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.

.github/workflows/stockfish.yml

@@ -17,7 +17,6 @@ jobs:
     env:
       COMPILER: ${{ matrix.config.compiler }}
       COMP: ${{ matrix.config.comp }}
-      CXXFLAGS: "-Werror"
     strategy:
       matrix:
         config:
@@ -25,15 +24,12 @@ jobs:
           # run_expensive_tests: true
           # run_32bit_tests: true
           # run_64bit_tests: true
-          # run_armv8_tests: true
-          # run_armv7_tests: true
           - {
               name: "Ubuntu 20.04 GCC",
               os: ubuntu-20.04,
               compiler: g++,
               comp: gcc,
               run_expensive_tests: true,
-              run_32bit_tests: true,
               run_64bit_tests: true,
               shell: 'bash {0}'
             }
@@ -42,26 +38,9 @@ jobs:
               os: ubuntu-20.04,
               compiler: clang++,
               comp: clang,
-              run_32bit_tests: true,
               run_64bit_tests: true,
               shell: 'bash {0}'
             }
-          - {
-              name: "Ubuntu 20.04 NDK armv8",
-              os: ubuntu-20.04,
-              compiler: aarch64-linux-android21-clang++,
-              comp: ndk,
-              run_armv8_tests: true,
-              shell: 'bash {0}'
-            }
-          - {
-              name: "Ubuntu 20.04 NDK armv7",
-              os: ubuntu-20.04,
-              compiler: armv7a-linux-androideabi21-clang++,
-              comp: ndk,
-              run_armv7_tests: true,
-              shell: 'bash {0}'
-            }
           - {
               name: "MacOS 10.15 Apple Clang",
               os: macos-10.15,
@@ -88,26 +67,6 @@ jobs:
               msys_env: 'x86_64-gcc',
               shell: 'msys2 {0}'
             }
-          - {
-              name: "Windows 2022 Mingw-w64 GCC i686",
-              os: windows-2022,
-              compiler: g++,
-              comp: mingw,
-              run_32bit_tests: true,
-              msys_sys: 'mingw32',
-              msys_env: 'i686-gcc',
-              shell: 'msys2 {0}'
-            }
-          - {
-              name: "Windows 2022 Mingw-w64 Clang x86_64",
-              os: windows-2022,
-              compiler: clang++,
-              comp: clang,
-              run_64bit_tests: true,
-              msys_sys: 'clang64',
-              msys_env: 'clang-x86_64-clang',
-              shell: 'msys2 {0}'
-            }
 
     defaults:
       run:
@@ -122,7 +81,7 @@ jobs:
         if: runner.os == 'Linux'
         run: |
           sudo apt update
-          sudo apt install expect valgrind g++-multilib qemu-user
+          sudo apt install expect valgrind g++-multilib
 
       - name: Setup msys and install required packages
         if: runner.os == 'Windows'
@@ -142,7 +101,6 @@ jobs:
 
       - name: Check compiler
         run: |
-          export PATH=$PATH:$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin
           $COMPILER -v
 
       - name: Test help target
@@ -154,7 +112,7 @@ jobs:
       - name: Test debug x86-32 build
         if: ${{ matrix.config.run_32bit_tests }}
         run: |
-          export CXXFLAGS="-Werror -D_GLIBCXX_DEBUG"
+          export CXXFLAGS="-D_GLIBCXX_DEBUG"
           make clean
           make -j2 ARCH=x86-32 optimize=no debug=yes build
           ../tests/signature.sh $benchref
@@ -192,7 +150,7 @@ jobs:
       - name: Test debug x86-64-modern build
         if: ${{ matrix.config.run_64bit_tests }}
         run: |
-          export CXXFLAGS="-Werror -D_GLIBCXX_DEBUG"
+          export CXXFLAGS="-D_GLIBCXX_DEBUG"
           make clean
           make -j2 ARCH=x86-64-modern optimize=no debug=yes build
           ../tests/signature.sh $benchref
@@ -264,37 +222,6 @@ jobs:
           make clean
           make -j2 ARCH=x86-64-vnni256 build
 
