diff --git a/.gitignore b/.gitignore new file mode 100644 index 00000000..8981efca --- /dev/null +++ b/.gitignore @@ -0,0 +1,12 @@ +# Files from build +**/*.o +**/*.s +src/.depend + +# Built binary +src/stockfish* +src/-lstdc++.res + +# Neural network for the NNUE evaluation +**/*.nnue + diff --git a/.travis.yml b/.travis.yml index a59ea425..092c7f53 100644 --- a/.travis.yml +++ b/.travis.yml @@ -1,6 +1,5 @@ language: cpp -sudo: required -dist: xenial +dist: bionic matrix: include: @@ -8,7 +7,6 @@ matrix: compiler: gcc addons: apt: - sources: ['ubuntu-toolchain-r-test'] packages: ['g++-8', 'g++-8-multilib', 'g++-multilib', 'valgrind', 'expect', 'curl'] env: - COMPILER=g++-8 @@ -18,23 +16,23 @@ matrix: compiler: clang addons: apt: - sources: ['ubuntu-toolchain-r-test', 'llvm-toolchain-xenial-6.0'] - packages: ['clang-6.0', 'llvm-6.0-dev', 'g++-multilib', 'valgrind', 'expect', 'curl'] + packages: ['clang-10', 'llvm-10-dev', 'g++-multilib', 'valgrind', 'expect', 'curl'] env: - - COMPILER=clang++-6.0 + - COMPILER=clang++-10 - COMP=clang - - LDFLAGS=-fuse-ld=lld - os: osx + osx_image: xcode12 compiler: gcc env: - COMPILER=g++ - COMP=gcc - os: osx + osx_image: xcode12 compiler: clang env: - - COMPILER=clang++ V='Apple LLVM 9.4.1' # Apple LLVM version 9.1.0 (clang-902.0.39.2) + - COMPILER=clang++ - COMP=clang branches: @@ -45,30 +43,59 @@ before_script: - cd src script: + # Download net + - make net + # Obtain bench reference from git log - git log HEAD | grep "\b[Bb]ench[ :]\+[0-9]\{7\}" | head -n 1 | sed "s/[^0-9]*\([0-9]*\).*/\1/g" > git_sig - export benchref=$(cat git_sig) - echo "Reference bench:" $benchref - # + + # Compiler version string + - $COMPILER -v + + # test help target + - make help + # Verify bench number against various builds - - export CXXFLAGS=-Werror - - make clean && make -j2 ARCH=x86-64 optimize=no debug=yes build && ../tests/signature.sh $benchref - - make clean && make -j2 ARCH=x86-32 optimize=no debug=yes build && ../tests/signature.sh $benchref - - make clean && make -j2 ARCH=x86-32 build && ../tests/signature.sh $benchref + - export CXXFLAGS="-Werror -D_GLIBCXX_DEBUG" + - make clean && make -j2 ARCH=x86-64-modern optimize=no debug=yes build && ../tests/signature.sh $benchref + - export CXXFLAGS="-Werror" + - make clean && make -j2 ARCH=x86-64-modern build && ../tests/signature.sh $benchref + - make clean && make -j2 ARCH=x86-64-ssse3 build && ../tests/signature.sh $benchref + - make clean && make -j2 ARCH=x86-64-sse3-popcnt build && ../tests/signature.sh $benchref + - make clean && make -j2 ARCH=x86-64 build && ../tests/signature.sh $benchref + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=general-64 build && ../tests/signature.sh $benchref; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32 optimize=no debug=yes build && ../tests/signature.sh $benchref; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32-sse41-popcnt build && ../tests/signature.sh $benchref; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32-sse2 build && ../tests/signature.sh $benchref; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-32 build && ../tests/signature.sh $benchref; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=general-32 build && ../tests/signature.sh $benchref; fi + # workaround: exclude a custom version of llvm+clang, which doesn't find llvm-profdata on ubuntu + - if [[ "$TRAVIS_OS_NAME" != "linux" || "$COMP" == "gcc" ]]; then make clean && make -j2 ARCH=x86-64-modern profile-build && ../tests/signature.sh $benchref; fi + + # compile only for some more advanced architectures (might not run in travis) + - make clean && make -j2 ARCH=x86-64-avx2 build + - make clean && make -j2 ARCH=x86-64-bmi2 build + - make clean && make -j2 ARCH=x86-64-avx512 build + - make clean && make -j2 ARCH=x86-64-vnni512 build + - make clean && make -j2 ARCH=x86-64-vnni256 build # # Check perft and reproducible search + - make clean && make -j2 ARCH=x86-64-modern build - ../tests/perft.sh - ../tests/reprosearch.sh + # # Valgrind # - export CXXFLAGS="-O1 -fno-inline" - - if [ -x "$(command -v valgrind )" ]; then make clean && make -j2 ARCH=x86-64 debug=yes optimize=no build > /dev/null && ../tests/instrumented.sh --valgrind; fi + - if [ -x "$(command -v valgrind )" ]; then make clean && make -j2 ARCH=x86-64-modern debug=yes optimize=no build > /dev/null && ../tests/instrumented.sh --valgrind; fi - if [ -x "$(command -v valgrind )" ]; then ../tests/instrumented.sh --valgrind-thread; fi + # # Sanitizer # - # Use g++-8 as a proxy for having sanitizers, might need revision as they become available for more recent versions of clang/gcc - - if [[ "$COMPILER" == "g++-8" ]]; then make clean && make -j2 ARCH=x86-64 sanitize=undefined optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-undefined; fi - - if [[ "$COMPILER" == "g++-8" ]]; then make clean && make -j2 ARCH=x86-64 sanitize=thread optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-thread; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-64-modern sanitize=undefined optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-undefined; fi + - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make clean && make -j2 ARCH=x86-64-modern sanitize=thread optimize=no debug=yes build > /dev/null && ../tests/instrumented.sh --sanitizer-thread; fi diff --git a/AUTHORS b/AUTHORS index 8317f545..662cc6ec 100644 --- a/AUTHORS +++ b/AUTHORS @@ -1,10 +1,17 @@ -# List of authors for Stockfish, updated for version 10 +# List of authors for Stockfish, as of August 4, 2020 +# Founders of the Stockfish project and fishtest infrastructure Tord Romstad (romstad) Marco Costalba (mcostalba) Joona Kiiski (zamar) Gary Linscott (glinscott) +# Authors and inventors of NNUE, training, NNUE port +Yu Nasu (ynasu87) +Motohiro Isozaki (yaneurao) +Hisayori Noda (nodchip) + +# all other authors of the code in alphabetical order Aditya (absimaldata) Adrian Petrescu (apetresc) Ajith Chandy Jose (ajithcj) @@ -12,37 +19,53 @@ Alain Savard (Rocky640) Alayan Feh (Alayan-stk-2) Alexander Kure Alexander Pagel (Lolligerhans) +Alfredo Menezes (lonfom169) Ali AlZhrani (Cooffe) Andrew Grant (AndyGrant) Andrey Neporada (nepal) Andy Duplain +Antoine Champion (antoinechampion) Aram Tumanian (atumanian) Arjun