class ClippedReLU {
   public:
    // Input/output type
-    using InputType = typename PreviousLayer::OutputType;
+    using InputType = std::int32_t;
    using OutputType = std::uint8_t;
    static_assert(std::is_same<InputType, std::int32_t>::value, "");
    // Number of input/output dimensions
-    static constexpr IndexType kInputDimensions =
+    static constexpr IndexType InputDimensions = InDims;
-        PreviousLayer::kOutputDimensions;
+    static constexpr IndexType OutputDimensions = InputDimensions;
-    static constexpr IndexType kOutputDimensions = kInputDimensions;
+    static constexpr IndexType PaddedOutputDimensions =
        ceil_to_multiple<IndexType>(OutputDimensions, 32);
-    // Size of forward propagation buffer used in this layer
+    using OutputBuffer = OutputType[PaddedOutputDimensions];
    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() {
+    static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
-      std::uint32_t hash_value = 0x538D24C7u;
+      std::uint32_t hashValue = 0x538D24C7u;
-      hash_value += PreviousLayer::GetHashValue();
+      hashValue += prevHash;
-      return hash_value;
+      return hashValue;
    }
    // Read network parameters
-    bool ReadParameters(std::istream& stream) {
+    bool read_parameters(std::istream&) {
-      return previous_layer_.ReadParameters(stream);
+      return true;
    }
    // Write network parameters
    bool write_parameters(std::ostream&) const {
      return true;
    }
    // Forward propagation
-    const OutputType* Propagate(
+    const OutputType* propagate(
-        const TransformedFeatureType* transformed_features, char* buffer) const {
+        const InputType* input, OutputType* output) 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;
+      if constexpr (InputDimensions % SimdWidth == 0) {
-      const __m256i kZero = _mm256_setzero_si256();
+        constexpr IndexType NumChunks = InputDimensions / SimdWidth;
-      const __m256i kOffsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+        const __m256i Zero = _mm256_setzero_si256();
-      const auto in = reinterpret_cast<const __m256i*>(input);
+        const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
-      const auto out = reinterpret_cast<__m256i*>(output);
+        const auto in = reinterpret_cast<const __m256i*>(input);
-      for (IndexType i = 0; i < kNumChunks; ++i) {
+        const auto out = reinterpret_cast<__m256i*>(output);
-        const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
+        for (IndexType i = 0; i < NumChunks; ++i) {
-
+          const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+              _mm256_load_si256(&in[i * 4 + 0]),
-          // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
+              _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
-          //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
+          const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
-          //       even though alignas is specified.
+              _mm256_load_si256(&in[i * 4 + 2]),
-          _mm256_loadu_si256
+              _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
-  #else
+          _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
-          _mm256_load_si256
+              _mm256_packs_epi16(words0, words1), Zero), Offsets));
-  #endif
+        }
-
+      } else {
-          (&in[i * 4 + 0]),
+        constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
-
+        const __m128i Zero = _mm_setzero_si128();
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+        const auto in = reinterpret_cast<const __m128i*>(input);
-          _mm256_loadu_si256
+        const auto out = reinterpret_cast<__m128i*>(output);
-  #else
+        for (IndexType i = 0; i < NumChunks; ++i) {
-          _mm256_load_si256
+          const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
-  #endif
+              _mm_load_si128(&in[i * 4 + 0]),
-
+              _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
-          (&in[i * 4 + 1])), kWeightScaleBits);
+          const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
-        const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
+              _mm_load_si128(&in[i * 4 + 2]),
-
+              _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+          const __m128i packedbytes = _mm_packs_epi16(words0, words1);
-          _mm256_loadu_si256
+          _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
-  #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;
+      constexpr IndexType Start =
        InputDimensions % SimdWidth == 0
        ? InputDimensions / SimdWidth * SimdWidth
        : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
-  #elif defined(USE_SSSE3)
+  #elif defined(USE_SSE2)
-      constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
+      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
  #ifdef USE_SSE41
-      const __m128i kZero = _mm_setzero_si128();
+      const __m128i Zero = _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) {
+      for (IndexType i = 0; i < NumChunks; ++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);
+            _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
        const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
            _mm_load_si128(&in[i * 4 + 2]),
-            _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits);
+            _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
        const __m128i packedbytes = _mm_packs_epi16(words0, words1);
        _mm_store_si128(&out[i],
  #ifdef USE_SSE41
-          _mm_max_epi8(packedbytes, kZero)
+          _mm_max_epi8(packedbytes, Zero)
  #else
          _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
  #endif
        );
      }
-      constexpr IndexType kStart = kNumChunks * kSimdWidth;
+      constexpr IndexType Start = NumChunks * SimdWidth;
  #elif defined(USE_MMX)
      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
      const __m64 k0x80s = _mm_set1_pi8(-128);
      const auto in = reinterpret_cast<const __m64*>(input);
      const auto out = reinterpret_cast<__m64*>(output);
      for (IndexType i = 0; i < NumChunks; ++i) {
        const __m64 words0 = _mm_srai_pi16(
            _mm_packs_pi32(in[i * 4 + 0], in[i * 4 + 1]),
            WeightScaleBits);
        const __m64 words1 = _mm_srai_pi16(
            _mm_packs_pi32(in[i * 4 + 2], in[i * 4 + 3]),
            WeightScaleBits);
        const __m64 packedbytes = _mm_packs_pi16(words0, words1);
        out[i] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
      }
      _mm_empty();
      constexpr IndexType Start = NumChunks * SimdWidth;
  #elif defined(USE_NEON)
-      constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2);
+      constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
-      const int8x8_t kZero = {0};
+      const int8x8_t Zero = {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) {
+      for (IndexType i = 0; i < NumChunks; ++i) {
        int16x8_t shifted;
        const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
-        pack[0] = vqshrn_n_s32(in[i * 2 + 0], kWeightScaleBits);
+        pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits);
-        pack[1] = vqshrn_n_s32(in[i * 2 + 1], kWeightScaleBits);
+        pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits);
-        out[i] = vmax_s8(vqmovn_s16(shifted), kZero);
+        out[i] = vmax_s8(vqmovn_s16(shifted), Zero);
      }
-      constexpr IndexType kStart = kNumChunks * (kSimdWidth / 2);
+      constexpr IndexType Start = NumChunks * (SimdWidth / 2);
  #else
-      constexpr IndexType kStart = 0;
+      constexpr IndexType Start = 0;
  #endif
-      for (IndexType i = kStart; i < kInputDimensions; ++i) {
+      for (IndexType i = Start; i < InputDimensions; ++i) {
        output[i] = static_cast<OutputType>(
-            std::max(0, std::min(127, input[i] >> kWeightScaleBits)));
+            std::max(0, std::min(127, input[i] >> WeightScaleBits)));
      }
      return output;
    }
   private:
    PreviousLayer previous_layer_;
  };
-}  // namespace Eval::NNUE::Layers
+}  // namespace Stockfish::Eval::NNUE::Layers
 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
@@ -1,68 +0,0 @@
 /*
  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
@@ -0,0 +1,120 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
  Copyright (C) 2004-2022 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_SQR_CLIPPED_RELU_H_INCLUDED
 #define NNUE_LAYERS_SQR_CLIPPED_RELU_H_INCLUDED
 #include "../nnue_common.h"
 namespace Stockfish::Eval::NNUE::Layers {
  // Clipped ReLU
  template <IndexType InDims>
  class SqrClippedReLU {
   public:
    // Input/output type
    using InputType = std::int32_t;
    using OutputType = std::uint8_t;
    // Number of input/output dimensions
    static constexpr IndexType InputDimensions = InDims;
    static constexpr IndexType OutputDimensions = InputDimensions;
    static constexpr IndexType PaddedOutputDimensions =
        ceil_to_multiple<IndexType>(OutputDimensions, 32);
    using OutputBuffer = OutputType[PaddedOutputDimensions];
    // Hash value embedded in the evaluation file
    static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
      std::uint32_t hashValue = 0x538D24C7u;
      hashValue += prevHash;
      return hashValue;
    }
    // Read network parameters
    bool read_parameters(std::istream&) {
      return true;
    }
    // Write network parameters
    bool write_parameters(std::ostream&) const {
      return true;
    }
    // Forward propagation
    const OutputType* propagate(
        const InputType* input, OutputType* output) const {
  #if defined(USE_SSE2)
      constexpr IndexType NumChunks = InputDimensions / 16;
  #ifdef USE_SSE41
      const __m128i Zero = _mm_setzero_si128();
  #else
      const __m128i k0x80s = _mm_set1_epi8(-128);
  #endif
      static_assert(WeightScaleBits == 6);
      const auto in = reinterpret_cast<const __m128i*>(input);
      const auto out = reinterpret_cast<__m128i*>(output);
      for (IndexType i = 0; i < NumChunks; ++i) {
        __m128i words0 = _mm_packs_epi32(
            _mm_load_si128(&in[i * 4 + 0]),
            _mm_load_si128(&in[i * 4 + 1]));
        __m128i words1 = _mm_packs_epi32(
            _mm_load_si128(&in[i * 4 + 2]),
            _mm_load_si128(&in[i * 4 + 3]));
        // Not sure if
        words0 = _mm_srli_epi16(_mm_mulhi_epi16(words0, words0), 3);
        words1 = _mm_srli_epi16(_mm_mulhi_epi16(words1, words1), 3);
        const __m128i packedbytes = _mm_packs_epi16(words0, words1);
        _mm_store_si128(&out[i],
  #ifdef USE_SSE41
          _mm_max_epi8(packedbytes, Zero)
  #else
          _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
  #endif
        );
      }
      constexpr IndexType Start = NumChunks * 16;
  #else
      constexpr IndexType Start = 0;
  #endif
      for (IndexType i = Start; i < InputDimensions; ++i) {
        output[i] = static_cast<OutputType>(
            // realy should be /127 but we need to make it fast
            // needs to be accounted for in the trainer
            std::max(0ll, std::min(127ll, (((long long)input[i] * input[i]) >> (2 * WeightScaleBits)) / 128)));
      }
      return output;
    }
  };
 }  // namespace Stockfish::Eval::NNUE::Layers
 #endif // NNUE_LAYERS_SQR_CLIPPED_RELU_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,17 +23,15 @@
 #include "nnue_architecture.h"
-namespace Eval::NNUE {
+namespace Stockfish::Eval::NNUE {
  // Class that holds the result of affine transformation of input features
-  struct alignas(32) Accumulator {
+  struct alignas(CacheLineSize) Accumulator {
-    std::int16_t
+    std::int16_t accumulation[2][TransformedFeatureDimensions];
-        accumulation[2][kRefreshTriggers.size()][kTransformedFeatureDimensions];
+    std::int32_t psqtAccumulation[2][PSQTBuckets];
-    Value score;
+    bool computed[2];
    bool computed_accumulation;
    bool computed_score;
  };
-}  // namespace Eval::NNUE
+}  // namespace Stockfish::Eval::NNUE
 #endif // NNUE_ACCUMULATOR_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -21,18 +21,118 @@
 #ifndef NNUE_ARCHITECTURE_H_INCLUDED
 #define NNUE_ARCHITECTURE_H_INCLUDED
-// Defines the network structure
+#include <memory>
 #include "architectures/halfkp_256x2-32-32.h"
-namespace Eval::NNUE {
+#include "nnue_common.h"
-  static_assert(kTransformedFeatureDimensions % kMaxSimdWidth == 0, "");
+#include "features/half_ka_v2_hm.h"
  static_assert(Network::kOutputDimensions == 1, "");
  static_assert(std::is_same<Network::OutputType, std::int32_t>::value, "");
-  // Trigger for full calculation instead of difference calculation
+#include "layers/affine_transform.h"
-  constexpr auto kRefreshTriggers = RawFeatures::kRefreshTriggers;
+#include "layers/clipped_relu.h"
 #include "layers/sqr_clipped_relu.h"
-}  // namespace Eval::NNUE
+#include "../misc.h"
 namespace Stockfish::Eval::NNUE {
 // Input features used in evaluation function
 using FeatureSet = Features::HalfKAv2_hm;
 // Number of input feature dimensions after conversion
 constexpr IndexType TransformedFeatureDimensions = 1024;
 constexpr IndexType PSQTBuckets = 8;
 constexpr IndexType LayerStacks = 8;
 struct Network
 {
  static constexpr int FC_0_OUTPUTS = 15;
  static constexpr int FC_1_OUTPUTS = 32;
  Layers::AffineTransform<TransformedFeatureDimensions, FC_0_OUTPUTS + 1> fc_0;
  Layers::SqrClippedReLU<FC_0_OUTPUTS + 1> ac_sqr_0;
  Layers::ClippedReLU<FC_0_OUTPUTS + 1> ac_0;
  Layers::AffineTransform<FC_0_OUTPUTS * 2, FC_1_OUTPUTS> fc_1;
  Layers::ClippedReLU<FC_1_OUTPUTS> ac_1;
  Layers::AffineTransform<FC_1_OUTPUTS, 1> fc_2;
  // Hash value embedded in the evaluation file
  static constexpr std::uint32_t get_hash_value() {
    // input slice hash
    std::uint32_t hashValue = 0xEC42E90Du;
    hashValue ^= TransformedFeatureDimensions * 2;
    hashValue = decltype(fc_0)::get_hash_value(hashValue);
    hashValue = decltype(ac_0)::get_hash_value(hashValue);
    hashValue = decltype(fc_1)::get_hash_value(hashValue);
    hashValue = decltype(ac_1)::get_hash_value(hashValue);
    hashValue = decltype(fc_2)::get_hash_value(hashValue);
    return hashValue;
  }
  // Read network parameters
  bool read_parameters(std::istream& stream) {
    if (!fc_0.read_parameters(stream)) return false;
    if (!ac_0.read_parameters(stream)) return false;
    if (!fc_1.read_parameters(stream)) return false;
    if (!ac_1.read_parameters(stream)) return false;
    if (!fc_2.read_parameters(stream)) return false;
    return true;
  }
  // Read network parameters
  bool write_parameters(std::ostream& stream) const {
    if (!fc_0.write_parameters(stream)) return false;
    if (!ac_0.write_parameters(stream)) return false;
    if (!fc_1.write_parameters(stream)) return false;
    if (!ac_1.write_parameters(stream)) return false;
    if (!fc_2.write_parameters(stream)) return false;
    return true;
  }
  std::int32_t propagate(const TransformedFeatureType* transformedFeatures)
  {
    struct alignas(CacheLineSize) Buffer
    {
      alignas(CacheLineSize) decltype(fc_0)::OutputBuffer fc_0_out;
      alignas(CacheLineSize) decltype(ac_sqr_0)::OutputType ac_sqr_0_out[ceil_to_multiple<IndexType>(FC_0_OUTPUTS * 2, 32)];
      alignas(CacheLineSize) decltype(ac_0)::OutputBuffer ac_0_out;
      alignas(CacheLineSize) decltype(fc_1)::OutputBuffer fc_1_out;
      alignas(CacheLineSize) decltype(ac_1)::OutputBuffer ac_1_out;
      alignas(CacheLineSize) decltype(fc_2)::OutputBuffer fc_2_out;
      Buffer()
      {
          std::memset(this, 0, sizeof(*this));
      }
    };
 #if defined(__clang__) && (__APPLE__)
    // workaround for a bug reported with xcode 12
    static thread_local auto tlsBuffer = std::make_unique<Buffer>();
    // Access TLS only once, cache result.
    Buffer& buffer = *tlsBuffer;
 #else
    alignas(CacheLineSize) static thread_local Buffer buffer;
 #endif
    fc_0.propagate(transformedFeatures, buffer.fc_0_out);
    ac_sqr_0.propagate(buffer.fc_0_out, buffer.ac_sqr_0_out);
    ac_0.propagate(buffer.fc_0_out, buffer.ac_0_out);
    std::memcpy(buffer.ac_sqr_0_out + FC_0_OUTPUTS, buffer.ac_0_out, FC_0_OUTPUTS * sizeof(decltype(ac_0)::OutputType));
    fc_1.propagate(buffer.ac_sqr_0_out, buffer.fc_1_out);
    ac_1.propagate(buffer.fc_1_out, buffer.ac_1_out);
    fc_2.propagate(buffer.ac_1_out, buffer.fc_2_out);
    // buffer.fc_0_out[FC_0_OUTPUTS] is such that 1.0 is equal to 127*(1<<WeightScaleBits) in quantized form
    // but we want 1.0 to be equal to 600*OutputScale
    std::int32_t fwdOut = int(buffer.fc_0_out[FC_0_OUTPUTS]) * (600*OutputScale) / (127*(1<<WeightScaleBits));
    std::int32_t outputValue = buffer.fc_2_out[0] + fwdOut;
    return outputValue;
  }
 };
 }  // namespace Stockfish::Eval::NNUE
 #endif // #ifndef NNUE_ARCHITECTURE_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -21,6 +21,13 @@
 #ifndef NNUE_COMMON_H_INCLUDED
 #define NNUE_COMMON_H_INCLUDED
 #include "../types.h"
 #include <cstring>
 #include <iostream>
 #include "../misc.h"  // for IsLittleEndian
 #if defined(USE_AVX2)
 #include <immintrin.h>
@@ -33,34 +40,40 @@
 #elif defined(USE_SSE2)
 #include <emmintrin.h>
 #elif defined(USE_MMX)
 #include <mmintrin.h>
 #elif defined(USE_NEON)
 #include <arm_neon.h>
 #endif
-namespace Eval::NNUE {
+namespace Stockfish::Eval::NNUE {
  // Version of the evaluation file
-  constexpr std::uint32_t kVersion = 0x7AF32F16u;
+  constexpr std::uint32_t Version = 0x7AF32F20u;
  // Constant used in evaluation value calculation
-  constexpr int FV_SCALE = 16;
+  constexpr int OutputScale = 16;
-  constexpr int kWeightScaleBits = 6;
+  constexpr int WeightScaleBits = 6;
  // Size of cache line (in bytes)
-  constexpr std::size_t kCacheLineSize = 64;
+  constexpr std::size_t CacheLineSize = 64;
  // SIMD width (in bytes)
  #if defined(USE_AVX2)
-  constexpr std::size_t kSimdWidth = 32;
+  constexpr std::size_t SimdWidth = 32;
  #elif defined(USE_SSE2)
-  constexpr std::size_t kSimdWidth = 16;
+  constexpr std::size_t SimdWidth = 16;
  #elif defined(USE_MMX)
  constexpr std::size_t SimdWidth = 8;
  #elif defined(USE_NEON)
-  constexpr std::size_t kSimdWidth = 16;
+  constexpr std::size_t SimdWidth = 16;
  #endif
-  constexpr std::size_t kMaxSimdWidth = 32;
+  constexpr std::size_t MaxSimdWidth = 32;
  // Type of input feature after conversion
  using TransformedFeatureType = std::uint8_t;
@@ -68,10 +81,86 @@ namespace Eval::NNUE {
  // Round n up to be a multiple of base
  template <typename IntType>
-  constexpr IntType CeilToMultiple(IntType n, IntType base) {
+  constexpr IntType ceil_to_multiple(IntType n, IntType base) {
-    return (n + base - 1) / base * base;
+      return (n + base - 1) / base * base;
  }
-}  // namespace Eval::NNUE
+  // read_little_endian() is our utility to read an integer (signed or unsigned, any size)
  // from a stream in little-endian order. We swap the byte order after the read if
  // necessary to return a result with the byte ordering of the compiling machine.
  template <typename IntType>
  inline IntType read_little_endian(std::istream& stream) {
      IntType result;
      if (IsLittleEndian)
          stream.read(reinterpret_cast<char*>(&result), sizeof(IntType));
      else
      {
          std::uint8_t u[sizeof(IntType)];
          typename std::make_unsigned<IntType>::type v = 0;
          stream.read(reinterpret_cast<char*>(u), sizeof(IntType));
          for (std::size_t i = 0; i < sizeof(IntType); ++i)
              v = (v << 8) | u[sizeof(IntType) - i - 1];
          std::memcpy(&result, &v, sizeof(IntType));
      }
      return result;
  }
  // write_little_endian() is our utility to write an integer (signed or unsigned, any size)
  // to a stream in little-endian order. We swap the byte order before the write if
  // necessary to always write in little endian order, independently of the byte
  // ordering of the compiling machine.
  template <typename IntType>
  inline void write_little_endian(std::ostream& stream, IntType value) {
      if (IsLittleEndian)
          stream.write(reinterpret_cast<const char*>(&value), sizeof(IntType));
      else
      {
          std::uint8_t u[sizeof(IntType)];
          typename std::make_unsigned<IntType>::type v = value;
          std::size_t i = 0;
          // if constexpr to silence the warning about shift by 8
          if constexpr (sizeof(IntType) > 1)
          {
            for (; i + 1 < sizeof(IntType); ++i)
            {
                u[i] = (std::uint8_t)v;
                v >>= 8;
            }
          }
          u[i] = (std::uint8_t)v;
          stream.write(reinterpret_cast<char*>(u), sizeof(IntType));
      }
  }
  // read_little_endian(s, out, N) : read integers in bulk from a little indian stream.
  // This reads N integers from stream s and put them in array out.
  template <typename IntType>
  inline void read_little_endian(std::istream& stream, IntType* out, std::size_t count) {
      if (IsLittleEndian)
          stream.read(reinterpret_cast<char*>(out), sizeof(IntType) * count);
      else
          for (std::size_t i = 0; i < count; ++i)
              out[i] = read_little_endian<IntType>(stream);
  }
  // write_little_endian(s, values, N) : write integers in bulk to a little indian stream.
  // This takes N integers from array values and writes them on stream s.
  template <typename IntType>
  inline void write_little_endian(std::ostream& stream, const IntType* values, std::size_t count) {
      if (IsLittleEndian)
          stream.write(reinterpret_cast<const char*>(values), sizeof(IntType) * count);
      else
          for (std::size_t i = 0; i < count; ++i)
              write_little_endian<IntType>(stream, values[i]);
  }
 }  // namespace Stockfish::Eval::NNUE
 #endif // #ifndef NNUE_COMMON_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,333 +23,569 @@
 #include "nnue_common.h"
 #include "nnue_architecture.h"
-#include "features/index_list.h"
+
 #include "../misc.h"
 #include "../position.h"
 #include <cstring> // std::memset()
-namespace Eval::NNUE {
+namespace Stockfish::Eval::NNUE {
  using BiasType       = std::int16_t;
  using WeightType     = std::int16_t;
  using PSQTWeightType = std::int32_t;
  // If vector instructions are enabled, we update and refresh the
  // accumulator tile by tile such that each tile fits in the CPU's
  // vector registers.
  #define VECTOR
  static_assert(PSQTBuckets % 8 == 0,
    "Per feature PSQT values cannot be processed at granularity lower than 8 at a time.");
  #ifdef USE_AVX512
  typedef __m512i vec_t;
  typedef __m256i psqt_vec_t;
  #define vec_load(a) _mm512_load_si512(a)
  #define vec_store(a,b) _mm512_store_si512(a,b)
  #define vec_add_16(a,b) _mm512_add_epi16(a,b)
  #define vec_sub_16(a,b) _mm512_sub_epi16(a,b)
  #define vec_mul_16(a,b) _mm512_mullo_epi16(a,b)
  #define vec_zero() _mm512_setzero_epi32()
  #define vec_set_16(a) _mm512_set1_epi16(a)
  #define vec_max_16(a,b) _mm512_max_epi16(a,b)
  #define vec_min_16(a,b) _mm512_min_epi16(a,b)
  inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
    vec_t compacted = _mm512_packs_epi16(_mm512_srli_epi16(a,7),_mm512_srli_epi16(b,7));
    return _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), compacted);
  }
  #define vec_load_psqt(a) _mm256_load_si256(a)
  #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
  #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
  #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
  #define vec_zero_psqt() _mm256_setzero_si256()
  #define NumRegistersSIMD 32
  #define MaxChunkSize 64
  #elif USE_AVX2
  typedef __m256i vec_t;
  typedef __m256i psqt_vec_t;
  #define vec_load(a) _mm256_load_si256(a)
  #define vec_store(a,b) _mm256_store_si256(a,b)
  #define vec_add_16(a,b) _mm256_add_epi16(a,b)
  #define vec_sub_16(a,b) _mm256_sub_epi16(a,b)
  #define vec_mul_16(a,b) _mm256_mullo_epi16(a,b)
  #define vec_zero() _mm256_setzero_si256()
  #define vec_set_16(a) _mm256_set1_epi16(a)
  #define vec_max_16(a,b) _mm256_max_epi16(a,b)
  #define vec_min_16(a,b) _mm256_min_epi16(a,b)
  inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
    vec_t compacted = _mm256_packs_epi16(_mm256_srli_epi16(a,7), _mm256_srli_epi16(b,7));
    return _mm256_permute4x64_epi64(compacted, 0b11011000);
  }
  #define vec_load_psqt(a) _mm256_load_si256(a)
  #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
  #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
  #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
  #define vec_zero_psqt() _mm256_setzero_si256()
  #define NumRegistersSIMD 16
  #define MaxChunkSize 32
  #elif USE_SSE2
  typedef __m128i vec_t;
  typedef __m128i psqt_vec_t;
  #define vec_load(a) (*(a))
  #define vec_store(a,b) *(a)=(b)
  #define vec_add_16(a,b) _mm_add_epi16(a,b)
  #define vec_sub_16(a,b) _mm_sub_epi16(a,b)
  #define vec_mul_16(a,b) _mm_mullo_epi16(a,b)
  #define vec_zero() _mm_setzero_si128()
  #define vec_set_16(a) _mm_set1_epi16(a)
  #define vec_max_16(a,b) _mm_max_epi16(a,b)
  #define vec_min_16(a,b) _mm_min_epi16(a,b)
  #define vec_msb_pack_16(a,b) _mm_packs_epi16(_mm_srli_epi16(a,7),_mm_srli_epi16(b,7))
  #define vec_load_psqt(a) (*(a))
  #define vec_store_psqt(a,b) *(a)=(b)
  #define vec_add_psqt_32(a,b) _mm_add_epi32(a,b)
  #define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b)
  #define vec_zero_psqt() _mm_setzero_si128()
  #define NumRegistersSIMD (Is64Bit ? 16 : 8)
  #define MaxChunkSize 16
  #elif USE_MMX
  typedef __m64 vec_t;
  typedef __m64 psqt_vec_t;
  #define vec_load(a) (*(a))
  #define vec_store(a,b) *(a)=(b)
  #define vec_add_16(a,b) _mm_add_pi16(a,b)
  #define vec_sub_16(a,b) _mm_sub_pi16(a,b)
  #define vec_mul_16(a,b) _mm_mullo_pi16(a,b)
  #define vec_zero() _mm_setzero_si64()
  #define vec_set_16(a) _mm_set1_pi16(a)
  inline vec_t vec_max_16(vec_t a,vec_t b){
    vec_t comparison = _mm_cmpgt_pi16(a,b);
    return _mm_or_si64(_mm_and_si64(comparison, a), _mm_andnot_si64(comparison, b));
  }
  inline vec_t vec_min_16(vec_t a,vec_t b){
    vec_t comparison = _mm_cmpgt_pi16(a,b);
    return _mm_or_si64(_mm_and_si64(comparison, b), _mm_andnot_si64(comparison, a));
  }
  #define vec_msb_pack_16(a,b) _mm_packs_pi16(_mm_srli_pi16(a,7),_mm_srli_pi16(b,7))
  #define vec_load_psqt(a) (*(a))
  #define vec_store_psqt(a,b) *(a)=(b)
  #define vec_add_psqt_32(a,b) _mm_add_pi32(a,b)
  #define vec_sub_psqt_32(a,b) _mm_sub_pi32(a,b)
  #define vec_zero_psqt() _mm_setzero_si64()
  #define vec_cleanup() _mm_empty()
  #define NumRegistersSIMD 8
  #define MaxChunkSize 8
  #elif USE_NEON
  typedef int16x8_t vec_t;
  typedef int32x4_t psqt_vec_t;
  #define vec_load(a) (*(a))
  #define vec_store(a,b) *(a)=(b)
  #define vec_add_16(a,b) vaddq_s16(a,b)
  #define vec_sub_16(a,b) vsubq_s16(a,b)
  #define vec_mul_16(a,b) vmulq_s16(a,b)
  #define vec_zero() vec_t{0}
  #define vec_set_16(a) vdupq_n_s16(a)
  #define vec_max_16(a,b) vmaxq_s16(a,b)
  #define vec_min_16(a,b) vminq_s16(a,b)
  inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
    const int8x8_t shifta = vshrn_n_s16(a, 7);
    const int8x8_t shiftb = vshrn_n_s16(b, 7);
    const int8x16_t compacted = vcombine_s8(shifta,shiftb);
    return *reinterpret_cast<const vec_t*> (&compacted);
  }
  #define vec_load_psqt(a) (*(a))
  #define vec_store_psqt(a,b) *(a)=(b)
  #define vec_add_psqt_32(a,b) vaddq_s32(a,b)
  #define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
  #define vec_zero_psqt() psqt_vec_t{0}
  #define NumRegistersSIMD 16
  #define MaxChunkSize 16
  #else
  #undef VECTOR
  #endif
  #ifdef VECTOR
      // Compute optimal SIMD register count for feature transformer accumulation.
      // We use __m* types as template arguments, which causes GCC to emit warnings
      // about losing some attribute information. This is irrelevant to us as we
      // only take their size, so the following pragma are harmless.
      #if defined(__GNUC__)
      #pragma GCC diagnostic push
      #pragma GCC diagnostic ignored "-Wignored-attributes"
      #endif
      template <typename SIMDRegisterType,
                typename LaneType,
                int      NumLanes,
                int      MaxRegisters>
      static constexpr int BestRegisterCount()
      {
          #define RegisterSize  sizeof(SIMDRegisterType)
          #define LaneSize      sizeof(LaneType)
          static_assert(RegisterSize >= LaneSize);
          static_assert(MaxRegisters <= NumRegistersSIMD);
          static_assert(MaxRegisters > 0);
          static_assert(NumRegistersSIMD > 0);
          static_assert(RegisterSize % LaneSize == 0);
          static_assert((NumLanes * LaneSize) % RegisterSize == 0);
          const int ideal = (NumLanes * LaneSize) / RegisterSize;
          if (ideal <= MaxRegisters)
            return ideal;
          // Look for the largest divisor of the ideal register count that is smaller than MaxRegisters
          for (int divisor = MaxRegisters; divisor > 1; --divisor)
            if (ideal % divisor == 0)
              return divisor;
          return 1;
      }
      static constexpr int NumRegs     = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>();
      static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>();
      #if defined(__GNUC__)
      #pragma GCC diagnostic pop
      #endif
  #endif
  // Input feature converter
  class FeatureTransformer {
   private:
    // Number of output dimensions for one side
-    static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
+    static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
    #ifdef VECTOR
    static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
    static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4;
    static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
    static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets");
    #endif
   public:
    // Output type
    using OutputType = TransformedFeatureType;
    // Number of input/output dimensions
-    static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
+    static constexpr IndexType InputDimensions = FeatureSet::Dimensions;
-    static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
+    static constexpr IndexType OutputDimensions = HalfDimensions;
    // Size of forward propagation buffer
-    static constexpr std::size_t kBufferSize =
+    static constexpr std::size_t BufferSize =
-        kOutputDimensions * sizeof(OutputType);
+        OutputDimensions * sizeof(OutputType);
    // Hash value embedded in the evaluation file
-    static constexpr std::uint32_t GetHashValue() {
+    static constexpr std::uint32_t get_hash_value() {
-      return RawFeatures::kHashValue ^ kOutputDimensions;
+      return FeatureSet::HashValue ^ (OutputDimensions * 2);
    }
    // Read network parameters
-    bool ReadParameters(std::istream& stream) {
+    bool read_parameters(std::istream& stream) {
-      stream.read(reinterpret_cast<char*>(biases_),
+
-                  kHalfDimensions * sizeof(BiasType));
+      read_little_endian<BiasType      >(stream, biases     , HalfDimensions                  );
-      stream.read(reinterpret_cast<char*>(weights_),
+      read_little_endian<WeightType    >(stream, weights    , HalfDimensions * InputDimensions);
-                  kHalfDimensions * kInputDimensions * sizeof(WeightType));
+      read_little_endian<PSQTWeightType>(stream, psqtWeights, PSQTBuckets    * InputDimensions);
      return !stream.fail();
    }
-    // Proceed with the difference calculation if possible
+    // Write network parameters
-    bool UpdateAccumulatorIfPossible(const Position& pos) const {
+    bool write_parameters(std::ostream& stream) const {
-      const auto now = pos.state();
+
-      if (now->accumulator.computed_accumulation) {
+      write_little_endian<BiasType      >(stream, biases     , HalfDimensions                  );
-        return true;
+      write_little_endian<WeightType    >(stream, weights    , HalfDimensions * InputDimensions);
-      }
+      write_little_endian<PSQTWeightType>(stream, psqtWeights, PSQTBuckets    * InputDimensions);
-      const auto prev = now->previous;
+
-      if (prev && prev->accumulator.computed_accumulation) {
+      return !stream.fail();
        UpdateAccumulator(pos);
        return true;
      }
      return false;
    }
    // Convert input features
-    void Transform(const Position& pos, OutputType* output, bool refresh) const {
+    std::int32_t transform(const Position& pos, OutputType* output, int bucket) const {
-      if (refresh || !UpdateAccumulatorIfPossible(pos)) {
+      update_accumulator(pos, WHITE);
-        RefreshAccumulator(pos);
+      update_accumulator(pos, BLACK);
      }
      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 auto& accumulation = pos.state()->accumulator.accumulation;
-        const IndexType offset = kHalfDimensions * p;
+      const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
-  #if defined(USE_AVX2)
+      const auto psqt = (
-        auto out = reinterpret_cast<__m256i*>(&output[offset]);
+            psqtAccumulation[perspectives[0]][bucket]
-        for (IndexType j = 0; j < kNumChunks; ++j) {
+          - psqtAccumulation[perspectives[1]][bucket]
-          __m256i sum0 =
+        ) / 2;
  #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*>(
+      for (IndexType p = 0; p < 2; ++p)
-              accumulation[perspectives[p]][0])[j * 2 + 0]);
+      {
-          __m256i sum1 =
+          const IndexType offset = (HalfDimensions / 2) * p;
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+#if defined(VECTOR)
            _mm256_loadu_si256
  #else
            _mm256_load_si256
  #endif
-            (&reinterpret_cast<const __m256i*>(
+          constexpr IndexType OutputChunkSize = MaxChunkSize;
-              accumulation[perspectives[p]][0])[j * 2 + 1]);
+          static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
          constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
-  #if defined(__MINGW32__) || defined(__MINGW64__)
+          vec_t Zero = vec_zero();
-          _mm256_storeu_si256
+          vec_t One = vec_set_16(127);
  #else
          _mm256_store_si256
  #endif
-          (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
+          const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
-              _mm256_packs_epi16(sum0, sum1), kZero), kControl));
+          const vec_t* in1 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
-        }
+                vec_t* out = reinterpret_cast<      vec_t*>(output + offset);
-  #elif defined(USE_SSSE3)
+          for (IndexType j = 0; j < NumOutputChunks; j += 1)
-        auto out = reinterpret_cast<__m128i*>(&output[offset]);
+          {
-        for (IndexType j = 0; j < kNumChunks; ++j) {
+              const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero);
-          __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
+              const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero);
-              accumulation[perspectives[p]][0])[j * 2 + 0]);
+              const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero);
-          __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
+              const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero);
              accumulation[perspectives[p]][0])[j * 2 + 1]);
      const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
-          _mm_store_si128(&out[j],
+              const vec_t pa = vec_mul_16(sum0a, sum1a);
              const vec_t pb = vec_mul_16(sum0b, sum1b);
-  #ifdef USE_SSE41
+              out[j] = vec_msb_pack_16(pa, pb);
-            _mm_max_epi8(packedbytes, kZero)
+          }
  #else
            _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
  #endif
-          );
+#else
        }
-  #elif defined(USE_NEON)
+          for (IndexType j = 0; j < HalfDimensions / 2; ++j) {
-        const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
+              BiasType sum0 = accumulation[static_cast<int>(perspectives[p])][j + 0];
-        for (IndexType j = 0; j < kNumChunks; ++j) {
+              BiasType sum1 = accumulation[static_cast<int>(perspectives[p])][j + HalfDimensions / 2];
-          int16x8_t sum = reinterpret_cast<const int16x8_t*>(
+              sum0 = std::max<int>(0, std::min<int>(127, sum0));
-              accumulation[perspectives[p]][0])[j];
+              sum1 = std::max<int>(0, std::min<int>(127, sum1));
-          out[j] = vmax_s8(vqmovn_s16(sum), kZero);
+              output[offset + j] = static_cast<OutputType>(sum0 * sum1 / 128);
-        }
+          }
  #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
 #endif
      }
-    }
+
 #if defined(vec_cleanup)
      vec_cleanup();
 #endif
      return psqt;
   } // end of function transform()
   private:
-    // Calculate cumulative value without using difference calculation
+    void update_accumulator(const Position& pos, const Color perspective) const {
    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)
+      // The size must be enough to contain the largest possible update.