-      # armv8 tests
-
-      - name: Test armv8 build
-        if: ${{ matrix.config.run_armv8_tests }}
-        run: |
-          export PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH
-          export LDFLAGS="-static -Wno-unused-command-line-argument"
-          make clean
-          make -j2 ARCH=armv8 build
-          ../tests/signature.sh $benchref
-
-      # armv7 tests
-
-      - name: Test armv7 build
-        if: ${{ matrix.config.run_armv7_tests }}
-        run: |
-          export PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH
-          export LDFLAGS="-static -Wno-unused-command-line-argument"
-          make clean
-          make -j2 ARCH=armv7 build
-          ../tests/signature.sh $benchref
-
-      - name: Test armv7-neon build
-        if: ${{ matrix.config.run_armv7_tests }}
-        run: |
-          export PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH
-          export LDFLAGS="-static -Wno-unused-command-line-argument"
-          make clean
-          make -j2 ARCH=armv7-neon build
-          ../tests/signature.sh $benchref
-
       # Other tests
 
       - name: Check perft and search reproducibility

README.md

@@ -166,6 +166,41 @@ For developers the following non-standard commands might be of interest, mainly
   * #### 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
 

docs/binpack.md

@@ -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)
+```

docs/convert.md

@@ -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.

docs/generate_training_data.md

@@ -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.

docs/generate_training_data_nonpv.md

@@ -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.

docs/stats.md

@@ -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)

docs/transform.md

@@ -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.
+

docs/validate_training_data.md

@@ -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`).

script/README.md

@@ -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
+
+

script/extract_bin.py

@@ -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()

script/interleave_binpacks.py

@@ -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()

script/pgn_to_plain.py

@@ -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()

script/shuffle_binpack.py

@@ -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()

src/Makefile

@@ -44,6 +44,12 @@ else
 	EXE = stockfish
 endif
 
+### Establish the operating system name
+KERNEL = $(shell uname -s)
+ifeq ($(KERNEL),Linux)
+	OS = $(shell uname -o)
+endif
+
 ### Installation dir definitions
 PREFIX = /usr/local
 BINDIR = $(PREFIX)/bin
@@ -51,19 +57,30 @@ BINDIR = $(PREFIX)/bin
 ### Built-in benchmark for pgo-builds
 ifeq ($(SDE_PATH),)
 	PGOBENCH = $(WINE_PATH) ./$(EXE) bench
+	PGOGENSFEN = $(WINE_PATH) ./$(EXE) gensfen depth 3 loop 1000 sfen_format bin output_file_name $(PGO_TRAINING_DATA_FILE)
 else
 	PGOBENCH = $(SDE_PATH) -- $(WINE_PATH) ./$(EXE) bench
+	PGOGENSFEN = $(SDE_PATH) -- $(WINE_PATH) ./$(EXE) gensfen depth 3 loop 1000 sfen_format bin output_file_name $(PGO_TRAINING_DATA_FILE)
 endif
 
 ### Source and object files
 SRCS = benchmark.cpp bitbase.cpp bitboard.cpp endgame.cpp evaluate.cpp main.cpp \
 	material.cpp misc.cpp movegen.cpp movepick.cpp pawns.cpp position.cpp psqt.cpp \
 	search.cpp thread.cpp timeman.cpp tt.cpp uci.cpp ucioption.cpp tune.cpp syzygy/tbprobe.cpp \
-	nnue/evaluate_nnue.cpp nnue/features/half_ka_v2_hm.cpp
+	nnue/evaluate_nnue.cpp \
+	nnue/features/half_ka_v2_hm.cpp \
+	tools/validate_training_data.cpp \
+	tools/sfen_packer.cpp \
+	tools/training_data_generator.cpp \
+	tools/training_data_generator_nonpv.cpp \
+	tools/opening_book.cpp \
+	tools/convert.cpp \
+	tools/transform.cpp \
+	tools/stats.cpp
 
 OBJS = $(notdir $(SRCS:.cpp=.o))
 
-VPATH = syzygy:nnue:nnue/features
+VPATH = syzygy:nnue:nnue/features:eval:extra:tools
 
 ### ==========================================================================
 ### Section 2. High-level Configuration
@@ -351,8 +368,9 @@ ifeq ($(MAKELEVEL),0)
        export ENV_LDFLAGS := $(LDFLAGS)
 endif
 