Temurnikar Auguste Pop Balint Pfliegel Ben Koshy (BKSpurgeon) Bill Henry (VoyagerOne) +Bojun Guo (noobpwnftw, Nooby) braich -Bojun Guo (noobpwnftw) -Brian Sheppard (SapphireBrand) +Brian Sheppard (SapphireBrand, briansheppard-toast) +Bruno de Melo Costa (BM123499) Bryan Cross (crossbr) +candirufish +Chess13234 Chris Cain (ceebo) -Dan Schmidt +Dale Weiler (graphitemaster) +Dan Schmidt (dfannius) +Daniel Axtens (daxtens) Daniel Dugovic (ddugovic) -Dariusz Orzechowski +Dariusz Orzechowski (dorzechowski) David Zar Daylen Yang (daylen) +Deshawn Mohan-Smith (GoldenRare) +Dieter Dobbelaere (ddobbelaere) DiscanX -Eelco de Groot +Dominik Schlösser (domschl) +double-beep +Eduardo Cáceres (eduherminio) +Eelco de Groot (KingDefender) Elvin Liu (solarlight2) erbsenzaehler Ernesto Gatti +Linmiao Xu (linrock) Fabian Beuke (madnight) Fabian Fichter (ianfab) +Fanael Linithien (Fanael) fanon Fauzi Akram Dabat (FauziAkram) Felix Wittmann gamander +Gary Heckman (gheckman) +George Sobala (gsobala) gguliash Gian-Carlo Pascutto (gcp) Gontran Lemaire (gonlem) @@ -60,41 +83,50 @@ Jacques B. (Timshel) Jan Ondruš (hxim) Jared Kish (Kurtbusch) Jarrod Torriero (DU-jdto) -Jean Gauthier (QuaisBla) +Jean Gauthier (OuaisBla) Jean-Francois Romang (jromang) +Jekaa Jerry Donald Watson (jerrydonaldwatson) +jjoshua2 Jonathan Calovski (Mysseno) -Jonathan D. (SFisGOD) +Jonathan Buladas Dumale (SFisGOD) Joost VandeVondele (vondele) Jörg Oster (joergoster) Joseph Ellis (jhellis3) Joseph R. Prostko jundery -Justin Blanchard +Justin Blanchard (UncombedCoconut) Kelly Wilson Ken Takusagawa kinderchocolate Kiran Panditrao (Krgp) Kojirion +Krystian Kuzniarek (kuzkry) Leonardo Ljubičić (ICCF World Champion) Leonid Pechenik (lp--) -Linus Arver +Linus Arver (listx) loco-loco Lub van den Berg (ElbertoOne) Luca Brivio (lucabrivio) Lucas Braesch (lucasart) Lyudmil Antonov (lantonov) Maciej Żenczykowski (zenczykowski) -Matthew Lai (matthewlai) -Matthew Sullivan +Malcolm Campbell (xoto10) Mark Tenzer (31m059) +marotear +Matt Ginsberg (mattginsberg) +Matthew Lai (matthewlai) +Matthew Sullivan (Matt14916) +Maxim Molchanov (Maxim) +Michael An (man) Michael Byrne (MichaelB7) -Michael Stembera (mstembera) Michael Chaly (Vizvezdenec) +Michael Stembera (mstembera) +Michael Whiteley (protonspring) Michel Van den Bergh (vdbergh) Miguel Lahoz (miguel-l) Mikael Bäckman (mbootsector) -Michael Whiteley (protonspring) +Mira Miroslav Fontán (Hexik) Moez Jellouli (MJZ1977) Mohammed Li (tthsqe12) @@ -102,14 +134,21 @@ Nathan Rugg (nmrugg) Nick Pelling (nickpelling) Nicklas Persson (NicklasPersson) Niklas Fiekas (niklasf) +Nikolay Kostov (NikolayIT) +Nguyen Pham (nguyenpham) +Norman Schmidt (FireFather) +notruck Ondrej Mosnáček (WOnder93) Oskar Werkelin Ahlin Pablo Vazquez +Panthee Pascal Romaret Pasquale Pigazzini (ppigazzini) Patrick Jansen (mibere) pellanda Peter Zsifkovits (CoffeeOne) +Praveen Kumar Tummala (praveentml) +Rahul Dsilva (silversolver1) Ralph Stößer (Ralph Stoesser) Raminder Singh renouve @@ -117,24 +156,39 @@ Reuven Peleg Richard Lloyd Rodrigo Exterckötter Tjäder Ron Britvich (Britvich) -Ronald de Man (syzygy1) +Ronald de Man (syzygy1, syzygy) +rqs Ryan Schmitt Ryan Takker +Sami Kiminki (skiminki) Sebastian Buchwald (UniQP) Sergei Antonov (saproj) +Sergei Ivanov (svivanov72) +Sergio Vieri (sergiovieri) sf-x -shane31 -Steinar Gunderson (sesse) +Shane Booth (shane31) +Shawn Varghese (xXH4CKST3RXx) +Siad Daboul (Topologist) Stefan Geschwentner (locutus2) Stefano Cardanobile (Stefano80) +Steinar Gunderson (sesse) Stéphane Nicolet (snicolet) Thanar2 thaspel +theo77186 +Tom Truscott Tom Vijlbrief (tomtor) -Torsten Franz (torfranz) +Tomasz Sobczyk (Sopel97) +Torsten Franz (torfranz, tfranzer) +Tracey Emery (basepr1me) +tttak +Unai Corzo (unaiic) Uri Blass (uriblass) -Vince Negri +Vince Negri (cuddlestmonkey) +zz4032 -# Additionally, we acknowledge the authors of fishtest, -# an essential framework for the development of Stockfish: + +# Additionally, we acknowledge the authors and maintainers of fishtest, +# an amazing and essential framework for the development of Stockfish! +# # https://github.com/glinscott/fishtest/blob/master/AUTHORS diff --git a/Readme.md b/README.md similarity index 52% rename from Readme.md rename to README.md index fcdd30a1..15153740 100644 --- a/Readme.md +++ b/README.md @@ -4,11 +4,17 @@ [![Build Status](https://ci.appveyor.com/api/projects/status/github/official-stockfish/Stockfish?branch=master&svg=true)](https://ci.appveyor.com/project/mcostalba/stockfish/branch/master) [Stockfish](https://stockfishchess.org) is a free, powerful UCI chess engine -derived from Glaurung 2.1. It is not a complete chess program and requires a -UCI-compatible GUI (e.g. XBoard with PolyGlot, Scid, Cute Chess, eboard, Arena, -Sigma Chess, Shredder, Chess Partner or Fritz) in order to be used comfortably. -Read the documentation for your GUI of choice for information about how to use -Stockfish with it. +derived from Glaurung 2.1. Stockfish is not a complete chess program and requires a +UCI-compatible graphical user interface (GUI) (e.g. XBoard with PolyGlot, Scid, +Cute Chess, eboard, Arena, Sigma Chess, Shredder, Chess Partner or Fritz) in order +to be used comfortably. Read the documentation for your GUI of choice for information +about how to use Stockfish with it. + +The Stockfish engine features two evaluation functions for chess, the classical +evaluation based on handcrafted terms, and the NNUE evaluation based on efficiently +updatable neural networks. The classical evaluation runs efficiently on almost all +CPU architectures, while the NNUE evaluation benefits from the vector +intrinsics available on most CPUs (sse2, avx2, neon, or similar). ## Files @@ -18,31 +24,25 @@ This distribution of Stockfish consists of the following files: * Readme.md, the file you are currently reading. * Copying.txt, a text file containing the GNU General Public License version 3. + + * AUTHORS, a text file with the list of authors for the project * src, a subdirectory containing the full source code, including a Makefile that can be used to compile Stockfish on Unix-like systems. + * a file with the .nnue extension, storing the neural network for the NNUE + evaluation. Binary distributions will have this file