-          auto accumulation = reinterpret_cast<__m256i*>(
+      // That might depend on the feature set and generally relies on the
-              &accumulator.accumulation[perspective][i][0]);
+      // feature set's update cost calculation to be correct and never
-          auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+      // allow updates with more added/removed features than MaxActiveDimensions.
          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)
+  #ifdef VECTOR
-          auto accumulation = reinterpret_cast<__m128i*>(
+      // Gcc-10.2 unnecessarily spills AVX2 registers if this array
-              &accumulator.accumulation[perspective][i][0]);
+      // is defined in the VECTOR code below, once in each branch
-          auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+      vec_t acc[NumRegs];
-          constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+      psqt_vec_t psqt[NumPsqtRegs];
          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
-        }
+      // Look for a usable accumulator of an earlier position. We keep track
      // of the estimated gain in terms of features to be added/subtracted.
      StateInfo *st = pos.state(), *next = nullptr;
      int gain = FeatureSet::refresh_cost(pos);
      while (st->previous && !st->accumulator.computed[perspective])
      {
        // This governs when a full feature refresh is needed and how many
        // updates are better than just one full refresh.
        if (   FeatureSet::requires_refresh(st, perspective)
            || (gain -= FeatureSet::update_cost(st) + 1) < 0)
          break;
        next = st;
        st = st->previous;
      }
-      accumulator.computed_accumulation = true;
+      if (st->accumulator.computed[perspective])
-      accumulator.computed_score = false;
+      {
-    }
+        if (next == nullptr)
          return;
-    // Calculate cumulative value using difference calculation
+        // Update incrementally in two steps. First, we update the "next"
-    void UpdateAccumulator(const Position& pos) const {
+        // accumulator. Then, we update the current accumulator (pos.state()).
      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)
+        // Gather all features to be updated.
-        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        const Square ksq = pos.square<KING>(perspective);
-        auto accumulation = reinterpret_cast<__m256i*>(
+        FeatureSet::IndexList removed[2], added[2];
-            &accumulator.accumulation[perspective][i][0]);
+        FeatureSet::append_changed_indices(
          ksq, next->dirtyPiece, perspective, removed[0], added[0]);
        for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
          FeatureSet::append_changed_indices(
            ksq, st2->dirtyPiece, perspective, removed[1], added[1]);
-  #elif defined(USE_SSE2)
+        // Mark the accumulators as computed.
-        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        next->accumulator.computed[perspective] = true;
-        auto accumulation = reinterpret_cast<__m128i*>(
+        pos.state()->accumulator.computed[perspective] = true;
            &accumulator.accumulation[perspective][i][0]);
-  #elif defined(USE_NEON)
+        // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
-        constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
+        StateInfo *states_to_update[3] =
-        auto accumulation = reinterpret_cast<int16x8_t*>(
+          { next, next == pos.state() ? nullptr : pos.state(), nullptr };
-            &accumulator.accumulation[perspective][i][0]);
+  #ifdef VECTOR
-  #endif
+        for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
        {
          // Load accumulator
          auto accTile = reinterpret_cast<vec_t*>(
            &st->accumulator.accumulation[perspective][j * TileHeight]);
          for (IndexType k = 0; k < NumRegs; ++k)
            acc[k] = vec_load(&accTile[k]);
          for (IndexType i = 0; states_to_update[i]; ++i)
          {
            // Difference calculation for the deactivated features
            for (const auto index : removed[i])
            {
              const IndexType offset = HalfDimensions * index + j * TileHeight;
              auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
              for (IndexType k = 0; k < NumRegs; ++k)
                acc[k] = vec_sub_16(acc[k], column[k]);
            }
            // Difference calculation for the activated features
            for (const auto index : added[i])
            {
              const IndexType offset = HalfDimensions * index + j * TileHeight;
              auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
              for (IndexType k = 0; k < NumRegs; ++k)
                acc[k] = vec_add_16(acc[k], column[k]);
            }
            // Store accumulator
            accTile = reinterpret_cast<vec_t*>(
              &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]);
            for (IndexType k = 0; k < NumRegs; ++k)
              vec_store(&accTile[k], acc[k]);
          }
        }
        for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
        {
          // Load accumulator
          auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
            &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
            psqt[k] = vec_load_psqt(&accTilePsqt[k]);
          for (IndexType i = 0; states_to_update[i]; ++i)
          {
            // Difference calculation for the deactivated features
            for (const auto index : removed[i])
            {
              const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
              auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
              for (std::size_t k = 0; k < NumPsqtRegs; ++k)
                psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
            }
            // Difference calculation for the activated features
            for (const auto index : added[i])
            {
              const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
              auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
              for (std::size_t k = 0; k < NumPsqtRegs; ++k)
                psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
            }
            // Store accumulator
            accTilePsqt = reinterpret_cast<psqt_vec_t*>(
              &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
            for (std::size_t k = 0; k < NumPsqtRegs; ++k)
              vec_store_psqt(&accTilePsqt[k], psqt[k]);
          }
        }
  #else
        for (IndexType i = 0; states_to_update[i]; ++i)
        {
          std::memcpy(states_to_update[i]->accumulator.accumulation[perspective],
              st->accumulator.accumulation[perspective],
              HalfDimensions * sizeof(BiasType));
          for (std::size_t k = 0; k < PSQTBuckets; ++k)
            states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k];
          st = states_to_update[i];
        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]) {
+          for (const auto index : removed[i])
-            const IndexType offset = kHalfDimensions * index;
+          {
            const IndexType offset = HalfDimensions * index;
-  #if defined(USE_AVX2)
+            for (IndexType j = 0; j < HalfDimensions; ++j)
-            auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+              st->accumulator.accumulation[perspective][j] -= weights[offset + j];
            for (IndexType j = 0; j < kNumChunks; ++j) {
              accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
            }
-  #elif defined(USE_SSE2)
+            for (std::size_t k = 0; k < PSQTBuckets; ++k)
-            auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+              st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k];
-            for (IndexType j = 0; j < kNumChunks; ++j) {
+          }
              accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
            }
-  #elif defined(USE_NEON)
+          // Difference calculation for the activated features
-            auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
+          for (const auto index : added[i])
-            for (IndexType j = 0; j < kNumChunks; ++j) {
+          {
-              accumulation[j] = vsubq_s16(accumulation[j], column[j]);
+            const IndexType offset = HalfDimensions * index;
            }
-  #else
+            for (IndexType j = 0; j < HalfDimensions; ++j)
-            for (IndexType j = 0; j < kHalfDimensions; ++j) {
+              st->accumulator.accumulation[perspective][j] += weights[offset + j];
              accumulator.accumulation[perspective][i][j] -=
                  weights_[offset + j];
            }
  #endif
            for (std::size_t k = 0; k < PSQTBuckets; ++k)
              st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
          }
        }
-        { // Difference calculation for the activated features
+  #endif
-          for (const auto index : added_indices[perspective]) {
+      }
-            const IndexType offset = kHalfDimensions * index;
+      else
      {
        // Refresh the accumulator
        auto& accumulator = pos.state()->accumulator;
        accumulator.computed[perspective] = true;
        FeatureSet::IndexList active;
        FeatureSet::append_active_indices(pos, perspective, active);
-  #if defined(USE_AVX2)
+  #ifdef VECTOR
-            auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
+        for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
-            for (IndexType j = 0; j < kNumChunks; ++j) {
+        {
-              accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
+          auto biasesTile = reinterpret_cast<const vec_t*>(
-            }
+              &biases[j * TileHeight]);
          for (IndexType k = 0; k < NumRegs; ++k)
            acc[k] = biasesTile[k];
-  #elif defined(USE_SSE2)
+          for (const auto index : active)
-            auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
+          {
-            for (IndexType j = 0; j < kNumChunks; ++j) {
+            const IndexType offset = HalfDimensions * index + j * TileHeight;
-              accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
+            auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
            }
-  #elif defined(USE_NEON)
+            for (unsigned k = 0; k < NumRegs; ++k)
-            auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
+              acc[k] = vec_add_16(acc[k], column[k]);
-            for (IndexType j = 0; j < kNumChunks; ++j) {
+          }
-              accumulation[j] = vaddq_s16(accumulation[j], column[j]);
+
-            }
+          auto accTile = reinterpret_cast<vec_t*>(
              &accumulator.accumulation[perspective][j * TileHeight]);
          for (unsigned k = 0; k < NumRegs; k++)
            vec_store(&accTile[k], acc[k]);
        }
        for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
        {
          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
            psqt[k] = vec_zero_psqt();
          for (const auto index : active)
          {
            const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
            auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
            for (std::size_t k = 0; k < NumPsqtRegs; ++k)
              psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
          }
          auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
            &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
            vec_store_psqt(&accTilePsqt[k], psqt[k]);
        }
  #else
-            for (IndexType j = 0; j < kHalfDimensions; ++j) {
+        std::memcpy(accumulator.accumulation[perspective], biases,
-              accumulator.accumulation[perspective][i][j] +=
+            HalfDimensions * sizeof(BiasType));
                  weights_[offset + j];
            }
  #endif
-          }
+        for (std::size_t k = 0; k < PSQTBuckets; ++k)
          accumulator.psqtAccumulation[perspective][k] = 0;
        for (const auto index : active)
        {
          const IndexType offset = HalfDimensions * index;
          for (IndexType j = 0; j < HalfDimensions; ++j)
            accumulator.accumulation[perspective][j] += weights[offset + j];
          for (std::size_t k = 0; k < PSQTBuckets; ++k)
            accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
        }
  #endif
      }
-      accumulator.computed_accumulation = true;
+  #if defined(USE_MMX)
-      accumulator.computed_score = false;
+      _mm_empty();
  #endif
    }
-    using BiasType = std::int16_t;
+    alignas(CacheLineSize) BiasType biases[HalfDimensions];
-    using WeightType = std::int16_t;
+    alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
-
+    alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
    alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
    alignas(kCacheLineSize)
        WeightType weights_[kHalfDimensions * kInputDimensions];
  };
-}  // namespace Eval::NNUE
+}  // namespace Stockfish::Eval::NNUE
 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,35 +24,38 @@
 #include "position.h"
 #include "thread.h"
 namespace Stockfish {
 namespace {
  #define V Value
  #define S(mg, eg) make_score(mg, eg)
  // Pawn penalties
-  constexpr Score Backward      = S( 9, 24);
+  constexpr Score Backward      = S( 6, 19);
-  constexpr Score Doubled       = S(11, 56);
+  constexpr Score Doubled       = S(11, 51);
-  constexpr Score Isolated      = S( 5, 15);
+  constexpr Score DoubledEarly  = S(17,  7);
-  constexpr Score WeakLever     = S( 0, 56);
+  constexpr Score Isolated      = S( 1, 20);
-  constexpr Score WeakUnopposed = S(13, 27);
+  constexpr Score WeakLever     = S( 2, 57);
  constexpr Score WeakUnopposed = S(15, 18);
  // Bonus for blocked pawns at 5th or 6th rank
-  constexpr Score BlockedPawn[2] = { S(-11, -4), S(-3, 4) };
+  constexpr Score BlockedPawn[2] = { S(-19, -8), S(-7, 3) };
  constexpr Score BlockedStorm[RANK_NB] = {
-    S(0, 0), S(0, 0), S(76, 78), S(-10, 15), S(-7, 10), S(-4, 6), S(-1, 2)
+    S(0, 0), S(0, 0), S(64, 75), S(-3, 14), S(-12, 19), S(-7, 4), S(-10, 5)
  };
  // Connected pawn bonus
-  constexpr int Connected[RANK_NB] = { 0, 7, 8, 12, 29, 48, 86 };
+  constexpr int Connected[RANK_NB] = { 0, 3, 7, 7, 15, 54, 86 };
  // Strength of pawn shelter for our king by [distance from edge][rank].
  // RANK_1 = 0 is used for files where we have no pawn, or pawn is behind our king.
  constexpr Value ShelterStrength[int(FILE_NB) / 2][RANK_NB] = {
-    { V( -6), V( 81), V( 93), V( 58), V( 39), V( 18), V(  25) },
+    { V(-2), V(85), V(95), V(53), V(39), V(23), V(25) },
-    { V(-43), V( 61), V( 35), V(-49), V(-29), V(-11), V( -63) },
+    { V(-55), V(64), V(32), V(-55), V(-30), V(-11), V(-61) },
-    { V(-10), V( 75), V( 23), V( -2), V( 32), V(  3), V( -45) },
+    { V(-11), V(75), V(19), V(-6), V(26), V(9), V(-47) },
-    { V(-39), V(-13), V(-29), V(-52), V(-48), V(-67), V(-166) }
+    { V(-41), V(-11), V(-27), V(-58), V(-42), V(-66), V(-163) }
  };
  // Danger of enemy pawns moving toward our king by [distance from edge][rank].
@@ -60,12 +63,18 @@ namespace {
  // is behind our king. Note that UnblockedStorm[0][1-2] accommodate opponent pawn
  // on edge, likely blocked by our king.
  constexpr Value UnblockedStorm[int(FILE_NB) / 2][RANK_NB] = {
-    { V( 85), V(-289), V(-166), V(97), V(50), V( 45), V( 50) },
+    { V(94), V(-280), V(-170), V(90), V(59), V(47), V(53) },
-    { V( 46), V( -25), V( 122), V(45), V(37), V(-10), V( 20) },
+    { V(43), V(-17), V(128), V(39), V(26), V(-17), V(15) },
-    { V( -6), V(  51), V( 168), V(34), V(-2), V(-22), V(-14) },
+    { V(-9), V(62), V(170), V(34), V(-5), V(-20), V(-11) },
-    { V(-15), V( -11), V( 101), V( 4), V(11), V(-15), V(-29) }
+    { V(-27), V(-19), V(106), V(10), V(2), V(-13), V(-24) }
  };
  // KingOnFile[semi-open Us][semi-open Them] contains bonuses/penalties
  // for king when the king is on a semi-open or open file.
  constexpr Score KingOnFile[2][2] = {{ S(-18,11), S(-6,-3)  },
                                     {  S(  0, 0), S( 5,-4) }};
  #undef S
  #undef V
@@ -80,13 +89,14 @@ namespace {
    constexpr Color     Them = ~Us;
    constexpr Direction Up   = pawn_push(Us);
    constexpr Direction Down = -Up;
    Bitboard neighbours, stoppers, support, phalanx, opposed;
    Bitboard lever, leverPush, blocked;
    Square s;
    bool backward, passed, doubled;
    Score score = SCORE_ZERO;
-    const Square* pl = pos.squares<PAWN>(Us);
+    Bitboard b = pos.pieces(Us, PAWN);
    Bitboard ourPawns   = pos.pieces(  Us, PAWN);
    Bitboard theirPawns = pos.pieces(Them, PAWN);
@@ -99,8 +109,10 @@ namespace {
    e->blockedCount += popcount(shift<Up>(ourPawns) & (theirPawns | doubleAttackThem));
    // Loop through all pawns of the current color and score each pawn
-    while ((s = *pl++) != SQ_NONE)
+    while (b)
    {
        s = pop_lsb(b);
        assert(pos.piece_on(s) == make_piece(Us, PAWN));
        Rank r = relative_rank(Us, s);
@@ -116,6 +128,13 @@ namespace {
        phalanx    = neighbours & rank_bb(s);
        support    = neighbours & rank_bb(s - Up);
        if (doubled)
        {
            // Additional doubled penalty if none of their pawns is fixed
            if (!(ourPawns & shift<Down>(theirPawns | pawn_attacks_bb<Them>(theirPawns))))
                score -= DoubledEarly;
        }
        // A pawn is backward when it is behind all pawns of the same color on
        // the adjacent files and cannot safely advance.
        backward =  !(neighbours & forward_ranks_bb(Them, s + Up))
@@ -147,7 +166,7 @@ namespace {
        if (support | phalanx)
        {
            int v =  Connected[r] * (2 + bool(phalanx) - bool(opposed))
-                   + 21 * popcount(support);
+                   + 22 * popcount(support);
            score += make_score(v, v * (r - 2) / 4);
        }
@@ -165,14 +184,14 @@ namespace {
        else if (backward)
            score -=  Backward
-                    + WeakUnopposed * !opposed;
+                    + WeakUnopposed * !opposed * bool(~(FileABB | FileHBB) & s);
        if (!support)
            score -=  Doubled * doubled
                    + WeakLever * more_than_one(lever);
-        if (blocked && r > RANK_4)
+        if (blocked && r >= RANK_5)
-            score += BlockedPawn[r-4];
+            score += BlockedPawn[r - RANK_5];
    }
    return score;
@@ -219,7 +238,7 @@ Score Entry::evaluate_shelter(const Position& pos, Square ksq) const {
  Score bonus = make_score(5, 5);
-  File center = Utility::clamp(file_of(ksq), FILE_B, FILE_G);
+  File center = std::clamp(file_of(ksq), FILE_B, FILE_G);
  for (File f = File(center - 1); f <= File(center + 1); ++f)
  {
      b = ourPawns & file_bb(f);
@@ -237,6 +256,9 @@ Score Entry::evaluate_shelter(const Position& pos, Square ksq) const {
          bonus -= make_score(UnblockedStorm[d][theirRank], 0);
  }
  // King On File
  bonus -= KingOnFile[pos.is_on_semiopen_file(Us, ksq)][pos.is_on_semiopen_file(Them, ksq)];
  return bonus;
 }
@@ -269,7 +291,7 @@ Score Entry::do_king_safety(const Position& pos) {
  if (pawns & attacks_bb<KING>(ksq))
      minPawnDist = 1;
  else while (pawns)
-      minPawnDist = std::min(minPawnDist, distance(ksq, pop_lsb(&pawns)));
+      minPawnDist = std::min(minPawnDist, distance(ksq, pop_lsb(pawns)));
  return shelter - make_score(0, 16 * minPawnDist);
 }
@@ -279,3 +301,5 @@ template Score Entry::do_king_safety<WHITE>(const Position& pos);
 template Score Entry::do_king_safety<BLACK>(const Position& pos);
 } // namespace Pawns
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,7 +23,7 @@
 #include "position.h"
 #include "types.h"
-namespace Pawns {
+namespace Stockfish::Pawns {
 /// Pawns::Entry contains various information about a pawn structure. A lookup
 /// to the pawn hash table (performed by calling the probe function) returns a
@@ -65,6 +65,6 @@ typedef HashTable<Entry, 131072> Table;
 Entry* probe(const Position& pos);
-} // namespace Pawns
+} // namespace Stockfish::Pawns
 #endif // #ifndef PAWNS_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,6 +23,8 @@
 #include <iomanip>
 #include <sstream>
 #include "nnue/evaluate_nnue.h"
 #include "bitboard.h"
 #include "misc.h"
 #include "movegen.h"
@@ -32,8 +34,13 @@
 #include "uci.h"
 #include "syzygy/tbprobe.h"
 #include "tools/packed_sfen.h"
 #include "tools/sfen_packer.h"
 using std::string;
 namespace Stockfish {
 namespace Zobrist {
  Key psq[PIECE_NB][SQUARE_NB];
@@ -71,12 +78,14 @@ std::ostream& operator<<(std::ostream& os, const Position& pos) {
     << std::setfill(' ') << std::dec << "\nCheckers: ";
  for (Bitboard b = pos.checkers(); b; )
-      os << UCI::square(pop_lsb(&b)) << " ";
+      os << UCI::square(pop_lsb(b)) << " ";
  if (    int(Tablebases::MaxCardinality) >= popcount(pos.pieces())
      && !pos.can_castle(ANY_CASTLING))
  {
      StateInfo st;
      ASSERT_ALIGNED(&st, Eval::NNUE::CacheLineSize);
      Position p;
      p.set(pos.fen(), pos.is_chess960(), &st, pos.this_thread());
      Tablebases::ProbeState s1, s2;
@@ -195,12 +204,8 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th
  std::memset(this, 0, sizeof(Position));
  std::memset(si, 0, sizeof(StateInfo));
  std::fill_n(&pieceList[0][0], sizeof(pieceList) / sizeof(Square), SQ_NONE);
  st = si;
  // Each piece on board gets a unique ID used to track the piece later
  PieceId piece_id, next_piece_id = PIECE_ID_ZERO;
  ss >> std::noskipws;
  // 1. Piece placement
@@ -212,21 +217,8 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th
      else if (token == '/')
          sq += 2 * SOUTH;
-      else if ((idx = PieceToChar.find(token)) != string::npos)
+      else if ((idx = PieceToChar.find(token)) != string::npos) {
-      {
+          put_piece(Piece(idx), sq);
          auto pc = Piece(idx);
          put_piece(pc, sq);
          if (Eval::useNNUE)
          {
              // Kings get a fixed ID, other pieces get ID in order of placement
              piece_id =
                (idx == W_KING) ? PIECE_ID_WKING :
                (idx == B_KING) ? PIECE_ID_BKING :
                next_piece_id++;
              evalList.put_piece(piece_id, sq, pc);
          }
          ++sq;
      }
  }
@@ -318,7 +310,7 @@ void Position::set_castling_right(Color c, Square rfrom) {
  Square kto = relative_square(c, cr & KING_SIDE ? SQ_G1 : SQ_C1);
  Square rto = relative_square(c, cr & KING_SIDE ? SQ_F1 : SQ_D1);
-  castlingPath[cr] =   (between_bb(rfrom, rto) | between_bb(kfrom, kto) | rto | kto)
+  castlingPath[cr] =   (between_bb(rfrom, rto) | between_bb(kfrom, kto))
                    & ~(kfrom | rfrom);
 }
@@ -357,7 +349,7 @@ void Position::set_state(StateInfo* si) const {
  for (Bitboard b = pieces(); b; )
  {
-      Square s = pop_lsb(&b);
+      Square s = pop_lsb(b);
      Piece pc = piece_on(s);
      si->key ^= Zobrist::psq[pc][s];
@@ -408,7 +400,7 @@ Position& Position::set(const string& code, Color c, StateInfo* si) {
 /// Position::fen() returns a FEN representation of the position. In case of
 /// Chess960 the Shredder-FEN notation is used. This is mainly a debugging function.
-const string Position::fen() const {
+string Position::fen() const {
  int emptyCnt;
  std::ostringstream ss;
@@ -474,7 +466,7 @@ Bitboard Position::slider_blockers(Bitboard sliders, Square s, Bitboard& pinners
  while (snipers)
  {
-    Square sniperSq = pop_lsb(&snipers);
+    Square sniperSq = pop_lsb(snipers);
    Bitboard b = between_bb(s, sniperSq) & occupancy;
    if (b && !more_than_one(b))
@@ -518,7 +510,7 @@ bool Position::legal(Move m) const {
  // En passant captures are a tricky special case. Because they are rather
  // uncommon, we do it simply by testing whether the king is attacked after
  // the move is made.
-  if (type_of(m) == ENPASSANT)
+  if (type_of(m) == EN_PASSANT)
  {
      Square ksq = square<KING>(us);
      Square capsq = to - pawn_push(us);
@@ -546,22 +538,20 @@ bool Position::legal(Move m) const {
          if (attackers_to(s) & pieces(~us))
              return false;
-      // In case of Chess960, verify that when moving the castling rook we do
+      // In case of Chess960, verify if the Rook blocks some checks
      // not discover some hidden checker.
      // For instance an enemy queen in SQ_A1 when castling rook is in SQ_B1.
-      return   !chess960
+      return !chess960 || !(blockers_for_king(us) & to_sq(m));
            || !(attacks_bb<ROOK>(to, pieces() ^ to_sq(m)) & pieces(~us, ROOK, QUEEN));
  }
  // If the moving piece is a king, check whether the destination square is
  // attacked by the opponent.
  if (type_of(piece_on(from)) == KING)
-      return !(attackers_to(to) & pieces(~us));
+      return !(attackers_to(to, pieces() ^ from) & pieces(~us));
  // A non-king move is legal if and only if it is not pinned or it
  // is moving along the ray towards or away from the king.
-  return   !(blockers_for_king(us) & from)
+  return !(blockers_for_king(us) & from)
-        ||  aligned(from, to, square<KING>(us));
+      || aligned(from, to, square<KING>(us));
 }
@@ -577,8 +567,10 @@ bool Position::pseudo_legal(const Move m) const {
  Piece pc = moved_piece(m);
  // Use a slower but simpler function for uncommon cases
  // yet we skip the legality check of MoveList<LEGAL>().
  if (type_of(m) != NORMAL)
-      return MoveList<LEGAL>(*this).contains(m);
+      return checkers() ? MoveList<    EVASIONS>(*this).contains(m)
                        : MoveList<NON_EVASIONS>(*this).contains(m);
  // Is not a promotion, so promotion piece must be empty
  if (promotion_type(m) - KNIGHT != NO_PIECE_TYPE)
@@ -623,8 +615,8 @@ bool Position::pseudo_legal(const Move m) const {
          if (more_than_one(checkers()))
              return false;
-          // Our move must be a blocking evasion or a capture of the checking piece
+          // Our move must be a blocking interposition or a capture of the checking piece
-          if (!((between_bb(lsb(checkers()), square<KING>(us)) | checkers()) & to))
+          if (!(between_bb(square<KING>(us), lsb(checkers())) & to))
              return false;
      }
      // In case of king moves under check we have to remove king so as to catch
@@ -668,7 +660,7 @@ bool Position::gives_check(Move m) const {
  // of direct checks and ordinary discovered check, so the only case we
  // need to handle is the unusual case of a discovered check through
  // the captured pawn.
-  case ENPASSANT:
+  case EN_PASSANT:
  {
      Square capsq = make_square(file_of(to), rank_of(from));
      Bitboard b = (pieces() ^ from ^ capsq) | to;
@@ -676,19 +668,15 @@ bool Position::gives_check(Move m) const {
      return  (attacks_bb<  ROOK>(square<KING>(~sideToMove), b) & pieces(sideToMove, QUEEN, ROOK))
            | (attacks_bb<BISHOP>(square<KING>(~sideToMove), b) & pieces(sideToMove, QUEEN, BISHOP));
  }
-  case CASTLING:
+  default: //CASTLING
  {
-      Square kfrom = from;
+      // Castling is encoded as 'king captures the rook'
-      Square rfrom = to; // Castling is encoded as 'king captures the rook'
+      Square ksq = square<KING>(~sideToMove);
-      Square kto = relative_square(sideToMove, rfrom > kfrom ? SQ_G1 : SQ_C1);
+      Square rto = relative_square(sideToMove, to > from ? SQ_F1 : SQ_D1);
      Square rto = relative_square(sideToMove, rfrom > kfrom ? SQ_F1 : SQ_D1);
-      return   (attacks_bb<ROOK>(rto) & square<KING>(~sideToMove))
+      return   (attacks_bb<ROOK>(rto) & ksq)
-            && (attacks_bb<ROOK>(rto, (pieces() ^ kfrom ^ rfrom) | rto | kto) & square<KING>(~sideToMove));
+            && (attacks_bb<ROOK>(rto, pieces() ^ from ^ to) & ksq);
  }
  default:
      assert(false);
      return false;
  }
 }
@@ -719,10 +707,8 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
  ++st->pliesFromNull;
  // Used by NNUE
-  st->accumulator.computed_accumulation = false;
+  st->accumulator.computed[WHITE] = false;
-  st->accumulator.computed_score = false;
+  st->accumulator.computed[BLACK] = false;
  PieceId dp0 = PIECE_ID_NONE;
  PieceId dp1 = PIECE_ID_NONE;
  auto& dp = st->dirtyPiece;
  dp.dirty_num = 1;
@@ -731,7 +717,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
  Square from = from_sq(m);
  Square to = to_sq(m);
  Piece pc = piece_on(from);
-  Piece captured = type_of(m) == ENPASSANT ? make_piece(them, PAWN) : piece_on(to);
+  Piece captured = type_of(m) == EN_PASSANT ? make_piece(them, PAWN) : piece_on(to);
  assert(color_of(pc) == us);
  assert(captured == NO_PIECE || color_of(captured) == (type_of(m) != CASTLING ? them : us));
@@ -757,7 +743,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
      // update non-pawn material.
      if (type_of(captured) == PAWN)
      {
-          if (type_of(m) == ENPASSANT)
+          if (type_of(m) == EN_PASSANT)
          {
              capsq -= pawn_push(us);
@@ -773,20 +759,18 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
      else
          st->nonPawnMaterial[them] -= PieceValue[MG][captured];
-      if (Eval::useNNUE)
+      if (Eval::NNUE::useNNUE != Eval::NNUE::UseNNUEMode::False)
      {
-          dp.dirty_num = 2; // 2 pieces moved
+          dp.dirty_num = 2;  // 1 piece moved, 1 piece captured
-          dp1 = piece_id_on(capsq);
+          dp.piece[1] = captured;
-          dp.pieceId[1] = dp1;
+          dp.from[1] = capsq;
-          dp.old_piece[1] = evalList.piece_with_id(dp1);
+          dp.to[1] = SQ_NONE;
          evalList.put_piece(dp1, capsq, NO_PIECE);
          dp.new_piece[1] = evalList.piece_with_id(dp1);
      }
      // Update board and piece lists
      remove_piece(capsq);
-      if (type_of(m) == ENPASSANT)
+      if (type_of(m) == EN_PASSANT)
          board[capsq] = NO_PIECE;
      // Update material hash key and prefetch access to materialTable
@@ -819,13 +803,11 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
  // Move the piece. The tricky Chess960 castling is handled earlier
  if (type_of(m) != CASTLING)
  {
-      if (Eval::useNNUE)
+      if (Eval::NNUE::useNNUE != Eval::NNUE::UseNNUEMode::False)
      {
-          dp0 = piece_id_on(from);
+          dp.piece[0] = pc;
-          dp.pieceId[0] = dp0;
+          dp.from[0] = from;
-          dp.old_piece[0] = evalList.piece_with_id(dp0);
+          dp.to[0] = to;
          evalList.put_piece(dp0, to, pc);
          dp.new_piece[0] = evalList.piece_with_id(dp0);
      }
      move_piece(from, to);
@@ -834,7 +816,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
  // If the moving piece is a pawn do some special extra work
  if (type_of(pc) == PAWN)
  {
-      // Set en-passant square if the moved pawn can be captured
+      // Set en passant square if the moved pawn can be captured
      if (   (int(to) ^ int(from)) == 16
          && (pawn_attacks_bb(us, to - pawn_push(us)) & pieces(them, PAWN)))
      {
@@ -852,11 +834,14 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) {
          remove_piece(to);
          put_piece(promotion, to);
-          if (Eval::useNNUE)
+          if (Eval::NNUE::useNNUE != Eval::NNUE::UseNNUEMode::False)
          {
-              dp0 = piece_id_on(to);
+              // Promoting pawn to SQ_NONE, promoted piece from SQ_NONE
-              evalList.put_piece(dp0, to, promotion);
+              dp.to[0] = SQ_NONE;
-              dp.new_piece[0] = evalList.piece_with_id(dp0);
+              dp.piece[dp.dirty_num] = promotion;
              dp.from[dp.dirty_num] = SQ_NONE;
              dp.to[dp.dirty_num] = to;
              dp.dirty_num++;
          }
          // Update hash keys
@@ -950,17 +935,11 @@ void Position::undo_move(Move m) {
  {
      move_piece(to, from); // Put the piece back at the source square
      if (Eval::useNNUE)
      {
          PieceId dp0 = st->dirtyPiece.pieceId[0];
          evalList.put_piece(dp0, from, pc);
      }
      if (st->capturedPiece)
      {
          Square capsq = to;
-          if (type_of(m) == ENPASSANT)
+          if (type_of(m) == EN_PASSANT)
          {
              capsq -= pawn_push(us);
@@ -972,14 +951,6 @@ void Position::undo_move(Move m) {
          }
          put_piece(st->capturedPiece, capsq); // Restore the captured piece
          if (Eval::useNNUE)
          {
              PieceId dp1 = st->dirtyPiece.pieceId[1];
              assert(evalList.piece_with_id(dp1).from[WHITE] == PS_NONE);
              assert(evalList.piece_with_id(dp1).from[BLACK] == PS_NONE);
              evalList.put_piece(dp1, capsq, st->capturedPiece);
          }
      }
  }
@@ -1001,32 +972,16 @@ void Position::do_castling(Color us, Square from, Square& to, Square& rfrom, Squ
  rto = relative_square(us, kingSide ? SQ_F1 : SQ_D1);
  to = relative_square(us, kingSide ? SQ_G1 : SQ_C1);
-  if (Eval::useNNUE)
+  if (Do && Eval::NNUE::useNNUE != Eval::NNUE::UseNNUEMode::False)
  {
      PieceId dp0, dp1;
      auto& dp = st->dirtyPiece;
-      dp.dirty_num = 2; // 2 pieces moved
+      dp.piece[0] = make_piece(us, KING);
-
+      dp.from[0] = from;
-      if (Do)
+      dp.to[0] = to;
-      {
+      dp.piece[1] = make_piece(us, ROOK);
-          dp0 = piece_id_on(from);
+      dp.from[1] = rfrom;
-          dp1 = piece_id_on(rfrom);
+      dp.to[1] = rto;
-          dp.pieceId[0] = dp0;
+      dp.dirty_num = 2;
          dp.old_piece[0] = evalList.piece_with_id(dp0);
          evalList.put_piece(dp0, to, make_piece(us, KING));
          dp.new_piece[0] = evalList.piece_with_id(dp0);
          dp.pieceId[1] = dp1;
          dp.old_piece[1] = evalList.piece_with_id(dp1);
          evalList.put_piece(dp1, rto, make_piece(us, ROOK));
          dp.new_piece[1] = evalList.piece_with_id(dp1);
      }
      else
      {
          dp0 = piece_id_on(to);
          dp1 = piece_id_on(rto);
          evalList.put_piece(dp0, from, make_piece(us, KING));
          evalList.put_piece(dp1, rfrom, make_piece(us, ROOK));
      }
  }
  // Remove both pieces first since squares could overlap in Chess960
@@ -1038,7 +993,7 @@ void Position::do_castling(Color us, Square from, Square& to, Square& rfrom, Squ
 }
-/// Position::do(undo)_null_move() is used to do(undo) a "null move": it flips
+/// Position::do_null_move() is used to do a "null move": it flips
 /// the side to move without executing any move on the board.
 void Position::do_null_move(StateInfo& newSt) {
@@ -1046,17 +1001,17 @@ void Position::do_null_move(StateInfo& newSt) {
  assert(!checkers());
  assert(&newSt != st);
-  if (Eval::useNNUE)
+  std::memcpy(&newSt, st, offsetof(StateInfo, accumulator));
  {
      std::memcpy(&newSt, st, sizeof(StateInfo));
      st->accumulator.computed_score = false;
  }
  else
      std::memcpy(&newSt, st, offsetof(StateInfo, accumulator));
  newSt.previous = st;
  st = &newSt;
  // Used by NNUE
  st->dirtyPiece.dirty_num = 0;
  st->dirtyPiece.piece[0] = NO_PIECE; // Avoid checks in UpdateAccumulator()
  st->accumulator.computed[WHITE] = false;
  st->accumulator.computed[BLACK] = false;
  if (st->epSquare != SQ_NONE)
  {
      st->key ^= Zobrist::enpassant[file_of(st->epSquare)];
@@ -1064,9 +1019,9 @@ void Position::do_null_move(StateInfo& newSt) {
  }
  st->key ^= Zobrist::side;
  prefetch(TT.first_entry(st->key));
  ++st->rule50;
  prefetch(TT.first_entry(key()));
  st->pliesFromNull = 0;
  sideToMove = ~sideToMove;
@@ -1078,6 +1033,9 @@ void Position::do_null_move(StateInfo& newSt) {
  assert(pos_is_ok());
 }
 /// Position::undo_null_move() must be used to undo a "null move"
 void Position::undo_null_move() {
  assert(!checkers());
@@ -1089,7 +1047,7 @@ void Position::undo_null_move() {
 /// Position::key_after() computes the new hash key after the given move. Needed
 /// for speculative prefetch. It doesn't recognize special moves like castling,
-/// en-passant and promotions.