-CXXFLAGS = $(ENV_CXXFLAGS) -Wall -Wcast-qual -fno-exceptions -std=c++17 $(EXTRACXXFLAGS)
-DEPENDFLAGS = $(ENV_DEPENDFLAGS) -std=c++17
+ADDITIONAL_INCLUDE_DIRECTORIES = -I.
+CXXFLAGS = $(ENV_CXXFLAGS) -Wall -Wcast-qual -fno-exceptions -std=c++17 $(ADDITIONAL_INCLUDE_DIRECTORIES) $(EXTRACXXFLAGS)
+DEPENDFLAGS = $(ENV_DEPENDFLAGS) -std=c++17 $(ADDITIONAL_INCLUDE_DIRECTORIES)
 LDFLAGS = $(ENV_LDFLAGS) $(EXTRALDFLAGS)
 
 ifeq ($(COMP),)
@@ -362,7 +380,7 @@ endif
 ifeq ($(COMP),gcc)
 	comp=gcc
 	CXX=g++
-	CXXFLAGS += -pedantic -Wextra -Wshadow
+	CXXFLAGS += -pedantic -Wextra -Wshadow -lstdc++fs
 
 	ifeq ($(arch),$(filter $(arch),armv7 armv8))
 		ifeq ($(OS),Android)
@@ -441,8 +459,8 @@ ifeq ($(COMP),clang)
 endif
 
 ifeq ($(KERNEL),Darwin)
-	CXXFLAGS += -mmacosx-version-min=10.14
-	LDFLAGS += -mmacosx-version-min=10.14
+	CXXFLAGS += -mmacosx-version-min=10.15
+	LDFLAGS += -mmacosx-version-min=10.15
 	ifneq ($(arch),any)
 		CXXFLAGS += -arch $(arch)
 		LDFLAGS += -arch $(arch)
@@ -518,13 +536,13 @@ ifneq ($(comp),mingw)
 		# Haiku has pthreads in its libroot, so only link it in on other platforms
 		ifneq ($(KERNEL),Haiku)
 			ifneq ($(COMP),ndk)
-				LDFLAGS += -lpthread
-			endif
+			LDFLAGS += -lpthread
 		endif
 	endif
 endif
+endif
 
-### 3.2.1 Debugging
+### 3.2.2 Debugging
 ifeq ($(debug),no)
 	CXXFLAGS += -DNDEBUG
 else
@@ -661,7 +679,7 @@ ifeq ($(neon),yes)
 		CXXFLAGS += -mfpu=neon
 	endif
 	endif
-	endif
+endif
 endif
 
 ### 3.7 pext
@@ -803,6 +821,7 @@ profile-build: net config-sanity objclean profileclean
 	@echo ""
 	@echo "Step 2/4. Running benchmark for pgo-build ..."
 	$(PGOBENCH) 2>&1 | tail -n 4
+	$(PGOGENSFEN) 2>&1 | tail -n 4
 	@echo ""
 	@echo "Step 3/4. Building optimized executable ..."
 	$(MAKE) ARCH=$(ARCH) COMP=$(COMP) objclean
@@ -849,12 +868,12 @@ net:
 
 # clean binaries and objects
 objclean:
-	@rm -f stockfish stockfish.exe *.o ./syzygy/*.o ./nnue/*.o ./nnue/features/*.o
+	@rm -f $(EXE) *.o ./syzygy/*.o ./nnue/*.o ./nnue/features/*.o ./tools/*.o ./extra/*.o ./eval/*.o
 
 # clean auxiliary profiling files
 profileclean:
 	@rm -rf profdir
-	@rm -f bench.txt *.gcda *.gcno ./syzygy/*.gcda ./nnue/*.gcda ./nnue/features/*.gcda *.s
+	@rm -f bench.txt *.gcda *.gcno ./syzygy/*.gcda ./nnue/*.gcda ./nnue/features/*.gcda *.s ./tools/*.gcda ./extra/*.gcda ./eval/*.gcda
 	@rm -f stockfish.profdata *.profraw
 	@rm -f stockfish.*args*
 	@rm -f stockfish.*lt*

src/evaluate.cpp

@@ -27,6 +27,8 @@
 #include <streambuf>
 #include <vector>
 
+#include "nnue/evaluate_nnue.h"
+
 #include "bitboard.h"
 #include "evaluate.h"
 #include "material.h"
@@ -37,7 +39,6 @@
 #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
@@ -53,15 +54,28 @@
   const unsigned int         gEmbeddedNNUESize = 1;
 #endif
 