embedded. -## UCI parameters +## UCI options Currently, Stockfish has the following UCI options: - * #### Debug Log File - Write all communication to and from the engine into a text file. - - * #### Contempt - A positive value for contempt favors middle game positions and avoids draws. - - * #### Analysis Contempt - By default, contempt is set to prefer the side to move. Set this option to "White" - or "Black" to analyse with contempt for that side, or "Off" to disable contempt. - * #### Threads The number of CPU threads used for searching a position. For best performance, set this equal to the number of CPU cores available. * #### Hash - The size of the hash table in MB. + The size of the hash table in MB. It is recommended to set Hash after setting Threads. * #### Clear Hash Clear the hash table. @@ -54,10 +54,27 @@ Currently, Stockfish has the following UCI options: Output the N best lines (principal variations, PVs) when searching. Leave at 1 for best performance. - * #### Skill Level - Lower the Skill Level in order to make Stockfish play weaker (see also UCI_LimitStrength). - Internally, MultiPV is enabled, and with a certain probability depending on the Skill Level a - weaker move will be played. + * #### Use NNUE + Toggle between the NNUE and classical evaluation functions. If set to "true", + the network parameters must be available to load from file (see also EvalFile), + if they are not embedded in the binary. + + * #### EvalFile + The name of the file of the NNUE evaluation parameters. Depending on the GUI the + filename might have to include the full path to the folder/directory that contains the file. + Other locations, such as the directory that contains the binary and the working directory, + are also searched. + + * #### UCI_AnalyseMode + An option handled by your GUI. + + * #### UCI_Chess960 + An option handled by your GUI. If true, Stockfish will play Chess960. + + * #### UCI_ShowWDL + If enabled, show approximate WDL statistics as part of the engine output. + These WDL numbers model expected game outcomes for a given evaluation and + game ply for engine self-play at fishtest LTC conditions (60+0.6s per game). * #### UCI_LimitStrength Enable weaker play aiming for an Elo rating as set by UCI_Elo. This option overrides Skill Level. @@ -66,26 +83,10 @@ Currently, Stockfish has the following UCI options: If enabled by UCI_LimitStrength, aim for an engine strength of the given Elo. This Elo rating has been calibrated at a time control of 60s+0.6s and anchored to CCRL 40/4. - * #### Move Overhead - Assume a time delay of x ms due to network and GUI overheads. This is useful to - avoid losses on time in those cases. - - * #### Minimum Thinking Time - Search for at least x ms per move. - - * #### Slow Mover - Lower values will make Stockfish take less time in games, higher values will - make it think longer. - - * #### nodestime - Tells the engine to use nodes searched instead of wall time to account for - elapsed time. Useful for engine testing. - - * #### UCI_Chess960 - An option handled by your GUI. If true, Stockfish will play Chess960. - - * #### UCI_AnalyseMode - An option handled by your GUI. + * #### Skill Level + Lower the Skill Level in order to make Stockfish play weaker (see also UCI_LimitStrength). + Internally, MultiPV is enabled, and with a certain probability depending on the Skill Level a + weaker move will be played. * #### SyzygyPath Path to the folders/directories storing the Syzygy tablebase files. Multiple @@ -101,8 +102,8 @@ Currently, Stockfish has the following UCI options: * #### SyzygyProbeDepth Minimum remaining search depth for which a position is probed. Set this option - to a higher value to probe less agressively if you experience too much slowdown - (in terms of nps) due to TB probing. + to a higher value to probe less aggressively if you experience too much slowdown + (in terms of nps) due to tablebase probing. * #### Syzygy50MoveRule Disable to let fifty-move rule draws detected by Syzygy tablebase probes count @@ -112,13 +113,70 @@ Currently, Stockfish has the following UCI options: Limit Syzygy tablebase probing to positions with at most this many pieces left (including kings and pawns). + * #### Contempt + A positive value for contempt favors middle game positions and avoids draws, + effective for the classical evaluation only. -## What to expect from Syzygybases? + * #### Analysis Contempt + By default, contempt is set to prefer the side to move. Set this option to "White" + or "Black" to analyse with contempt for that side, or "Off" to disable contempt. + + * #### Move Overhead + Assume a time delay of x ms due to network and GUI overheads. This is useful to + avoid losses on time in those cases. + + * #### Slow Mover + Lower values will make Stockfish take less time in games, higher values will + make it think longer. + + * #### nodestime + Tells the engine to use nodes searched instead of wall time to account for + elapsed time. Useful for engine testing. + + * #### Debug Log File + Write all communication to and from the engine into a text file. + +## A note on classical evaluation versus NNUE evaluation + +Both approaches assign a value to a position that is used in alpha-beta (PVS) search +to find the best move. The classical evaluation computes this value as a function +of various chess concepts, handcrafted by experts, tested and tuned using fishtest. +The NNUE evaluation computes this value with a neural network based on basic +inputs (e.g. piece positions only). The network is optimized and trained +on the evaluations of millions of positions at moderate search depth. + +The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward. +It can be evaluated efficiently on CPUs, and exploits the fact that only parts +of the neural network need to be updated after a typical chess move. +[The nodchip repository](https://github.com/nodchip/Stockfish) provides additional +tools to train and develop the NNUE networks. On CPUs supporting modern vector instructions +(avx2 and similar), the NNUE evaluation results in much stronger playing strength, even +if the nodes per second computed by the engine is somewhat lower (roughly 80% of nps +is typical). + +Notes: + +1) the NNUE evaluation