+/// en passant and promotions.
 Key Position::key_after(Move m) const {
@@ -1114,7 +1072,7 @@ bool Position::see_ge(Move m, Value threshold) const {
  assert(is_ok(m));
-  // Only deal with normal moves, assume others pass a simple see
+  // Only deal with normal moves, assume others pass a simple SEE
  if (type_of(m) != NORMAL)
      return VALUE_ZERO >= threshold;
@@ -1128,8 +1086,9 @@ bool Position::see_ge(Move m, Value threshold) const {
  if (swap <= 0)
      return true;
  assert(color_of(piece_on(from)) == sideToMove);
  Bitboard occupied = pieces() ^ from ^ to;
-  Color stm = color_of(piece_on(from));
+  Color stm = sideToMove;
  Bitboard attackers = attackers_to(to, occupied);
  Bitboard stmAttackers, bb;
  int res = 1;
@@ -1143,10 +1102,10 @@ bool Position::see_ge(Move m, Value threshold) const {
      if (!(stmAttackers = attackers & pieces(stm)))
          break;
-      // Don't allow pinned pieces to attack (except the king) as long as
+      // Don't allow pinned pieces to attack as long as there are
-      // there are pinners on their original square.
+      // pinners on their original square.
-      if (st->pinners[~stm] & occupied)
+      if (pinners(~stm) & occupied)
-          stmAttackers &= ~st->blockersForKing[stm];
+          stmAttackers &= ~blockers_for_king(stm);
      if (!stmAttackers)
          break;
@@ -1160,7 +1119,7 @@ bool Position::see_ge(Move m, Value threshold) const {
          if ((swap = PawnValueMg - swap) < res)
              break;
-          occupied ^= lsb(bb);
+          occupied ^= least_significant_square_bb(bb);
          attackers |= attacks_bb<BISHOP>(to, occupied) & pieces(BISHOP, QUEEN);
      }
@@ -1169,7 +1128,7 @@ bool Position::see_ge(Move m, Value threshold) const {
          if ((swap = KnightValueMg - swap) < res)
              break;
-          occupied ^= lsb(bb);
+          occupied ^= least_significant_square_bb(bb);
      }
      else if ((bb = stmAttackers & pieces(BISHOP)))
@@ -1177,7 +1136,7 @@ bool Position::see_ge(Move m, Value threshold) const {
          if ((swap = BishopValueMg - swap) < res)
              break;
-          occupied ^= lsb(bb);
+          occupied ^= least_significant_square_bb(bb);
          attackers |= attacks_bb<BISHOP>(to, occupied) & pieces(BISHOP, QUEEN);
      }
@@ -1186,7 +1145,7 @@ bool Position::see_ge(Move m, Value threshold) const {
          if ((swap = RookValueMg - swap) < res)
              break;
-          occupied ^= lsb(bb);
+          occupied ^= least_significant_square_bb(bb);
          attackers |= attacks_bb<ROOK>(to, occupied) & pieces(ROOK, QUEEN);
      }
@@ -1195,7 +1154,7 @@ bool Position::see_ge(Move m, Value threshold) const {
          if ((swap = QueenValueMg - swap) < res)
              break;
-          occupied ^= lsb(bb);
+          occupied ^= least_significant_square_bb(bb);
          attackers |=  (attacks_bb<BISHOP>(to, occupied) & pieces(BISHOP, QUEEN))
                      | (attacks_bb<ROOK  >(to, occupied) & pieces(ROOK  , QUEEN));
      }
@@ -1224,6 +1183,22 @@ bool Position::is_draw(int ply) const {
 }
 /// Position::is_fifty_move_draw() returns true if a game can be claimed
 /// by a fifty-move draw rule.
 bool Position::is_fifty_move_draw() const {
  return (st->rule50 > 99 && (!checkers() || MoveList<LEGAL>(*this).size()));
 }
 /// Position::is_three_fold_repetition() returns true if there is 3-fold repetition.
 bool Position::is_three_fold_repetition() const {
  return st->repetition < 0;
 }
 // Position::has_repeated() tests whether there has been at least one repetition
 // of positions since the last capture or pawn move.
@@ -1269,7 +1244,7 @@ bool Position::has_game_cycle(int ply) const {
          Square s1 = from_sq(move);
          Square s2 = to_sq(move);
-          if (!(between_bb(s1, s2) & pieces()))
+          if (!((between_bb(s1, s2) ^ s2) & pieces()))
          {
              if (ply > i)
                  return true;
@@ -1366,21 +1341,17 @@ bool Position::pos_is_ok() const {
              assert(0 && "pos_is_ok: Bitboards");
  StateInfo si = *st;
  ASSERT_ALIGNED(&si, Eval::NNUE::CacheLineSize);
  set_state(&si);
  if (std::memcmp(&si, st, sizeof(StateInfo)))
      assert(0 && "pos_is_ok: State");
  for (Piece pc : Pieces)
  {
      if (   pieceCount[pc] != popcount(pieces(color_of(pc), type_of(pc)))
          || pieceCount[pc] != std::count(board, board + SQUARE_NB, pc))
          assert(0 && "pos_is_ok: Pieces");
      for (int i = 0; i < pieceCount[pc]; ++i)
          if (board[pieceList[pc][i]] != pc || index[pieceList[pc][i]] != i)
              assert(0 && "pos_is_ok: Index");
  }
  for (Color c : { WHITE, BLACK })
      for (CastlingRights cr : {c & KING_SIDE, c & QUEEN_SIDE})
      {
@@ -1395,3 +1366,19 @@ bool Position::pos_is_ok() const {
  return true;
 }
 // Add a function that directly unpacks for speed. It's pretty tough.
 // Write it by combining packer::unpack() and Position::set().
 // If there is a problem with the passed phase and there is an error, non-zero is returned.
 int Position::set_from_packed_sfen(const Tools::PackedSfen& sfen , StateInfo* si, Thread* th, bool frc)
 {
  return Tools::set_from_packed_sfen(*this, sfen, si, th, frc);
 }
 // Get the packed sfen. Returns to the buffer specified in the argument.
 void Position::sfen_pack(Tools::PackedSfen& sfen, bool resetCastlingRights)
 {
  sfen = Tools::sfen_pack(*this, resetCastlingRights);
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -26,10 +26,15 @@
 #include "bitboard.h"
 #include "evaluate.h"
 #include "psqt.h"
 #include "types.h"
 #include "nnue/nnue_accumulator.h"
 #include "tools/packed_sfen.h"
 #include "tools/sfen_packer.h"
 namespace Stockfish {
 /// StateInfo struct stores information needed to restore a Position object to
 /// its previous state when we retract a move. Whenever a move is made on the
@@ -49,11 +54,11 @@ struct StateInfo {
  // Not copied when making a move (will be recomputed anyhow)
  Key        key;
  Bitboard   checkersBB;
  Piece      capturedPiece;
  StateInfo* previous;
  Bitboard   blockersForKing[COLOR_NB];
  Bitboard   pinners[COLOR_NB];
  Bitboard   checkSquares[PIECE_TYPE_NB];
  Piece      capturedPiece;
  int        repetition;
  // Used by NNUE
@@ -86,7 +91,7 @@ public:
  // FEN string input/output
  Position& set(const std::string& fenStr, bool isChess960, StateInfo* si, Thread* th);
  Position& set(const std::string& code, Color c, StateInfo* si);
-  const std::string fen() const;
+  std::string fen() const;
  // Position representation
  Bitboard pieces(PieceType pt) const;
@@ -99,7 +104,6 @@ public:
  bool empty(Square s) const;
  template<PieceType Pt> int count(Color c) const;
  template<PieceType Pt> int count() const;
  template<PieceType Pt> const Square* squares(Color c) const;
  template<PieceType Pt> Square square(Color c) const;
  bool is_on_semiopen_file(Color c, Square s) const;
@@ -113,12 +117,13 @@ public:
  Bitboard checkers() const;
  Bitboard blockers_for_king(Color c) const;
  Bitboard check_squares(PieceType pt) const;
-  bool is_discovery_check_on_king(Color c, Move m) const;
+  Bitboard pinners(Color c) const;
  // Attacks to/from a given square
  Bitboard attackers_to(Square s) const;
  Bitboard attackers_to(Square s, Bitboard occupied) const;
  Bitboard slider_blockers(Bitboard sliders, Square s, Bitboard& pinners) const;
  template<PieceType Pt> Bitboard attacks_by(Color c) const;
  // Properties of moves
  bool legal(Move m) const;
@@ -126,7 +131,6 @@ public:
  bool capture(Move m) const;
  bool capture_or_promotion(Move m) const;
  bool gives_check(Move m) const;
  bool advanced_pawn_push(Move m) const;
  Piece moved_piece(Move m) const;
  Piece captured_piece() const;
@@ -157,6 +161,8 @@ public:
  bool is_chess960() const;
  Thread* this_thread() const;
  bool is_draw(int ply) const;
  bool is_fifty_move_draw() const;
  bool is_three_fold_repetition() const;
  bool has_game_cycle(int ply) const;
  bool has_repeated() const;
  int rule50_count() const;
@@ -170,7 +176,31 @@ public:
  // Used by NNUE
  StateInfo* state() const;
-  const EvalList* eval_list() const;
+
  // --sfenization helper
  friend int Tools::set_from_packed_sfen(Position& pos, const Tools::PackedSfen& sfen, StateInfo* si, Thread* th, bool frc);
  // Get the packed sfen. Returns to the buffer specified in the argument.
  // Do not include gamePly in pack.
  void sfen_pack(Tools::PackedSfen& sfen, bool resetCastlingRights);
  // It is slow to go through sfen, so I made a function to set packed sfen directly.
  // Equivalent to pos.set(sfen_unpack(data),si,th);.
  // If there is a problem with the passed phase and there is an error, non-zero is returned.
  // PackedSfen does not include gamePly so it cannot be restored. If you want to set it, specify it with an argument.
  int set_from_packed_sfen(const Tools::PackedSfen& sfen, StateInfo* si, Thread* th, bool frc);
  void clear() { std::memset(this, 0, sizeof(Position)); }
  // Give the board, hand piece, and turn, and return the sfen.
  //static std::string sfen_from_rawdata(Piece board[81], Hand hands[2], Color turn, int gamePly);
  // Returns the position of the ball on the c side.
  Square king_square(Color c) const { return lsb(pieces(c, KING)); }
  void put_piece(Piece pc, Square s);
  void remove_piece(Square s);
 private:
  // Initialization helpers (used while setting up a position)
@@ -179,40 +209,26 @@ private:
  void set_check_info(StateInfo* si) const;
  // Other helpers
  void put_piece(Piece pc, Square s);
  void remove_piece(Square s);
  void move_piece(Square from, Square to);
  template<bool Do>
  void do_castling(Color us, Square from, Square& to, Square& rfrom, Square& rto);
  // ID of a piece on a given square
  PieceId piece_id_on(Square sq) const;
  // Data members
  Piece board[SQUARE_NB];
  Bitboard byTypeBB[PIECE_TYPE_NB];
  Bitboard byColorBB[COLOR_NB];
  int pieceCount[PIECE_NB];
  Square pieceList[PIECE_NB][16];
  int index[SQUARE_NB];
  int castlingRightsMask[SQUARE_NB];
  Square castlingRookSquare[CASTLING_RIGHT_NB];
  Bitboard castlingPath[CASTLING_RIGHT_NB];
  Thread* thisThread;
  StateInfo* st;
  int gamePly;
  Color sideToMove;
  Score psq;
  Thread* thisThread;
  StateInfo* st;
  bool chess960;
  // List of pieces used in NNUE evaluation function
  EvalList evalList;
 };
 namespace PSQT {
  extern Score psq[PIECE_NB][SQUARE_NB];
 }
 extern std::ostream& operator<<(std::ostream& os, const Position& pos);
 inline Color Position::side_to_move() const {
@@ -260,13 +276,9 @@ template<PieceType Pt> inline int Position::count() const {
  return count<Pt>(WHITE) + count<Pt>(BLACK);
 }
 template<PieceType Pt> inline const Square* Position::squares(Color c) const {
  return pieceList[make_piece(c, Pt)];
 }
 template<PieceType Pt> inline Square Position::square(Color c) const {
-  assert(pieceCount[make_piece(c, Pt)] == 1);
+  assert(count<Pt>(c) == 1);
-  return squares<Pt>(c)[0];
+  return lsb(pieces(c, Pt));
 }
 inline Square Position::ep_square() const {
@@ -301,6 +313,22 @@ inline Bitboard Position::attackers_to(Square s) const {
  return attackers_to(s, pieces());
 }
 template<PieceType Pt>
 inline Bitboard Position::attacks_by(Color c) const {
  if constexpr (Pt == PAWN)
      return c == WHITE ? pawn_attacks_bb<WHITE>(pieces(WHITE, PAWN))
                        : pawn_attacks_bb<BLACK>(pieces(BLACK, PAWN));
  else
  {
      Bitboard threats = 0;
      Bitboard attackers = pieces(c, Pt);
      while (attackers)
          threats |= attacks_bb<Pt>(pop_lsb(attackers), pieces());
      return threats;
  }
 }
 inline Bitboard Position::checkers() const {
  return st->checkersBB;
 }
@@ -309,29 +337,25 @@ inline Bitboard Position::blockers_for_king(Color c) const {
  return st->blockersForKing[c];
 }
 inline Bitboard Position::pinners(Color c) const {
  return st->pinners[c];
 }
 inline Bitboard Position::check_squares(PieceType pt) const {
  return st->checkSquares[pt];
 }
 inline bool Position::is_discovery_check_on_king(Color c, Move m) const {
  return st->blockersForKing[c] & from_sq(m);
 }
 inline bool Position::pawn_passed(Color c, Square s) const {
  return !(pieces(~c, PAWN) & passed_pawn_span(c, s));
 }
 inline bool Position::advanced_pawn_push(Move m) const {
  return   type_of(moved_piece(m)) == PAWN
        && relative_rank(sideToMove, to_sq(m)) > RANK_5;
 }
 inline int Position::pawns_on_same_color_squares(Color c, Square s) const {
  return popcount(pieces(c, PAWN) & ((DarkSquares & s) ? DarkSquares : ~DarkSquares));
 }
 inline Key Position::key() const {
-  return st->key;
+  return st->rule50 < 14 ? st->key
                         : st->key ^ make_key((st->rule50 - 14) / 8);
 }
 inline Key Position::pawn_key() const {
@@ -380,7 +404,7 @@ inline bool Position::capture_or_promotion(Move m) const {
 inline bool Position::capture(Move m) const {
  assert(is_ok(m));
  // Castling is encoded as "king captures rook"
-  return (!empty(to_sq(m)) && type_of(m) != CASTLING) || type_of(m) == ENPASSANT;
+  return (!empty(to_sq(m)) && type_of(m) != CASTLING) || type_of(m) == EN_PASSANT;
 }
 inline Piece Position::captured_piece() const {
@@ -396,35 +420,25 @@ inline void Position::put_piece(Piece pc, Square s) {
  board[s] = pc;
  byTypeBB[ALL_PIECES] |= byTypeBB[type_of(pc)] |= s;
  byColorBB[color_of(pc)] |= s;
-  index[s] = pieceCount[pc]++;
+  pieceCount[pc]++;
  pieceList[pc][index[s]] = s;
  pieceCount[make_piece(color_of(pc), ALL_PIECES)]++;
  psq += PSQT::psq[pc][s];
 }
 inline void Position::remove_piece(Square s) {
  // WARNING: This is not a reversible operation. If we remove a piece in
  // do_move() and then replace it in undo_move() we will put it at the end of
  // the list and not in its original place, it means index[] and pieceList[]
  // are not invariant to a do_move() + undo_move() sequence.
  Piece pc = board[s];
  byTypeBB[ALL_PIECES] ^= s;
  byTypeBB[type_of(pc)] ^= s;
  byColorBB[color_of(pc)] ^= s;
-  /* board[s] = NO_PIECE;  Not needed, overwritten by the capturing one */
+  board[s] = NO_PIECE;
-  Square lastSquare = pieceList[pc][--pieceCount[pc]];
+  pieceCount[pc]--;
  index[lastSquare] = index[s];
  pieceList[pc][index[lastSquare]] = lastSquare;
  pieceList[pc][pieceCount[pc]] = SQ_NONE;
  pieceCount[make_piece(color_of(pc), ALL_PIECES)]--;
  psq -= PSQT::psq[pc][s];
 }
 inline void Position::move_piece(Square from, Square to) {
  // index[from] is not updated and becomes stale. This works as long as index[]
  // is accessed just by known occupied squares.
  Piece pc = board[from];
  Bitboard fromTo = from | to;
  byTypeBB[ALL_PIECES] ^= fromTo;
@@ -432,8 +446,6 @@ inline void Position::move_piece(Square from, Square to) {
  byColorBB[color_of(pc)] ^= fromTo;
  board[from] = NO_PIECE;
  board[to] = pc;
  index[to] = index[from];
  pieceList[pc][index[to]] = to;
  psq += PSQT::psq[pc][to] - PSQT::psq[pc][from];
 }
@@ -446,20 +458,8 @@ inline StateInfo* Position::state() const {
  return st;
 }
-inline const EvalList* Position::eval_list() const {
+static const char* const StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
-  return &evalList;
+} // namespace Stockfish
 }
 inline PieceId Position::piece_id_on(Square sq) const
 {
  assert(piece_on(sq) != NO_PIECE);
  PieceId pid = evalList.piece_id_list[sq];
  assert(is_ok(pid));
  return pid;
 }
 #endif // #ifndef POSITION_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -16,19 +16,23 @@
  along with this program.  If not, see <http://www.gnu.org/licenses/>.
 */
 #include "psqt.h"
 #include <algorithm>
 #include "types.h"
 #include "bitboard.h"
 #include "types.h"
-namespace PSQT {
+namespace Stockfish {
-#define S(mg, eg) make_score(mg, eg)
+namespace
 {
-// Bonus[PieceType][Square / 2] contains Piece-Square scores. For each piece
+auto constexpr S = make_score;
-// type on a given square a (middlegame, endgame) score pair is assigned. Table
+
-// is defined for files A..D and white side: it is symmetric for black side and
+// 'Bonus' contains Piece-Square parameters.
-// second half of the files.
+// Scores are explicit for files A to D, implicitly mirrored for E to H.
 constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = {
  { },
  { },
@@ -43,14 +47,14 @@ constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = {
   { S(-201,-100), S(-83,-88), S(-56,-56), S(-26,-17) }
  },
  { // Bishop
-   { S(-53,-57), S( -5,-30), S( -8,-37), S(-23,-12) },
+   { S(-37,-40), S(-4 ,-21), S( -6,-26), S(-16, -8) },
-   { S(-15,-37), S(  8,-13), S( 19,-17), S(  4,  1) },
+   { S(-11,-26), S(  6, -9), S( 13,-12), S(  3,  1) },
-   { S( -7,-16), S( 21, -1), S( -5, -2), S( 17, 10) },
+   { S(-5 ,-11), S( 15, -1), S( -4, -1), S( 12,  7) },
-   { S( -5,-20), S( 11, -6), S( 25,  0), S( 39, 17) },
+   { S(-4 ,-14), S(  8, -4), S( 18,  0), S( 27, 12) },
-   { S(-12,-17), S( 29, -1), S( 22,-14), S( 31, 15) },
+   { S(-8 ,-12), S( 20, -1), S( 15,-10), S( 22, 11) },
-   { S(-16,-30), S(  6,  6), S(  1,  4), S( 11,  6) },
+   { S(-11,-21), S(  4,  4), S(  1,  3), S(  8,  4) },
-   { S(-17,-31), S(-14,-20), S(  5, -1), S(  0,  1) },
+   { S(-12,-22), S(-10,-14), S(  4, -1), S(  0,  1) },
-   { S(-48,-46), S(  1,-42), S(-14,-37), S(-23,-24) }
+   { S(-34,-32), S(  1,-29), S(-10,-26), S(-16,-17) }
  },
  { // Rook
   { S(-31, -9), S(-20,-13), S(-14,-10), S(-5, -9) },
@@ -64,13 +68,13 @@ constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = {
  },
  { // Queen
   { S( 3,-69), S(-5,-57), S(-5,-47), S( 4,-26) },
-   { S(-3,-55), S( 5,-31), S( 8,-22), S(12, -4) },
+   { S(-3,-54), S( 5,-31), S( 8,-22), S(12, -4) },
   { S(-3,-39), S( 6,-18), S(13, -9), S( 7,  3) },
   { S( 4,-23), S( 5, -3), S( 9, 13), S( 8, 24) },
   { S( 0,-29), S(14, -6), S(12,  9), S( 5, 21) },
-   { S(-4,-38), S(10,-18), S( 6,-12), S( 8,  1) },
+   { S(-4,-38), S(10,-18), S( 6,-11), S( 8,  1) },
   { S(-5,-50), S( 6,-27), S(10,-24), S( 8, -8) },
-   { S(-2,-75), S(-2,-52), S( 1,-43), S(-2,-36) }
+   { S(-2,-74), S(-2,-52), S( 1,-43), S(-2,-34) }
  },
  { // King
   { S(271,  1), S(327, 45), S(271, 85), S(198, 76) },
@@ -87,19 +91,22 @@ constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = {
 constexpr Score PBonus[RANK_NB][FILE_NB] =
  { // Pawn (asymmetric distribution)
   { },
-   { S(  3,-10), S(  3, -6), S( 10, 10), S( 19,  0), S( 16, 14), S( 19,  7), S(  7, -5), S( -5,-19) },
+   { S(  2, -8), S(  4, -6), S( 11,  9), S( 18,  5), S( 16, 16), S( 21,  6), S(  9, -6), S( -3,-18) },
-   { S( -9,-10), S(-15,-10), S( 11,-10), S( 15,  4), S( 32,  4), S( 22,  3), S(  5, -6), S(-22, -4) },
+   { S( -9, -9), S(-15, -7), S( 11,-10), S( 15,  5), S( 31,  2), S( 23,  3), S(  6, -8), S(-20, -5) },
-   { S( -4,  6), S(-23, -2), S(  6, -8), S( 20, -4), S( 40,-13), S( 17,-12), S(  4,-10), S( -8, -9) },
+   { S( -3,  7), S(-20,  1), S(  8, -8), S( 19, -2), S( 39,-14), S( 17,-13), S(  2,-11), S( -5, -6) },
-   { S( 13, 10), S(  0,  5), S(-13,  4), S(  1, -5), S( 11, -5), S( -2, -5), S(-13, 14), S(  5,  9) },
+   { S( 11, 12), S( -4,  6), S(-11,  2), S(  2, -6), S( 11, -5), S(  0, -4), S(-12, 14), S(  5,  9) },
-   { S(  5, 28), S(-12, 20), S( -7, 21), S( 22, 28), S( -8, 30), S( -5,  7), S(-15,  6), S( -8, 13) },
+   { S(  3, 27), S(-11, 18), S( -6, 19), S( 22, 29), S( -8, 30), S( -5,  9), S(-14,  8), S(-11, 14) },
-   { S( -7,  0), S(  7,-11), S( -3, 12), S(-13, 21), S(  5, 25), S(-16, 19), S( 10,  4), S( -8,  7) }
+   { S( -7, -1), S(  6,-14), S( -2, 13), S(-11, 22), S(  4, 24), S(-14, 17), S( 10,  7), S( -9,  7) }
  };
-#undef S
+} // namespace
 namespace PSQT
 {
 Score psq[PIECE_NB][SQUARE_NB];
 // PSQT::init() initializes piece-square tables: the white halves of the tables are
 // copied from Bonus[] and PBonus[], adding the piece value, then the black halves of
 // the tables are initialized by flipping and changing the sign of the white scores.
@@ -107,16 +114,18 @@ void init() {
  for (Piece pc : {W_PAWN, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING})
  {
-      Score score = make_score(PieceValue[MG][pc], PieceValue[EG][pc]);
+    Score score = make_score(PieceValue[MG][pc], PieceValue[EG][pc]);
-      for (Square s = SQ_A1; s <= SQ_H8; ++s)
+    for (Square s = SQ_A1; s <= SQ_H8; ++s)
-      {
+    {
-          File f = File(edge_distance(file_of(s)));
+      File f = File(edge_distance(file_of(s)));
-          psq[ pc][s] = score + (type_of(pc) == PAWN ? PBonus[rank_of(s)][file_of(s)]
+      psq[ pc][s] = score + (type_of(pc) == PAWN ? PBonus[rank_of(s)][file_of(s)]
-                                                     : Bonus[pc][rank_of(s)][f]);
+                                                 : Bonus[pc][rank_of(s)][f]);
-          psq[~pc][flip_rank(s)] = -psq[pc][s];
+      psq[~pc][flip_rank(s)] = -psq[pc][s];
-      }
+    }
  }
 }
 } // namespace PSQT
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -16,30 +16,23 @@
  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
+#ifndef PSQT_H_INCLUDED
-#define NNUE_FEATURES_COMMON_H_INCLUDED
+#define PSQT_H_INCLUDED
 #include "../../evaluate.h"
 #include "../nnue_common.h"
-namespace Eval::NNUE::Features {
+#include "types.h"
  class IndexList;
-  template <typename... FeatureTypes>
+namespace Stockfish::PSQT
-  class FeatureSet;
+{
-  // Trigger to perform full calculations instead of difference only
+extern Score psq[PIECE_NB][SQUARE_NB];
  enum class TriggerEvent {
    kFriendKingMoved // calculate full evaluation when own king moves
  };
-  enum class Side {
+// Fill psqt array from a set of internally linked parameters
-    kFriend // side to move
+extern void init();
  };
-}  // namespace Eval::NNUE::Features
+} // namespace Stockfish::PSQT
-#endif // #ifndef NNUE_FEATURES_COMMON_H_INCLUDED
+
 #endif // PSQT_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,6 +24,9 @@
 #include "misc.h"
 #include "movepick.h"
 #include "types.h"
 #include "uci.h"
 namespace Stockfish {
 class Position;
@@ -32,6 +35,7 @@ namespace Search {
 /// Threshold used for countermoves based pruning
 constexpr int CounterMovePruneThreshold = 0;
 extern bool prune_at_shallow_depth;
 /// Stack struct keeps track of the information we need to remember from nodes
 /// shallower and deeper in the tree during the search. Each search thread has
@@ -48,6 +52,10 @@ struct Stack {
  int statScore;
  int moveCount;
  bool inCheck;
  bool ttPv;
  bool ttHit;
  int doubleExtensions;
  int cutoffCnt;
 };
@@ -67,9 +75,9 @@ struct RootMove {
  Value score = -VALUE_INFINITE;
  Value previousScore = -VALUE_INFINITE;
  Value averageScore = -VALUE_INFINITE;
  int selDepth = 0;
  int tbRank = 0;
  int bestMoveCount = 0;
  Value tbScore;
  std::vector<Move> pv;
 };
@@ -86,6 +94,7 @@ struct LimitsType {
    time[WHITE] = time[BLACK] = inc[WHITE] = inc[BLACK] = npmsec = movetime = TimePoint(0);
    movestogo = depth = mate = perft = infinite = 0;
    nodes = 0;
    silent = false;
  }
  bool use_time_management() const {
@@ -96,6 +105,9 @@ struct LimitsType {
  TimePoint time[COLOR_NB], inc[COLOR_NB], npmsec, movetime, startTime;
  int movestogo, depth, mate, perft, infinite;
  int64_t nodes;
  // Silent mode that does not output to the screen (for continuous self-play in process)
  // Do not output PV at this time.
  bool silent;
 };
 extern LimitsType Limits;
@@ -103,6 +115,36 @@ extern LimitsType Limits;
 void init();
 void clear();
-} // namespace Search
+// A pair of reader and evaluation value. Returned by Tools::search(),Tools::qsearch().
 using ValueAndPV = std::pair<Value, std::vector<Move>>;
 ValueAndPV qsearch(Position& pos);
 ValueAndPV search(Position& pos, int depth_, size_t multiPV = 1, uint64_t nodesLimit = 0);
 namespace MCTS {
  struct MctsContinuation {
    std::uint64_t numVisits;
    Value value;
    float actionValue;
    std::vector<Move> pv;
  };
  ValueAndPV search_mcts(
    Position& pos,
    std::uint64_t nodes,
    Depth leafDepth,
    float explorationFactor);
  std::vector<MctsContinuation> search_mcts_multipv(
    Position& pos,
    std::uint64_t numPlayouts,
    Depth leafDepth,
    float explorationFactor);
 }
 }
 } // namespace Stockfish
 #endif // #ifndef SEARCH_H_INCLUDED
@@ -0,0 +1,387 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
  Copyright (C) 2004-2022 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/>.
 */
 #ifndef STOCKFISH_SIMD_H_INCLUDED
 #define STOCKFISH_SIMD_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_MMX)
 # include <mmintrin.h>
 #elif defined(USE_NEON)
 # include <arm_neon.h>
 #endif
 // The inline asm is only safe for GCC, where it is necessary to get good codegen.
 // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101693
 // Clang does fine without it.
 // Play around here: https://godbolt.org/z/7EWqrYq51
 #if (defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER))
 #define USE_INLINE_ASM
 #endif
 // Use either the AVX512 or AVX-VNNI version of the VNNI instructions.