-
 using namespace std;
 
 namespace Stockfish {
 
 namespace Eval {
 
-  bool useNNUE;
-  string currentEvalFileName = "None";
+  namespace NNUE {
+    string currentEvalFileName = "None";
+    UseNNUEMode useNNUE;
+
+    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;
+
+      return UseNNUEMode::False;
+    }
+  }
 
   /// 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"
@@ -73,8 +87,8 @@ namespace Eval {
 
   void NNUE::init() {
 
-    useNNUE = Options["Use NNUE"];
-    if (!useNNUE)
+    useNNUE = nnue_mode_from_option(Options["Use NNUE"]);
+    if (useNNUE == UseNNUEMode::False)
         return;
 
     string eval_file = string(Options["EvalFile"]);
@@ -124,7 +138,7 @@ namespace Eval {
     if (eval_file.empty())
         eval_file = EvalFileDefaultName;
 
-    if (useNNUE && currentEvalFileName != eval_file)
+    if (useNNUE != UseNNUEMode::False && currentEvalFileName != eval_file)
     {
 
         string msg1 = "If the UCI option \"Use NNUE\" is set to true, network evaluation parameters compatible with the engine must be available.";
@@ -142,7 +156,7 @@ namespace Eval {
         exit(EXIT_FAILURE);
     }
 
-    if (useNNUE)
+    if (useNNUE != UseNNUEMode::False)
         sync_cout << "info string NNUE evaluation using " << eval_file << " enabled" << sync_endl;
     else
         sync_cout << "info string classical evaluation enabled" << sync_endl;
@@ -1082,7 +1096,19 @@ make_v:
 
 Value Eval::evaluate(const Position& pos) {
 
+  pos.this_thread()->on_eval();
+
   Value v;
+
+  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) &&
@@ -1090,14 +1116,14 @@ Value Eval::evaluate(const Position& pos) {
 
   // 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 (!useNNUE || useClassical)
+  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 (useNNUE && !useClassical)
+  if (NNUE::useNNUE != NNUE::UseNNUEMode::False && !useClassical)
   {
        Value nnue     = NNUE::evaluate(pos, true);     // NNUE
        int scale      = 1080 + 110 * pos.non_pawn_material() / 5120;
@@ -1171,14 +1197,14 @@ std::string Eval::trace(Position& pos) {
      << "|      Total | " << Term(TOTAL)
      << "+------------+-------------+-------------+-------------+\n";
 
-  if (Eval::useNNUE)
+  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 (Eval::useNNUE)
+  if (NNUE::useNNUE != NNUE::UseNNUEMode::False)
   {
       v = NNUE::evaluate(pos, false);
       v = pos.side_to_move() == WHITE ? v : -v;
@@ -1188,7 +1214,7 @@ std::string Eval::trace(Position& pos) {
   v = evaluate(pos);
   v = pos.side_to_move() == WHITE ? v : -v;
   ss << "Final evaluation       " << to_cp(v) << " (white side)";
-  if (Eval::useNNUE)
+  if (NNUE::useNNUE != NNUE::UseNNUEMode::False)
      ss << " [with scaled NNUE, hybrid, ...]";
   ss << "\n";
 

src/evaluate.h

@@ -33,15 +33,21 @@ namespace Eval {
   std::string trace(Position& pos);
   Value evaluate(const Position& pos);
 
-  extern bool useNNUE;
-  extern std::string currentEvalFileName;
-
   // The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
   // for the build process (profile-build and fishtest) to work. Do not change the
   // name of the macro, as it is used in the Makefile.
   #define EvalFileDefaultName   "nn-3c0aa92af1da.nnue"
 
   namespace NNUE {
+    enum struct UseNNUEMode
+    {
+      False,
+      True,
+      Pure
+    };
+
+    extern UseNNUEMode useNNUE;
+    extern std::string currentEvalFileName;
 
     std::string trace(Position& pos);
     Value evaluate(const Position& pos, bool adjusted = false);

src/main.cpp

@@ -18,6 +18,8 @@
 
 #include <iostream>
 
+#include "nnue/evaluate_nnue.h"
+
 #include "bitboard.h"
 #include "endgame.h"
 #include "position.h"

src/misc.cpp

@@ -65,6 +65,8 @@ 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
@@ -669,4 +671,30 @@ void init(int argc, char* argv[]) {
 