depends on the Stockfish binary and the network parameter +file (see the EvalFile UCI option). Not every parameter file is compatible with a given +Stockfish binary, but the default value of the EvalFile UCI option is the name of a network +that is guaranteed to be compatible with that binary. + +2) to use the NNUE evaluation, the additional data file with neural network parameters +needs to be available. Normally, this file is already embedded in the binary or it +can be downloaded. The filename for the default (recommended) net can be found as the default +value of the `EvalFile` UCI option, with the format `nn-[SHA256 first 12 digits].nnue` +(for instance, `nn-c157e0a5755b.nnue`). This file can be downloaded from +``` +https://tests.stockfishchess.org/api/nn/[filename] +``` +replacing `[filename]` as needed. + +## What to expect from the Syzygy tablebases? If the engine is searching a position that is not in the tablebases (e.g. a position with 8 pieces), it will access the tablebases during the search. -If the engine reports a very large score (typically 153.xx), this means -that it has found a winning line into a tablebase position. +If the engine reports a very large score (typically 153.xx), this means +it has found a winning line into a tablebase position. If the engine is given a position to search that is in the tablebases, it will use the tablebases at the beginning of the search to preselect all @@ -126,14 +184,14 @@ good moves, i.e. all moves that preserve the win or preserve the draw while taking into account the 50-move rule. It will then perform a search only on those moves. **The engine will not move immediately**, unless there is only a single good move. **The engine likely -will not report a mate score even if the position is known to be won.** +will not report a mate score, even if the position is known to be won.** It is therefore clear that this behaviour is not identical to what one might be used to with Nalimov tablebases. There are technical reasons for this difference, the main technical reason being that Nalimov tablebases use the -DTM metric (distance-to-mate), while Syzygybases use a variation of the +DTM metric (distance-to-mate), while the Syzygy tablebases use a variation of the DTZ metric (distance-to-zero, zero meaning any move that resets the 50-move -counter). This special metric is one of the reasons that Syzygybases are +counter). This special metric is one of the reasons that the Syzygy tablebases are more compact than Nalimov tablebases, while still storing all information needed for optimal play and in addition being able to take into account the 50-move rule. @@ -167,20 +225,59 @@ setting additional environment variables. ```mpirun``` should forward stdin/stdo to ```rank 0``` only (e.g. ```srun --input=0 --output=0```). Refer to your MPI documentation for more info. +## Large Pages + +Stockfish supports large pages on Linux and Windows. Large pages make +the hash access more efficient, improving the engine speed, especially +on large hash sizes. Typical increases are 5..10% in terms of nodes per +second, but speed increases up to 30% have been measured. The support is +automatic. Stockfish attempts to use large pages when available and +will fall back to regular memory allocation when this is not the case. + +### Support on Linux + +Large page support on Linux is obtained by the Linux kernel +transparent huge pages functionality. Typically, transparent huge pages +are already enabled, and no configuration is needed. + +### Support on Windows + +The use of large pages requires "Lock Pages in Memory" privilege. See +[Enable the Lock Pages in Memory Option (Windows)](https://docs.microsoft.com/en-us/sql/database-engine/configure-windows/enable-the-lock-pages-in-memory-option-windows) +on how to enable this privilege, then run [RAMMap](https://docs.microsoft.com/en-us/sysinternals/downloads/rammap) +to double-check that large pages are used. We suggest that you reboot +your computer after you have enabled large pages, because long Windows +sessions suffer from memory fragmentation, which may prevent Stockfish +from getting large pages: a fresh session is better in this regard. + ## Compiling Stockfish yourself from the sources -On Unix-like systems, it should be possible to compile Stockfish -directly from the source code with the included Makefile. +Stockfish has support for 32 or 64-bit CPUs, certain hardware +instructions, big-endian machines such as Power PC, and other platforms. -Stockfish has support for 32 or 64-bit CPUs, the hardware POPCNT -instruction, big-endian machines such as Power PC, and other platforms. +On Unix-like systems, it should be easy to compile Stockfish +directly from the source code with the included Makefile in the folder +`src`. In general it is recommended to run `make help` to see a list of make +targets with corresponding descriptions. -In general it is recommended to run `make help` to see a list of make -targets with corresponding descriptions. When not using the Makefile to -compile (for instance with Microsoft MSVC) you need to manually -set/unset some switches in the compiler command line; see file *types.h* -for a quick reference. +``` + cd src + make help + make net + make build ARCH=x86-64-modern +``` +When not using the Makefile to compile (for instance, with Microsoft MSVC) you +need to manually set/unset some switches in the compiler command line; see +file *types.h* for a quick reference. + +When reporting an issue or a bug, please tell us which version and +compiler you used to create your executable. These informations can +be found by typing the following commands in a console: + +``` + ./stockfish compiler +``` ## Understanding the code base and participating in the project @@ -190,12 +287,12 @@ community effort. There are a few ways to help contribute to its growth. ### Donating hardware Improving Stockfish requires a massive amount of testing. You can donate -your hardware resources by installing the [Fishtest Worker](https://github.com/glinscott/fishtest/wiki/Running-the-worker) -and view the current tests on [Fishtest](http://tests.stockfishchess.org/tests). +your hardware resources by installing the [Fishtest