 #if defined(USE_AVXVNNI)
 #define VNNI_PREFIX "%{vex%} "
 #else
 #define VNNI_PREFIX ""
 #endif
 namespace Stockfish::Simd {
 #if defined (USE_AVX512)
    [[maybe_unused]] static int m512_hadd(__m512i sum, int bias) {
      return _mm512_reduce_add_epi32(sum) + bias;
    }
    /*
      Parameters:
        sum0 = [zmm0.i128[0], zmm0.i128[1], zmm0.i128[2], zmm0.i128[3]]
        sum1 = [zmm1.i128[0], zmm1.i128[1], zmm1.i128[2], zmm1.i128[3]]
        sum2 = [zmm2.i128[0], zmm2.i128[1], zmm2.i128[2], zmm2.i128[3]]
        sum3 = [zmm3.i128[0], zmm3.i128[1], zmm3.i128[2], zmm3.i128[3]]
      Returns:
        ret = [
          reduce_add_epi32(zmm0.i128[0]), reduce_add_epi32(zmm1.i128[0]), reduce_add_epi32(zmm2.i128[0]), reduce_add_epi32(zmm3.i128[0]),
          reduce_add_epi32(zmm0.i128[1]), reduce_add_epi32(zmm1.i128[1]), reduce_add_epi32(zmm2.i128[1]), reduce_add_epi32(zmm3.i128[1]),
          reduce_add_epi32(zmm0.i128[2]), reduce_add_epi32(zmm1.i128[2]), reduce_add_epi32(zmm2.i128[2]), reduce_add_epi32(zmm3.i128[2]),
          reduce_add_epi32(zmm0.i128[3]), reduce_add_epi32(zmm1.i128[3]), reduce_add_epi32(zmm2.i128[3]), reduce_add_epi32(zmm3.i128[3])
        ]
    */
    [[maybe_unused]] static __m512i m512_hadd128x16_interleave(
        __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) {
      __m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
      __m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
      __m512i sum23a = _mm512_unpacklo_epi32(sum2, sum3);
      __m512i sum23b = _mm512_unpackhi_epi32(sum2, sum3);
      __m512i sum01 = _mm512_add_epi32(sum01a, sum01b);
      __m512i sum23 = _mm512_add_epi32(sum23a, sum23b);
      __m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
      __m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
      return _mm512_add_epi32(sum0123a, sum0123b);
    }
    [[maybe_unused]] static __m128i m512_haddx4(
        __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
        __m128i bias) {
      __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
      __m256i sum256lo = _mm512_castsi512_si256(sum);
      __m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
      sum256lo = _mm256_add_epi32(sum256lo, sum256hi);
      __m128i sum128lo = _mm256_castsi256_si128(sum256lo);
      __m128i sum128hi = _mm256_extracti128_si256(sum256lo, 1);
      return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
    }
    [[maybe_unused]] static void m512_add_dpbusd_epi32(
        __m512i& acc,
        __m512i a,
        __m512i b) {
 # if defined (USE_VNNI)
 #   if defined (USE_INLINE_ASM)
      asm(
        "vpdpbusd %[b], %[a], %[acc]\n\t"
        : [acc]"+v"(acc)
        : [a]"v"(a), [b]"vm"(b)
      );
 #   else
      acc = _mm512_dpbusd_epi32(acc, a, b);
 #   endif
 # else
 #   if defined (USE_INLINE_ASM)
      __m512i tmp = _mm512_maddubs_epi16(a, b);
      asm(
          "vpmaddwd    %[tmp], %[ones], %[tmp]\n\t"
          "vpaddd      %[acc], %[tmp], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp]"+&v"(tmp)
          : [ones]"v"(_mm512_set1_epi16(1))
      );
 #   else
      __m512i product0 = _mm512_maddubs_epi16(a, b);
      product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
      acc = _mm512_add_epi32(acc, product0);
 #   endif
 # endif
    }
    [[maybe_unused]] static void m512_add_dpbusd_epi32x2(
        __m512i& acc,
        __m512i a0, __m512i b0,
        __m512i a1, __m512i b1) {
 # if defined (USE_VNNI)
 #   if defined (USE_INLINE_ASM)
      asm(
        "vpdpbusd %[b0], %[a0], %[acc]\n\t"
        "vpdpbusd %[b1], %[a1], %[acc]\n\t"
        : [acc]"+v"(acc)
        : [a0]"v"(a0), [b0]"vm"(b0), [a1]"v"(a1), [b1]"vm"(b1)
      );
 #   else
      acc = _mm512_dpbusd_epi32(acc, a0, b0);
      acc = _mm512_dpbusd_epi32(acc, a1, b1);
 #   endif
 # else
 #   if defined (USE_INLINE_ASM)
      __m512i tmp0 = _mm512_maddubs_epi16(a0, b0);
      __m512i tmp1 = _mm512_maddubs_epi16(a1, b1);
      asm(
          "vpaddsw     %[tmp0], %[tmp1], %[tmp0]\n\t"
          "vpmaddwd    %[tmp0], %[ones], %[tmp0]\n\t"
          "vpaddd      %[acc], %[tmp0], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp0]"+&v"(tmp0)
          : [tmp1]"v"(tmp1), [ones]"v"(_mm512_set1_epi16(1))
      );
 #   else
      __m512i product0 = _mm512_maddubs_epi16(a0, b0);
      __m512i product1 = _mm512_maddubs_epi16(a1, b1);
      product0 = _mm512_adds_epi16(product0, product1);
      product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
      acc = _mm512_add_epi32(acc, product0);
 #   endif
 # endif
    }
 #endif
 #if defined (USE_AVX2)
    [[maybe_unused]] static int m256_hadd(__m256i sum, int bias) {
      __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
      sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
      sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
      return _mm_cvtsi128_si32(sum128) + bias;
    }
    [[maybe_unused]] static __m128i m256_haddx4(
        __m256i sum0, __m256i sum1, __m256i sum2, __m256i sum3,
        __m128i bias) {
      sum0 = _mm256_hadd_epi32(sum0, sum1);
      sum2 = _mm256_hadd_epi32(sum2, sum3);
      sum0 = _mm256_hadd_epi32(sum0, sum2);
      __m128i sum128lo = _mm256_castsi256_si128(sum0);
      __m128i sum128hi = _mm256_extracti128_si256(sum0, 1);
      return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
    }
    [[maybe_unused]] static void m256_add_dpbusd_epi32(
        __m256i& acc,
        __m256i a,
        __m256i b) {
 # if defined (USE_VNNI)
 #   if defined (USE_INLINE_ASM)
      asm(
        VNNI_PREFIX "vpdpbusd %[b], %[a], %[acc]\n\t"
        : [acc]"+v"(acc)
        : [a]"v"(a), [b]"vm"(b)
      );
 #   else
      acc = _mm256_dpbusd_epi32(acc, a, b);
 #   endif
 # else
 #   if defined (USE_INLINE_ASM)
      __m256i tmp = _mm256_maddubs_epi16(a, b);
      asm(
          "vpmaddwd    %[tmp], %[ones], %[tmp]\n\t"
          "vpaddd      %[acc], %[tmp], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp]"+&v"(tmp)
          : [ones]"v"(_mm256_set1_epi16(1))
      );
 #   else
      __m256i product0 = _mm256_maddubs_epi16(a, b);
      product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
      acc = _mm256_add_epi32(acc, product0);
 #   endif
 # endif
    }
    [[maybe_unused]] static void m256_add_dpbusd_epi32x2(
        __m256i& acc,
        __m256i a0, __m256i b0,
        __m256i a1, __m256i b1) {
 # if defined (USE_VNNI)
 #   if defined (USE_INLINE_ASM)
      asm(
        VNNI_PREFIX "vpdpbusd %[b0], %[a0], %[acc]\n\t"
        VNNI_PREFIX "vpdpbusd %[b1], %[a1], %[acc]\n\t"
        : [acc]"+v"(acc)
        : [a0]"v"(a0), [b0]"vm"(b0), [a1]"v"(a1), [b1]"vm"(b1)
      );
 #   else
      acc = _mm256_dpbusd_epi32(acc, a0, b0);
      acc = _mm256_dpbusd_epi32(acc, a1, b1);
 #   endif
 # else
 #   if defined (USE_INLINE_ASM)
      __m256i tmp0 = _mm256_maddubs_epi16(a0, b0);
      __m256i tmp1 = _mm256_maddubs_epi16(a1, b1);
      asm(
          "vpaddsw     %[tmp0], %[tmp1], %[tmp0]\n\t"
          "vpmaddwd    %[tmp0], %[ones], %[tmp0]\n\t"
          "vpaddd      %[acc], %[tmp0], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp0]"+&v"(tmp0)
          : [tmp1]"v"(tmp1), [ones]"v"(_mm256_set1_epi16(1))
      );
 #   else
      __m256i product0 = _mm256_maddubs_epi16(a0, b0);
      __m256i product1 = _mm256_maddubs_epi16(a1, b1);
      product0 = _mm256_adds_epi16(product0, product1);
      product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
      acc = _mm256_add_epi32(acc, product0);
 #   endif
 # endif
    }
 #endif
 #if defined (USE_SSSE3)
    [[maybe_unused]] static int m128_hadd(__m128i sum, int bias) {
      sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
      sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
      return _mm_cvtsi128_si32(sum) + bias;
    }
    [[maybe_unused]] static __m128i m128_haddx4(
        __m128i sum0, __m128i sum1, __m128i sum2, __m128i sum3,
        __m128i bias) {
      sum0 = _mm_hadd_epi32(sum0, sum1);
      sum2 = _mm_hadd_epi32(sum2, sum3);
      sum0 = _mm_hadd_epi32(sum0, sum2);
      return _mm_add_epi32(sum0, bias);
    }
    [[maybe_unused]] static void m128_add_dpbusd_epi32(
        __m128i& acc,
        __m128i a,
        __m128i b) {
 #   if defined (USE_INLINE_ASM)
      __m128i tmp = _mm_maddubs_epi16(a, b);
      asm(
          "pmaddwd    %[ones], %[tmp]\n\t"
          "paddd      %[tmp], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp]"+&v"(tmp)
          : [ones]"v"(_mm_set1_epi16(1))
      );
 #   else
      __m128i product0 = _mm_maddubs_epi16(a, b);
      product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
      acc = _mm_add_epi32(acc, product0);
 #   endif
    }
    [[maybe_unused]] static void m128_add_dpbusd_epi32x2(
        __m128i& acc,
        __m128i a0, __m128i b0,
        __m128i a1, __m128i b1) {
 #   if defined (USE_INLINE_ASM)
      __m128i tmp0 = _mm_maddubs_epi16(a0, b0);
      __m128i tmp1 = _mm_maddubs_epi16(a1, b1);
      asm(
          "paddsw     %[tmp1], %[tmp0]\n\t"
          "pmaddwd    %[ones], %[tmp0]\n\t"
          "paddd      %[tmp0], %[acc]\n\t"
          : [acc]"+v"(acc), [tmp0]"+&v"(tmp0)
          : [tmp1]"v"(tmp1), [ones]"v"(_mm_set1_epi16(1))
      );
 #   else
      __m128i product0 = _mm_maddubs_epi16(a0, b0);
      __m128i product1 = _mm_maddubs_epi16(a1, b1);
      product0 = _mm_adds_epi16(product0, product1);
      product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
      acc = _mm_add_epi32(acc, product0);
 #   endif
    }
 #endif
 #if defined (USE_NEON)
    [[maybe_unused]] static int neon_m128_reduce_add_epi32(int32x4_t s) {
 #   if USE_NEON >= 8
      return vaddvq_s32(s);
 #   else
      return s[0] + s[1] + s[2] + s[3];
 #   endif
    }
    [[maybe_unused]] static int neon_m128_hadd(int32x4_t sum, int bias) {
      return neon_m128_reduce_add_epi32(sum) + bias;
    }
    [[maybe_unused]] static int32x4_t neon_m128_haddx4(
        int32x4_t sum0, int32x4_t sum1, int32x4_t sum2, int32x4_t sum3,
        int32x4_t bias) {
      int32x4_t hsums {
        neon_m128_reduce_add_epi32(sum0),
        neon_m128_reduce_add_epi32(sum1),
        neon_m128_reduce_add_epi32(sum2),
        neon_m128_reduce_add_epi32(sum3)
      };
      return vaddq_s32(hsums, bias);
    }
    [[maybe_unused]] static void neon_m128_add_dpbusd_epi32x2(
        int32x4_t& acc,
        int8x8_t a0, int8x8_t b0,
        int8x8_t a1, int8x8_t b1) {
      int16x8_t product = vmull_s8(a0, b0);
      product = vmlal_s8(product, a1, b1);
      acc = vpadalq_s16(acc, product);
    }
 #endif
 }
 #endif // STOCKFISH_SIMD_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -50,9 +50,11 @@
 #include <windows.h>
 #endif
-using namespace Tablebases;
+using namespace Stockfish::Tablebases;
-int Tablebases::MaxCardinality;
+int Stockfish::Tablebases::MaxCardinality = 0;
 namespace Stockfish {
 namespace {
@@ -103,9 +105,6 @@ template<> inline void swap_endian<uint8_t>(uint8_t&) {}
 template<typename T, int LE> T number(void* addr)
 {
    static const union { uint32_t i; char c[4]; } Le = { 0x01020304 };
    static const bool IsLittleEndian = (Le.c[0] == 4);
    T v;
    if ((uintptr_t)addr & (alignof(T) - 1)) // Unaligned pointer (very rare)
@@ -190,7 +189,8 @@ public:
        std::stringstream ss(Paths);
        std::string path;
-        while (std::getline(ss, path, SepChar)) {
+        while (std::getline(ss, path, SepChar))
        {
            fname = path + "/" + f;
            std::ifstream::open(fname);
            if (is_open())
@@ -223,7 +223,9 @@ public:
        *mapping = statbuf.st_size;
        *baseAddress = mmap(nullptr, statbuf.st_size, PROT_READ, MAP_SHARED, fd, 0);
 #if defined(MADV_RANDOM)
        madvise(*baseAddress, statbuf.st_size, MADV_RANDOM);
 #endif
        ::close(fd);
        if (*baseAddress == MAP_FAILED)
@@ -563,7 +565,8 @@ int decompress_pairs(PairsData* d, uint64_t idx) {
    int buf64Size = 64;
    Sym sym;
-    while (true) {
+    while (true)
    {
        int len = 0; // This is the symbol length - d->min_sym_len
        // Now get the symbol length. For any symbol s64 of length l right-padded
@@ -601,8 +604,8 @@ int decompress_pairs(PairsData* d, uint64_t idx) {
    // We binary-search for our value recursively expanding into the left and
    // right child symbols until we reach a leaf node where symlen[sym] + 1 == 1
    // that will store the value we need.
-    while (d->symlen[sym]) {
+    while (d->symlen[sym])
-
+    {
        Sym left = d->btree[sym].get<LR::Left>();
        // If a symbol contains 36 sub-symbols (d->symlen[sym] + 1 = 36) and
@@ -707,7 +710,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
        leadPawns = b = pos.pieces(color_of(pc), PAWN);
        do
-            squares[size++] = pop_lsb(&b) ^ flipSquares;
+            squares[size++] = pop_lsb(b) ^ flipSquares;
        while (b);
        leadPawnsCnt = size;
@@ -727,7 +730,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
    // directly map them to the correct color and square.
    b = pos.pieces() ^ leadPawns;
    do {
-        Square s = pop_lsb(&b);
+        Square s = pop_lsb(b);
        squares[size] = s ^ flipSquares;
        pieces[size++] = Piece(pos.piece_on(s) ^ flipColor);
    } while (b);
@@ -758,7 +761,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
    if (entry->hasPawns) {
        idx = LeadPawnIdx[leadPawnsCnt][squares[0]];
-        std::sort(squares + 1, squares + leadPawnsCnt, pawns_comp);
+        std::stable_sort(squares + 1, squares + leadPawnsCnt, pawns_comp);
        for (int i = 1; i < leadPawnsCnt; ++i)
            idx += Binomial[i][MapPawns[squares[i]]];
@@ -766,7 +769,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
        goto encode_remaining; // With pawns we have finished special treatments
    }
-    // In positions withouth pawns, we further flip the squares to ensure leading
+    // In positions without pawns, we further flip the squares to ensure leading
    // piece is below RANK_5.
    if (rank_of(squares[0]) > RANK_4)
        for (int i = 0; i < size; ++i)
@@ -809,7 +812,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
    // Rs "together" in 62 * 61 / 2 ways (we divide by 2 because rooks can be
    // swapped and still get the same position.)
    //
-    // In case we have at least 3 unique pieces (inlcuded kings) we encode them
+    // In case we have at least 3 unique pieces (included kings) we encode them
    // together.
    if (entry->hasUniquePieces) {
@@ -824,7 +827,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu
                   + (squares[1] - adjust1)) * 62
                   +  squares[2] - adjust2;
-        // First piece is on a1-h8 diagonal, second below: map this occurence to
+        // First piece is on a1-h8 diagonal, second below: map this occurrence to
        // 6 to differentiate from the above case, rank_of() maps a1-d4 diagonal
        // to 0...3 and finally MapB1H1H7[] maps the b1-h1-h7 triangle to 0..27.
        else if (off_A1H8(squares[1]))
@@ -854,12 +857,12 @@ encode_remaining:
    idx *= d->groupIdx[0];
    Square* groupSq = squares + d->groupLen[0];
-    // Encode remainig pawns then pieces according to square, in ascending order
+    // Encode remaining pawns then pieces according to square, in ascending order
    bool remainingPawns = entry->hasPawns && entry->pawnCount[1];
    while (d->groupLen[++next])
    {
-        std::sort(groupSq, groupSq + d->groupLen[next]);
+        std::stable_sort(groupSq, groupSq + d->groupLen[next]);
        uint64_t n = 0;
        // Map down a square if "comes later" than a square in the previous
@@ -882,7 +885,7 @@ encode_remaining:
 // Group together pieces that will be encoded together. The general rule is that
 // a group contains pieces of same type and color. The exception is the leading
-// group that, in case of positions withouth pawns, can be formed by 3 different
+// group that, in case of positions without pawns, can be formed by 3 different
 // pieces (default) or by the king pair when there is not a unique piece apart
 // from the kings. When there are pawns, pawns are always first in pieces[].
 //
@@ -914,7 +917,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) {
    //
    // This ensures unique encoding for the whole position. The order of the
    // groups is a per-table parameter and could not follow the canonical leading
-    // pawns/pieces -> remainig pawns -> remaining pieces. In particular the
+    // pawns/pieces -> remaining pawns -> remaining pieces. In particular the
    // first group is at order[0] position and the remaining pawns, when present,
    // are at order[1] position.
    bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides
@@ -934,7 +937,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) {
            d->groupIdx[1] = idx;
            idx *= Binomial[d->groupLen[1]][48 - d->groupLen[0]];
        }
-        else // Remainig pieces
+        else // Remaining pieces
        {
            d->groupIdx[next] = idx;
            idx *= Binomial[d->groupLen[next]][freeSquares];
@@ -944,7 +947,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) {
    d->groupIdx[n] = idx;
 }
-// In Recursive Pairing each symbol represents a pair of childern symbols. So
+// In Recursive Pairing each symbol represents a pair of children symbols. So
 // read d->btree[] symbols data and expand each one in his left and right child
 // symbol until reaching the leafs that represent the symbol value.
 uint8_t set_symlen(PairsData* d, Sym s, std::vector<bool>& visited) {
@@ -998,7 +1001,7 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) {
    // so that d->lowestSym[i] >= d->lowestSym[i+1] (when read as LittleEndian).
    // Starting from this we compute a base64[] table indexed by symbol length
    // and containing 64 bit values so that d->base64[i] >= d->base64[i+1].
-    // See http://www.eecs.harvard.edu/~michaelm/E210/huffman.pdf
+    // See https://en.wikipedia.org/wiki/Huffman_coding
    for (int i = d->base64.size() - 2; i >= 0; --i) {
        d->base64[i] = (d->base64[i + 1] + number<Sym, LittleEndian>(&d->lowestSym[i])
                                         - number<Sym, LittleEndian>(&d->lowestSym[i + 1])) / 2;
@@ -1139,7 +1142,7 @@ void* mapped(TBTable<Type>& e, const Position& pos) {
    if (e.ready.load(std::memory_order_acquire))
        return e.baseAddress; // Could be nullptr if file does not exist
-    std::unique_lock<std::mutex> lk(mutex);
+    std::scoped_lock<std::mutex> lk(mutex);
    if (e.ready.load(std::memory_order_relaxed)) // Recheck under lock
        return e.baseAddress;
@@ -1287,7 +1290,7 @@ void Tablebases::init(const std::string& paths) {
    for (auto s : diagonal)
        MapA1D1D4[s] = code++;
-    // MapKK[] encodes all the 461 possible legal positions of two kings where
+    // MapKK[] encodes all the 462 possible legal positions of two kings where
    // the first is in the a1-d1-d4 triangle. If the first king is on the a1-d4
    // diagonal, the other one shall not to be above the a1-h8 diagonal.
    std::vector<std::pair<int, Square>> bothOnDiagonal;
@@ -1314,7 +1317,7 @@ void Tablebases::init(const std::string& paths) {
    for (auto p : bothOnDiagonal)
        MapKK[p.first][p.second] = code++;
-    // Binomial[] stores the Binomial Coefficents using Pascal rule. There
+    // Binomial[] stores the Binomial Coefficients using Pascal rule. There
    // are Binomial[k][n] ways to choose k elements from a set of n elements.
    Binomial[0][0] = 1;
@@ -1334,7 +1337,7 @@ void Tablebases::init(const std::string& paths) {
    for (int leadPawnsCnt = 1; leadPawnsCnt <= 5; ++leadPawnsCnt)
        for (File f = FILE_A; f <= FILE_D; ++f)
        {
-            // Restart the index at every file because TB table is splitted
+            // Restart the index at every file because TB table is split
            // by file, so we can reuse the same index for different files.
            int idx = 0;
@@ -1438,7 +1441,7 @@ WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) {
 // If n = 100 immediately after a capture or pawn move, then the position
 // is also certainly a win, and during the whole phase until the next
 // capture or pawn move, the inequality to be preserved is
-// dtz + 50-movecounter <= 100.
+// dtz + 50-move-counter <= 100.
 //
 // In short, if a move is available resulting in dtz + 50-move-counter <= 99,
 // then do not accept moves leading to dtz + 50-move-counter == 100.
@@ -1533,6 +1536,14 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) {
            WDLScore wdl = -probe_wdl(pos, &result);
            dtz = dtz_before_zeroing(wdl);
        }
        else if (pos.is_draw(1))
        {
            // In case a root move leads to a draw by repetition or
            // 50-move rule, we set dtz to zero. Note: since we are
            // only 1 ply from the root, this must be a true 3-fold
            // repetition inside the game history.
            dtz = 0;
        }
        else
        {
            // Otherwise, take dtz for the new position and correct by 1 ply
@@ -1583,6 +1594,7 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) {
    ProbeState result;
    StateInfo st;
    WDLScore wdl;
    bool rule50 = Options["Syzygy50MoveRule"];
@@ -1591,7 +1603,10 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) {
    {
        pos.do_move(m.pv[0], st);
-        WDLScore wdl = -probe_wdl(pos, &result);
+        if (pos.is_draw(1))
            wdl = WDLDraw;
        else
            wdl = -probe_wdl(pos, &result);
        pos.undo_move(m.pv[0]);
@@ -1608,3 +1623,5 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) {
    return true;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,7 +23,7 @@
 #include "../search.h"
-namespace Tablebases {
+namespace Stockfish::Tablebases {
 enum WDLScore {
    WDLLoss        = -2, // Loss
@@ -38,7 +38,7 @@ enum WDLScore {
 // Possible states after a probing operation
 enum ProbeState {
    FAIL              =  0, // Probe failed (missing file table)
-    OK                =  1, // Probe succesful
+    OK                =  1, // Probe successful
    CHANGE_STM        = -1, // DTZ should check the other side
    ZEROING_BEST_MOVE =  2  // Best move zeroes DTZ (capture or pawn move)
 };
@@ -73,6 +73,6 @@ inline std::ostream& operator<<(std::ostream& os, const ProbeState v) {
    return os;
 }
-}
+} // namespace Stockfish::Tablebases
 #endif
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -26,15 +26,18 @@
 #include "syzygy/tbprobe.h"
 #include "tt.h"
 namespace Stockfish {
 ThreadPool Threads; // Global object
 /// Thread constructor launches the thread and waits until it goes to sleep
 /// in idle_loop(). Note that 'searching' and 'exit' should be already set.
-Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) {
+Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this), maxNodes(0) {
  wait_for_search_finished();
  wait_for_worker_finished();
 }
@@ -51,32 +54,21 @@ Thread::~Thread() {
 }
 /// Thread::bestMoveCount(Move move) return best move counter for the given root move
 int Thread::best_move_count(Move move) const {
  auto rm = std::find(rootMoves.begin() + pvIdx,
                      rootMoves.begin() + pvLast, move);
  return rm != rootMoves.begin() + pvLast ? rm->bestMoveCount : 0;
 }
 /// Thread::clear() reset histories, usually before a new game
 void Thread::clear() {
  counterMoves.fill(MOVE_NONE);
  mainHistory.fill(0);
  lowPlyHistory.fill(0);
  captureHistory.fill(0);
-
+  previousDepth = 0;
  for (bool inCheck : { false, true })
      for (StatsType c : { NoCaptures, Captures })
      {
          for (auto& to : continuationHistory[inCheck][c])
                for (auto& h : to)
-                      h->fill(0);
+                      h->fill(-71);
          continuationHistory[inCheck][c][NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1);
      }
 }
@@ -91,6 +83,14 @@ void Thread::start_searching() {
  cv.notify_one(); // Wake up the thread in idle_loop()
 }
 void Thread::execute_with_worker(std::function<void(Thread&)> t)
 {
  std::lock_guard<std::mutex> lk(mutex);
  worker = std::move(t);
  searching = true;
  cv.notify_one(); // Wake up the thread in idle_loop()
 }
 /// Thread::wait_for_search_finished() blocks on the condition variable
 /// until the thread has finished searching.
@@ -102,6 +102,12 @@ void Thread::wait_for_search_finished() {
 }
 void Thread::wait_for_worker_finished() {
  std::unique_lock<std::mutex> lk(mutex);
  cv.wait(lk, [&]{ return !searching; });
 }
 /// Thread::idle_loop() is where the thread is parked, blocked on the
 /// condition variable, when it has no work to do.
@@ -119,15 +125,25 @@ void Thread::idle_loop() {
  {
      std::unique_lock<std::mutex> lk(mutex);
      searching = false;
      worker = nullptr;
      cv.notify_one(); // Wake up anyone waiting for search finished
      cv.wait(lk, [&]{ return searching; });
      if (exit)
          return;
      auto wrk = std::move(worker);
      lk.unlock();
-      search();
+      if (wrk)
      {
        wrk(*this);
      }
      else
      {
        search();
      }
  }
 }
@@ -137,14 +153,16 @@ void Thread::idle_loop() {
 void ThreadPool::set(size_t requested) {
-  if (size() > 0) { // destroy any existing thread(s)
+  if (size() > 0)   // destroy any existing thread(s)
  {
      main()->wait_for_search_finished();
      while (size() > 0)
          delete back(), pop_back();
  }
-  if (requested > 0) { // create new thread(s)
+  if (requested > 0)   // create new thread(s)
  {
      push_back(new MainThread(0));
      while (size() < requested)
@@ -169,9 +187,17 @@ void ThreadPool::clear() {
  main()->callsCnt = 0;
  main()->bestPreviousScore = VALUE_INFINITE;
  main()->bestPreviousAverageScore = VALUE_INFINITE;
  main()->previousTimeReduction = 1.0;
 }
 void ThreadPool::execute_with_workers(const std::function<void(Thread&)>& worker)
 {
  for(Thread* th : *this)
  {
    th->execute_with_worker(worker);
  }
 }
 /// ThreadPool::start_thinking() wakes up main thread waiting in idle_loop() and
 /// returns immediately. Main thread will wake up other threads and start the search.
@@ -192,9 +218,6 @@ void ThreadPool::start_thinking(Position& pos, StateListPtr& states,
          || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
          rootMoves.emplace_back(m);
  if (!rootMoves.empty())
      Tablebases::rank_root_moves(pos, rootMoves);
  // After ownership transfer 'states' becomes empty, so if we stop the search
  // and call 'go' again without setting a new position states.get() == NULL.
  assert(states.get() || setupStates.get());
@@ -204,21 +227,18 @@ void ThreadPool::start_thinking(Position& pos, StateListPtr& states,
  // We use Position::set() to set root position across threads. But there are
  // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot
-  // be deduced from a fen string, so set() clears them and to not lose the info
+  // be deduced from a fen string, so set() clears them and they are set from
-  // we need to backup and later restore setupStates->back(). Note that setupStates
+  // setupStates->back() later. The rootState is per thread, earlier states are shared
-  // is shared by threads but is accessed in read-only mode.
+  // since they are read-only.
  StateInfo tmp = setupStates->back();
  for (Thread* th : *this)
  {
      th->nodes = th->tbHits = th->nmpMinPly = th->bestMoveChanges = 0;
      th->rootDepth = th->completedDepth = 0;
      th->rootMoves = rootMoves;
-      th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th);
+      th->rootPos.set(pos.fen(), pos.is_chess960(), &th->rootState, th);
      th->rootState = setupStates->back();
  }
  setupStates->back() = tmp;
  main()->start_searching();
 }
@@ -238,16 +258,16 @@ Thread* ThreadPool::get_best_thread() const {
        votes[th->rootMoves[0].pv[0]] +=
            (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth);
-          if (abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY)
+        if (abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY)
-          {
+        {
-              // Make sure we pick the shortest mate / TB conversion or stave off mate the longest
+            // Make sure we pick the shortest mate / TB conversion or stave off mate the longest
-              if (th->rootMoves[0].score > bestThread->rootMoves[0].score)
+            if (th->rootMoves[0].score > bestThread->rootMoves[0].score)
-                  bestThread = th;
+                bestThread = th;
-          }
+        }
-          else if (   th->rootMoves[0].score >= VALUE_TB_WIN_IN_MAX_PLY
+        else if (   th->rootMoves[0].score >= VALUE_TB_WIN_IN_MAX_PLY
-                   || (   th->rootMoves[0].score > VALUE_TB_LOSS_IN_MAX_PLY
+                 || (   th->rootMoves[0].score > VALUE_TB_LOSS_IN_MAX_PLY
-                       && votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]]))
+                     && votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]]))
-              bestThread = th;
+            bestThread = th;
    }
    return bestThread;
@@ -272,3 +292,12 @@ void ThreadPool::wait_for_search_finished() const {
        if (th != front())
            th->wait_for_search_finished();
 }
 void ThreadPool::wait_for_workers_finished() const {
    for (Thread* th : *this)
        th->wait_for_worker_finished();
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,6 +24,7 @@
 #include <mutex>
 #include <thread>
 #include <vector>
 #include <functional>
 #include "material.h"
 #include "movepick.h"
@@ -32,47 +33,83 @@
 #include "search.h"
 #include "thread_win32_osx.h"
 namespace Stockfish {
 /// Thread class keeps together all the thread-related stuff. We use
 /// per-thread pawn and material hash tables so that once we get a
 /// pointer to an entry its life time is unlimited and we don't have
 /// to care about someone changing the entry under our feet.
 namespace Detail {
  template <typename T>
  struct TypeIdentity {
    using Type = T;
  };
 }
 class Thread {
  std::mutex mutex;
  std::condition_variable cv;
  size_t idx;
  bool exit = false, searching = true; // Set before starting std::thread
  std::function<void(Thread&)> worker;
  std::function<void(Position&)> on_eval_callback;
  NativeThread stdThread;
 public:
  explicit Thread(size_t);
  virtual ~Thread();
  virtual void search();
  // The function object to be executed is taken by value to remove
  // the need for separate lvalue and rvalue overloads.
  // The worker thread needs to have ownership of the task
  // to be executed because otherwise there's no way to manage its lifetime.
  virtual void execute_with_worker(std::function<void(Thread&)> t);
  void clear();
  void idle_loop();
  void start_searching();
  void wait_for_search_finished();
-  int best_move_count(Move move) const;
+  size_t id() const { return idx; }
  void wait_for_worker_finished();
  template <typename FuncT>
  void set_eval_callback(FuncT&& f) { on_eval_callback = std::forward<FuncT>(f); }
  void clear_eval_callback() { on_eval_callback = nullptr; }
  void on_eval() { if (on_eval_callback) on_eval_callback(rootPos); }
  Pawns::Table pawnsTable;
  Material::Table materialTable;
  size_t pvIdx, pvLast;
-  uint64_t ttHitAverage;
+  RunningAverage complexityAverage;
  std::atomic<uint64_t> nodes, tbHits, bestMoveChanges;
  int selDepth, nmpMinPly;
  Color nmpColor;
-  std::atomic<uint64_t> nodes, tbHits, bestMoveChanges;
+  Value bestValue, optimism[COLOR_NB];
  uint64_t maxNodes;
  Position rootPos;
  StateInfo rootState;
  Search::RootMoves rootMoves;
-  Depth rootDepth, completedDepth;
+  Depth rootDepth, completedDepth, depth, previousDepth;
  Value rootDelta;
  CounterMoveHistory counterMoves;
  ButterflyHistory mainHistory;
  LowPlyHistory lowPlyHistory;
  CapturePieceToHistory captureHistory;
  ContinuationHistory continuationHistory[2][2];
-  Score contempt;
+  Score trend;
  int failedHighCnt;
  bool rootInTB;
  int Cardinality;
  bool UseRule50;
  Depth ProbeDepth;
 };
@@ -87,6 +124,7 @@ struct MainThread : public Thread {
  double previousTimeReduction;
  Value bestPreviousScore;
  Value bestPreviousAverageScore;
  Value iterValue[4];
  int callsCnt;
  bool stopOnPonderhit;
@@ -100,6 +138,61 @@ struct MainThread : public Thread {
 struct ThreadPool : public std::vector<Thread*> {
  // Each thread gets its own copy of the `worker` function object.
  // This means that each worker thread will have exclusive access
  // to the state of the `worker` function object.
  void execute_with_workers(const std::function<void(Thread&)>& worker);
  template <typename IndexT, typename FuncT>
  void for_each_index_with_workers(
    IndexT begin,
    typename Detail::TypeIdentity<IndexT>::Type end,
    FuncT func)
  {
    // This value must outlive the function call.
    // It's fairly safe if we make it static
    // because for_each_index_with_workers
    // is not reentrant nor thread safe.
    static std::atomic<IndexT> i_atomic;
    i_atomic.store(begin);
    execute_with_workers(
      [end, func](Thread& th) mutable {
        for(;;) {
          const auto i = i_atomic.fetch_add(1);
          if (i >= end)
            break;
          func(th, i);
        }
      });
  }
  template <typename IndexT, typename FuncT>
  void for_each_index_chunk_with_workers(
    IndexT begin,
    typename Detail::TypeIdentity<IndexT>::Type end,
    FuncT func)
  {
    // This value must outlive the function call.
    // It's fairly safe if we make it static
    // because for_each_index_with_workers
    // is not reentrant nor thread safe.
    const IndexT size = end - begin;
    const IndexT chunk_size = (size + this->size()) / this->size();
    execute_with_workers(
      [chunk_size, end, func](Thread& th) mutable {
        const IndexT thread_id = th.id();
        const IndexT offset = chunk_size * thread_id;
        if (offset >= end)
          return;
        const IndexT count = offset + chunk_size > end ? end - offset : chunk_size;
        func(th, offset, count);
      });
  }
  void start_thinking(Position&, StateListPtr&, const Search::LimitsType&, bool = false);
  void clear();
  void set(size_t);
@@ -110,6 +203,7 @@ struct ThreadPool : public std::vector<Thread*> {
  Thread* get_best_thread() const;
  void start_searching();
  void wait_for_search_finished() const;
  void wait_for_workers_finished() const;
  std::atomic_bool stop, increaseDepth;
@@ -127,4 +221,6 @@ private:
 extern ThreadPool Threads;
 } // namespace Stockfish
 #endif // #ifndef THREAD_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -27,10 +27,12 @@
 /// The implementation calls pthread_create() with the stack size parameter
 /// equal to the linux 8MB default, on platforms that support it.
-#if defined(__APPLE__) || defined(__MINGW32__) || defined(__MINGW64__)
+#if defined(__APPLE__) || defined(__MINGW32__) || defined(__MINGW64__) || defined(USE_PTHREADS)
 #include <pthread.h>
 namespace Stockfish {
 static const size_t TH_STACK_SIZE = 8 * 1024 * 1024;
 template <class T, class P = std::pair<T*, void(T::*)()>>
@@ -57,10 +59,16 @@ public:
  void join() { pthread_join(thread, NULL); }
 };
 } // namespace Stockfish
 #else // Default case: use STL classes
 namespace Stockfish {
 typedef std::thread NativeThread;
 } // namespace Stockfish
 #endif
 #endif // #ifndef THREAD_WIN32_OSX_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,6 +24,8 @@
 #include "timeman.h"
 #include "uci.h"
 namespace Stockfish {
 TimeManagement Time; // Our global time management object
@@ -38,9 +40,9 @@ void TimeManagement::init(Search::LimitsType& limits, Color us, int ply) {
  TimePoint slowMover       = TimePoint(Options["Slow Mover"]);
  TimePoint npmsec          = TimePoint(Options["nodestime"]);
-  // opt_scale is a percentage of available time to use for the current move.
+  // optScale is a percentage of available time to use for the current move.
-  // max_scale is a multiplier applied to optimumTime.
+  // maxScale is a multiplier applied to optimumTime.
-  double opt_scale, max_scale;
+  double optScale, maxScale;
  // If we have to play in 'nodes as time' mode, then convert from time
  // to nodes, and use resulting values in time management formulas.