 } // 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

src/misc.h

@@ -19,12 +19,21 @@
 #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"
 
@@ -44,6 +53,20 @@ 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() {
@@ -138,6 +161,219 @@ private:
   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:
@@ -181,6 +417,19 @@ inline int64_t sigmoid(int64_t t, int64_t x0,
 /// 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;
@@ -192,7 +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()); }
 
@@ -200,8 +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;
@@ -216,6 +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
@@ -226,6 +591,111 @@ 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[]);
 

src/movegen.h

@@ -68,6 +68,9 @@ 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;
 };

src/nnue/nnue_common.h

@@ -21,6 +21,8 @@
 #ifndef NNUE_COMMON_H_INCLUDED
 #define NNUE_COMMON_H_INCLUDED
 
+#include "../types.h"
+
 #include <cstring>
 #include <iostream>
 

src/nnue/nnue_feature_transformer.h

@@ -24,6 +24,9 @@
 #include "nnue_common.h"
 #include "nnue_architecture.h"
 
+#include "../misc.h"
+#include "../position.h"
+
 #include <cstring> // std::memset()
 
 namespace Stockfish::Eval::NNUE {

src/position.cpp

@@ -23,6 +23,8 @@
 #include <iomanip>
 #include <sstream>
 
+#include "nnue/evaluate_nnue.h"
+
 #include "bitboard.h"
 #include "misc.h"
 #include "movegen.h"
@@ -32,6 +34,9 @@
 #include "uci.h"
 #include "syzygy/tbprobe.h"
 
+#include "tools/packed_sfen.h"
+#include "tools/sfen_packer.h"
+
 using std::string;
 
 namespace Stockfish {
@@ -754,7 +759,7 @@ 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;  // 1 piece moved, 1 piece captured
           dp.piece[1] = captured;
@@ -798,7 +803,7 @@ 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)
       {
           dp.piece[0] = pc;
           dp.from[0] = from;
@@ -829,7 +834,7 @@ 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)
           {
               // Promoting pawn to SQ_NONE, promoted piece from SQ_NONE
               dp.to[0] = SQ_NONE;
@@ -967,7 +972,7 @@ 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 (Do && Eval::useNNUE)
+  if (Do && Eval::NNUE::useNNUE != Eval::NNUE::UseNNUEMode::False)
   {
       auto& dp = st->dirtyPiece;
       dp.piece[0] = make_piece(us, KING);
@@ -1001,6 +1006,7 @@ void Position::do_null_move(StateInfo& newSt) {
   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;
@@ -1177,6 +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.
 
@@ -1345,4 +1367,18 @@ 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

src/position.h

@@ -31,6 +31,9 @@
 
 #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
@@ -126,6 +129,7 @@ public:
   bool legal(Move m) const;
   bool pseudo_legal(const Move m) const;
   bool capture(Move m) const;
+  bool capture_or_promotion(Move m) const;
   bool gives_check(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;
@@ -171,6 +177,28 @@ public:
   // Used by NNUE
   StateInfo* state() 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);
 
@@ -368,6 +396,11 @@ inline bool Position::is_chess960() const {
   return chess960;
 }
 
+inline bool Position::capture_or_promotion(Move m) const {
+  assert(is_ok(m));
+  return type_of(m) != NORMAL ? type_of(m) != CASTLING : !empty(to_sq(m));
+}
+
 inline bool Position::capture(Move m) const {
   assert(is_ok(m));
   // Castling is encoded as "king captures rook"
@@ -425,6 +458,8 @@ inline StateInfo* Position::state() const {
   return st;
 }
 
+static const char* const StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1";
+
 } // namespace Stockfish
 
 #endif // #ifndef POSITION_H_INCLUDED

src/search.h

@@ -24,6 +24,7 @@
 #include "misc.h"
 #include "movepick.h"
 #include "types.h"
+#include "uci.h"
 
 namespace Stockfish {
 
@@ -34,6 +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
@@ -92,6 +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 {
@@ -102,6 +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;
@@ -109,7 +115,35 @@ 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
 

src/syzygy/tbprobe.cpp

@@ -52,7 +52,7 @@
 
 using namespace Stockfish::Tablebases;
 
-int Stockfish::Tablebases::MaxCardinality;
+int Stockfish::Tablebases::MaxCardinality = 0;
 
 namespace Stockfish {
 

src/thread.cpp

@@ -34,9 +34,10 @@ 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();
 }
 
 
@@ -82,6 +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.
@@ -93,6 +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.
 