Worker](https://github.com/glinscott/fishtest/wiki/Running-the-worker:-overview) +and view the current tests on [Fishtest](https://tests.stockfishchess.org/tests). ### Improving the code -If you want to help improve the code, there are several valuable ressources: +If you want to help improve the code, there are several valuable resources: * [In this wiki,](https://www.chessprogramming.org) many techniques used in Stockfish are explained with a lot of background information. @@ -206,8 +303,9 @@ generic rather than being focused on Stockfish's precise implementation. Nevertheless, a helpful resource. * The latest source can always be found on [GitHub](https://github.com/official-stockfish/Stockfish). -Discussions about Stockfish take place in the [FishCooking](https://groups.google.com/forum/#!forum/fishcooking) -group and engine testing is done on [Fishtest](http://tests.stockfishchess.org/tests). +Discussions about Stockfish take place these days mainly in the [FishCooking](https://groups.google.com/forum/#!forum/fishcooking) +group and on the [Stockfish Discord channel](https://discord.gg/nv8gDtt). +The engine testing is done on [Fishtest](https://tests.stockfishchess.org/tests). If you want to help improve Stockfish, please read this [guideline](https://github.com/glinscott/fishtest/wiki/Creating-my-first-test) first, where the basics of Stockfish development are explained. @@ -215,16 +313,17 @@ first, where the basics of Stockfish development are explained. ## Terms of use Stockfish is free, and distributed under the **GNU General Public License version 3** -(GPL v3). Essentially, this means that you are free to do almost exactly +(GPL v3). Essentially, this means you are free to do almost exactly what you want with the program, including distributing it among your -friends, making it available for download from your web site, selling +friends, making it available for download from your website, selling it (either by itself or as part of some bigger software package), or using it as the starting point for a software project of your own. The only real limitation is that whenever you distribute Stockfish in -some way, you must always include the full source code, or a pointer -to where the source code can be found. If you make any changes to the -source code, these changes must also be made available under the GPL. +some way, you MUST always include the full source code, or a pointer +to where the source code can be found, to generate the exact binary +you are distributing. If you make any changes to the source code, +these changes must also be made available under the GPL. For full details, read the copy of the GPL v3 found in the file named *Copying.txt*. diff --git a/Top CPU Contributors.txt b/Top CPU Contributors.txt index e882aa43..f5347ea1 100644 --- a/Top CPU Contributors.txt +++ b/Top CPU Contributors.txt @@ -1,146 +1,189 @@ -Contributors with >10,000 CPU hours as of November 4, 2018 +Contributors to Fishtest with >10,000 CPU hours, as of Feb 15, 2021. Thank you! -Username CPU Hours Games played -noobpwnftw 3730975 292309380 -mibere 535242 43333774 -crunchy 375564 29121434 -cw 371664 28748719 -fastgm 318178 22283584 -JojoM 295354 20958931 -dew 215476 17079219 -ctoks 214031 17312035 -glinscott 204517 13932027 -bking_US 187568 12233168 -velislav 168404 13336219 -CSU_Dynasty 168069 14417712 -Thanar 162373 13842179 -spams 149531 10940322 -Fisherman 141137 12099359 -drabel 134441 11180178 -leszek 133658 9812120 -marrco 133566 10115202 -sqrt2 128420 10022279 -vdbergh 123230 9200516 -tvijlbrief 123007 9498831 -vdv 120381 8555423 -malala 117291 8126488 -dsmith 114010 7622414 -BrunoBanani 104938 7448565 -CoffeeOne 100042 4593596 -Data 94621 8433010 -mgrabiak 92248 7787406 -bcross 89440 8506568 -brabos 81868 6647613 -BRAVONE 80811 5341681 -psk 77195 6156031 -nordlandia 74833 6231930 -robal 72818 5969856 -TueRens 72523 6383294 -sterni1971 71049 5647590 -sunu 65855 5360884 -mhoram 65034 5192880 -davar 64794 5457564 -nssy 64607 5371952 -Pking_cda 64499 5704075 -biffhero 63557 5480444 -teddybaer 62147 5585620 -solarlight 61278 5402642 -ElbertoOne 60156 5504304 -jromang 58854 4704502 -dv8silencer 57421 3961325 -tinker 56039 4204914 -Freja 50331 3808121 -renouve 50318 3544864 -robnjr 47504 4131742 -grandphish2 47377 4110003 -eva42 46857 4075716 -ttruscott 46802 3811534 -finfish 46244 3481661 -rap 46201 3219490 -ronaldjerum 45641 3964331 -xoto 44998 4170431 -gvreuls 44359 3902234 -bigpen0r 41780 3448224 -Bobo1239 40767 3657490 -Antihistamine 39218 2792761 -mhunt 38991 2697512 -racerschmacer 38929 3756111 -VoyagerOne 35896 3378887 -homyur 35561 3012398 -rkl 33217 2978536 -pb00067 33034 2803485 -speedycpu 32043 2531964 -SC 31954 2848432 -EthanOConnor 31638 2143255 -oryx 30962 2899534 -gri 30108 2429137 -csnodgrass 29396 2808611 -Garf 28887 2873564 -Pyafue 28885 1986098 -jkiiski 28014 1923255 -slakovv 27017 2031279 -Prcuvu 26300 2307154 -hyperbolic.tom 26248 2200777 -jbwiebe 25663 2129063 -anst 25525 2279159 -Patrick_G 24222 1835674 -nabildanial 23524 1586321 -achambord 23495 1942546 -Sharaf_DG 22975 1790697 -chriswk 22876 1947731 -ncfish1 22689 1830009 -cuistot 22201 1383031 -Zirie 21171 1493227 -Isidor 20634 1736219 -JanErik 20596 1791991 -xor12 20535 1819280 -team-oh 20364 1653708 -nesoneg 20264 1493435 -dex 20110 1682756 -rstoesser 19802 1335177 -Vizvezdenec 19750 1695579 -eastorwest 19531 1841839 -sg4032 18913 1720157 -horst.prack 18425 1708197 -cisco2015 18408 1793774 -ianh2105 18133 1668562 -MazeOfGalious 18022 1644593 -ville 17900 1539130 -j3corre 17607 975954 -eudhan 17502 1424648 -jmdana 17351 1287546 -iisiraider 17175 1118788 -jundery 17172 1115855 -wei 16852 1822582 -SFTUser 16635 1363975 -purplefishies 16621 1106850 -DragonLord 16599 1252348 -chris 15274 1575333 -IgorLeMasson 15201 1364148 -dju 15074 914278 -Flopzee 14700 1331632 -OssumOpossum 14149 1029265 -enedene 13762 935618 -ako027ako 13442 1250249 -AdrianSA 13324 924980 -bpfliegel 13318 886523 -Nikolay.IT 13260 1155612 -jpulman 12776 854815 -joster 12438 988413 -fatmurphy 12015 901134 -Nesa92 11711 1132245 -Adrian.Schmidt123 11542 898699 -modolief 11228 926456 -Dark_wizzie 11214 1017910 -mschmidt 10973 818594 -Andrew Grant 10780 947859 -infinity 10762 746397 -SapphireBrand 10692 1024604 -Thomas A. Anderson 10553 