@@ -66,6 +68,9 @@ void TimeManagement::init(Search::LimitsType& limits, Color us, int ply) {
  TimePoint timeLeft =  std::max(TimePoint(1),
      limits.time[us] + limits.inc[us] * (mtg - 1) - moveOverhead * (2 + mtg));
  // Use extra time with larger increments
  double optExtra = std::clamp(1.0 + 12.0 * limits.inc[us] / limits.time[us], 1.0, 1.12);
  // A user may scale time usage by setting UCI option "Slow Mover"
  // Default is 100 and changing this value will probably lose elo.
  timeLeft = slowMover * timeLeft / 100;
@@ -75,23 +80,26 @@ void TimeManagement::init(Search::LimitsType& limits, Color us, int ply) {
  // game time for the current move, so also cap to 20% of available game time.
  if (limits.movestogo == 0)
  {
-      opt_scale = std::min(0.008 + std::pow(ply + 3.0, 0.5) / 250.0,
+      optScale = std::min(0.0084 + std::pow(ply + 3.0, 0.5) * 0.0042,
-                           0.2 * limits.time[us] / double(timeLeft));
+                           0.2 * limits.time[us] / double(timeLeft))
-      max_scale = std::min(7.0, 4.0 + ply / 12.0);
+                 * optExtra;
      maxScale = std::min(7.0, 4.0 + ply / 12.0);
  }
  // x moves in y seconds (+ z increment)
  else
  {
-      opt_scale = std::min((0.8 + ply / 128.0) / mtg,
+      optScale = std::min((0.88 + ply / 116.4) / mtg,
-                            0.8 * limits.time[us] / double(timeLeft));
+                            0.88 * limits.time[us] / double(timeLeft));
-      max_scale = std::min(6.3, 1.5 + 0.11 * mtg);
+      maxScale = std::min(6.3, 1.5 + 0.11 * mtg);
  }
  // Never use more than 80% of the available time for this move
-  optimumTime = TimePoint(opt_scale * timeLeft);
+  optimumTime = TimePoint(optScale * timeLeft);
-  maximumTime = TimePoint(std::min(0.8 * limits.time[us] - moveOverhead, max_scale * optimumTime));
+  maximumTime = TimePoint(std::min(0.8 * limits.time[us] - moveOverhead, maxScale * optimumTime));
  if (Options["Ponder"])
      optimumTime += optimumTime / 4;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -23,6 +23,8 @@
 #include "search.h"
 #include "thread.h"
 namespace Stockfish {
 /// The TimeManagement class computes the optimal time to think depending on
 /// the maximum available time, the game move number and other parameters.
@@ -44,4 +46,6 @@ private:
 extern TimeManagement Time;
 } // namespace Stockfish
 #endif // #ifndef TIMEMAN_H_INCLUDED
@@ -0,0 +1,815 @@
 #include "convert.h"
 #include "uci.h"
 #include "misc.h"
 #include "thread.h"
 #include "position.h"
 #include "tt.h"
 #include "extra/nnue_data_binpack_format.h"
 #include "nnue/evaluate_nnue.h"
 #include "syzygy/tbprobe.h"
 #include <sstream>
 #include <fstream>
 #include <unordered_set>
 #include <iomanip>
 #include <list>
 #include <cmath>    // std::exp(),std::pow(),std::log()
 #include <cstring>  // memcpy()
 #include <memory>
 #include <limits>
 #include <optional>
 #include <chrono>
 #include <random>
 #include <regex>
 #include <filesystem>
 using namespace std;
 namespace sys = std::filesystem;
 namespace Stockfish::Tools
 {
    bool fen_is_ok(Position& pos, std::string input_fen) {
        std::string pos_fen = pos.fen();
        std::istringstream ss_input(input_fen);
        std::istringstream ss_pos(pos_fen);
        // example : "2r4r/4kpp1/nb1np3/p2p3p/B2P1BP1/PP6/4NPKP/2R1R3 w - h6 0 24"
        //       --> "2r4r/4kpp1/nb1np3/p2p3p/B2P1BP1/PP6/4NPKP/2R1R3"
        std::string str_input, str_pos;
        ss_input >> str_input;
        ss_pos >> str_pos;
        // Only compare "Piece placement field" between input_fen and pos.fen().
        return str_input == str_pos;
    }
    void convert_bin(
        const vector<string>& filenames,
        const string& output_file_name,
        const int ply_minimum,
        const int ply_maximum,
        const int interpolate_eval,
        const int src_score_min_value,
        const int src_score_max_value,
        const int dest_score_min_value,
        const int dest_score_max_value,
        const bool check_invalid_fen,
        const bool check_illegal_move)
    {
        std::cout << "check_invalid_fen=" << check_invalid_fen << std::endl;
        std::cout << "check_illegal_move=" << check_illegal_move << std::endl;
        std::fstream fs;
        uint64_t data_size = 0;
        uint64_t filtered_size = 0;
        uint64_t filtered_size_fen = 0;
        uint64_t filtered_size_move = 0;
        uint64_t filtered_size_ply = 0;
        auto th = Threads.main();
        auto& tpos = th->rootPos;
        // convert plain rag to packed sfenvalue for Yaneura king
        fs.open(output_file_name, ios::app | ios::binary);
        StateListPtr states;
        for (auto filename : filenames) {
            std::cout << "convert " << filename << " ... ";
            std::string line;
            ifstream ifs;
            ifs.open(filename);
            PackedSfenValue p;
            data_size = 0;
            filtered_size = 0;
            filtered_size_fen = 0;
            filtered_size_move = 0;
            filtered_size_ply = 0;
            p.gamePly = 1; // Not included in apery format. Should be initialized
            bool ignore_flag_fen = false;
            bool ignore_flag_move = false;
            bool ignore_flag_ply = false;
            while (std::getline(ifs, line)) {
                std::stringstream ss(line);
                std::string token;
                std::string value;
                ss >> token;
                if (token == "fen") {
                    states = StateListPtr(new std::deque<StateInfo>(1)); // Drop old and create a new one
                    std::string input_fen = line.substr(4);
                    tpos.set(input_fen, false, &states->back(), Threads.main());
                    if (check_invalid_fen && !fen_is_ok(tpos, input_fen)) {
                        ignore_flag_fen = true;
                        filtered_size_fen++;
                    }
                    else {
                        tpos.sfen_pack(p.sfen, false);
                    }
                }
                else if (token == "move") {
                    ss >> value;
                    Move move = UCI::to_move(tpos, value);
                    if (check_illegal_move && move == MOVE_NONE) {
                        ignore_flag_move = true;
                        filtered_size_move++;
                    }
                    else {
                        p.move = move;
                    }
                }
                else if (token == "score") {
                    double score;
                    ss >> score;
                    // Training Formula ?Issue #71 ?nodchip/Stockfish https://github.com/nodchip/Stockfish/issues/71
                    // Normalize to [0.0, 1.0].
                    score = (score - src_score_min_value) / (src_score_max_value - src_score_min_value);
                    // Scale to [dest_score_min_value, dest_score_max_value].
                    score = score * (dest_score_max_value - dest_score_min_value) + dest_score_min_value;
                    p.score = std::clamp((int32_t)std::round(score), -(int32_t)VALUE_MATE, (int32_t)VALUE_MATE);
                }
                else if (token == "ply") {
                    int temp;
                    ss >> temp;
                    if (temp < ply_minimum || temp > ply_maximum) {
                        ignore_flag_ply = true;
                        filtered_size_ply++;
                    }
                    p.gamePly = uint16_t(temp); // No cast here?
                    if (interpolate_eval != 0) {
                        p.score = min(3000, interpolate_eval * temp);
                    }
                }
                else if (token == "result") {
                    int temp;
                    ss >> temp;
                    p.game_result = int8_t(temp); // Do you need a cast here?
                    if (interpolate_eval) {
                        p.score = p.score * p.game_result;
                    }
                }
                else if (token == "e") {
                    if (!(ignore_flag_fen || ignore_flag_move || ignore_flag_ply)) {
                        fs.write((char*)&p, sizeof(PackedSfenValue));
                        data_size += 1;
                        // debug
                        // std::cout<<tpos<<std::endl;
                        // std::cout<<p.score<<","<<int(p.gamePly)<<","<<int(p.game_result)<<std::endl;
                    }
                    else {
                        filtered_size++;
                    }
                    ignore_flag_fen = false;
                    ignore_flag_move = false;
                    ignore_flag_ply = false;
                }
            }
            std::cout << "done " << data_size << " parsed " << filtered_size << " is filtered"
                << " (invalid fen:" << filtered_size_fen << ", illegal move:" << filtered_size_move << ", invalid ply:" << filtered_size_ply << ")" << std::endl;
            ifs.close();
        }
        std::cout << "all done" << std::endl;
        fs.close();
    }
    static inline void ltrim(std::string& s) {
        s.erase(s.begin(), std::find_if(s.begin(), s.end(), [](int ch) {
            return !std::isspace(ch);
            }));
    }
    static inline void rtrim(std::string& s) {
        s.erase(std::find_if(s.rbegin(), s.rend(), [](int ch) {
            return !std::isspace(ch);
            }).base(), s.end());
    }
    static inline void trim(std::string& s) {
        ltrim(s);
        rtrim(s);
    }
    int parse_game_result_from_pgn_extract(std::string result) {
        // White Win
        if (result == "\"1-0\"") {
            return 1;
        }
        // Black Win
        else if (result == "\"0-1\"") {
            return -1;
        }
        // Draw
        else {
            return 0;
        }
    }
    // 0.25 -->  0.25 * PawnValueEg
    // #-4  --> -mate_in(4)
    // #3   -->  mate_in(3)
    // -M4  --> -mate_in(4)
    // +M3  -->  mate_in(3)
    Value parse_score_from_pgn_extract(std::string eval, bool& success) {
        success = true;
        if (eval.substr(0, 1) == "#") {
            if (eval.substr(1, 1) == "-") {
                return -mate_in(stoi(eval.substr(2, eval.length() - 2)));
            }
            else {
                return mate_in(stoi(eval.substr(1, eval.length() - 1)));
            }
        }
        else if (eval.substr(0, 2) == "-M") {
            //std::cout << "eval=" << eval << std::endl;
            return -mate_in(stoi(eval.substr(2, eval.length() - 2)));
        }
        else if (eval.substr(0, 2) == "+M") {
            //std::cout << "eval=" << eval << std::endl;
            return mate_in(stoi(eval.substr(2, eval.length() - 2)));
        }
        else {
            char* endptr;
            double value = strtod(eval.c_str(), &endptr);
            if (*endptr != '\0') {
                success = false;
                return VALUE_ZERO;
            }
            else {
                return Value(value * static_cast<double>(PawnValueEg));
            }
        }
    }
    // for Debug
    //#define DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT
    bool is_like_fen(std::string fen) {
        int count_space = std::count(fen.cbegin(), fen.cend(), ' ');
        int count_slash = std::count(fen.cbegin(), fen.cend(), '/');
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
        //std::cout << "count_space=" << count_space << std::endl;
        //std::cout << "count_slash=" << count_slash << std::endl;
 #endif
        return count_space == 5 && count_slash == 7;
    }
    void convert_bin_from_pgn_extract(
        const vector<string>& filenames,
        const string& output_file_name,
        const bool pgn_eval_side_to_move,
        const bool convert_no_eval_fens_as_score_zero)
    {
        std::cout << "pgn_eval_side_to_move=" << pgn_eval_side_to_move << std::endl;
        std::cout << "convert_no_eval_fens_as_score_zero=" << convert_no_eval_fens_as_score_zero << std::endl;
        auto th = Threads.main();
        auto& pos = th->rootPos;
        std::fstream ofs;
        ofs.open(output_file_name, ios::out | ios::binary);
        int game_count = 0;
        int fen_count = 0;
        for (auto filename : filenames) {
            std::cout << now_string() << " convert " << filename << std::endl;
            ifstream ifs;
            ifs.open(filename);
            int game_result = 0;
            std::string line;
            while (std::getline(ifs, line)) {
                if (line.empty()) {
                    continue;
                }
                else if (line.substr(0, 1) == "[") {
                    std::regex pattern_result(R"(\[Result (.+?)\])");
                    std::smatch match;
                    // example: [Result "1-0"]
                    if (std::regex_search(line, match, pattern_result)) {
                        game_result = parse_game_result_from_pgn_extract(match.str(1));
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
                        std::cout << "game_result=" << game_result << std::endl;
 #endif
                        game_count++;
                        if (game_count % 10000 == 0) {
                            std::cout << now_string() << " game_count=" << game_count << ", fen_count=" << fen_count << std::endl;
                        }
                    }
                    continue;
                }
                else {
                    int gamePly = 1;
                    auto itr = line.cbegin();
                    while (true) {
                        gamePly++;
                        PackedSfenValue psv;
                        memset((char*)&psv, 0, sizeof(PackedSfenValue));
                        // fen
                        {
                            bool fen_found = false;
                            while (!fen_found) {
                                std::regex pattern_bracket(R"(\{(.+?)\})");
                                std::smatch match;
                                if (!std::regex_search(itr, line.cend(), match, pattern_bracket)) {
                                    break;
                                }
                                itr += match.position(0) + match.length(0) - 1;
                                std::string str_fen = match.str(1);
                                trim(str_fen);
                                if (is_like_fen(str_fen)) {
                                    fen_found = true;
                                    StateInfo si;
                                    pos.set(str_fen, false, &si, th);
                                    pos.sfen_pack(psv.sfen, false);
                                }
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
                                std::cout << "str_fen=" << str_fen << std::endl;
                                std::cout << "fen_found=" << fen_found << std::endl;
 #endif
                            }
                            if (!fen_found) {
                                break;
                            }
                        }
                        // move
                        {
                            std::regex pattern_move(R"(\}(.+?)\{)");
                            std::smatch match;
                            if (!std::regex_search(itr, line.cend(), match, pattern_move)) {
                                break;
                            }
                            itr += match.position(0) + match.length(0) - 1;
                            std::string str_move = match.str(1);
                            trim(str_move);
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
                            std::cout << "str_move=" << str_move << std::endl;
 #endif
                            psv.move = UCI::to_move(pos, str_move);
                        }
                        // eval
                        bool eval_found = false;
                        {
                            std::regex pattern_bracket(R"(\{(.+?)\})");
                            std::smatch match;
                            if (!std::regex_search(itr, line.cend(), match, pattern_bracket)) {
                                break;
                            }
                            std::string str_eval_clk = match.str(1);
                            trim(str_eval_clk);
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
                            std::cout << "str_eval_clk=" << str_eval_clk << std::endl;
 #endif
                            // example: { [%eval 0.25] [%clk 0:10:00] }
                            // example: { [%eval #-4] [%clk 0:10:00] }
                            // example: { [%eval #3] [%clk 0:10:00] }
                            // example: { +0.71/22 1.2s }
                            // example: { -M4/7 0.003s }
                            // example: { M3/245 0.017s }
                            // example: { +M1/245 0.010s, White mates }
                            // example: { 0.60 }
                            // example: { book }
                            // example: { rnbqkb1r/pp3ppp/2p1pn2/3p4/2PP4/2N2N2/PP2PPPP/R1BQKB1R w KQkq - 0 5 }
                            // Considering the absence of eval
                            if (!is_like_fen(str_eval_clk)) {
                                itr += match.position(0) + match.length(0) - 1;
                                if (str_eval_clk != "book") {
                                    std::regex pattern_eval1(R"(\[\%eval (.+?)\])");
                                    std::regex pattern_eval2(R"((.+?)\/)");
                                    std::string str_eval;
                                    if (std::regex_search(str_eval_clk, match, pattern_eval1) ||
                                        std::regex_search(str_eval_clk, match, pattern_eval2)) {
                                        str_eval = match.str(1);
                                        trim(str_eval);
                                    }
                                    else {
                                        str_eval = str_eval_clk;
                                    }
                                    bool success = false;
                                    Value value = parse_score_from_pgn_extract(str_eval, success);
                                    if (success) {
                                        eval_found = true;
                                        psv.score = std::clamp(value, -VALUE_MATE, VALUE_MATE);
                                    }
 #if defined(DEBUG_CONVERT_BIN_FROM_PGN_EXTRACT)
                                    std::cout << "str_eval=" << str_eval << std::endl;
                                    std::cout << "success=" << success << ", psv.score=" << psv.score << std::endl;
 #endif
                                }
                            }
                        }
                        // write
                        if (eval_found || convert_no_eval_fens_as_score_zero) {
                            if (!eval_found && convert_no_eval_fens_as_score_zero) {
                                psv.score = 0;
                            }
                            psv.gamePly = gamePly;
                            psv.game_result = game_result;
                            if (pos.side_to_move() == BLACK) {
                                if (!pgn_eval_side_to_move) {
                                    psv.score *= -1;
                                }
                                psv.game_result *= -1;
                            }
                            ofs.write((char*)&psv, sizeof(PackedSfenValue));
                            fen_count++;
                        }
                    }
                    game_result = 0;
                }
            }
        }
        std::cout << now_string() << " game_count=" << game_count << ", fen_count=" << fen_count << std::endl;
        std::cout << now_string() << " all done" << std::endl;
        ofs.close();
    }
    void convert_plain(
        const vector<string>& filenames,
        const string& output_file_name)
    {
        Position tpos;
        std::ofstream ofs;
        ofs.open(output_file_name, ios::app);
        auto th = Threads.main();
        for (auto filename : filenames) {
            std::cout << "convert " << filename << " ... ";
            // Just convert packedsfenvalue to text
            std::fstream fs;
            fs.open(filename, ios::in | ios::binary);
            PackedSfenValue p;
            while (true)
            {
                if (fs.read((char*)&p, sizeof(PackedSfenValue))) {
                    StateInfo si;
                    tpos.set_from_packed_sfen(p.sfen, &si, th, false);
                    // write as plain text
                    ofs << "fen " << tpos.fen() << std::endl;
                    ofs << "move " << UCI::move(Move(p.move), false) << std::endl;
                    ofs << "score " << p.score << std::endl;
                    ofs << "ply " << int(p.gamePly) << std::endl;
                    ofs << "result " << int(p.game_result) << std::endl;
                    ofs << "e" << std::endl;
                }
                else {
                    break;
                }
            }
            fs.close();
            std::cout << "done" << std::endl;
        }
        ofs.close();
        std::cout << "all done" << std::endl;
    }
    static inline const std::string plain_extension = ".plain";
    static inline const std::string bin_extension = ".bin";
    static inline const std::string binpack_extension = ".binpack";
    static bool file_exists(const std::string& name)
    {
        std::ifstream f(name);
        return f.good();
    }
    static bool ends_with(const std::string& lhs, const std::string& end)
    {
        if (end.size() > lhs.size()) return false;
        return std::equal(end.rbegin(), end.rend(), lhs.rbegin());
    }
    static bool is_convert_of_type(
        const std::string& input_path,
        const std::string& output_path,
        const std::string& expected_input_extension,
        const std::string& expected_output_extension)
    {
        return ends_with(input_path, expected_input_extension)
            && ends_with(output_path, expected_output_extension);
    }
    using ConvertFunctionType = void(std::string inputPath, std::string outputPath, std::ios_base::openmode om, bool validate);
    static ConvertFunctionType* get_convert_function(const std::string& input_path, const std::string& output_path)
    {
        if (is_convert_of_type(input_path, output_path, plain_extension, bin_extension))
            return binpack::convertPlainToBin;
        if (is_convert_of_type(input_path, output_path, plain_extension, binpack_extension))
            return binpack::convertPlainToBinpack;
        if (is_convert_of_type(input_path, output_path, bin_extension, plain_extension))
            return binpack::convertBinToPlain;
        if (is_convert_of_type(input_path, output_path, bin_extension, binpack_extension))
            return binpack::convertBinToBinpack;
        if (is_convert_of_type(input_path, output_path, binpack_extension, plain_extension))
            return binpack::convertBinpackToPlain;
        if (is_convert_of_type(input_path, output_path, binpack_extension, bin_extension))
            return binpack::convertBinpackToBin;
        return nullptr;
    }
    static void convert(const std::string& input_path, const std::string& output_path, std::ios_base::openmode om, bool validate)
    {
        if(!file_exists(input_path))
        {
            std::cerr << "Input file does not exist.\n";
            return;
        }
        auto func = get_convert_function(input_path, output_path);
        if (func != nullptr)
        {
            func(input_path, output_path, om, validate);
        }
        else
        {
            std::cerr << "Conversion between files of these types is not supported.\n";
        }
    }
    static void convert(const std::vector<std::string>& args)
    {
        if (args.size() < 2 || args.size() > 4)
        {
            std::cerr << "Invalid arguments.\n";
            std::cerr << "Usage: convert from_path to_path [append] [validate]\n";
            return;
        }
        const bool append = std::find(args.begin() + 2, args.end(), "append") != args.end();
        const bool validate = std::find(args.begin() + 2, args.end(), "validate") != args.end();
        const std::ios_base::openmode openmode =
            append
            ? std::ios_base::app
            : std::ios_base::trunc;
        convert(args[0], args[1], openmode, validate);
    }
    void convert(istringstream& is)
    {
        std::vector<std::string> args;
        while (true)
        {
            std::string token = "";
            is >> token;
            if (token == "")
                break;
            args.push_back(token);
        }
        convert(args);
    }
    static void append_files_from_dir(
        std::vector<std::string>& filenames,
        const std::string& base_dir,
        const std::string& target_dir)
    {
        string kif_base_dir = Path::combine(base_dir, target_dir);
        sys::path p(kif_base_dir); // Origin of enumeration
        std::for_each(sys::directory_iterator(p), sys::directory_iterator(),
            [&](const sys::path& path) {
                if (sys::is_regular_file(path))
                    filenames.push_back(Path::combine(target_dir, path.filename().generic_string()));
            });
    }
    static void rebase_files(
        std::vector<std::string>& filenames,
        const std::string& base_dir)
    {
        for (auto& file : filenames)
        {
            file = Path::combine(base_dir, file);
        }
    }
    void convert_bin_from_pgn_extract(std::istringstream& is)
    {
        std::vector<std::string> filenames;
        string base_dir;
        string target_dir;
        bool pgn_eval_side_to_move = false;
        bool convert_no_eval_fens_as_score_zero = false;
        string output_file_name = "shuffled_sfen.bin";
        while (true)
        {
            string option;
            is >> option;
            if (option == "")
                break;
            if (option == "targetdir") is >> target_dir;
            else if (option == "targetfile")
            {
                std::string filename;
                is >> filename;
                filenames.push_back(filename);
            }
            else if (option == "basedir")   is >> base_dir;
            else if (option == "pgn_eval_side_to_move") is >> pgn_eval_side_to_move;
            else if (option == "convert_no_eval_fens_as_score_zero") is >> convert_no_eval_fens_as_score_zero;
            else if (option == "output_file_name") is >> output_file_name;
            else
            {
                cout << "Unknown option: " << option << ". Ignoring.\n";
            }
        }
        if (!target_dir.empty())
        {
            append_files_from_dir(filenames, base_dir, target_dir);
        }
        rebase_files(filenames, base_dir);
        Eval::NNUE::init();
        cout << "convert_bin_from_pgn-extract.." << endl;
        convert_bin_from_pgn_extract(
            filenames,
            output_file_name,
            pgn_eval_side_to_move,
            convert_no_eval_fens_as_score_zero);
    }
    void convert_bin(std::istringstream& is)
    {
        std::vector<std::string> filenames;
        string base_dir;
        string target_dir;
        int ply_minimum = 0;
        int ply_maximum = 114514;
        bool interpolate_eval = 0;
        bool check_invalid_fen = false;
        bool check_illegal_move = false;
        bool pgn_eval_side_to_move = false;
        bool convert_no_eval_fens_as_score_zero = false;
        double src_score_min_value = 0.0;
        double src_score_max_value = 1.0;
        double dest_score_min_value = 0.0;
        double dest_score_max_value = 1.0;
        string output_file_name = "shuffled_sfen.bin";
        while (true)
        {
            string option;
            is >> option;
            if (option == "")
                break;
            if (option == "targetdir") is >> target_dir;
            else if (option == "targetfile")
            {
                std::string filename;
                is >> filename;
                filenames.push_back(filename);
            }
            else if (option == "basedir")   is >> base_dir;
            else if (option == "ply_minimum") is >> ply_minimum;
            else if (option == "ply_maximum") is >> ply_maximum;
            else if (option == "interpolate_eval") is >> interpolate_eval;
            else if (option == "check_invalid_fen") is >> check_invalid_fen;
            else if (option == "check_illegal_move") is >> check_illegal_move;
            else if (option == "pgn_eval_side_to_move") is >> pgn_eval_side_to_move;
            else if (option == "convert_no_eval_fens_as_score_zero") is >> convert_no_eval_fens_as_score_zero;
            else if (option == "src_score_min_value") is >> src_score_min_value;
            else if (option == "src_score_max_value") is >> src_score_max_value;
            else if (option == "dest_score_min_value") is >> dest_score_min_value;
            else if (option == "dest_score_max_value") is >> dest_score_max_value;
            else if (option == "output_file_name") is >> output_file_name;
            else
            {
                cout << "Unknown option: " << option << ". Ignoring.\n";
            }
        }
        if (!target_dir.empty())
        {
            append_files_from_dir(filenames, base_dir, target_dir);
        }
        rebase_files(filenames, base_dir);
        Eval::NNUE::init();
        cout << "convert_bin.." << endl;
            convert_bin(
                filenames,
                output_file_name,
                ply_minimum,
                ply_maximum,
                interpolate_eval,
                src_score_min_value,
                src_score_max_value,
                dest_score_min_value,
                dest_score_max_value,
                check_invalid_fen,
                check_illegal_move
            );
    }
    void convert_plain(std::istringstream& is)
    {
        std::vector<std::string> filenames;
        string base_dir;
        string target_dir;
        string output_file_name = "shuffled_sfen.bin";
        while (true)
        {
            string option;
            is >> option;
            if (option == "")
                break;
            if (option == "targetdir") is >> target_dir;
            else if (option == "targetfile")
            {
                std::string filename;
                is >> filename;
                filenames.push_back(filename);
            }
            else if (option == "basedir")   is >> base_dir;
            else if (option == "output_file_name") is >> output_file_name;
            else
            {
                cout << "Unknown option: " << option << ". Ignoring.\n";
            }
        }
        if (!target_dir.empty())
        {
            append_files_from_dir(filenames, base_dir, target_dir);
        }
        rebase_files(filenames, base_dir);
        Eval::NNUE::init();
        cout << "convert_plain.." << endl;
        convert_plain(filenames, output_file_name);
    }
 }
@@ -0,0 +1,18 @@
 #ifndef _CONVERT_H_
 #define _CONVERT_H_
 #include <vector>
 #include <string>
 #include <sstream>
 namespace Stockfish::Tools {
    void convert(std::istringstream& is);
    void convert_bin_from_pgn_extract(std::istringstream& is);
    void convert_bin(std::istringstream& is);
    void convert_plain(std::istringstream& is);
 }
 #endif
@@ -0,0 +1,43 @@
 #include "opening_book.h"
 #include <fstream>
 namespace Stockfish::Tools {
    EpdOpeningBook::EpdOpeningBook(const std::string& file, PRNG& prng) :
        OpeningBook(file)
    {
        std::ifstream in(file);
        if (!in)
        {
            return;
        }
        std::string line;
        while (std::getline(in, line))
        {
            if (line.empty())
                continue;
            fens.emplace_back(line);
        }
        Algo::shuffle(fens, prng);
    }
    static bool ends_with(const std::string& lhs, const std::string& end)
    {
        if (end.size() > lhs.size()) return false;
        return std::equal(end.rbegin(), end.rend(), lhs.rbegin());
    }
    std::unique_ptr<OpeningBook> open_opening_book(const std::string& filename, PRNG& prng)
    {
        if (ends_with(filename, ".epd"))
            return std::make_unique<EpdOpeningBook>(filename, prng);
        return nullptr;
    }
 }
@@ -0,0 +1,60 @@
 #ifndef LEARN_OPENING_BOOK_H
 #define LEARN_OPENING_BOOK_H
 #include "misc.h"
 #include "position.h"
 #include "thread.h"
 #include <vector>
 #include <random>
 #include <optional>
 #include <string>
 #include <cstdint>
 #include <memory>
 #include <mutex>
 namespace Stockfish::Tools {
    struct OpeningBook {
        const std::string& next_fen()
        {
            assert(fens.size() > 0);
            std::unique_lock lock(mutex);
            auto& fen = fens[current_index++];
            if (current_index >= fens.size())
                current_index = 0;
            return fen;
        }
        std::size_t size() const { return fens.size(); }
        const std::string& get_filename() const { return filename; }
    protected:
        OpeningBook(const std::string& file) :
            filename(file),
            current_index(0)
        {
        }
        std::mutex mutex;
        std::string filename;
        std::vector<std::string> fens;
        std::size_t current_index;
    };
    struct EpdOpeningBook : OpeningBook {
        EpdOpeningBook(const std::string& file, PRNG& prng);
    };
    std::unique_ptr<OpeningBook> open_opening_book(const std::string& filename, PRNG& prng);
 }
 #endif
@@ -0,0 +1,46 @@
 #ifndef _PACKED_SFEN_H_
 #define _PACKED_SFEN_H_
 #include <vector>
 #include <cstdint>
 namespace Stockfish::Tools {
    // packed sfen
    struct PackedSfen { std::uint8_t data[32]; };
    // Structure in which PackedSfen and evaluation value are integrated
    // If you write different contents for each option, it will be a problem when reusing the teacher game
    // For the time being, write all the following members regardless of the options.
    struct PackedSfenValue
    {
        // phase
        PackedSfen sfen;
        // Evaluation value returned from Tools::search()
        std::int16_t score;
        // PV first move
        // Used when finding the match rate with the teacher
        std::uint16_t move;
        // Trouble of the phase from the initial phase.
        std::uint16_t gamePly;
        // 1 if the player on this side ultimately wins the game. -1 if you are losing.
        // 0 if a draw is reached.
        // The draw is in the teacher position generation command gensfen,
        // Only write if LEARN_GENSFEN_DRAW_RESULT is enabled.
        std::int8_t game_result;
        // When exchanging the file that wrote the teacher aspect with other people
        //Because this structure size is not fixed, pad it so that it is 40 bytes in any environment.
        std::uint8_t padding;
        // 32 + 2 + 2 + 2 + 1 + 1 = 40bytes
    };
    // Phase array: PSVector stands for packed sfen vector.
    using PSVector = std::vector<PackedSfenValue>;
 }
 #endif
@@ -0,0 +1,389 @@
 #include "sfen_packer.h"
 #include "packed_sfen.h"
 #include "misc.h"
 #include "position.h"
 #include <sstream>
 #include <fstream>
 #include <cstring> // std::memset()
 using namespace std;
 namespace Stockfish::Tools {
    // Class that handles bitstream
    // useful when doing aspect encoding
    struct BitStream
    {
        // Set the memory to store the data in advance.
        // Assume that memory is cleared to 0.
        void set_data(std::uint8_t* data_) { data = data_; reset(); }
        // Get the pointer passed in set_data().
        uint8_t* get_data() const { return data; }
        // Get the cursor.
        int get_cursor() const { return bit_cursor; }
        // reset the cursor
        void reset() { bit_cursor = 0; }
        // Write 1bit to the stream.
        // If b is non-zero, write out 1. If 0, write 0.
        void write_one_bit(int b)
        {
            if (b)
                data[bit_cursor / 8] |= 1 << (bit_cursor & 7);
            ++bit_cursor;
        }
        // Get 1 bit from the stream.
        int read_one_bit()
        {
            int b = (data[bit_cursor / 8] >> (bit_cursor & 7)) & 1;
            ++bit_cursor;
            return b;
        }
        // write n bits of data
        // Data shall be written out from the lower order of d.
        void write_n_bit(int d, int n)
        {
            for (int i = 0; i <n; ++i)
                write_one_bit(d & (1 << i));
        }
        // read n bits of data
        // Reverse conversion of write_n_bit().
        int read_n_bit(int n)
        {
            int result = 0;
            for (int i = 0; i < n; ++i)
                result |= read_one_bit() ? (1 << i) : 0;
            return result;
        }
    private:
        // Next bit position to read/write.
        int bit_cursor;
        // data entity
        std::uint8_t* data;
    };
    // Class for compressing/decompressing sfen
    // sfen can be packed to 256bit (32bytes) by Huffman coding.
    // This is proven by mini. The above is Huffman coding.
    //
    // Internal format = 1-bit turn + 7-bit king position *2 + piece on board (Huffman coding) + hand piece (Huffman coding)
    // Side to move (White = 0, Black = 1) (1bit)
    // White King Position (6 bits)
    // Black King Position (6 bits)
    // Huffman Encoding of the board
    // Castling availability (1 bit x 4)
    // En passant square (1 or 1 + 6 bits)
    // Rule 50 (6 bits)
    // Game play (8 bits)
    //
    // TODO(someone): Rename SFEN to FEN.
    //
    struct SfenPacker
    {
        void pack(const Position& pos, bool resetCastlingRights);
        // sfen packed by pack() (256bit = 32bytes)
        // Or sfen to decode with unpack()
        uint8_t *data; // uint8_t[32];
        BitStream stream;
        // Output the board pieces to stream.
        void write_board_piece_to_stream(Piece pc);
        // Read one board piece from stream
        Piece read_board_piece_from_stream();
    };
    // Huffman coding
    // * is simplified from mini encoding to make conversion easier.
    //
    // Huffman Encoding
    //
    // Empty  xxxxxxx0
    // Pawn   xxxxx001 + 1 bit (Color)
    // Knight xxxxx011 + 1 bit (Color)
    // Bishop xxxxx101 + 1 bit (Color)
    // Rook   xxxxx111 + 1 bit (Color)
    // Queen   xxxx1001 + 1 bit (Color)
    //
    // Worst case:
    // - 32 empty squares    32 bits
    // - 30 pieces           150 bits
    // - 2 kings             12 bits
    // - castling rights     4 bits
    // - ep square           7 bits
    // - rule50              7 bits
    // - game ply            16 bits
    // - TOTAL               228 bits < 256 bits
    struct HuffmanedPiece
    {
        int code; // how it will be coded
        int bits; // How many bits do you have
    };
    constexpr HuffmanedPiece huffman_table[] =
    {
        {0b0000,1}, // NO_PIECE
        {0b0001,4}, // PAWN
        {0b0011,4}, // KNIGHT
        {0b0101,4}, // BISHOP
        {0b0111,4}, // ROOK
        {0b1001,4}, // QUEEN
    };
    // Pack sfen and store in data[32].
    void SfenPacker::pack(const Position& pos, bool resetCastlingRights)
    {
        memset(data, 0, 32 /* 256bit */);
        stream.set_data(data);
        // turn
        // Side to move.
        stream.write_one_bit((int)(pos.side_to_move()));
        // 7-bit positions for leading and trailing balls
        // White king and black king, 6 bits for each.
        for(auto c: Colors)
            stream.write_n_bit(pos.king_square(c), 6);
        // Write the pieces on the board other than the kings.
        for (Rank r = RANK_8; r >= RANK_1; --r)
        {
            for (File f = FILE_A; f <= FILE_H; ++f)
            {
                Piece pc = pos.piece_on(make_square(f, r));
                if (type_of(pc) == KING)
                    continue;
                write_board_piece_to_stream(pc);
            }
        }
        if (resetCastlingRights)
        {
            stream.write_n_bit(0, 4);
        }
        else
        {
            stream.write_one_bit(pos.can_castle(WHITE_OO));
            stream.write_one_bit(pos.can_castle(WHITE_OOO));
            stream.write_one_bit(pos.can_castle(BLACK_OO));
            stream.write_one_bit(pos.can_castle(BLACK_OOO));
        }
        if (pos.ep_square() == SQ_NONE) {
            stream.write_one_bit(0);
        }
        else {
            stream.write_one_bit(1);
            stream.write_n_bit(static_cast<int>(pos.ep_square()), 6);
        }
        stream.write_n_bit(pos.state()->rule50, 6);
        const int fm = 1 + (pos.game_ply()-(pos.side_to_move() == BLACK)) / 2;
        stream.write_n_bit(fm, 8);
        // Write high bits of half move. This is a fix for the
        // limited range of half move counter.