@@ -110,15 +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();
+      }
   }
 }
 
@@ -166,6 +191,13 @@ void ThreadPool::clear() {
   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.
@@ -186,9 +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());
@@ -264,4 +293,11 @@ void ThreadPool::wait_for_search_finished() const {
             th->wait_for_search_finished();
 }
 
+
+void ThreadPool::wait_for_workers_finished() const {
+
+    for (Thread* th : *this)
+        th->wait_for_worker_finished();
+}
+
 } // namespace Stockfish

src/thread.h

@@ -24,6 +24,7 @@
 #include <mutex>
 #include <thread>
 #include <vector>
+#include <functional>
 
 #include "material.h"
 #include "movepick.h"
@@ -39,24 +40,51 @@ namespace Stockfish {
 /// 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();
   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;
@@ -65,6 +93,7 @@ public:
   int selDepth, nmpMinPly;
   Color nmpColor;
   Value bestValue, optimism[COLOR_NB];
+  uint64_t maxNodes;
 
   Position rootPos;
   StateInfo rootState;
@@ -76,6 +105,11 @@ public:
   CapturePieceToHistory captureHistory;
   ContinuationHistory continuationHistory[2][2];
   Score trend;
+  int failedHighCnt;
+  bool rootInTB;
+  int Cardinality;
+  bool UseRule50;
+  Depth ProbeDepth;
 };
 
 
@@ -104,6 +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);
@@ -114,6 +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;
 

src/tools/convert.cpp

@@ -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);
+    }
+}

src/tools/convert.h

@@ -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

src/tools/opening_book.cpp

@@ -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;
+    }
+
+}

src/tools/opening_book.h

@@ -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

src/tools/packed_sfen.h

@@ -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

src/tools/sfen_packer.cpp

@@ -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;
+    }
+}

src/tools/sfen_packer.h

@@ -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

src/tools/sfen_reader.h

@@ -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;
+    };
+}

src/tools/sfen_stream.h

@@ -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

src/tools/sfen_writer.h

@@ -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;
+    };
+}

src/tools/stats.h

@@ -0,0 +1,12 @@
+#ifndef _STATS_H_
+#define _STATS_H_
+
+#include <sstream>
+
+namespace Stockfish::Tools::Stats {
+
+    void gather_statistics(std::istringstream& is);
+
+}
+
+#endif

src/tools/training_data_generator.cpp

@@ -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;
+    }
+}

src/tools/training_data_generator.h

@@ -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

src/tools/training_data_generator_nonpv.cpp

@@ -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;
+    }
+}

src/tools/training_data_generator_nonpv.h

@@ -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

src/tools/transform.h

@@ -0,0 +1,12 @@
+#ifndef _TRANSFORM_H_
+#define _TRANSFORM_H_
+
+#include <sstream>
+
+namespace Stockfish::Tools {
+
+    void transform(std::istringstream& is);
+
+}
+
+#endif

src/tools/validate_training_data.cpp

@@ -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);
+    }
+}

src/tools/validate_training_data.h

@@ -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

src/tt.cpp

@@ -30,11 +30,16 @@ 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;
@@ -119,6 +124,11 @@ 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
 

src/tt.h

@@ -92,6 +92,8 @@ public:
     return &table[mul_hi64(key, clusterCount)].entry[0];
   }
 
+  static bool enable_transposition_table;
+
 private:
   friend struct TTEntry;
 

src/types.h

@@ -137,6 +137,8 @@ enum Color {
   WHITE, BLACK, COLOR_NB = 2
 };
 
+constexpr Color Colors[2] = { WHITE, BLACK };
+
 enum CastlingRights {
   NO_CASTLING,
   WHITE_OO,
@@ -449,6 +451,11 @@ constexpr Square to_sq(Move m) {
   return Square(m & 0x3F);
 }
 
+// Return relative square when turning the board 180 degrees
+constexpr Square rotate180(Square sq) {
+    return (Square)(sq ^ 0x3F);
+}
+
 constexpr int from_to(Move m) {
   return m & 0xFFF;
 }

src/uci.cpp

@@ -22,15 +22,23 @@
 #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;
 
@@ -40,10 +48,6 @@ 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
@@ -96,7 +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;
 
@@ -110,10 +114,7 @@ namespace {
     while (is >> token)
         value += (value.empty() ? "" : " ") + token;
 
-    if (Options.count(name))
-        Options[name] = value;
-    else
-        sync_cout << "No such option: " << name << sync_endl;
+    UCI::setoption(name, value);
   }
 
 
@@ -182,7 +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
     }
@@ -197,6 +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) {
@@ -213,7 +226,7 @@ namespace {
      double b = (((bs[0] * m + bs[1]) * m + bs[2]) * m) + bs[3];
 
      // Transform eval to centipawns with limited range
-     double x = std::clamp(double(100 * v) / PawnValueEg, -2000.0, 2000.0);
+     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)));
@@ -221,6 +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
@@ -264,7 +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();
@@ -285,6 +360,26 @@ void UCI::loop(int argc, char* argv[]) {
               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)."

src/uci.h

@@ -75,6 +75,7 @@ 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
 

src/ucioption.cpp

@@ -21,6 +21,7 @@
 #include <ostream>
 #include <sstream>
 
+#include "nnue/evaluate_nnue.h"
 #include "evaluate.h"
 #include "misc.h"
 #include "search.h"
@@ -45,6 +46,12 @@ void on_threads(const Option& o) { Threads.set(size_t(o)); }
 void on_tb_path(const Option& o) { Tablebases::init(o); }
 void on_use_NNUE(const Option& ) { Eval::NNUE::init(); }
 void on_eval_file(const Option& ) { Eval::NNUE::init(); }
+void on_prune_at_shallow_depth(const Option& o) {
+    Search::prune_at_shallow_depth = o;
+}
+void on_enable_transposition_table(const Option& o) {
+    TranspositionTable::enable_transposition_table = o;
+}
 
 /// Our case insensitive less() function as required by UCI protocol
 bool CaseInsensitiveLess::operator() (const string& s1, const string& s2) const {
@@ -79,8 +86,22 @@ void init(OptionsMap& o) {
   o["SyzygyProbeDepth"]      << Option(1, 1, 100);
   o["Syzygy50MoveRule"]      << Option(true);
   o["SyzygyProbeLimit"]      << Option(7, 0, 7);
-  o["Use NNUE"]              << Option(true, on_use_NNUE);
+  o["Use NNUE"]              << Option("true var true var false var pure", "true", on_use_NNUE);
   o["EvalFile"]              << Option(EvalFileDefaultName, on_eval_file);
+  // When the evaluation function is loaded at the ucinewgame timing, it is necessary to convert the new evaluation function.
+  // I want to hit the test eval convert command, but there is no new evaluation function
+  // It ends abnormally before executing this command.
+  // Therefore, with this hidden option, you can suppress the loading of the evaluation function when ucinewgame,
+  // Hit the test eval convert command.
+  o["SkipLoadingEval"]       << Option(false);
+  // When learning the evaluation function, you can change the folder to save the evaluation function.
+  // Evalsave by default. This folder shall be prepared in advance.
+  // Automatically create a folder under this folder like "0/", "1/", ... and save the evaluation function file there.
+  o["EvalSaveDir"] << Option("evalsave");
+  // Prune at shallow depth on PV nodes. False is recommended when using fixed depth search.
+  o["PruneAtShallowDepth"] << Option(true, on_prune_at_shallow_depth);
+  // Enable transposition table.
+  o["EnableTranspositionTable"] << Option(true, on_enable_transposition_table);
 }
 