        // This is backwards compatibile.
        stream.write_n_bit(fm >> 8, 8);
        // Write the highest bit of rule50 at the end. This is a backwards
        // compatibile fix for rule50 having only 6 bits stored.
        // This bit is just ignored by the old parsers.
        stream.write_n_bit(pos.state()->rule50 >> 6, 1);
        assert(stream.get_cursor() <= 256);
    }
    // Output the board pieces to stream.
    void SfenPacker::write_board_piece_to_stream(Piece pc)
    {
        // piece type
        PieceType pr = type_of(pc);
        auto c = huffman_table[pr];
        stream.write_n_bit(c.code, c.bits);
        if (pc == NO_PIECE)
            return;
        // first and second flag
        stream.write_one_bit(color_of(pc));
    }
    // Read one board piece from stream
    Piece SfenPacker::read_board_piece_from_stream()
    {
        PieceType pr = NO_PIECE_TYPE;
        int code = 0, bits = 0;
        while (true)
        {
            code |= stream.read_one_bit() << bits;
            ++bits;
            assert(bits <= 6);
            for (pr = NO_PIECE_TYPE; pr <KING; ++pr)
                if (huffman_table[pr].code == code
                    && huffman_table[pr].bits == bits)
                    goto Found;
        }
    Found:;
        if (pr == NO_PIECE_TYPE)
            return NO_PIECE;
        // first and second flag
        Color c = (Color)stream.read_one_bit();
        return make_piece(c, pr);
    }
    int set_from_packed_sfen(Position& pos, const PackedSfen& sfen, StateInfo* si, Thread* th, bool frc)
    {
        SfenPacker packer;
        auto& stream = packer.stream;
        // TODO: separate streams for writing and reading. Here we actually have to
        // const_cast which is not safe in the long run.
        stream.set_data(const_cast<uint8_t*>(reinterpret_cast<const uint8_t*>(&sfen)));
        pos.clear();
        std::memset(si, 0, sizeof(StateInfo));
        si->accumulator.computed[WHITE] = false;
        si->accumulator.computed[BLACK] = false;
        pos.st = si;
        // Active color
        pos.sideToMove = (Color)stream.read_one_bit();
        // First the position of the ball
        for (auto c : Colors)
            pos.board[stream.read_n_bit(6)] = make_piece(c, KING);
        // Piece placement
        for (Rank r = RANK_8; r >= RANK_1; --r)
        {
            for (File f = FILE_A; f <= FILE_H; ++f)
            {
                auto sq = make_square(f, r);
                // it seems there are already balls
                Piece pc;
                if (type_of(pos.board[sq]) != KING)
                {
                    assert(pos.board[sq] == NO_PIECE);
                    pc = packer.read_board_piece_from_stream();
                }
                else
                {
                    pc = pos.board[sq];
                    // put_piece() will catch ASSERT unless you remove it all.
                    pos.board[sq] = NO_PIECE;
                }
                // There may be no pieces, so skip in that case.
                if (pc == NO_PIECE)
                    continue;
                pos.put_piece(Piece(pc), sq);
                if (stream.get_cursor()> 256)
                    return 1;
            }
        }
        // Castling availability.
        pos.st->castlingRights = 0;
        if (stream.read_one_bit()) {
            Square rsq;
            for (rsq = relative_square(WHITE, SQ_H1); pos.piece_on(rsq) != W_ROOK; --rsq) {}
            pos.set_castling_right(WHITE, rsq);
        }
        if (stream.read_one_bit()) {
            Square rsq;
            for (rsq = relative_square(WHITE, SQ_A1); pos.piece_on(rsq) != W_ROOK; ++rsq) {}
            pos.set_castling_right(WHITE, rsq);
        }
        if (stream.read_one_bit()) {
            Square rsq;
            for (rsq = relative_square(BLACK, SQ_H1); pos.piece_on(rsq) != B_ROOK; --rsq) {}
            pos.set_castling_right(BLACK, rsq);
        }
        if (stream.read_one_bit()) {
            Square rsq;
            for (rsq = relative_square(BLACK, SQ_A1); pos.piece_on(rsq) != B_ROOK; ++rsq) {}
            pos.set_castling_right(BLACK, rsq);
        }
        // En passant square. Ignore if no pawn capture is possible
        if (stream.read_one_bit()) {
            Square ep_square = static_cast<Square>(stream.read_n_bit(6));
            pos.st->epSquare = ep_square;
            if (!(pos.attackers_to(pos.st->epSquare) & pos.pieces(pos.sideToMove, PAWN))
                || !(pos.pieces(~pos.sideToMove, PAWN) & (pos.st->epSquare + pawn_push(~pos.sideToMove))))
                pos.st->epSquare = SQ_NONE;
        }
        else {
            pos.st->epSquare = SQ_NONE;
        }
        // Halfmove clock
        pos.st->rule50 = stream.read_n_bit(6);
        // Fullmove number
        pos.gamePly = stream.read_n_bit(8);
        // Read the highest bit of rule50. This was added as a fix for rule50
        // counter having only 6 bits stored.
        // In older entries this will just be a zero bit.
        pos.gamePly |= stream.read_n_bit(8) << 8;
        // Read the highest bit of rule50. This was added as a fix for rule50
        // counter having only 6 bits stored.
        // In older entries this will just be a zero bit.
        pos.st->rule50 |= stream.read_n_bit(1) << 6;
        // Convert from fullmove starting from 1 to gamePly starting from 0,
        // handle also common incorrect FEN with fullmove = 0.
        pos.gamePly = std::max(2 * (pos.gamePly - 1), 0) + (pos.sideToMove == BLACK);
        assert(stream.get_cursor() <= 256);
        pos.chess960 = frc;
        pos.thisThread = th;
        pos.set_state(pos.st);
        assert(pos.pos_is_ok());
        return 0;
    }
    PackedSfen sfen_pack(Position& pos, bool resetCastlingRights)
    {
        PackedSfen sfen;
        SfenPacker sp;
        sp.data = (uint8_t*)&sfen;
        sp.pack(pos, resetCastlingRights);
        return sfen;
    }
 }
@@ -0,0 +1,22 @@
 #ifndef _SFEN_PACKER_H_
 #define _SFEN_PACKER_H_
 #include "types.h"
 #include "packed_sfen.h"
 #include <cstdint>
 namespace Stockfish {
    class Position;
    struct StateInfo;
    class Thread;
 }
 namespace Stockfish::Tools {
    int set_from_packed_sfen(Position& pos, const PackedSfen& sfen, StateInfo* si, Thread* th, bool frc);
    PackedSfen sfen_pack(Position& pos, bool resetCastlingRights);
 }
 #endif
@@ -0,0 +1,352 @@
 #include "sfen_stream.h"
 #include "packed_sfen.h"
 #include "misc.h"
 #include <string>
 #include <vector>
 #include <deque>
 #include <memory>
 #include <mutex>
 #include <list>
 #include <atomic>
 #include <optional>
 #include <iostream>
 #include <cstdint>
 #include <thread>
 #include <functional>
 namespace Stockfish::Tools{
    enum struct SfenReaderMode
    {
        Sequential,
        Cyclic
    };
    // Sfen reader
    struct SfenReader
    {
        // Number of phases buffered by each thread 0.1M phases. 4M phase at 40HT
        static constexpr size_t DEFAULT_THREAD_BUFFER_SIZE = 10 * 1000;
        // Buffer for reading files (If this is made larger,
        // the shuffle becomes larger and the phases may vary.
        // If it is too large, the memory consumption will increase.
        // SFEN_READ_SIZE is a multiple of THREAD_BUFFER_SIZE.
        static constexpr const size_t DEFAULT_SFEN_READ_SIZE = 1000 * 1000 * 10;
        // Do not use std::random_device().
        // Because it always the same integers on MinGW.
        SfenReader(
            const std::vector<std::string>& filenames_,
            bool do_shuffle,
            SfenReaderMode mode_,
            int thread_num,
            const std::string& seed,
            size_t read_size = DEFAULT_SFEN_READ_SIZE,
            size_t buffer_size = DEFAULT_THREAD_BUFFER_SIZE
        ) :
            filenames(filenames_.begin(), filenames_.end()),
            mode(mode_),
            // Due to the implementation of waiting for buffer empty a bit
            // the read size must be at least twice the buffer size.
            sfen_read_size(std::max(read_size, buffer_size * 2)),
            thread_buffer_size(buffer_size),
            prng(seed)
        {
            packed_sfens.resize(thread_num);
            total_read = 0;
            end_of_files = false;
            shuffle = do_shuffle;
            stop_flag = false;
            num_buffers_in_pool.store(0);
            file_worker_thread = std::thread([&] {
                this->file_read_worker();
            });
        }
        ~SfenReader()
        {
            stop_flag = true;
            if (file_worker_thread.joinable())
                file_worker_thread.join();
        }
        // Load the phase for calculation such as mse.
        PSVector read_some(uint64_t count, uint64_t count_tries, std::function<bool(const PackedSfenValue&)> do_take)
        {
            PSVector psv;
            psv.reserve(count);
            for (uint64_t i = 0; i < count_tries; ++i)
            {
                PackedSfenValue ps;
                if (!read_to_thread_buffer(0, ps))
                {
                    std::cout << "ERROR (sfen_reader): Reading failed." << std::endl;
                    return psv;
                }
                if (do_take(ps))
                {
                    psv.push_back(ps);
                    if (psv.size() >= count)
                        break;
                }
            }
            return psv;
        }
        // [ASYNC] Thread returns one aspect. Otherwise returns false.
        bool read_to_thread_buffer(size_t thread_id, PackedSfenValue& ps)
        {
            // If there are any positions left in the thread buffer
            // then retrieve one and return it.
            auto& thread_ps = packed_sfens[thread_id];
            // Fill the read buffer if there is no remaining buffer,
            // but if it doesn't even exist, finish.
            // If the buffer is empty, fill it.
            if ((thread_ps == nullptr || thread_ps->empty())
                && !read_to_thread_buffer_impl(thread_id))
                return false;
            // read_to_thread_buffer_impl() returned true,
            // Since the filling of the thread buffer with the
            // phase has been completed successfully
            // thread_ps->rbegin() is alive.
            ps = thread_ps->back();
            thread_ps->pop_back();
            // If you've run out of buffers, call delete yourself to free this buffer.
            if (thread_ps->empty())
            {
                thread_ps.reset();
            }
            return true;
        }
        // [ASYNC] Read some aspects into thread buffer.
        bool read_to_thread_buffer_impl(size_t thread_id)
        {
            while (true)
            {
                {
                    std::unique_lock<std::mutex> lk(mutex);
                    // If you can fill from the file buffer, that's fine.
                    if (packed_sfens_pool.size() != 0)
                    {
                        // It seems that filling is possible, so fill and finish.
                        packed_sfens[thread_id] = std::move(packed_sfens_pool.front());
                        packed_sfens_pool.pop_front();
                        num_buffers_in_pool.fetch_sub(1);
                        total_read += thread_buffer_size;
                        return true;
                    }
                }
                // The file to read is already gone. No more use.
                if (end_of_files)
                    return false;
                // Waiting for file worker to fill packed_sfens_pool.
                // The mutex isn't locked, so it should fill up soon.
                // Poor man's condition variable.
                sleep(1);
            }
        }
        void file_read_worker()
        {
            std::string currentFilename;
            uint64_t numEntriesReadFromCurrentFile = 0;
            auto open_next_file = [&]() {
                // no more
                for(;;)
                {
                    sfen_input_stream.reset();
                    if (filenames.empty())
                        return false;
                    // Get the next file name.
                    currentFilename = filenames.front();
                    filenames.pop_front();
                    numEntriesReadFromCurrentFile = 0;
                    sfen_input_stream = open_sfen_input_file(currentFilename);
                    auto out = sync_region_cout.new_region();
                    if (sfen_input_stream == nullptr)
                    {
                        out << "INFO (sfen_reader): File does not exist: " << currentFilename << '\n';
                    }
                    else
                    {
                        out << "INFO (sfen_reader): Opened file for reading: " << currentFilename << '\n';
                        // in case the file is empty or was deleted.
                        if (sfen_input_stream->eof())
                        {
                            out << "  - File empty, nothing to read.\n";
                        }
                        else
                        {
                            return true;
                        }
                    }
                }
            };
            if (sfen_input_stream == nullptr && !open_next_file())
            {
                auto out = sync_region_cout.new_region();
                out << "INFO (sfen_reader): End of files." << std::endl;
                end_of_files = true;
                return;
            }
            // We want to set the `end_of_files` only after we read everything AND copy to the buffer pool.
            bool local_end_of_files = false;
            while (!local_end_of_files)
            {
                // Wait for the buffer to run out.
                // This size() is read only, so you don't need to lock it.
                while (!stop_flag && num_buffers_in_pool.load() >= sfen_read_size / thread_buffer_size)
                    sleep(100);
                if (stop_flag)
                    return;
                PSVector sfens;
                sfens.reserve(sfen_read_size);
                // Read from the file into the file buffer.
                while (sfens.size() < sfen_read_size)
                {
                    std::optional<PackedSfenValue> p = sfen_input_stream->next();
                    if (p.has_value())
                    {
                        sfens.push_back(*p);
                        ++numEntriesReadFromCurrentFile;
                    }
                    else
                    {
                        if (mode == SfenReaderMode::Cyclic
                            && numEntriesReadFromCurrentFile > 0)
                        {
                            // The file contained data so we add it again to the end of the queue.
                            filenames.emplace_back(currentFilename);
                        }
                        if(!open_next_file())
                        {
                            // There was no next file. Abort.
                            auto out = sync_region_cout.new_region();
                            out << "INFO (sfen_reader): End of files." << std::endl;
                            local_end_of_files = true;
                            break;
                        }
                    }
                }
                // Shuffle the read phase data.
                if (shuffle)
                {
                    Algo::shuffle(sfens, prng);
                }
                std::vector<std::unique_ptr<PSVector>> buffers;
                for (size_t offset = 0; offset < sfens.size(); offset += thread_buffer_size)
                {
                    const size_t count =
                        offset + thread_buffer_size > sfens.size()
                        ? sfens.size() - offset
                        : thread_buffer_size;
                    // Delete this pointer on the receiving side.
                    auto buf = std::make_unique<PSVector>();
                    buf->resize(count);
                    memcpy(
                        buf->data(),
                        &sfens[offset],
                        sizeof(PackedSfenValue) * count);
                    buffers.emplace_back(std::move(buf));
                }
                {
                    std::unique_lock<std::mutex> lk(mutex);
                    // The mutex lock is required because the%
                    // contents of packed_sfens_pool are changed.
                    for (auto& buf : buffers)
                    {
                        num_buffers_in_pool.fetch_add(1);
                        packed_sfens_pool.emplace_back(std::move(buf));
                    }
                }
            }
            end_of_files = true;
        }
    protected:
        // worker thread reading file in background
        std::thread file_worker_thread;
        // sfen files
        std::deque<std::string> filenames;
        std::atomic<bool> stop_flag;
        // number of phases read (file to memory buffer)
        std::atomic<uint64_t> total_read;
        // Do not shuffle when reading the phase.
        bool shuffle;
        SfenReaderMode mode;
        size_t sfen_read_size;
        size_t thread_buffer_size;
        // Random number to shuffle when reading the phase
        PRNG prng;
        // Did you read the files and reached the end?
        std::atomic<bool> end_of_files;
        // handle of sfen file
        std::unique_ptr<BasicSfenInputStream> sfen_input_stream;
        // sfen for each thread
        // (When the thread is used up, the thread should call delete to release it.)
        std::vector<std::unique_ptr<PSVector>> packed_sfens;
        // Mutex when accessing packed_sfens_pool
        std::mutex mutex;
        // pool of sfen. The worker thread read from the file is added here.
        // Each worker thread fills its own packed_sfens[thread_id] from here.
        // * Lock and access the mutex.
        std::list<std::unique_ptr<PSVector>> packed_sfens_pool;
        std::atomic<size_t> num_buffers_in_pool;
    };
 }
@@ -0,0 +1,222 @@
 #ifndef _SFEN_STREAM_H_
 #define _SFEN_STREAM_H_
 #include "packed_sfen.h"
 #include "extra/nnue_data_binpack_format.h"
 #include <optional>
 #include <fstream>
 #include <string>
 #include <memory>
 namespace Stockfish::Tools {
    enum struct SfenOutputType
    {
        Bin,
        Binpack
    };
    static bool ends_with(const std::string& lhs, const std::string& end)
    {
        if (end.size() > lhs.size()) return false;
        return std::equal(end.rbegin(), end.rend(), lhs.rbegin());
    }
    static bool has_extension(const std::string& filename, const std::string& extension)
    {
        return ends_with(filename, "." + extension);
    }
    static std::string filename_with_extension(const std::string& filename, const std::string& ext)
    {
        if (ends_with(filename, ext))
        {
            return filename;
        }
        else
        {
            return filename + "." + ext;
        }
    }
    struct BasicSfenInputStream
    {
        virtual std::optional<PackedSfenValue> next() = 0;
        virtual bool eof() const = 0;
        virtual ~BasicSfenInputStream() {}
    };
    struct BinSfenInputStream : BasicSfenInputStream
    {
        static constexpr auto openmode = std::ios::in | std::ios::binary;
        static inline const std::string extension = "bin";
        BinSfenInputStream(std::string filename) :
            m_stream(filename, openmode),
            m_eof(!m_stream)
        {
        }
        std::optional<PackedSfenValue> next() override
        {
            PackedSfenValue e;
            if(m_stream.read(reinterpret_cast<char*>(&e), sizeof(PackedSfenValue)))
            {
                return e;
            }
            else
            {
                m_eof = true;
                return std::nullopt;
            }
        }
        bool eof() const override
        {
            return m_eof;
        }
        ~BinSfenInputStream() override {}
    private:
        std::fstream m_stream;
        bool m_eof;
    };
    struct BinpackSfenInputStream : BasicSfenInputStream
    {
        static constexpr auto openmode = std::ios::in | std::ios::binary;
        static inline const std::string extension = "binpack";
        BinpackSfenInputStream(std::string filename) :
            m_stream(filename, openmode),
            m_eof(!m_stream.hasNext())
        {
        }
        std::optional<PackedSfenValue> next() override
        {
            static_assert(sizeof(binpack::nodchip::PackedSfenValue) == sizeof(PackedSfenValue));
            if (!m_stream.hasNext())
            {
                m_eof = true;
                return std::nullopt;
            }
            auto training_data_entry = m_stream.next();
            auto v = binpack::trainingDataEntryToPackedSfenValue(training_data_entry);
            PackedSfenValue psv;
            // same layout, different types. One is from generic library.
            std::memcpy(&psv, &v, sizeof(PackedSfenValue));
            return psv;
        }
        bool eof() const override
        {
            return m_eof;
        }
        ~BinpackSfenInputStream() override {}
    private:
        binpack::CompressedTrainingDataEntryReader m_stream;
        bool m_eof;
    };
    struct BasicSfenOutputStream
    {
        virtual void write(const PSVector& sfens) = 0;
        virtual ~BasicSfenOutputStream() {}
    };
    struct BinSfenOutputStream : BasicSfenOutputStream
    {
        static constexpr auto openmode = std::ios::out | std::ios::binary | std::ios::app;
        static inline const std::string extension = "bin";
        BinSfenOutputStream(std::string filename) :
            m_stream(filename_with_extension(filename, extension), openmode)
        {
        }
        void write(const PSVector& sfens) override
        {
            m_stream.write(reinterpret_cast<const char*>(sfens.data()), sizeof(PackedSfenValue) * sfens.size());
        }
        ~BinSfenOutputStream() override {}
    private:
        std::fstream m_stream;
    };
    struct BinpackSfenOutputStream : BasicSfenOutputStream
    {
        static constexpr auto openmode = std::ios::out | std::ios::binary | std::ios::app;
        static inline const std::string extension = "binpack";
        BinpackSfenOutputStream(std::string filename) :
            m_stream(filename_with_extension(filename, extension), openmode)
        {
        }
        void write(const PSVector& sfens) override
        {
            static_assert(sizeof(binpack::nodchip::PackedSfenValue) == sizeof(PackedSfenValue));
            for(auto& sfen : sfens)
            {
                // The library uses a type that's different but layout-compatibile.
                binpack::nodchip::PackedSfenValue e;
                std::memcpy(&e, &sfen, sizeof(binpack::nodchip::PackedSfenValue));
                m_stream.addTrainingDataEntry(binpack::packedSfenValueToTrainingDataEntry(e));
            }
        }
        ~BinpackSfenOutputStream() override {}
    private:
        binpack::CompressedTrainingDataEntryWriter m_stream;
    };
    inline std::unique_ptr<BasicSfenInputStream> open_sfen_input_file(const std::string& filename)
    {
        if (has_extension(filename, BinSfenInputStream::extension))
            return std::make_unique<BinSfenInputStream>(filename);
        else if (has_extension(filename, BinpackSfenInputStream::extension))
            return std::make_unique<BinpackSfenInputStream>(filename);
        return nullptr;
    }
    inline std::unique_ptr<BasicSfenOutputStream> create_new_sfen_output(const std::string& filename, SfenOutputType sfen_output_type)
    {
        switch(sfen_output_type)
        {
            case SfenOutputType::Bin:
                return std::make_unique<BinSfenOutputStream>(filename);
            case SfenOutputType::Binpack:
                return std::make_unique<BinpackSfenOutputStream>(filename);
        }
        assert(false);
        return nullptr;
    }
    inline std::unique_ptr<BasicSfenOutputStream> create_new_sfen_output(const std::string& filename)
    {
        if (has_extension(filename, BinSfenOutputStream::extension))
            return std::make_unique<BinSfenOutputStream>(filename);
        else if (has_extension(filename, BinpackSfenOutputStream::extension))
            return std::make_unique<BinpackSfenOutputStream>(filename);
        return nullptr;
    }
 }
 #endif
@@ -0,0 +1,203 @@
 #include "packed_sfen.h"
 #include "sfen_stream.h"
 #include "misc.h"
 #include "extra/nnue_data_binpack_format.h"
 #include "syzygy/tbprobe.h"
 #include <cstring>
 #include <fstream>
 #include <limits>
 #include <list>
 #include <memory>
 #include <optional>
 #include <shared_mutex>
 #include <thread>
 #include <atomic>
 namespace Stockfish::Tools {
    // Helper class for exporting Sfen
    struct SfenWriter
    {
        // Amount of sfens required to flush the buffer.
        static constexpr size_t SFEN_WRITE_SIZE = 5000;
        // File name to write and number of threads to create
        SfenWriter(std::string filename_, int thread_num, uint64_t save_count, SfenOutputType sfen_output_type)
        {
            sfen_buffers_pool.reserve((size_t)thread_num * 10);
            sfen_buffers.resize(thread_num);
            auto out = sync_region_cout.new_region();
            out << "INFO (sfen_writer): Creating new data file at " << filename_ << std::endl;
            sfen_format = sfen_output_type;
            output_file_stream = create_new_sfen_output(filename_, sfen_format);
            filename = filename_;
            save_every = save_count;
            finished = false;
            file_worker_thread = std::thread([&] { this->file_write_worker(); });
        }
        ~SfenWriter()
        {
            flush();
            finished = true;
            file_worker_thread.join();
            output_file_stream.reset();
 #if !defined(NDEBUG)
            {
                // All buffers should be empty since file_worker_thread
                // should have written everything before exiting.
                for (const auto& p : sfen_buffers) { assert(p == nullptr); (void)p ; }
                assert(sfen_buffers_pool.empty());
            }
 #endif
        }
        void write(size_t thread_id, const PackedSfenValue& psv)
        {
            // We have a buffer for each thread and add it there.
            // If the buffer overflows, write it to a file.
            // This buffer is prepared for each thread.
            auto& buf = sfen_buffers[thread_id];
            // Secure since there is no buf at the first time
            // and immediately after writing the thread buffer.
            if (!buf)
            {
                buf = std::make_unique<PSVector>();
                buf->reserve(SFEN_WRITE_SIZE);
            }
            // Buffer is exclusive to this thread.
            // There is no need for a critical section.
            buf->push_back(psv);
            if (buf->size() >= SFEN_WRITE_SIZE)
            {
                // If you load it in sfen_buffers_pool, the worker will do the rest.
                // Critical section since sfen_buffers_pool is shared among threads.
                std::unique_lock<std::mutex> lk(mutex);
                sfen_buffers_pool.emplace_back(std::move(buf));
            }
        }
        void flush()
        {
            for (size_t i = 0; i < sfen_buffers.size(); ++i)
            {
                flush(i);
            }
        }
        // Move what remains in the buffer for your thread to a buffer for writing to a file.
        void flush(size_t thread_id)
        {
            std::unique_lock<std::mutex> lk(mutex);
            auto& buf = sfen_buffers[thread_id];
            // There is a case that buf==nullptr, so that check is necessary.
            if (buf && buf->size() != 0)
            {
                sfen_buffers_pool.emplace_back(std::move(buf));
            }
        }
        // Dedicated thread to write to file
        void file_write_worker()
        {
            while (!finished || sfen_buffers_pool.size())
            {
                std::vector<std::unique_ptr<PSVector>> buffers;
                {
                    std::unique_lock<std::mutex> lk(mutex);
                    // Atomically swap take the filled buffers and
                    // create a new buffer pool for threads to fill.
                    buffers = std::move(sfen_buffers_pool);
                    sfen_buffers_pool = std::vector<std::unique_ptr<PSVector>>();
                }
                if (!buffers.size())
                {
                    // Poor man's condition variable.
                    sleep(100);
                }
                else
                {
                    for (auto& buf : buffers)
                    {
                        output_file_stream->write(*buf);
                        sfen_write_count += buf->size();
                        // Add the processed number here, and if it exceeds save_every,
                        // change the file name and reset this counter.
                        sfen_write_count_current_file += buf->size();
                        if (sfen_write_count_current_file >= save_every)
                        {
                            sfen_write_count_current_file = 0;
                            // Sequential number attached to the file
                            int n = (int)(sfen_write_count / save_every);
                            // Rename the file and open it again.
                            // Add ios::app in consideration of overwriting.
                            // (Depending on the operation, it may not be necessary.)
                            std::string new_filename = filename + "_" + std::to_string(n);
                            output_file_stream = create_new_sfen_output(new_filename, sfen_format);
                            auto out = sync_region_cout.new_region();
                            out << "INFO (sfen_writer): Creating new data file at " << new_filename << std::endl;
                        }
                    }
                }
            }
        }
    private:
        std::unique_ptr<BasicSfenOutputStream> output_file_stream;
        // A new net is saved after every save_every sfens are processed.
        uint64_t save_every = std::numeric_limits<uint64_t>::max();
        // File name passed in the constructor
        std::string filename;
        // Thread to write to the file
        std::thread file_worker_thread;
        // Flag that all threads have finished
        std::atomic<bool> finished;
        SfenOutputType sfen_format;
        // buffer before writing to file
        // sfen_buffers is the buffer for each thread
        // sfen_buffers_pool is a buffer for writing.
        // After loading the phase in the former buffer by SFEN_WRITE_SIZE,
        // transfer it to the latter.
        std::vector<std::unique_ptr<PSVector>> sfen_buffers;
        std::vector<std::unique_ptr<PSVector>> sfen_buffers_pool;
        // Mutex required to access sfen_buffers_pool
        std::mutex mutex;
        // Number of sfens written in total, and the
        // number of sfens written in the current file.
        uint64_t sfen_write_count = 0;
        uint64_t sfen_write_count_current_file = 0;
    };
 }
@@ -0,0 +1,12 @@
 #ifndef _STATS_H_
 #define _STATS_H_
 #include <sstream>
 namespace Stockfish::Tools::Stats {
    void gather_statistics(std::istringstream& is);
 }
 #endif
@@ -0,0 +1,952 @@
 #include "training_data_generator.h"
 #include "sfen_writer.h"
 #include "packed_sfen.h"
 #include "opening_book.h"
 #include "misc.h"
 #include "position.h"
 #include "thread.h"
 #include "tt.h"
 #include "uci.h"
 #include "extra/nnue_data_binpack_format.h"
 #include "nnue/evaluate_nnue.h"
 #include "syzygy/tbprobe.h"
 #include <atomic>
 #include <chrono>
 #include <climits>
 #include <cmath>
 #include <cstring>
 #include <fstream>
 #include <iomanip>
 #include <limits>
 #include <list>
 #include <memory>
 #include <optional>
 #include <random>
 #include <shared_mutex>
 #include <sstream>
 #include <unordered_set>
 using namespace std;
 namespace Stockfish::Tools
 {
    // Class to generate sfen with multiple threads
    struct TrainingDataGenerator
    {
        struct Params
        {
            // Min and max depths for search during gensfen
            int search_depth_min = 3;
            int search_depth_max = -1;
            // Number of the nodes to be searched.
            // 0 represents no limits.
            uint64_t nodes = 0;
            // Upper limit of evaluation value of generated situation
            int eval_limit = 3000;
            // minimum ply with random move
            // maximum ply with random move
            // Number of random moves in one station
            int random_move_minply = 1;
            int random_move_maxply = 24;
            int random_move_count = 5;
            // Move kings with a probability of 1/N when randomly moving like Apery software.
            // When you move the king again, there is a 1/N chance that it will randomly moved
            // once in the opponent's turn.
            // Apery has N=2. Specifying 0 here disables this function.
            int random_move_like_apery = 0;
            // For when using multi pv instead of random move.
            // random_multi_pv is the number of candidates for MultiPV.
            // When adopting the move of the candidate move, the difference
            // between the evaluation value of the move of the 1st place
            // and the evaluation value of the move of the Nth place is.
            // Must be in the range random_multi_pv_diff.
            // random_multi_pv_depth is the search depth for MultiPV.
            int random_multi_pv = 0;
            int random_multi_pv_diff = 32000;
            int random_multi_pv_depth = -1;
            uint64_t random_multi_pv_nodes = 0;
            // The minimum and maximum ply (number of steps from
            // the initial phase) of the sfens to write out.
            int write_minply = 16;
            int write_maxply = 400;
            uint64_t save_every = std::numeric_limits<uint64_t>::max();
            std::string output_file_name = "training_data";
            SfenOutputType sfen_format = SfenOutputType::Binpack;
            std::string seed;
            bool write_out_draw_game_in_training_data_generation = true;
            bool detect_draw_by_consecutive_low_score = true;
            bool detect_draw_by_insufficient_mating_material = true;
            uint64_t num_threads;
            std::string book;
            void enforce_constraints()
            {
                search_depth_max = std::max(search_depth_min, search_depth_max);
                // Limit the maximum to a one-stop score. (Otherwise you might not end the loop)
                eval_limit = std::min(eval_limit, (int)mate_in(2));
                save_every = std::max(save_every, REPORT_STATS_EVERY);
                num_threads = Options["Threads"];
                if (random_multi_pv_depth == -1)
                    random_multi_pv_depth = search_depth_max;
                if (random_multi_pv_nodes == 0)
                    random_multi_pv_nodes = nodes;
            }
        };
        // Hash to limit the export of identical sfens
        static constexpr uint64_t GENSFEN_HASH_SIZE = 64 * 1024 * 1024;
        // It must be 2**N because it will be used as the mask to calculate hash_index.
        static_assert((GENSFEN_HASH_SIZE& (GENSFEN_HASH_SIZE - 1)) == 0);
        static constexpr uint64_t REPORT_DOT_EVERY = 5000;
        static constexpr uint64_t REPORT_STATS_EVERY = 200000;
        static_assert(REPORT_STATS_EVERY % REPORT_DOT_EVERY == 0);
        TrainingDataGenerator(
            const Params& prm
        ) :
            params(prm),
            sfen_writer(prm.output_file_name, prm.num_threads, prm.save_every, prm.sfen_format)
        {
            hash.resize(GENSFEN_HASH_SIZE);
            prngs.reserve(prm.num_threads);
            auto seed = prm.seed;
            for (uint64_t i = 0; i < prm.num_threads; ++i)
            {
                prngs.emplace_back(seed);
                seed = prngs.back().next_random_seed();
            }
            if (!prm.book.empty())
            {
                opening_book = open_opening_book(prm.book, prngs[0]);
                if (opening_book == nullptr)
                {
                    std::cout << "WARNING: Failed to open opening book " << prm.book << ". Falling back to startpos.\n";
                }
            }
            // Output seed to veryfy by the user if it's not identical by chance.
            std::cout << prngs[0] << std::endl;
        }
        void generate(uint64_t limit, uint64_t limit_seconds);
    private:
        Params params;
        std::vector<PRNG> prngs;
        std::mutex stats_mutex;
        TimePoint last_stats_report_time;
        // sfen exporter
        SfenWriter sfen_writer;
        SynchronizedRegionLogger::Region out;
        vector<Key> hash; // 64MB*sizeof(HASH_KEY) = 512MB
        std::unique_ptr<OpeningBook> opening_book;
        static void set_gensfen_search_limits();
        void generate_worker(
            Thread& th,
            std::atomic<uint64_t>& counter,
            uint64_t limit,
            uint64_t limit_seconds);
        bool was_seen_before(const Position& pos);
        optional<int8_t> get_current_game_result(
            Position& pos,
            const vector<int>& move_hist_scores) const;
        vector<uint8_t> generate_random_move_flags(PRNG& prng);
        optional<Move> choose_random_move(
            PRNG& prng,
            Position& pos,
            std::vector<uint8_t>& random_move_flag,
            int ply,
            int& random_move_c);
        bool commit_psv(
            Thread& th,
            PSVector& sfens,
            int8_t lastTurnIsWin,
            std::atomic<uint64_t>& counter,
            uint64_t limit,
            Color result_color);
        void initial_report();
        void report(uint64_t done, uint64_t new_done);
        void maybe_report(uint64_t done);
    };
    void TrainingDataGenerator::set_gensfen_search_limits()
    {
        // About Search::Limits
        // Be careful because this member variable is global and affects other threads.
        auto& limits = Search::Limits;
        // Make the search equivalent to the "go infinite" command. (Because it is troublesome if time management is done)
        limits.infinite = true;
        // Since PV is an obstacle when displayed, erase it.
        limits.silent = true;
        // If you use this, it will be compared with the accumulated nodes of each thread. Therefore, do not use it.
        limits.nodes = 0;
        // depth is also processed by the one passed as an argument of Tools::search().
        limits.depth = 0;
    }
    void TrainingDataGenerator::generate(uint64_t limit, uint64_t limit_seconds)
    {
        last_stats_report_time = 0;
        initial_report();
        set_gensfen_search_limits();
        std::atomic<uint64_t> counter{0};
        Threads.execute_with_workers([&counter, limit, limit_seconds, this](Thread& th) {
            generate_worker(th, counter, limit, limit_seconds);
        });
        Threads.wait_for_workers_finished();
        sfen_writer.flush();
        report(counter.load(), counter.load() % REPORT_STATS_EVERY + 1);
        std::cout << std::endl;
    }
    void TrainingDataGenerator::generate_worker(
        Thread& th,
        std::atomic<uint64_t>& counter,
        uint64_t limit,
        uint64_t limit_seconds)
    {
        // For the time being, it will be treated as a draw
        // at the maximum number of steps to write.