 
@@ -134,7 +155,7 @@ Option::operator double() const {
 }
 
 Option::operator std::string() const {
-  assert(type == "string");
+  assert(type == "check" || type == "spin" || type == "combo" || type == "button" || type == "string");
   return currentValue;
 }
 

tests/instrumented.sh

@@ -16,6 +16,9 @@ case $1 in
     exeprefix='valgrind --error-exitcode=42 --errors-for-leak-kinds=all --leak-check=full'
     postfix='1>/dev/null'
     threads="1"
+    bench_depth=5
+    go_depth=10
+    tt_size=16
   ;;
   --valgrind-thread)
     echo "valgrind-thread testing started"
@@ -23,6 +26,9 @@ case $1 in
     exeprefix='valgrind --fair-sched=try --error-exitcode=42'
     postfix='1>/dev/null'
     threads="2"
+    bench_depth=5
+    go_depth=10
+    tt_size=16
   ;;
   --sanitizer-undefined)
     echo "sanitizer-undefined testing started"
@@ -30,6 +36,9 @@ case $1 in
     exeprefix=''
     postfix='2>&1 | grep -A50 "runtime error:"'
     threads="1"
+    bench_depth=8
+    go_depth=20
+    tt_size=128
   ;;
   --sanitizer-thread)
     echo "sanitizer-thread testing started"
@@ -37,6 +46,9 @@ case $1 in
     exeprefix=''
     postfix='2>&1 | grep -A50 "WARNING: ThreadSanitizer:"'
     threads="2"
+    bench_depth=8
+    go_depth=20
+    tt_size=128
 
 cat << EOF > tsan.supp
 race:Stockfish::TTEntry::move
@@ -70,7 +82,7 @@ for args in "eval" \
             "go depth 10" \
             "go movetime 1000" \
             "go wtime 8000 btime 8000 winc 500 binc 500" \
-            "bench 128 $threads 8 default depth"
+            "bench $tt_size $threads $bench_depth default depth"
 do
 
    echo "$prefix $exeprefix ./stockfish $args $postfix"
@@ -121,7 +133,7 @@ cat << EOF > syzygy.exp
  send "uci\n"
  send "setoption name SyzygyPath value ../tests/syzygy/\n"
  expect "info string Found 35 tablebases" {} timeout {exit 1}
- send "bench 128 1 8 default depth\n"
+ send "bench $tt_size 1 $bench_depth default depth\n"
  send "quit\n"
  expect eof
 
@@ -130,7 +142,57 @@ cat << EOF > syzygy.exp
  exit \$value
 EOF
 
-for exp in game.exp syzygy.exp
+# generate_training_data testing 01
+cat << EOF > data_generation01.exp
+ set timeout 240
+ spawn $exeprefix ./stockfish
+
+ send "uci\n"
+ expect "uciok"
+
+ send "setoption name Threads value $threads\n"
+ send "setoption name Use NNUE value false\n"
+ send "isready\n"
+ send "generate_training_data depth 3 count 100 keep_draws 1 eval_limit 32000 output_file_name training_data/training_data.bin output_format bin\n"
+ expect "INFO: Gensfen finished."
+ send "convert_plain targetfile training_data/training_data.bin output_file_name training_data.txt\n"
+ expect "all done"
+ send "generate_training_data depth 3 count 100 keep_draws 1 eval_limit 32000 output_file_name training_data/training_data.binpack output_format binpack\n"
+ expect "INFO: Gensfen finished."
+
+ send "quit\n"
+ expect eof
+
+ # return error code of the spawned program, useful for valgrind
+ lassign [wait] pid spawnid os_error_flag value
+ exit \$value
+EOF
+
+# generate_training_data testing 02
+cat << EOF > data_generation02.exp
+ set timeout 240
+ spawn $exeprefix ./stockfish
+
+ send "uci\n"
+ expect "uciok"
+
+ send "setoption name Threads value $threads\n"
+ send "setoption name Use NNUE value true\n"
+ send "isready\n"
+ send "generate_training_data depth 4 count 50 keep_draws 1 eval_limit 32000 output_file_name validation_data/validation_data.bin output_format bin\n"
+ expect "INFO: Gensfen finished."
+ send "generate_training_data depth 4 count 50 keep_draws 1 eval_limit 32000 output_file_name validation_data/validation_data.binpack output_format binpack\n"
+ expect "INFO: Gensfen finished."
+
+ send "quit\n"
+ expect eof
+
+ # return error code of the spawned program, useful for valgrind
+ lassign [wait] pid spawnid os_error_flag value
+ exit \$value
+EOF
+
+for exp in game.exp data_generation01.exp data_generation02.exp
 do
 
   echo "$prefix expect $exp $postfix"