        // Maximum StateInfo + Search PV to advance to leaf buffer
        std::vector<StateInfo, AlignedAllocator<StateInfo>> states(
            params.write_maxply + MAX_PLY /* == search_depth_min + α */);
        StateInfo si;
        auto& prng = prngs[th.id()];
        // end flag
        bool quit = false;
        const auto start_time = now();
        const bool frc = Options["UCI_Chess960"];
        // repeat until the specified number of times
        while (!quit)
        {
            if (static_cast<uint64_t>(now() - start_time) / 1000 >= limit_seconds)
            {
                break;
            }
            // It is necessary to set a dependent thread for Position.
            // When parallelizing, Threads (since this is a vector<Thread*>,
            // Do the same for up to Threads[0]...Threads[thread_num-1].
            auto& pos = th.rootPos;
            if (opening_book != nullptr)
            {
                auto& fen = opening_book->next_fen();
                pos.set(fen, frc, &si, &th);
            }
            else
            {
                pos.set(StartFEN, frc, &si, &th);
            }
            int resign_counter = 0;
            bool should_resign = prng.rand(10) > 1;
            // Vector for holding the sfens in the current simulated game.
            PSVector packed_sfens;
            packed_sfens.reserve(params.write_maxply + MAX_PLY);
            // Precomputed flags. Used internally by choose_random_move.
            vector<uint8_t> random_move_flag = generate_random_move_flags(prng);
            // A counter that keeps track of the number of random moves
            // When random_move_minply == -1, random moves are
            // performed continuously, so use it at this time.
            // Used internally by choose_random_move.
            int actual_random_move_count = 0;
            // Save history of move scores for adjudication
            vector<int> move_hist_scores;
            auto flush_psv = [&](int8_t result) {
                quit = commit_psv(th, packed_sfens, result, counter, limit, pos.side_to_move());
            };
            for (int ply = 0; ; ++ply)
            {
                // Current search depth
                const int depth = params.search_depth_min + (int)prng.rand(params.search_depth_max - params.search_depth_min + 1);
                // Starting search calls init_for_search
                auto [search_value, search_pv] = Search::search(pos, depth, 1, params.nodes);
                // This has to be performed after search because it needs to know
                // rootMoves which are filled in init_for_search.
                const auto result = get_current_game_result(pos, move_hist_scores);
                if (result.has_value())
                {
                    flush_psv(result.value());
                    break;
                }
                // Always adjudivate by eval limit.
                // Also because of this we don't have to check for TB/MATE scores
                if (abs(search_value) >= params.eval_limit)
                {
                    resign_counter++;
                    if ((should_resign && resign_counter >= 4) || abs(search_value) >= VALUE_KNOWN_WIN) {
                        flush_psv((search_value >= params.eval_limit) ? 1 : -1);
                        break;
                    }
                }
                else
                {
                    resign_counter = 0;
                }
                // In case there is no PV and the game was not ended here
                // there is nothing we can do, we can't continue the game,
                // we don't know the result, so discard this game.
                if (search_pv.empty())
                {
                    break;
                }
                // Save the move score for adjudication.
                move_hist_scores.push_back(search_value);
                // Discard stuff before write_minply is reached
                // because it can harm training due to overfitting.
                // Initial positions would be too common.
                if (ply >= params.write_minply && !was_seen_before(pos))
                {
                    auto& psv = packed_sfens.emplace_back();
                    // Here we only write the position data.
                    // Result is added after the whole game is done.
                    pos.sfen_pack(psv.sfen, pos.is_chess960());
                    psv.score = search_value;
                    psv.move = search_pv[0];
                    psv.gamePly = ply;
                }
                // Update the next move according to best search result or random move.
                auto random_move = choose_random_move(prng, pos, random_move_flag, ply, actual_random_move_count);
                const Move next_move = random_move.has_value() ? *random_move : search_pv[0];
                // We don't have the whole game yet, but it ended,
                // so the writing process ends and the next game starts.
                // This shouldn't really happen.
                if (!is_ok(next_move))
                {
                    break;
                }
                // Do move.
                pos.do_move(next_move, states[ply]);
            }
        }
    }
    bool TrainingDataGenerator::was_seen_before(const Position& pos)
    {
        // Look into the position hashtable to see if the same
        // position was seen before.
        // This is a good heuristic to exlude already seen
        // positions without many false positives.
        auto key = pos.key();
        auto hash_index = (size_t)(key & (GENSFEN_HASH_SIZE - 1));
        auto old_key = hash[hash_index];
        if (key == old_key)
        {
            return true;
        }
        else
        {
            // Replace with the current key.
            hash[hash_index] = key;
            return false;
        }
    }
    optional<int8_t> TrainingDataGenerator::get_current_game_result(
        Position& pos,
        const vector<int>& move_hist_scores) const
    {
        // Variables for draw adjudication.
        // Todo: Make this as an option.
        // start the adjudication when ply reaches this value
        constexpr int adj_draw_ply = 80;
        // 4 move scores for each side have to be checked
        constexpr int adj_draw_cnt = 8;
        // move score in CP
        constexpr int adj_draw_score = 0;
        // For the time being, it will be treated as a
        // draw at the maximum number of steps to write.
        const int ply = move_hist_scores.size();
        // has it reached the max length or is a draw by fifty-move rule
        // or by 3-fold repetition
        if (ply >= params.write_maxply
            || pos.is_fifty_move_draw()
            || pos.is_three_fold_repetition())
        {
            return 0;
        }
        if(pos.this_thread()->rootMoves.empty())
        {
            // If there is no legal move
            return pos.checkers()
                ? -1 /* mate */
                : 0 /* stalemate */;
        }
        // Adjudicate game to a draw if the last 4 scores of each engine is 0.
        if (params.detect_draw_by_consecutive_low_score)
        {
            if (ply >= adj_draw_ply)
            {
                int num_cons_plies_within_draw_score = 0;
                bool is_adj_draw = false;
                for (auto it = move_hist_scores.rbegin();
                    it != move_hist_scores.rend(); ++it)
                {
                    if (abs(*it) <= adj_draw_score)
                    {
                        num_cons_plies_within_draw_score++;
                    }
                    else
                    {
                        // Draw scores must happen on consecutive plies
                        break;
                    }
                    if (num_cons_plies_within_draw_score >= adj_draw_cnt)
                    {
                        is_adj_draw = true;
                        break;
                    }
                }
                if (is_adj_draw)
                {
                    return 0;
                }
            }
        }
        // Draw by insufficient mating material
        if (params.detect_draw_by_insufficient_mating_material)
        {
            if (pos.count<ALL_PIECES>() <= 4)
            {
                int num_pieces = pos.count<ALL_PIECES>();
                // (1) KvK
                if (num_pieces == 2)
                {
                    return 0;
                }
                // (2) KvK + 1 minor piece
                if (num_pieces == 3)
                {
                    int minor_pc = pos.count<BISHOP>(WHITE) + pos.count<KNIGHT>(WHITE) +
                        pos.count<BISHOP>(BLACK) + pos.count<KNIGHT>(BLACK);
                    if (minor_pc == 1)
                    {
                        return 0;
                    }
                }
                // (3) KBvKB, bishops of the same color
                else if (num_pieces == 4)
                {
                    if (pos.count<BISHOP>(WHITE) == 1 && pos.count<BISHOP>(BLACK) == 1)
                    {
                        // Color of bishops is black.
                        if ((pos.pieces(WHITE, BISHOP) & DarkSquares)
                            && (pos.pieces(BLACK, BISHOP) & DarkSquares))
                        {
                            return 0;
                        }
                        // Color of bishops is white.
                        if ((pos.pieces(WHITE, BISHOP) & ~DarkSquares)
                            && (pos.pieces(BLACK, BISHOP) & ~DarkSquares))
                        {
                            return 0;
                        }
                    }
                }
            }
        }
        return nullopt;
    }
    vector<uint8_t> TrainingDataGenerator::generate_random_move_flags(PRNG& prng)
    {
        vector<uint8_t> random_move_flag;
        // Depending on random move selection parameters setup
        // the array of flags that indicates whether a random move
        // be taken at a given ply.
        // Make an array like a[0] = 0 ,a[1] = 1, ...
        // Fisher-Yates shuffle and take out the first N items.
        // Actually, I only want N pieces, so I only need
        // to shuffle the first N pieces with Fisher-Yates.
        vector<int> a;
        a.reserve((size_t)params.random_move_maxply);
        // random_move_minply ,random_move_maxply is specified by 1 origin,
        // Note that we are handling 0 origin here.
        for (int i = std::max(params.random_move_minply - 1, 0); i < params.random_move_maxply; ++i)
        {
            a.push_back(i);
        }
        // In case of Apery random move, insert() may be called random_move_count times.
        // Reserve only the size considering it.
        random_move_flag.resize((size_t)params.random_move_maxply + params.random_move_count);
        // A random move that exceeds the size() of a[] cannot be applied, so limit it.
        for (int i = 0; i < std::min(params.random_move_count, (int)a.size()); ++i)
        {
            swap(a[i], a[prng.rand((uint64_t)a.size() - i) + i]);
            random_move_flag[a[i]] = true;
        }
        return random_move_flag;
    }
    optional<Move> TrainingDataGenerator::choose_random_move(
        PRNG& prng,
        Position& pos,
        std::vector<uint8_t>& random_move_flag,
        int ply,
        int& random_move_c)
    {
        optional<Move> random_move;
        // Randomly choose one from legal move
        if (
            // 1. Random move of random_move_count times from random_move_minply to random_move_maxply
            (params.random_move_minply != -1 && ply < (int)random_move_flag.size() && random_move_flag[ply]) ||
            // 2. A mode to perform random move of random_move_count times after leaving the startpos
            (params.random_move_minply == -1 && random_move_c < params.random_move_count))
        {
            ++random_move_c;
            // It's not a mate, so there should be one legal move...
            if (params.random_multi_pv == 0)
            {
                // Normal random move
                MoveList<LEGAL> list(pos);
                // I don't really know the goodness and badness of making this the Apery method.
                if (params.random_move_like_apery == 0
                    || prng.rand(params.random_move_like_apery) != 0)
                {
                    // Normally one move from legal move
                    random_move = list.at((size_t)prng.rand((uint64_t)list.size()));
                }
                else
                {
                    // if you can move the king, move the king
                    Move moves[8]; // Near 8
                    Move* p = &moves[0];
                    for (auto& m : list)
                    {
                        if (type_of(pos.moved_piece(m)) == KING)
                        {
                            *(p++) = m;
                        }
                    }
                    size_t n = p - &moves[0];
                    if (n != 0)
                    {
                        // move to move the king
                        random_move = moves[prng.rand(n)];
                        // In Apery method, at this time there is a 1/2 chance
                        // that the opponent will also move randomly
                        if (prng.rand(2) == 0)
                        {
                            // Is it a simple hack to add a "1" next to random_move_flag[ply]?
                            random_move_flag.insert(random_move_flag.begin() + ply + 1, 1, true);
                        }
                    }
                    else
                    {
                        // Normally one move from legal move
                        random_move = list.at((size_t)prng.rand((uint64_t)list.size()));
                    }
                }
            }
            else
            {
                Search::search(pos, params.random_multi_pv_depth, params.random_multi_pv, params.random_multi_pv_nodes);
                // Select one from the top N hands of root Moves
                auto& rm = pos.this_thread()->rootMoves;
                uint64_t s = min((uint64_t)rm.size(), (uint64_t)params.random_multi_pv);
                for (uint64_t i = 1; i < s; ++i)
                {
                    // The difference from the evaluation value of rm[0] must
                    // be within the range of random_multi_pv_diff.
                    // It can be assumed that rm[x].score is arranged in descending order.
                    if (rm[0].score > rm[i].score + params.random_multi_pv_diff)
                    {
                        s = i;
                        break;
                    }
                }
                random_move = rm[prng.rand(s)].pv[0];
            }
        }
        return random_move;
    }
    // Write out the phases loaded in sfens to a file.
    // result: win/loss in the next phase after the final phase in sfens
    // 1 when winning. -1 when losing. Pass 0 for a draw.
    // Return value: true if the specified number of
    // sfens has already been reached and the process ends.
    bool TrainingDataGenerator::commit_psv(
        Thread& th,
        PSVector& sfens,
        int8_t result,
        std::atomic<uint64_t>& counter,
        uint64_t limit,
        Color result_color)
    {
        if (!params.write_out_draw_game_in_training_data_generation && result == 0)
        {
            // We didn't write anything so why quit.
            return false;
        }
        auto side_to_move_from_sfen = [](auto& sfen){
            return (Color)(sfen.sfen.data[0] & 1);
        };
        // From the final stage (one step before) to the first stage, give information on the outcome of the game for each stage.
        // The phases stored in sfens are assumed to be continuous (in order).
        for (auto it = sfens.rbegin(); it != sfens.rend(); ++it)
        {
            // The side to move is packed as the lowest bit of the first byte
            const Color side_to_move = side_to_move_from_sfen(*it);
            it->game_result = side_to_move == result_color ? result : -result;
        }
        const bool frc = th.rootPos.is_chess960();
        // Write sfens in move order to make potential compression easier
        for (auto& sfen : sfens)
        {
            // Skip positions with castling bestmove in FRC so that we don't
            // need to support it in the trainer.
            if (frc && type_of((Move)sfen.move) == CASTLING)
            {
                continue;
            }
            // Return true if there is already enough data generated.
            const auto iter = counter.fetch_add(1);
            if (iter >= limit)
                return true;
            // because `iter` was done, now we do one more
            maybe_report(iter + 1);
            // Write out one sfen.
            sfen_writer.write(th.id(), sfen);
        }
        return false;
    }
    void TrainingDataGenerator::initial_report()
    {
        out = sync_region_cout.new_region();
        const auto now_time = now();
        out
            << '\n'
            << 0 << " sfens, "
            << "at " << now_string() << endl;
        last_stats_report_time = now_time;
        out = sync_region_cout.new_region();
    }
    void TrainingDataGenerator::report(uint64_t done, uint64_t new_done)
    {
        const auto now_time = now();
        const TimePoint elapsed = now_time - last_stats_report_time + 1;
        out
            << '\n'
            << done << " sfens, "
            << new_done * 1000 / elapsed << " sfens/second, "
            << "at " << now_string() << endl;
        last_stats_report_time = now_time;
        out = sync_region_cout.new_region();
    }
    void TrainingDataGenerator::maybe_report(uint64_t done)
    {
        if (done % REPORT_DOT_EVERY == 0)
        {
            std::lock_guard lock(stats_mutex);
            if (last_stats_report_time == 0)
            {
                last_stats_report_time = now();
                out = sync_region_cout.new_region();
            }
            if (done != 0)
            {
                out << '.';
                if (done % REPORT_STATS_EVERY == 0)
                {
                    report(done, REPORT_STATS_EVERY);
                }
            }
        }
    }
    // Command to generate a game record
    void generate_training_data(istringstream& is)
    {
        // Number of generated game records default = 8 billion phases (Ponanza specification)
        uint64_t loop_max = 8000000000UL;
        uint64_t time_max = 8000000000UL;
        TrainingDataGenerator::Params params;
        // Add a random number to the end of the file name.
        bool random_file_name = false;
        std::string sfen_format = "binpack";
        string token;
        while (true)
        {
            token = "";
            is >> token;
            if (token == "")
                break;
            if (token == "depth")
            {
                is >> params.search_depth_min;
                params.search_depth_max = params.search_depth_min;
            }
            else if (token == "min_depth")
                is >> params.search_depth_min;
            else if (token == "max_depth")
                is >> params.search_depth_max;
            else if (token == "nodes")
                is >> params.nodes;
            else if (token == "count")
                is >> loop_max;
            else if (token == "max_time_seconds")
                is >> time_max;
            else if (token == "max_time_minutes")
            {
                is >> time_max;
                time_max *= 60;
            }
            else if (token == "max_time_hours")
            {
                is >> time_max;
                time_max *= 3600;
            }
            else if (token == "output_file_name")
                is >> params.output_file_name;
            else if (token == "eval_limit")
                is >> params.eval_limit;
            else if (token == "random_move_min_ply")
                is >> params.random_move_minply;
            else if (token == "random_move_max_ply")
                is >> params.random_move_maxply;
            else if (token == "random_move_count")
                is >> params.random_move_count;
            else if (token == "random_move_like_apery")
                is >> params.random_move_like_apery;
            else if (token == "random_multi_pv")
                is >> params.random_multi_pv;
            else if (token == "random_multi_pv_diff")
                is >> params.random_multi_pv_diff;
            else if (token == "random_multi_pv_depth")
                is >> params.random_multi_pv_depth;
            else if (token == "random_multi_pv_nodes")
                is >> params.random_multi_pv_nodes;
            else if (token == "write_min_ply")
                is >> params.write_minply;
            else if (token == "write_max_ply")
                is >> params.write_maxply;
            else if (token == "save_every")
                is >> params.save_every;
            else if (token == "book")
                is >> params.book;
            else if (token == "random_file_name")
                is >> random_file_name;
            else if (token == "keep_draws")
                is >> params.write_out_draw_game_in_training_data_generation;
            else if (token == "adjudicate_draws_by_score")
                is >> params.detect_draw_by_consecutive_low_score;
            else if (token == "adjudicate_draws_by_insufficient_material")
                is >> params.detect_draw_by_insufficient_mating_material;
            else if (token == "data_format")
                is >> sfen_format;
            else if (token == "seed")
                is >> params.seed;
            else if (token == "set_recommended_uci_options")
            {
                UCI::setoption("Skill Level", "20");
                UCI::setoption("UCI_LimitStrength", "false");
                UCI::setoption("PruneAtShallowDepth", "false");
                UCI::setoption("EnableTranspositionTable", "true");
            }
            else
            {
                cout << "ERROR: Unknown option " << token << ". Exiting...\n";
                return;
            }
        }
        if (!sfen_format.empty())
        {
            if (sfen_format == "bin")
                params.sfen_format = SfenOutputType::Bin;
            else if (sfen_format == "binpack")
                params.sfen_format = SfenOutputType::Binpack;
            else
                cout << "WARNING: Unknown sfen format `" << sfen_format << "`. Using bin\n";
        }
        if (random_file_name)
        {
            // Give a random number to output_file_name at this point.
            // Do not use std::random_device().  Because it always the same integers on MinGW.
            PRNG r(params.seed);
            // Just in case, reassign the random numbers.
            for (int i = 0; i < 10; ++i)
                r.rand(1);
            auto to_hex = [](uint64_t u) {
                std::stringstream ss;
                ss << std::hex << u;
                return ss.str();
            };
            // I don't want to wear 64bit numbers by accident, so I'next_move going to make a 64bit number 2 just in case.
            params.output_file_name += "_" + to_hex(r.rand<uint64_t>()) + to_hex(r.rand<uint64_t>());
        }
        params.enforce_constraints();
        std::cout << "INFO: Executing generate_training_data command\n";
        std::cout << "INFO: Parameters:\n";
        std::cout
            << "  - search_depth_min       = " << params.search_depth_min << endl
            << "  - search_depth_max       = " << params.search_depth_max << endl
            << "  - nodes                  = " << params.nodes << endl
            << "  - count                  = " << loop_max << endl
            << "  - max_time_seconds       = " << time_max << endl
            << "  - eval_limit             = " << params.eval_limit << endl
            << "  - num threads (UCI)      = " << params.num_threads << endl
            << "  - random_move_min_ply    = " << params.random_move_minply << endl
            << "  - random_move_max_ply    = " << params.random_move_maxply << endl
            << "  - random_move_count      = " << params.random_move_count << endl
            << "  - random_move_like_apery = " << params.random_move_like_apery << endl
            << "  - random_multi_pv        = " << params.random_multi_pv << endl
            << "  - random_multi_pv_diff   = " << params.random_multi_pv_diff << endl
            << "  - random_multi_pv_depth  = " << params.random_multi_pv_depth << endl
            << "  - random_multi_pv_nodes  = " << params.random_multi_pv_nodes << endl
            << "  - write_min_ply          = " << params.write_minply << endl
            << "  - write_max_ply          = " << params.write_maxply << endl
            << "  - book                   = " << params.book << endl
            << "  - output_file_name       = " << params.output_file_name << endl
            << "  - save_every             = " << params.save_every << endl
            << "  - random_file_name       = " << random_file_name << endl
            << "  - write_drawn_games      = " << params.write_out_draw_game_in_training_data_generation << endl
            << "  - draw by low score      = " << params.detect_draw_by_consecutive_low_score << endl
            << "  - draw by insuff. mat.   = " << params.detect_draw_by_insufficient_mating_material << endl;
        // Show if the training data generator uses NNUE.
        Eval::NNUE::verify();
        Threads.main()->ponder = false;
        TrainingDataGenerator gensfen(params);
        gensfen.generate(loop_max, time_max);
        std::cout << "INFO: generate_training_data finished." << endl;
    }
 }
@@ -0,0 +1,14 @@
 #ifndef _GENSFEN_H_
 #define _GENSFEN_H_
 #include "position.h"
 #include <sstream>
 namespace Stockfish::Tools {
    // Automatic generation of teacher position
    void generate_training_data(std::istringstream& is);
 }
 #endif
@@ -0,0 +1,498 @@
 #include "training_data_generator_nonpv.h"
 #include "sfen_writer.h"
 #include "packed_sfen.h"
 #include "opening_book.h"
 #include "misc.h"
 #include "position.h"
 #include "thread.h"
 #include "tt.h"
 #include "uci.h"
 #include "extra/nnue_data_binpack_format.h"
 #include "nnue/evaluate_nnue.h"
 #include "syzygy/tbprobe.h"
 #include <atomic>
 #include <chrono>
 #include <climits>
 #include <cmath>
 #include <cstring>
 #include <fstream>
 #include <iomanip>
 #include <limits>
 #include <list>
 #include <memory>
 #include <optional>
 #include <random>
 #include <shared_mutex>
 #include <sstream>
 #include <unordered_set>
 using namespace std;
 namespace Stockfish::Tools
 {
    // Class to generate sfen with multiple threads
    struct TrainingDataGeneratorNonPv
    {
        struct Params
        {
            // The depth for search on the fens gathered during exploration
            int search_depth = 3;
            // the min/max number of nodes to use for exploration per ply
            int exploration_min_nodes = 5000;
            int exploration_max_nodes = 15000;
            // The pct of positions explored that are saved for rescoring
            float exploration_save_rate = 0.01;
            // Upper limit of evaluation value of generated situation
            int eval_limit = 4000;
            // the upper limit on evaluation during exploration selfplay
            int exploration_eval_limit = 4000;
            int exploration_max_ply = 200;
            int exploration_min_pieces = 8;
            std::string output_file_name = "training_data_nonpv";
            SfenOutputType sfen_format = SfenOutputType::Binpack;
            std::string seed;
            int num_threads;
            std::string book;
            bool smart_fen_skipping = false;
            void enforce_constraints()
            {
                // Limit the maximum to a one-stop score. (Otherwise you might not end the loop)
                eval_limit = std::min(eval_limit, (int)mate_in(2));
                exploration_eval_limit = std::min(eval_limit, (int)mate_in(2));
                exploration_min_nodes = std::max(100, exploration_min_nodes);
                exploration_max_nodes = std::max(exploration_min_nodes, exploration_max_nodes);
                num_threads = Options["Threads"];
            }
        };
        static constexpr uint64_t REPORT_DOT_EVERY = 5000;
        static constexpr uint64_t REPORT_STATS_EVERY = 200000;
        static_assert(REPORT_STATS_EVERY % REPORT_DOT_EVERY == 0);
        TrainingDataGeneratorNonPv(
            const Params& prm
        ) :
            params(prm),
            prng(prm.seed),
            sfen_writer(prm.output_file_name, prm.num_threads, std::numeric_limits<uint64_t>::max(), prm.sfen_format)
        {
            if (!prm.book.empty())
            {
                opening_book = open_opening_book(prm.book, prng);
                if (opening_book == nullptr)
                {
                    std::cout << "WARNING: Failed to open opening book " << prm.book << ". Falling back to startpos.\n";
                }
            }
            // Output seed to veryfy by the user if it's not identical by chance.
            std::cout << prng << std::endl;
        }
        void generate(uint64_t limit);
    private:
        Params params;
        PRNG prng;
        std::mutex stats_mutex;
        TimePoint last_stats_report_time;
        // sfen exporter
        SfenWriter sfen_writer;
        SynchronizedRegionLogger::Region out;
        std::unique_ptr<OpeningBook> opening_book;
        static void set_gensfen_search_limits();
        void generate_worker(
            Thread& th,
            std::atomic<uint64_t>& counter,
            uint64_t limit);
        bool commit_psv(
            Thread& th,
            PSVector& sfens,
            std::atomic<uint64_t>& counter,
            uint64_t limit);
        PSVector do_exploration(
            Thread& th,
            int count);
        void report(uint64_t done, uint64_t new_done);
        void maybe_report(uint64_t done);
    };
    void TrainingDataGeneratorNonPv::set_gensfen_search_limits()
    {
        // About Search::Limits
        // Be careful because this member variable is global and affects other threads.
        auto& limits = Search::Limits;
        // Make the search equivalent to the "go infinite" command. (Because it is troublesome if time management is done)
        limits.infinite = true;
        // Since PV is an obstacle when displayed, erase it.
        limits.silent = true;
        // If you use this, it will be compared with the accumulated nodes of each thread. Therefore, do not use it.
        limits.nodes = 0;
        // depth is also processed by the one passed as an argument of Tools::search().
        limits.depth = 0;
    }
    void TrainingDataGeneratorNonPv::generate(uint64_t limit)
    {
        last_stats_report_time = 0;
        set_gensfen_search_limits();
        std::atomic<uint64_t> counter{0};
        Threads.execute_with_workers([&counter, limit, this](Thread& th) {
            generate_worker(th, counter, limit);
        });
        Threads.wait_for_workers_finished();
        sfen_writer.flush();
        if (limit % REPORT_STATS_EVERY != 0)
        {
            report(limit, limit % REPORT_STATS_EVERY);
        }
        std::cout << std::endl;
    }
    PSVector TrainingDataGeneratorNonPv::do_exploration(
        Thread& th,
        int count)
    {
        constexpr int max_depth = 30;
        PSVector psv;
        std::vector<StateInfo, AlignedAllocator<StateInfo>> states(
            max_depth + MAX_PLY /* == search_depth_min + α */);
        th.set_eval_callback([this, &psv](Position& pos) {
            if ((double)prng.rand<uint64_t>() / std::numeric_limits<uint64_t>::max() < params.exploration_save_rate)
            {
                psv.emplace_back();
                pos.sfen_pack(psv.back().sfen, pos.is_chess960());
            }
        });
        auto& pos = th.rootPos;
        StateInfo si;
        const bool frc = Options["UCI_Chess960"];
        for (int i = 0; i < count; ++i)
        {
            if (opening_book != nullptr)
            {
                auto& fen = opening_book->next_fen();
                pos.set(fen, frc, &si, &th);
            }
            else
            {
                pos.set(StartFEN, frc, &si, &th);
            }
            for(int ply = 0; ply < params.exploration_max_ply; ++ply)
            {
                auto nodes = prng.rand(params.exploration_max_nodes - params.exploration_min_nodes + 1) + params.exploration_min_nodes;
                auto [search_value, search_pv] = Search::search(pos, max_depth, 1, nodes);
                if (search_pv.empty())
                {
                    break;
                }
                if (std::abs(search_value) > params.exploration_eval_limit)
                {
                    break;
                }
                pos.do_move(search_pv[0], states[ply]);
                if (popcount(pos.pieces()) < params.exploration_min_pieces)
                {
                    break;
                }
            }
        }
        th.clear_eval_callback();
        return psv;
    }
    void TrainingDataGeneratorNonPv::generate_worker(
        Thread& th,
        std::atomic<uint64_t>& counter,
        uint64_t limit)
    {
        constexpr int exploration_batch_size = 1;
        StateInfo si;
        PSVector psv;
        // end flag
        bool quit = false;
        const bool frc = Options["UCI_Chess960"];
        // repeat until the specified number of times
        while (!quit)
        {
            // It is necessary to set a dependent thread for Position.
            // When parallelizing, Threads (since this is a vector<Thread*>,
            // Do the same for up to Threads[0]...Threads[thread_num-1].
            auto& pos = th.rootPos;
            auto packed_sfens = do_exploration(th, exploration_batch_size);
            psv.clear();
            for (auto& ps : packed_sfens)
            {
                pos.set_from_packed_sfen(ps.sfen, &si, &th, frc);
                pos.state()->rule50 = 0;
                if (params.smart_fen_skipping && pos.checkers())
                {
                    continue;
                }
                auto [search_value, search_pv] = Search::search(pos, params.search_depth, 1);
                if (search_pv.empty())
                {
                    continue;
                }
                if (std::abs(search_value) > params.eval_limit)
                {
                    continue;
                }
                if (params.smart_fen_skipping && pos.capture_or_promotion(search_pv[0]))
                {
                    continue;
                }
                auto& new_ps = psv.emplace_back();
                pos.sfen_pack(new_ps.sfen, pos.is_chess960());
                new_ps.score = search_value;
                new_ps.move = search_pv[0];
                new_ps.gamePly = 1;
                new_ps.game_result = 0;
                new_ps.padding = 0;
            }
            quit = commit_psv(th, psv, counter, limit);
        }
    }
    // Write out the phases loaded in sfens to a file.
    // result: win/loss in the next phase after the final phase in sfens
    // 1 when winning. -1 when losing. Pass 0 for a draw.
    // Return value: true if the specified number of
    // sfens has already been reached and the process ends.
    bool TrainingDataGeneratorNonPv::commit_psv(
        Thread& th,
        PSVector& sfens,
        std::atomic<uint64_t>& counter,
        uint64_t limit)
    {
        const bool frc = th.rootPos.is_chess960();
        // Write sfens in move order to make potential compression easier
        for (auto& sfen : sfens)
        {
            // Skip positions with castling bestmove in FRC so that we don't
            // need to support it in the trainer.
            if (frc && type_of((Move)sfen.move) == CASTLING)
            {
                continue;
            }
            // Return true if there is already enough data generated.
            const auto iter = counter.fetch_add(1);
            if (iter >= limit)
                return true;
            // because `iter` was done, now we do one more
            maybe_report(iter + 1);
            // Write out one sfen.
            sfen_writer.write(th.id(), sfen);
        }
        return false;
    }
    void TrainingDataGeneratorNonPv::report(uint64_t done, uint64_t new_done)
    {
        const auto now_time = now();
        const TimePoint elapsed = now_time - last_stats_report_time + 1;
        out
            << endl
            << done << " sfens, "
            << new_done * 1000 / elapsed << " sfens/second, "
            << "at " << now_string() << sync_endl;
        last_stats_report_time = now_time;
        out = sync_region_cout.new_region();
    }
    void TrainingDataGeneratorNonPv::maybe_report(uint64_t done)
    {
        if (done % REPORT_DOT_EVERY == 0)
        {
            std::lock_guard lock(stats_mutex);
            if (last_stats_report_time == 0)
            {
                last_stats_report_time = now();
                out = sync_region_cout.new_region();
            }
            if (done != 0)
            {
                out << '.';
                if (done % REPORT_STATS_EVERY == 0)
                {
                    report(done, REPORT_STATS_EVERY);
                }
            }
        }
    }
    // Command to generate a game record
    void generate_training_data_nonpv(istringstream& is)
    {
        // Number of generated game records default = 8 billion phases (Ponanza specification)
        TrainingDataGeneratorNonPv::Params params;
        uint64_t count = 1'000'000;
        // Add a random number to the end of the file name.
        std::string sfen_format = "binpack";
        string token;
        while (true)
        {
            token = "";
            is >> token;
            if (token == "")
                break;
            if (token == "depth")
                is >> params.search_depth;
            else if (token == "count")
                is >> count;
            else if (token == "output_file")
                is >> params.output_file_name;
            else if (token == "exploration_eval_limit")
                is >> params.exploration_eval_limit;
            else if (token == "eval_limit")
                is >> params.eval_limit;
            else if (token == "exploration_min_nodes")
                is >> params.exploration_min_nodes;
            else if (token == "exploration_max_nodes")
                is >> params.exploration_max_nodes;
            else if (token == "exploration_min_pieces")
                is >> params.exploration_min_pieces;
            else if (token == "exploration_save_rate")
                is >> params.exploration_save_rate;
            else if (token == "book")
                is >> params.book;
            else if (token == "data_format")
                is >> sfen_format;
            else if (token == "seed")
                is >> params.seed;
            else if (token == "smart_fen_skipping")
                params.smart_fen_skipping = true;
            else if (token == "set_recommended_uci_options")
            {
                UCI::setoption("Skill Level", "20");
                UCI::setoption("UCI_LimitStrength", "false");
                UCI::setoption("PruneAtShallowDepth", "false");
                UCI::setoption("EnableTranspositionTable", "true");
            }
            else
            {
                cout << "ERROR: Unknown option " << token << ". Exiting...\n";
                return;
            }
        }
        if (!sfen_format.empty())
        {
            if (sfen_format == "bin")
                params.sfen_format = SfenOutputType::Bin;
            else if (sfen_format == "binpack")
                params.sfen_format = SfenOutputType::Binpack;
            else
                cout << "WARNING: Unknown sfen format `" << sfen_format << "`. Using bin\n";
        }
        params.enforce_constraints();
        std::cout << "INFO: Executing generate_training_data_nonpv command\n";
        std::cout << "INFO: Parameters:\n";
        std::cout
            << "  - search_depth           = " << params.search_depth << endl
            << "  - output_file            = " << params.output_file_name << endl
            << "  - exploration_eval_limit = " << params.exploration_eval_limit << endl
            << "  - eval_limit             = " << params.eval_limit << endl
            << "  - exploration_min_nodes  = " << params.exploration_min_nodes << endl
            << "  - exploration_max_nodes  = " << params.exploration_max_nodes << endl
            << "  - exploration_min_pieces = " << params.exploration_min_pieces << endl
            << "  - exploration_save_rate  = " << params.exploration_save_rate << endl
            << "  - book                   = " << params.book << endl
            << "  - data_format            = " << sfen_format << endl
            << "  - seed                   = " << params.seed << endl
            << "  - count                  = " << count << endl;
        // Show if the training data generator uses NNUE.
        Eval::NNUE::verify();
        Threads.main()->ponder = false;
        TrainingDataGeneratorNonPv gensfen(params);
        gensfen.generate(count);
        std::cout << "INFO: generate_training_data_nonpv finished." << endl;
    }
 }
@@ -0,0 +1,12 @@
 #ifndef _GENSFEN_NONPV_H_
 #define _GENSFEN_NONPV_H_
 #include <sstream>
 namespace Stockfish::Tools {
    // Automatic generation of teacher position
    void generate_training_data_nonpv(std::istringstream& is);
 }
 #endif
@@ -0,0 +1,12 @@
 #ifndef _TRANSFORM_H_
 #define _TRANSFORM_H_
 #include <sstream>
 namespace Stockfish::Tools {
    void transform(std::istringstream& is);
 }
 #endif
@@ -0,0 +1,122 @@
 #include "validate_training_data.h"
 #include "uci.h"
 #include "misc.h"
 #include "thread.h"
 #include "position.h"
 #include "tt.h"
 #include "extra/nnue_data_binpack_format.h"
 #include "nnue/evaluate_nnue.h"
 #include "syzygy/tbprobe.h"
 #include <sstream>
 #include <fstream>
 #include <unordered_set>
 #include <iomanip>
 #include <list>
 #include <cmath>    // std::exp(),std::pow(),std::log()
 #include <cstring>  // memcpy()
 #include <memory>
 #include <limits>
 #include <optional>
 #include <chrono>
 #include <random>
 #include <regex>
 #include <filesystem>
 using namespace std;
 namespace sys = std::filesystem;
 namespace Stockfish::Tools
 {
    static inline const std::string plain_extension = ".plain";
    static inline const std::string bin_extension = ".bin";
    static inline const std::string binpack_extension = ".binpack";
    static bool file_exists(const std::string& name)
    {
        std::ifstream f(name);
        return f.good();
    }
    static bool ends_with(const std::string& lhs, const std::string& end)
    {
        if (end.size() > lhs.size()) return false;
        return std::equal(end.rbegin(), end.rend(), lhs.rbegin());
    }
    static bool is_validation_of_type(
        const std::string& input_path,
        const std::string& expected_input_extension)
    {
        return ends_with(input_path, expected_input_extension);
    }
    using ValidateFunctionType = void(std::string inputPath);
    static ValidateFunctionType* get_validate_function(const std::string& input_path)
    {
        if (is_validation_of_type(input_path, plain_extension))
            return binpack::validatePlain;
        if (is_validation_of_type(input_path, bin_extension))
            return binpack::validateBin;
        if (is_validation_of_type(input_path, binpack_extension))
            return binpack::validateBinpack;
        return nullptr;
    }
    static void validate_training_data(const std::string& input_path)
    {
        if(!file_exists(input_path))
        {
            std::cerr << "Input file does not exist.\n";
            return;
        }
        auto func = get_validate_function(input_path);
        if (func != nullptr)
        {
            func(input_path);
        }
        else
        {
            std::cerr << "Validation of files of this type is not supported.\n";
        }
    }
    static void validate_training_data(const std::vector<std::string>& args)
    {
        if (args.size() != 1)
        {
            std::cerr << "Invalid arguments.\n";
            std::cerr << "Usage: validate in_path\n";
            return;
        }
        validate_training_data(args[0]);
    }
    void validate_training_data(istringstream& is)
    {
        std::vector<std::string> args;
        while (true)
        {
            std::string token = "";
            is >> token;
            if (token == "")
                break;
            args.push_back(token);
        }
        validate_training_data(args);
    }
 }
@@ -0,0 +1,12 @@
 #ifndef _VALIDATE_TRAINING_DATA_H_
 #define _VALIDATE_TRAINING_DATA_H_
 #include <vector>
 #include <string>
 #include <sstream>
 namespace Stockfish::Tools {
    void validate_training_data(std::istringstream& is);
 }
 #endif
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -26,29 +26,37 @@
 #include "tt.h"
 #include "uci.h"
 namespace Stockfish {
 TranspositionTable TT; // Our global transposition table
 bool TranspositionTable::enable_transposition_table = true;
 /// TTEntry::save() populates the TTEntry with a new node's data, possibly
 /// overwriting an old position. Update is not atomic and can be racy.
 void TTEntry::save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev) {
  if (!TranspositionTable::enable_transposition_table) {
      return;
  }
  // Preserve any existing move for the same position
  if (m || (uint16_t)k != key16)
      move16 = (uint16_t)m;
-  // Overwrite less valuable entries
+  // Overwrite less valuable entries (cheapest checks first)
-  if ((uint16_t)k != key16
+  if (   b == BOUND_EXACT
-      || d - DEPTH_OFFSET > depth8 - 4
+      || (uint16_t)k != key16
-      || b == BOUND_EXACT)
+      || d - DEPTH_OFFSET + 2 * pv > depth8 - 4)
  {
-      assert(d >= DEPTH_OFFSET);
+      assert(d > DEPTH_OFFSET);
      assert(d < 256 + DEPTH_OFFSET);
      key16     = (uint16_t)k;
      depth8    = (uint8_t)(d - DEPTH_OFFSET);
      genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b);
      value16   = (int16_t)v;
      eval16    = (int16_t)ev;
      genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b);
      depth8    = (uint8_t)(d - DEPTH_OFFSET);
  }
 }
@@ -61,11 +69,12 @@ void TranspositionTable::resize(size_t mbSize) {
  Threads.main()->wait_for_search_finished();
-  aligned_ttmem_free(mem);
+  aligned_large_pages_free(table);
  clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
-  table = static_cast<Cluster*>(aligned_ttmem_alloc(clusterCount * sizeof(Cluster), mem));
+
-  if (!mem)
+  table = static_cast<Cluster*>(aligned_large_pages_alloc(clusterCount * sizeof(Cluster)));
  if (!table)
  {
      std::cerr << "Failed to allocate " << mbSize
                << "MB for transposition table." << std::endl;
@@ -115,26 +124,32 @@ void TranspositionTable::clear() {
 TTEntry* TranspositionTable::probe(const Key key, bool& found) const {
  if (!enable_transposition_table) {
      found = false;
      return first_entry(0);
  }
  TTEntry* const tte = first_entry(key);
  const uint16_t key16 = (uint16_t)key;  // Use the low 16 bits as key inside the cluster
  for (int i = 0; i < ClusterSize; ++i)
-      if (!tte[i].key16 || tte[i].key16 == key16)
+      if (tte[i].key16 == key16 || !tte[i].depth8)
      {
-          tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & 0x7)); // Refresh
+          tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & (GENERATION_DELTA - 1))); // Refresh
-          return found = (bool)tte[i].key16, &tte[i];
+          return found = (bool)tte[i].depth8, &tte[i];
      }
  // Find an entry to be replaced according to the replacement strategy
  TTEntry* replace = tte;
  for (int i = 1; i < ClusterSize; ++i)
      // Due to our packed storage format for generation and its cyclic
-      // nature we add 263 (256 is the modulus plus 7 to keep the unrelated
+      // nature we add GENERATION_CYCLE (256 is the modulus, plus what
-      // lowest three bits from affecting the result) to calculate the entry
+      // is needed to keep the unrelated lowest n bits from affecting
-      // age correctly even after generation8 overflows into the next cycle.
+      // the result) to calculate the entry age correctly even after
-      if (  replace->depth8 - ((263 + generation8 - replace->genBound8) & 0xF8)
+      // generation8 overflows into the next cycle.
-          >   tte[i].depth8 - ((263 + generation8 -   tte[i].genBound8) & 0xF8))
+      if (  replace->depth8 - ((GENERATION_CYCLE + generation8 - replace->genBound8) & GENERATION_MASK)
          >   tte[i].depth8 - ((GENERATION_CYCLE + generation8 -   tte[i].genBound8) & GENERATION_MASK))
          replace = &tte[i];
  return found = false, replace;
@@ -149,7 +164,9 @@ int TranspositionTable::hashfull() const {
  int cnt = 0;
  for (int i = 0; i < 1000; ++i)
      for (int j = 0; j < ClusterSize; ++j)
-          cnt += (table[i].entry[j].genBound8 & 0xF8) == generation8;
+          cnt += table[i].entry[j].depth8 && (table[i].entry[j].genBound8 & GENERATION_MASK) == generation8;
  return cnt / ClusterSize;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -22,16 +22,18 @@
 #include "misc.h"
 #include "types.h"
 namespace Stockfish {
 /// TTEntry struct is the 10 bytes transposition table entry, defined as below:
 ///
 /// key        16 bit
-/// move       16 bit
+/// depth       8 bit
 /// value      16 bit
 /// eval value 16 bit
 /// generation  5 bit
 /// pv node     1 bit
 /// bound type  2 bit
-/// depth       8 bit
+/// move       16 bit
 /// value      16 bit
 /// eval value 16 bit
 struct TTEntry {
@@ -47,11 +49,11 @@ private:
  friend class TranspositionTable;
  uint16_t key16;
  uint8_t  depth8;
  uint8_t  genBound8;
  uint16_t move16;
  int16_t  value16;
  int16_t  eval16;
  uint8_t  genBound8;
  uint8_t  depth8;
 };
@@ -72,9 +74,15 @@ class TranspositionTable {
  static_assert(sizeof(Cluster) == 32, "Unexpected Cluster size");
  // Constants used to refresh the hash table periodically
  static constexpr unsigned GENERATION_BITS  = 3;                                // nb of bits reserved for other things
  static constexpr int      GENERATION_DELTA = (1 << GENERATION_BITS);           // increment for generation field
  static constexpr int      GENERATION_CYCLE = 255 + (1 << GENERATION_BITS);     // cycle length
  static constexpr int      GENERATION_MASK  = (0xFF << GENERATION_BITS) & 0xFF; // mask to pull out generation number
 public:
- ~TranspositionTable() { aligned_ttmem_free(mem); }
+ ~TranspositionTable() { aligned_large_pages_free(table); }
-  void new_search() { generation8 += 8; } // Lower 3 bits are used by PV flag and Bound
+  void new_search() { generation8 += GENERATION_DELTA; } // Lower bits are used for other things
  TTEntry* probe(const Key key, bool& found) const;
  int hashfull() const;
  void resize(size_t mbSize);
@@ -84,15 +92,18 @@ public:
    return &table[mul_hi64(key, clusterCount)].entry[0];
  }
  static bool enable_transposition_table;
 private:
  friend struct TTEntry;
  size_t clusterCount;
  Cluster* table;
  void* mem;
  uint8_t generation8; // Size must be not bigger than TTEntry::genBound8
 };
 extern TranspositionTable TT;
 } // namespace Stockfish
 #endif // #ifndef TT_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -26,9 +26,10 @@
 using std::string;
 namespace Stockfish {
 bool Tune::update_on_last;
 const UCI::Option* LastOption = nullptr;
 BoolConditions Conditions;
 static std::map<std::string, int> TuneResults;
 string Tune::next(string& names, bool pop) {
@@ -108,23 +109,7 @@ template<> void Tune::Entry<Score>::read_option() {
 template<> void Tune::Entry<Tune::PostUpdate>::init_option() {}
 template<> void Tune::Entry<Tune::PostUpdate>::read_option() { value(); }
-
+} // namespace Stockfish
 // Set binary conditions according to a probability that depends
 // on the corresponding parameter value.
 void BoolConditions::set() {
  static PRNG rng(now());
  static bool startup = true; // To workaround fishtest bench
  for (size_t i = 0; i < binary.size(); i++)
      binary[i] = !startup && (values[i] + int(rng.rand<unsigned>() % variance) > threshold);
  startup = false;
  for (size_t i = 0; i < binary.size(); i++)
      sync_cout << binary[i] << sync_endl;
 }
 // Init options with tuning session results instead of default values. Useful to
@@ -138,7 +123,11 @@ void BoolConditions::set() {
 #include <cmath>
 namespace Stockfish {
 void Tune::read_results() {
  /* ...insert your values here... */
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,6 +24,8 @@
 #include <type_traits>
 #include <vector>
 namespace Stockfish {
 typedef std::pair<int, int> Range; // Option's min-max values
 typedef Range (RangeFun) (int);
@@ -44,27 +46,6 @@ struct SetRange {
 #define SetDefaultRange SetRange(default_range)
 /// BoolConditions struct is used to tune boolean conditions in the
 /// code by toggling them on/off according to a probability that
 /// depends on the value of a tuned integer parameter: for high
 /// values of the parameter condition is always disabled, for low
 /// values is always enabled, otherwise it is enabled with a given
 /// probability that depnends on the parameter under tuning.
 struct BoolConditions {
  void init(size_t size) { values.resize(size, defaultValue), binary.resize(size, 0); }
  void set();
  std::vector<int> binary, values;
  int defaultValue = 465, variance = 40, threshold = 500;
  SetRange range = SetRange(0, 1000);
 };
 extern BoolConditions Conditions;
 inline void set_conditions() { Conditions.set(); }
 /// Tune class implements the 'magic' code that makes the setup of a fishtest
 /// tuning session as easy as it can be. Mainly you have just to remove const
 /// qualifiers from the variables you want to tune and flag them for tuning, so
@@ -103,7 +84,7 @@ class Tune {
  static Tune& instance() { static Tune t; return t; } // Singleton
-  // Use polymorphism to accomodate Entry of different types in the same vector
+  // Use polymorphism to accommodate Entry of different types in the same vector
  struct EntryBase {
    virtual ~EntryBase() = default;
    virtual void init_option() = 0;
@@ -130,9 +111,9 @@ class Tune {
    SetRange range;
  };
-  // Our facilty to fill the container, each Entry corresponds to a parameter to tune.
+  // Our facility to fill the container, each Entry corresponds to a parameter
-  // We use variadic templates to deal with an unspecified number of entries, each one
+  // to tune. We use variadic templates to deal with an unspecified number of
-  // of a possible different type.
+  // entries, each one of a possible different type.
  static std::string next(std::string& names, bool pop = true);
  int add(const SetRange&, std::string&&) { return 0; }
@@ -157,14 +138,6 @@ class Tune {
    return add(value, (next(names), std::move(names)), args...);
  }
  // Template specialization for BoolConditions
  template<typename... Args>
  int add(const SetRange& range, std::string&& names, BoolConditions& cond, Args&&... args) {
    for (size_t size = cond.values.size(), i = 0; i < size; i++)
        add(cond.range, next(names, i == size - 1) + "_" + std::to_string(i), cond.values[i]);
    return add(range, std::move(names), args...);
  }
  std::vector<std::unique_ptr<EntryBase>> list;
 public:
@@ -185,9 +158,6 @@ public:
 #define UPDATE_ON_LAST() bool UNIQUE(p, __LINE__) = Tune::update_on_last = true
-// Some macro to tune toggling of boolean conditions
+} // namespace Stockfish
 #define CONDITION(x) (Conditions.binary[__COUNTER__] || (x))
 #define TUNE_CONDITIONS() int UNIQUE(c, __LINE__) = (Conditions.init(__COUNTER__), 0); \
                          TUNE(Conditions, set_conditions)
 #endif // #ifndef TUNE_H_INCLUDED
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -57,6 +57,12 @@
 /// _WIN32             Building on Windows (any)
 /// _WIN64             Building on Windows 64 bit
 #if defined(__GNUC__ ) && (__GNUC__ < 9 || (__GNUC__ == 9 && __GNUC_MINOR__ <= 2)) && defined(_WIN32) && !defined(__clang__)
 #define ALIGNAS_ON_STACK_VARIABLES_BROKEN
 #endif
 #define ASSERT_ALIGNED(ptr, alignment) assert(reinterpret_cast<uintptr_t>(ptr) % alignment == 0)
 #if defined(_WIN64) && defined(_MSC_VER) // No Makefile used
 #  include <intrin.h> // Microsoft header for _BitScanForward64()
 #  define IS_64BIT
@@ -77,6 +83,8 @@
 #  define pext(b, m) 0
 #endif
 namespace Stockfish {
 #ifdef USE_POPCNT
 constexpr bool HasPopCnt = true;
 #else
@@ -107,7 +115,7 @@ constexpr int MAX_PLY   = 246;
 /// bit  6-11: origin square (from 0 to 63)
 /// bit 12-13: promotion piece type - 2 (from KNIGHT-2 to QUEEN-2)
 /// bit 14-15: special move flag: promotion (1), en passant (2), castling (3)
-/// NOTE: EN-PASSANT bit is set only when a pawn can be captured
+/// NOTE: en passant bit is set only when a pawn can be captured
 ///
 /// Special cases are MOVE_NONE and MOVE_NULL. We can sneak these in because in
 /// any normal move destination square is always different from origin square
@@ -121,7 +129,7 @@ enum Move : int {
 enum MoveType {
  NORMAL,
  PROMOTION = 1 << 14,
-  ENPASSANT = 2 << 14,
+  EN_PASSANT = 2 << 14,
  CASTLING  = 3 << 14
 };
@@ -129,6 +137,8 @@ enum Color {
  WHITE, BLACK, COLOR_NB = 2
 };
 constexpr Color Colors[2] = { WHITE, BLACK };
 enum CastlingRights {
  NO_CASTLING,
  WHITE_OO,
@@ -178,12 +188,11 @@ enum Value : int {
  VALUE_MATE_IN_MAX_PLY  =  VALUE_MATE - MAX_PLY,
  VALUE_MATED_IN_MAX_PLY = -VALUE_MATE_IN_MAX_PLY,
-  PawnValueMg   = 124,   PawnValueEg   = 206,
+  PawnValueMg   = 126,   PawnValueEg   = 208,
  KnightValueMg = 781,   KnightValueEg = 854,
  BishopValueMg = 825,   BishopValueEg = 915,
  RookValueMg   = 1276,  RookValueEg   = 1380,
  QueenValueMg  = 2538,  QueenValueEg  = 2682,
  Tempo = 28,
  MidgameLimit  = 15258, EndgameLimit  = 3915
 };
@@ -196,27 +205,11 @@ enum PieceType {
 enum Piece {
  NO_PIECE,
-  W_PAWN = 1, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING,
+  W_PAWN = PAWN,     W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING,
-  B_PAWN = 9, B_KNIGHT, B_BISHOP, B_ROOK, B_QUEEN, B_KING,
+  B_PAWN = PAWN + 8, B_KNIGHT, B_BISHOP, B_ROOK, B_QUEEN, B_KING,
  PIECE_NB = 16
 };
 // An ID used to track the pieces. Max. 32 pieces on board.
 enum PieceId {
  PIECE_ID_ZERO   = 0,
  PIECE_ID_KING   = 30,
  PIECE_ID_WKING  = 30,
  PIECE_ID_BKING  = 31,
  PIECE_ID_NONE   = 32
 };
 inline PieceId operator++(PieceId& d, int) {
  PieceId x = d;
  d = PieceId(int(d) + 1);
  return x;
 }
 constexpr Value PieceValue[PHASE_NB][PIECE_NB] = {
  { VALUE_ZERO, PawnValueMg, KnightValueMg, BishopValueMg, RookValueMg, QueenValueMg, VALUE_ZERO, VALUE_ZERO,
    VALUE_ZERO, PawnValueMg, KnightValueMg, BishopValueMg, RookValueMg, QueenValueMg, VALUE_ZERO, VALUE_ZERO },
@@ -232,7 +225,8 @@ enum : int {
  DEPTH_QS_RECAPTURES = -5,
  DEPTH_NONE   = -6,
-  DEPTH_OFFSET = DEPTH_NONE
+
  DEPTH_OFFSET = -7 // value used only for TT entry occupancy check
 };
 enum Square : int {
@@ -270,93 +264,20 @@ enum Rank : int {
  RANK_1, RANK_2, RANK_3, RANK_4, RANK_5, RANK_6, RANK_7, RANK_8, RANK_NB
 };
-// unique number for each piece type on each square
+// Keep track of what a move changes on the board (used by NNUE)
 enum PieceSquare : uint32_t {
  PS_NONE     =  0,
  PS_W_PAWN   =  1,
  PS_B_PAWN   =  1 * SQUARE_NB + 1,
  PS_W_KNIGHT =  2 * SQUARE_NB + 1,
  PS_B_KNIGHT =  3 * SQUARE_NB + 1,
  PS_W_BISHOP =  4 * SQUARE_NB + 1,
  PS_B_BISHOP =  5 * SQUARE_NB + 1,
  PS_W_ROOK   =  6 * SQUARE_NB + 1,
  PS_B_ROOK   =  7 * SQUARE_NB + 1,
  PS_W_QUEEN  =  8 * SQUARE_NB + 1,
  PS_B_QUEEN  =  9 * SQUARE_NB + 1,
  PS_W_KING   = 10 * SQUARE_NB + 1,
  PS_END      = PS_W_KING, // pieces without kings (pawns included)
  PS_B_KING   = 11 * SQUARE_NB + 1,
  PS_END2     = 12 * SQUARE_NB + 1
 };
 struct ExtPieceSquare {
  PieceSquare from[COLOR_NB];
 };
 // Array for finding the PieceSquare corresponding to the piece on the board
 extern ExtPieceSquare kpp_board_index[PIECE_NB];
 constexpr bool is_ok(PieceId pid);
 constexpr Square rotate180(Square sq);
 // Structure holding which tracked piece (PieceId) is where (PieceSquare)
 class EvalList {
 public:
  // Max. number of pieces without kings is 30 but must be a multiple of 4 in AVX2
  static const int MAX_LENGTH = 32;
  // Array that holds the piece id for the pieces on the board
  PieceId piece_id_list[SQUARE_NB];
  // List of pieces, separate from White and Black POV
  PieceSquare* piece_list_fw() const { return const_cast<PieceSquare*>(pieceListFw); }
  PieceSquare* piece_list_fb() const { return const_cast<PieceSquare*>(pieceListFb); }
  // Place the piece pc with piece_id on the square sq on the board
  void put_piece(PieceId piece_id, Square sq, Piece pc)
  {
      assert(is_ok(piece_id));
      if (pc != NO_PIECE)
      {
          pieceListFw[piece_id] = PieceSquare(kpp_board_index[pc].from[WHITE] + sq);
          pieceListFb[piece_id] = PieceSquare(kpp_board_index[pc].from[BLACK] + rotate180(sq));
          piece_id_list[sq] = piece_id;
      }
      else
      {
          pieceListFw[piece_id] = PS_NONE;
          pieceListFb[piece_id] = PS_NONE;
          piece_id_list[sq] = piece_id;
      }
  }
  // Convert the specified piece_id piece to ExtPieceSquare type and return it
  ExtPieceSquare piece_with_id(PieceId piece_id) const
  {
      ExtPieceSquare eps;
      eps.from[WHITE] = pieceListFw[piece_id];
      eps.from[BLACK] = pieceListFb[piece_id];
      return eps;
  }
 private:
  PieceSquare pieceListFw[MAX_LENGTH];
  PieceSquare pieceListFb[MAX_LENGTH];
 };
 // For differential evaluation of pieces that changed since last turn
 struct DirtyPiece {
  // Number of changed pieces
  int dirty_num;
-  // The ids of changed pieces, max. 2 pieces can change in one move
+  // Max 3 pieces can change in one move. A promotion with capture moves
-  PieceId pieceId[2];
+  // both the pawn and the captured piece to SQ_NONE and the piece promoted
  // to from SQ_NONE to the capture square.
  Piece piece[3];
-  // What changed from the piece with that piece number
+  // From and to squares, which may be SQ_NONE
-  ExtPieceSquare old_piece[2];
+  Square from[3];
-  ExtPieceSquare new_piece[2];
+  Square to[3];
 };
 /// Score enum stores a middlegame and an endgame value in a single integer (enum).
@@ -406,8 +327,6 @@ ENABLE_FULL_OPERATORS_ON(Value)
 ENABLE_FULL_OPERATORS_ON(Direction)
 ENABLE_INCR_OPERATORS_ON(Piece)
 ENABLE_INCR_OPERATORS_ON(PieceSquare)
 ENABLE_INCR_OPERATORS_ON(PieceId)
 ENABLE_INCR_OPERATORS_ON(PieceType)
 ENABLE_INCR_OPERATORS_ON(Square)
 ENABLE_INCR_OPERATORS_ON(File)
@@ -496,10 +415,6 @@ inline Color color_of(Piece pc) {
  return Color(pc >> 3);
 }
 constexpr bool is_ok(PieceId pid) {
  return pid < PIECE_ID_NONE;
 }
 constexpr bool is_ok(Square s) {
  return s >= SQ_A1 && s <= SQ_H8;
 }
@@ -538,11 +453,11 @@ constexpr Square to_sq(Move m) {
 // Return relative square when turning the board 180 degrees
 constexpr Square rotate180(Square sq) {
-  return (Square)(sq ^ 0x3F);
+    return (Square)(sq ^ 0x3F);
 }
 constexpr int from_to(Move m) {
- return m & 0xFFF;
+  return m & 0xFFF;
 }
 constexpr MoveType type_of(Move m) {
@@ -557,10 +472,6 @@ constexpr Move make_move(Square from, Square to) {
  return Move((from << 6) + to);
 }
 constexpr Move reverse_move(Move m) {
  return make_move(to_sq(m), from_sq(m));
 }
 template<MoveType T>
 constexpr Move make(Square from, Square to, PieceType pt = KNIGHT) {
  return Move(T + ((pt - KNIGHT) << 12) + (from << 6) + to);
@@ -575,6 +486,8 @@ constexpr Key make_key(uint64_t seed) {
  return seed * 6364136223846793005ULL + 1442695040888963407ULL;
 }
 } // namespace Stockfish
 #endif // #ifndef TYPES_H_INCLUDED
 #include "tune.h" // Global visibility to tuning setup
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -22,26 +22,32 @@
 #include <sstream>
 #include <string>
 #include "nnue/evaluate_nnue.h"
 #include "evaluate.h"
 #include "movegen.h"
 #include "position.h"
 #include "search.h"
 #include "syzygy/tbprobe.h"
 #include "thread.h"
 #include "timeman.h"
 #include "tt.h"
 #include "uci.h"
-#include "syzygy/tbprobe.h"
+
 #include "tools/validate_training_data.h"
 #include "tools/training_data_generator.h"
 #include "tools/training_data_generator_nonpv.h"
 #include "tools/convert.h"
 #include "tools/transform.h"
 #include "tools/stats.h"
 using namespace std;
 namespace Stockfish {
 extern vector<string> setup_bench(const Position&, istream&);
 namespace {
  // FEN string of the initial position, normal chess
  const char* StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
  // position() is called when engine receives the "position" UCI command.
  // The function sets up the position described in the given FEN string ("fen")
  // or the starting position ("startpos") and then makes the moves given in the
@@ -85,7 +91,7 @@ namespace {
    Position p;
    p.set(pos.fen(), Options["UCI_Chess960"], &states->back(), Threads.main());
-    Eval::verify_NNUE();
+    Eval::NNUE::verify();
    sync_cout << "\n" << Eval::trace(p) << sync_endl;
  }
@@ -94,7 +100,7 @@ namespace {
  // setoption() is called when engine receives the "setoption" UCI command. The
  // function updates the UCI option ("name") to the given value ("value").
-  void setoption(istringstream& is) {
+  void setoption_from_stream(istringstream& is) {
    string token, name, value;
@@ -108,10 +114,7 @@ namespace {
    while (is >> token)
        value += (value.empty() ? "" : " ") + token;
-    if (Options.count(name))
+    UCI::setoption(name, value);
        Options[name] = value;
    else
        sync_cout << "No such option: " << name << sync_endl;
  }
@@ -170,7 +173,7 @@ namespace {
        if (token == "go" || token == "eval")
        {
-            cerr << "\nPosition: " << cnt++ << '/' << num << endl;
+            cerr << "\nPosition: " << cnt++ << '/' << num << " (" << pos.fen() << ")" << endl;
            if (token == "go")
            {
               go(pos, is, states);
@@ -180,7 +183,7 @@ namespace {
            else
               trace_eval(pos);
        }
-        else if (token == "setoption")  setoption(is);
+        else if (token == "setoption")  setoption_from_stream(is);
        else if (token == "position")   position(pos, is, states);
        else if (token == "ucinewgame") { Search::clear(); elapsed = now(); } // Search::clear() may take some while
    }
@@ -195,6 +198,18 @@ namespace {
         << "\nNodes/second    : " << 1000 * nodes / elapsed << endl;
  }
 } // namespace
 namespace UCI {
  void setoption(const std::string& name, const std::string& value)
  {
      if (Options.count(name))
          Options[name] = value;
      else
          sync_cout << "No such option: " << name << sync_endl;
  }
  // The win rate model returns the probability (per mille) of winning given an eval
  // and a game-ply. The model fits rather accurately the LTC fishtest statistics.
  int win_rate_model(Value v, int ply) {
@@ -205,13 +220,13 @@ namespace {
     // Coefficients of a 3rd order polynomial fit based on fishtest data
     // for two parameters needed to transform eval to the argument of a
     // logistic function.
-     double as[] = {-8.24404295, 64.23892342, -95.73056462, 153.86478679};
+     double as[] = {-1.17202460e-01, 5.94729104e-01, 1.12065546e+01, 1.22606222e+02};
-     double bs[] = {-3.37154371, 28.44489198, -56.67657741,  72.05858751};
+     double bs[] = {-1.79066759,  11.30759193, -17.43677612,  36.47147479};
     double a = (((as[0] * m + as[1]) * m + as[2]) * m) + as[3];
     double b = (((bs[0] * m + bs[1]) * m + bs[2]) * m) + bs[3];
     // Transform eval to centipawns with limited range
-     double x = Utility::clamp(double(100 * v) / PawnValueEg, -1000.0, 1000.0);
+     double x = std::clamp(double(100 * v) / PawnValueEg, -1000.0, 1000.0);
     // Return win rate in per mille (rounded to nearest)
     return int(0.5 + 1000 / (1 + std::exp((a - x) / b)));
@@ -219,6 +234,68 @@ namespace {
 } // namespace
 // --------------------
 // Call qsearch(),search() directly for testing
 // --------------------
 void qsearch_cmd(Position& pos)
 {
  cout << "qsearch : ";
  auto pv = Search::qsearch(pos);
  cout << "Value = " << pv.first << " , " << UCI::value(pv.first) << " , PV = ";
  for (auto m : pv.second)
    cout << UCI::move(m, false) << " ";
  cout << endl;
 }
 void search_cmd(Position& pos, istringstream& is)
 {
  string token;
  int depth = 1;
  int multi_pv = (int)Options["MultiPV"];
  while (is >> token)
  {
    if (token == "depth")
      is >> depth;
    if (token == "multipv")
      is >> multi_pv;
  }
  cout << "search depth = " << depth << " , multi_pv = " << multi_pv << " : ";
  auto pv = Search::search(pos, depth, multi_pv);
  cout << "Value = " << pv.first << " , " << UCI::value(pv.first) << " , PV = ";
  for (auto m : pv.second)
    cout << UCI::move(m, false) << " ";
  cout << endl;
 }
 void search_mcts_cmd(Position& pos, istringstream& is)
 {
  string token;
  int nodes = 1000;
  int leafDepth = 3;
  float explorationFactor = 0.25f;
  while (is >> token)
  {
    if (token == "nodes")
      is >> nodes;
    if (token == "leaf_depth")
      is >> leafDepth;
    if (token == "exploration_factor")
      is >> explorationFactor;
  }
  cout << "search nodes = " << nodes << " , leaf_depth = " << leafDepth << " :\n";
  auto continuations = Search::MCTS::search_mcts_multipv(pos, nodes, leafDepth, explorationFactor);
  for (auto&& [numVisits, value, actionValue, pv] : continuations)
  {
    cout << "NumVisits = " << numVisits << " , Value = " << UCI::value(value) << " , ActionValue = " << actionValue << " , PV = ";
    for (auto m : pv)
      cout << UCI::move(m, false) << " ";
    cout << endl;
  }
  cout << endl;
 }
 /// UCI::loop() waits for a command from stdin, parses it and calls the appropriate
 /// function. Also intercepts EOF from stdin to ensure gracefully exiting if the
@@ -262,7 +339,7 @@ void UCI::loop(int argc, char* argv[]) {
                    << "\n"       << Options
                    << "\nuciok"  << sync_endl;
-      else if (token == "setoption")  setoption(is);
+      else if (token == "setoption")  setoption_from_stream(is);
      else if (token == "go")         go(pos, is, states);
      else if (token == "position")   position(pos, is, states);
      else if (token == "ucinewgame") Search::clear();
@@ -275,8 +352,42 @@ void UCI::loop(int argc, char* argv[]) {
      else if (token == "d")        sync_cout << pos << sync_endl;
      else if (token == "eval")     trace_eval(pos);
      else if (token == "compiler") sync_cout << compiler_info() << sync_endl;
-      else
+      else if (token == "export_net")
-          sync_cout << "Unknown command: " << cmd << sync_endl;
+      {
          std::optional<std::string> filename;
          std::string f;
          if (is >> skipws >> f)
              filename = f;
          Eval::NNUE::save_eval(filename);
      }
      else if (token == "generate_training_data") Tools::generate_training_data(is);
      else if (token == "generate_training_data_nonpv") Tools::generate_training_data_nonpv(is);
      else if (token == "convert") Tools::convert(is);
      else if (token == "validate_training_data") Tools::validate_training_data(is);
      else if (token == "convert_bin") Tools::convert_bin(is);
      else if (token == "convert_plain") Tools::convert_plain(is);
      else if (token == "convert_bin_from_pgn_extract") Tools::convert_bin_from_pgn_extract(is);
      else if (token == "transform") Tools::transform(is);
      else if (token == "gather_statistics") Tools::Stats::gather_statistics(is);
      // Command to call qsearch(),search() directly for testing
      else if (token == "qsearch") qsearch_cmd(pos);
      else if (token == "search_mcts") search_mcts_cmd(pos, is);
      else if (token == "search") search_cmd(pos, is);
      else if (token == "tasktest")
      {
        Threads.execute_with_workers([](auto& th) {
          std::cout << th.id() << '\n';
        });
      }
      else if (token == "--help" || token == "help" || token == "--license" || token == "license")
          sync_cout << "\nStockfish is a powerful chess engine and free software licensed under the GNU GPLv3."
                       "\nStockfish is normally used with a separate graphical user interface (GUI)."
                       "\nStockfish implements the universal chess interface (UCI) to exchange information."
                       "\nFor further information see https://github.com/official-stockfish/Stockfish#readme"
                       "\nor the corresponding README.md and Copying.txt files distributed with this program.\n" << sync_endl;
      else if (!token.empty() && token[0] != '#')
          sync_cout << "Unknown command: '" << cmd << "'. Type help for more information." << sync_endl;
  } while (token != "quit" && argc == 1); // Command line args are one-shot
 }
@@ -369,3 +480,5 @@ Move UCI::to_move(const Position& pos, string& str) {
  return MOVE_NONE;
 }
 } // namespace Stockfish
@@ -1,6 +1,6 @@
 /*
  Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2022 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
@@ -24,6 +24,8 @@
 #include "types.h"
 namespace Stockfish {
 class Position;
 namespace UCI {
@@ -73,9 +75,12 @@ std::string move(Move m, bool chess960);
 std::string pv(const Position& pos, Depth depth, Value alpha, Value beta);
 std::string wdl(Value v, int ply);
 Move to_move(const Position& pos, std::string& str);
 void setoption(const std::string& name, const std::string& value);
 } // namespace UCI
 extern UCI::OptionsMap Options;
 } // namespace Stockfish
 #endif // #ifndef UCI_H_INCLUDED
--- a/Show More
+++ b/Show More