736094 -basepi 10434 935168 -lantonov 10325 972610 -pgontarz 10294 878746 -Spprtr 10189 823246 -crocogoat 10115 1017325 -stocky 10083 718114 \ No newline at end of file +Username CPU Hours Games played +---------------------------------------------------- +noobpwnftw 23930906 1560559941 +dew 1169948 70333008 +mlang 957168 61657446 +mibere 703840 46867607 +tvijlbrief 517888 33379462 +JojoM 515404 30334272 +cw 443276 29385549 +crunchy 427035 27344275 +grandphish2 425794 26347253 +fastgm 414133 24519696 +gvreuls 377843 24708884 +CSU_Dynasty 338718 23030006 +Fisherman 326795 21820747 +TueRens 313730 19490246 +ctoks 298442 20052551 +velislav 270519 17355456 +bcross 241064 17196165 +glinscott 217799 13780820 +nordlandia 211692 13484886 +bking_US 198894 11876016 +drabel 191096 13129722 +leszek 189170 11446821 +mgrabiak 187153 12013300 +robal 181389 11539242 +Thanar 179852 12365359 +vdv 175274 9889046 +spams 157128 10319326 +marrco 150292 9401741 +sqrt2 147963 9724586 +CoffeeOne 137086 5022516 +vdbergh 137041 8926915 +malala 136182 8002293 +mhoram 132780 8398229 +xoto 124729 8652088 +davar 122092 7960001 +dsmith 122059 7570238 +Data 113305 8220352 +BrunoBanani 112960 7436849 +pemo 109598 5036441 +Dantist 106768 6431396 +MaZePallas 102741 6630419 +ElbertoOne 99028 7023771 +brabos 92118 6186135 +linrock 90903 6708639 +psk 89957 5984901 +sunu 88614 6020673 +sterni1971 86948 5613788 +Vizvezdenec 83761 5344740 +BRAVONE 81239 5054681 +nssy 76497 5259388 +cuistot 76366 4370584 +racerschmacer 75753 5442626 +teddybaer 75125 5407666 +Pking_cda 73776 5293873 +0x3C33 73133 4670293 +jromang 72117 5054915 +solarlight 70517 5028306 +dv8silencer 70287 3883992 +Bobo1239 68515 4652287 +manap 66273 4121774 +tinker 64321 4268390 +robnjr 57262 4053117 +Freja 56938 3733019 +ttruscott 56010 3680085 +rkl 54986 4150767 +renouve 53811 3501516 +finfish 51360 3370515 +eva42 51272 3599691 +rap 49985 3219146 +pb00067 49727 3298270 +amicic 49691 3042481 +ronaldjerum 47654 3240695 +bigpen0r 47278 3291647 +biffhero 46564 3111352 +VoyagerOne 45476 3452465 +eastorwest 45033 3071805 +speedycpu 43842 3003273 +jbwiebe 43305 2805433 +Antihistamine 41788 2761312 +mhunt 41735 2691355 +homyur 39893 2850481 +gri 39871 2515779 +oryx 38282 2944400 +Spprtr 38157 2470529 +SC 37290 2731014 +csnodgrass 36207 2688994 +jmdana 36157 2210661 +strelock 34716 2074055 +Garf 33800 2747562 +skiminki 33515 2055584 +EthanOConnor 33370 2090311 +slakovv 32915 2021889 +yurikvelo 32600 2255966 +Prcuvu 30377 2170122 +manapbk 30326 1770143 +anst 30301 2190091 +jkiiski 30136 1904470 +hyperbolic.tom 29840 2017394 +Pyafue 29650 1902349 +qurashee 27758 1509620 +OuaisBla 27636 1578800 +chriswk 26902 1868317 +achambord 26582 1767323 +Fifis 26376 1776853 +Patrick_G 26276 1801617 +yorkman 26193 1992080 +SFTUser 25182 1675689 +nabildanial 24942 1519409 +Sharaf_DG 24765 1786697 +ncfish1 24411 1520927 +agg177 23890 1395014 +JanErik 23408 1703875 +Isidor 23388 1680691 +Norabor 23164 1591830 +cisco2015 22895 1762069 +Zirie 22542 1472937 +team-oh 22272 1636708 +MazeOfGalious 21978 1629593 +sg4032 21945 1643065 +ianh2105 21725 1632562 +xor12 21628 1680365 +dex 21612 1467203 +nesoneg 21494 1463031 +jjoshua2 20997 1422689 +horst.prack 20878 1465656 +0xB00B1ES 20590 1208666 +sphinx 20515 1352368 +j3corre 20405 941444 +Adrian.Schmidt123 20316 1281436 +Ente 20017 1432602 +wei 19973 1745989 +rstoesser 19569 1293588 +eudhan 19274 1283717 +jundery 18445 1115855 +iisiraider 18247 1101015 +ville 17883 1384026 +chris 17698 1487385 +purplefishies 17595 1092533 +DMBK 17357 1279152 +DragonLord 17014 1162790 +dju 16515 929427 +IgorLeMasson 16064 1147232 +ako027ako 15671 1173203 +Nikolay.IT 15154 1068349 +Andrew Grant 15114 895539 +OssumOpossum 14857 1007129 +enedene 14476 905279 +bpfliegel 14298 884523 +jpulman 13982 870599 +joster 13794 950160 +Nesa92 13786 1114691 +crocogoat 13753 1114622 +Hjax 13535 915487 +Dark_wizzie 13422 1007152 +mpx86 12941 693640 +mabichito 12903 749391 +thijsk 12886 722107 +AdrianSA 12860 804972 +Flopzee 12698 894821 +fatmurphy 12547 853210 +scuzzi 12511 845761 +Karby 12429 735880 +SapphireBrand 12416 969604 +modolief 12386 896470 +pgontarz 12151 848794 +stocky 11954 699440 +mschmidt 11941 803401 +infinity 11470 727027 +torbjo 11395 729145 +Thomas A. Anderson 11372 732094 +d64 11263 789184 +Maxim 11129 804704 +snicolet 11106 869170 +MooTheCow 11008 694942 +savage84 10965 641068 +Rudolphous 10915 741268 +Wolfgang 10809 580032 +rpngn 10712 688203 +basepi 10637 744851 +michaelrpg 10409 735127 +dzjp 10343 732529 +ali-al-zhrani 10324 726502 +ols 10259 570669 +lbraesch 10252 647825 diff --git a/appveyor.yml b/appveyor.yml index 21f3bbe3..ab608409 100644 --- a/appveyor.yml +++ b/appveyor.yml @@ -4,10 +4,9 @@ clone_depth: 50 branches: only: - master - - appveyor # Operating system (build VM template) -os: Visual Studio 2017 +os: Visual Studio 2019 # Build platform, i.e. x86, x64, AnyCPU. This setting is optional. platform: @@ -36,8 +35,11 @@ before_build: $src = $src.Replace("\", "/") # Build CMakeLists.txt - $t = 'cmake_minimum_required(VERSION 3.8)', + $t = 'cmake_minimum_required(VERSION 3.17)', 'project(Stockfish)', + 'set(CMAKE_CXX_STANDARD 17)', + 'set(CMAKE_CXX_STANDARD_REQUIRED ON)', + 'set (CMAKE_CXX_EXTENSIONS OFF)', 'set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_SOURCE_DIR}/src)', 'set(source_files', $src, ')', 'add_executable(stockfish ${source_files})' @@ -51,13 +53,28 @@ before_build: $b = git log HEAD | sls "\b[Bb]ench[ :]+[0-9]{7}" | select -first 1 $bench = $b -match '\D+(\d+)' | % { $matches[1] } Write-Host "Reference bench:" $bench - $g = "Visual Studio 15 2017" - If (${env:PLATFORM} -eq 'x64') { $g = $g + ' Win64' } - cmake -G "${g}" . - Write-Host "Generated files for: " $g + $g = "Visual Studio 16 2019" + If (${env:PLATFORM} -eq 'x64') { $a = "x64" } + If (${env:PLATFORM} -eq 'x86') { $a = "Win32" } + cmake -G "${g}" -A ${a} . + Write-Host "Generated files for: " $g $a build_script: - cmake --build . --config %CONFIGURATION% -- /verbosity:minimal + - ps: | + # Download default NNUE net from fishtest + $nnuenet = Get-Content -Path src\evaluate.h | Select-String -CaseSensitive -Pattern "EvalFileDefaultName" | Select-String -CaseSensitive -Pattern "nn-[a-z0-9]{12}.nnue" + $dummy = $nnuenet -match "(?nn-[a-z0-9]{12}.nnue)" + $nnuenet = $Matches.nnuenet + Write-Host "Default net:" $nnuenet + $nnuedownloadurl = "https://tests.stockfishchess.org/api/nn/$nnuenet" + $nnuefilepath = "src\${env:CONFIGURATION}\$nnuenet" + if (Test-Path -Path $nnuefilepath) { + Write-Host "Already available." + } else { + Write-Host "Downloading $nnuedownloadurl to $nnuefilepath" + Invoke-WebRequest -Uri $nnuedownloadurl -OutFile $nnuefilepath + } before_test: - cd src/%CONFIGURATION% diff --git a/src/Makefile b/src/Makefile index bd51e4ea..a3613fbe 100644 --- a/src/Makefile +++ b/src/Makefile @@ -1,7 +1,5 @@ # Stockfish, a UCI chess playing engine derived from Glaurung 2.1 -# Copyright (C) 2004-2008 Tord Romstad (Glaurung author) -# Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad -# Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad +# Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) # # Stockfish is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by @@ -35,10 +33,15 @@ BINDIR = $(PREFIX)/bin ### Built-in benchmark for pgo-builds PGOBENCH = ./$(EXE) bench -### Object files -OBJS = benchmark.o bitbase.o bitboard.o cluster.o endgame.o evaluate.o main.o \ - material.o misc.o movegen.o movepick.o pawns.o position.o psqt.o \ - search.o thread.o timeman.o tt.o uci.o ucioption.o syzygy/tbprobe.o +### Source and object files +SRCS = benchmark.cpp bitbase.cpp bitboard.cpp cluster.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_kp.cpp + +OBJS = $(notdir $(SRCS:.cpp=.o)) + +VPATH = syzygy:nnue:nnue/features ### Establish the operating system name KERNEL = $(shell uname -s) @@ -50,7 +53,7 @@ endif ### Section 2. High-level Configuration ### ========================================================================== # -# flag --- Comp switch --- Description +# flag --- Comp switch --- Description # ---------------------------------------------------------------------------- # # debug = yes/no --- -DNDEBUG --- Enable/Disable debug mode @@ -62,8 +65,17 @@ endif # bits = 64/32 --- -DIS_64BIT --- 64-/32-bit operating system # prefetch = yes/no --- -DUSE_PREFETCH --- Use prefetch asm-instruction # popcnt = yes/no --- -DUSE_POPCNT --- Use popcnt asm-instruction -# sse = yes/no --- -msse --- Use Intel Streaming SIMD Extensions # pext = yes/no --- -DUSE_PEXT --- Use pext x86_64 asm-instruction +# sse = yes/no --- -msse --- Use Intel Streaming SIMD Extensions +# mmx = yes/no --- -mmmx --- Use Intel MMX instructions +# sse2 = yes/no --- -msse2 --- Use Intel Streaming SIMD Extensions 2 +# ssse3 = yes/no --- -mssse3 --- Use Intel Supplemental Streaming SIMD Extensions 3 +# sse41 = yes/no --- -msse4.1 --- Use Intel Streaming SIMD Extensions 4.1 +# avx2 = yes/no --- -mavx2 --- Use Intel Advanced Vector Extensions 2 +# avx512 = yes/no --- -mavx512bw --- Use Intel Advanced Vector Extensions 512 +# vnni256 = yes/no --- -mavx512vnni --- Use Intel Vector Neural Network Instructions 256 +# vnni512 = yes/no --- -mavx512vnni --- Use Intel Vector Neural Network Instructions 512 +# neon = yes/no --- -DUSE_NEON --- Use ARM SIMD architecture # mpi = yes/no --- -DUSE_MPI --- Use Message Passing Interface # # Note that Makefile is space sensitive, so when adding new architectures @@ -71,86 +83,221 @@ endif # at the end of the line for flag values. ### 2.1. General and architecture defaults + +ifeq ($(ARCH),) + ARCH = x86-64-modern + help_skip_sanity = yes +endif +# explicitly check for the list of supported architectures (as listed with make help), +# the user can override with `make ARCH=x86-32-vnni256 SUPPORTED_ARCH=true` +ifeq ($(ARCH), $(filter $(ARCH), \ + x86-64-vnni512 x86-64-vnni256 x86-64-avx512 x86-64-bmi2 x86-64-avx2 \ + x86-64-sse41-popcnt x86-64-modern x86-64-ssse3 x86-64-sse3-popcnt \ + x86-64 x86-32-sse41-popcnt x86-32-sse2 x86-32 ppc-64 ppc-32 \ + armv7 armv7-neon armv8 apple-silicon general-64 general-32)) + SUPPORTED_ARCH=true +else + SUPPORTED_ARCH=false +endif + optimize = yes debug = no sanitize = no -bits = 32 +bits = 64 prefetch = no popcnt = no -sse = no pext = no +sse = no +mmx = no +sse2 = no +ssse3 = no +sse41 = no +avx2 = no +avx512 = no +vnni256 = no +vnni512 = no +neon = no mpi = no +STRIP = strip ### 2.2 Architecture specific +ifeq ($(findstring x86,$(ARCH)),x86) + +# x86-32/64 + +ifeq ($(findstring x86-32,$(ARCH)),x86-32) + arch = i386 + bits = 32 + sse = yes + mmx = yes +else + arch = x86_64 + sse = yes + sse2 = yes +endif + +ifeq ($(findstring -sse,$(ARCH)),-sse) + sse = yes +endif + +ifeq ($(findstring -popcnt,$(ARCH)),-popcnt) + popcnt = yes +endif + +ifeq ($(findstring -mmx,$(ARCH)),-mmx) + mmx = yes +endif + +ifeq ($(findstring -sse2,$(ARCH)),-sse2) + sse = yes + sse2 = yes +endif + +ifeq ($(findstring -ssse3,$(ARCH)),-ssse3) + sse = yes + sse2 = yes + ssse3 = yes +endif + +ifeq ($(findstring -sse41,$(ARCH)),-sse41) + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes +endif + +ifeq ($(findstring -modern,$(ARCH)),-modern) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes +endif + +ifeq ($(findstring -avx2,$(ARCH)),-avx2) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes + avx2 = yes +endif + +ifeq ($(findstring -bmi2,$(ARCH)),-bmi2) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes + avx2 = yes + pext = yes +endif + +ifeq ($(findstring -avx512,$(ARCH)),-avx512) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes + avx2 = yes + pext = yes + avx512 = yes +endif + +ifeq ($(findstring -vnni256,$(ARCH)),-vnni256) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes + avx2 = yes + pext = yes + vnni256 = yes +endif + +ifeq ($(findstring -vnni512,$(ARCH)),-vnni512) + popcnt = yes + sse = yes + sse2 = yes + ssse3 = yes + sse41 = yes + avx2 = yes + pext = yes + avx512 = yes + vnni512 = yes +endif + +ifeq ($(sse),yes) + prefetch = yes +endif + +# 64-bit pext is not available on x86-32 +ifeq ($(bits),32) + pext = no +endif + +else + +# all other architectures + ifeq ($(ARCH),general-32) arch = any -endif - -ifeq ($(ARCH),x86-32-old) - arch = i386 -endif - -ifeq ($(ARCH),x86-32) - arch = i386 - prefetch = yes - sse = yes + bits = 32 endif ifeq ($(ARCH),general-64) arch = any - bits = 64 -endif - -ifeq ($(ARCH),x86-64) - arch = x86_64 - bits = 64 - prefetch = yes - sse = yes -endif - -ifeq ($(ARCH),x86-64-modern) - arch = x86_64 - bits = 64 - prefetch = yes - popcnt = yes - sse = yes -endif - -ifeq ($(ARCH),x86-64-bmi2) - arch = x86_64 - bits = 64 - prefetch = yes - popcnt = yes - sse = yes - pext = yes endif ifeq ($(ARCH),armv7) arch = armv7 prefetch = yes + bits = 32 +endif + +ifeq ($(ARCH),armv7-neon) + arch = armv7 + prefetch = yes + popcnt = yes + neon = yes + bits = 32 +endif + +ifeq ($(ARCH),armv8) + arch = armv8 + prefetch = yes + popcnt = yes + neon = yes +endif + +ifeq ($(ARCH),apple-silicon) + arch = arm64 + prefetch = yes + popcnt = yes + neon = yes endif ifeq ($(ARCH),ppc-32) arch = ppc + bits = 32 endif ifeq ($(ARCH),ppc-64) arch = ppc64 - bits = 64 popcnt = yes prefetch = yes endif +endif ### ========================================================================== -### Section 3. Low-level configuration +### Section 3. Low-level Configuration ### ========================================================================== ### 3.1 Selecting compiler (default = gcc) - -CXXFLAGS += -Wall -Wcast-qual -fno-exceptions -std=c++11 $(EXTRACXXFLAGS) -DEPENDFLAGS += -std=c++11 +CXXFLAGS += -Wall -Wcast-qual -fno-exceptions -std=c++17 $(EXTRACXXFLAGS) +DEPENDFLAGS += -std=c++17 LDFLAGS += $(EXTRALDFLAGS) ifeq ($(COMP),) @@ -162,7 +309,7 @@ ifeq ($(COMP),gcc) CXX=g++ CXXFLAGS += -pedantic -Wextra -Wshadow - ifeq ($(ARCH),armv7) + ifeq ($(arch),$(filter $(arch),armv7 armv8)) ifeq ($(OS),Android) CXXFLAGS += -m$(bits) LDFLAGS += -m$(bits) @@ -172,6 +319,10 @@ ifeq ($(COMP),gcc) LDFLAGS += -m$(bits) endif + ifeq ($(arch),$(filter $(arch),armv7)) + LDFLAGS += -latomic + endif + ifneq ($(KERNEL),Darwin) LDFLAGS += -Wl,--no-as-needed endif @@ -215,11 +366,13 @@ ifeq ($(COMP),clang) ifneq ($(KERNEL),Darwin) ifneq ($(KERNEL),OpenBSD) + ifneq ($(KERNEL),FreeBSD) LDFLAGS += -latomic endif endif + endif - ifeq ($(ARCH),armv7) + ifeq ($(arch),$(filter $(arch),armv7 armv8)) ifeq ($(OS),Android) CXXFLAGS += -m$(bits) LDFLAGS += -m$(bits) @@ -230,23 +383,40 @@ ifeq ($(COMP),clang) endif endif +ifeq ($(KERNEL),Darwin) + CXXFLAGS += -arch $(arch) -mmacosx-version-min=10.14 + LDFLAGS += -arch $(arch) -mmacosx-version-min=10.14 + XCRUN = xcrun +endif + +# To cross-compile for Android, NDK version r21 or later is recommended. +# In earlier NDK versions, you'll need to pass -fno-addrsig if using GNU binutils. +# Currently we don't know how to make PGO builds with the NDK yet. +ifeq ($(COMP),ndk) + CXXFLAGS += -stdlib=libc++ -fPIE + comp=clang + ifeq ($(arch),armv7) + CXX=armv7a-linux-androideabi16-clang++ + CXXFLAGS += -mthumb -march=armv7-a -mfloat-abi=softfp -mfpu=neon + STRIP=arm-linux-androideabi-strip + endif + ifeq ($(arch),armv8) + CXX=aarch64-linux-android21-clang++ + STRIP=aarch64-linux-android-strip + endif + LDFLAGS += -static-libstdc++ -pie -lm -latomic +endif + ifeq ($(comp),icc) profile_make = icc-profile-make profile_use = icc-profile-use -else -ifeq ($(comp),clang) +else ifeq ($(comp),clang) profile_make = clang-profile-make profile_use = clang-profile-use else profile_make = gcc-profile-make profile_use = gcc-profile-use endif -endif - -ifeq ($(KERNEL),Darwin) - CXXFLAGS += -arch $(arch) -mmacosx-version-min=10.9 - LDFLAGS += -arch $(arch) -mmacosx-version-min=10.9 -endif ### Travis CI script uses COMPILER to overwrite CXX ifdef COMPILER @@ -258,13 +428,26 @@ ifdef COMPCXX CXX=$(COMPCXX) endif +### Sometimes gcc is really clang +ifeq ($(COMP),gcc) + gccversion = $(shell $(CXX) --version) + gccisclang = $(findstring clang,$(gccversion)) + ifneq ($(gccisclang),) + profile_make = clang-profile-make + profile_use = clang-profile-use + endif +endif + ### On mingw use Windows threads, otherwise POSIX ifneq ($(comp),mingw) + CXXFLAGS += -DUSE_PTHREADS # On Android Bionic's C library comes with its own pthread implementation bundled in ifneq ($(OS),Android) # Haiku has pthreads in its libroot, so only link it in on other platforms ifneq ($(KERNEL),Haiku) - LDFLAGS += -lpthread + ifneq ($(COMP),ndk) + LDFLAGS += -lpthread + endif endif endif endif @@ -278,8 +461,8 @@ endif ### 3.2.2 Debugging with undefined behavior sanitizers ifneq ($(sanitize),no) - CXXFLAGS += -g3 -fsanitize=$(sanitize) -fuse-ld=gold - LDFLAGS += -fsanitize=$(sanitize) -fuse-ld=gold + CXXFLAGS += -g3 -fsanitize=$(sanitize) + LDFLAGS += -fsanitize=$(sanitize) endif ### 3.3 Optimization @@ -292,12 +475,16 @@ ifeq ($(optimize),yes) CXXFLAGS += -fno-gcse -mthumb -march=armv7-a -mfloat-abi=softfp endif endif - + ifeq ($(comp),$(filter $(comp),gcc clang icc)) ifeq ($(KERNEL),Darwin) CXXFLAGS += -mdynamic-no-pic endif endif + + ifeq ($(comp),clang) + CXXFLAGS += -fexperimental-new-pass-manager + endif endif ### 3.4 Bits @@ -309,7 +496,6 @@ endif ifeq ($(prefetch),yes) ifeq ($(sse),yes) CXXFLAGS += -msse - DEPENDFLAGS += -msse endif else CXXFLAGS += -DNO_PREFETCH @@ -317,7 +503,7 @@ endif ### 3.6 popcnt ifeq ($(popcnt),yes) - ifeq ($(arch),ppc64) + ifeq ($(arch),$(filter $(arch),ppc64 armv7 armv8 arm64)) CXXFLAGS += -DUSE_POPCNT else ifeq ($(comp),icc) CXXFLAGS += -msse3 -DUSE_POPCNT @@ -326,28 +512,121 @@ ifeq ($(popcnt),yes) endif endif + +ifeq ($(avx2),yes) + CXXFLAGS += -DUSE_AVX2 + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mavx2 + endif +endif + +ifeq ($(avx512),yes) + CXXFLAGS += -DUSE_AVX512 + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mavx512f -mavx512bw + endif +endif + +ifeq ($(vnni256),yes) + CXXFLAGS += -DUSE_VNNI + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mavx512f -mavx512bw -mavx512vnni -mavx512dq -mavx512vl -mprefer-vector-width=256 + endif +endif + +ifeq ($(vnni512),yes) + CXXFLAGS += -DUSE_VNNI + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mavx512vnni -mavx512dq -mavx512vl + endif +endif + +ifeq ($(sse41),yes) + CXXFLAGS += -DUSE_SSE41 + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -msse4.1 + endif +endif + +ifeq ($(ssse3),yes) + CXXFLAGS += -DUSE_SSSE3 + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mssse3 + endif +endif + +ifeq ($(sse2),yes) + CXXFLAGS += -DUSE_SSE2 + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -msse2 + endif +endif + +ifeq ($(mmx),yes) + CXXFLAGS += -DUSE_MMX + ifeq ($(comp),$(filter $(comp),gcc clang mingw)) + CXXFLAGS += -mmmx + endif +endif + +ifeq ($(neon),yes) + CXXFLAGS += -DUSE_NEON + ifeq ($(KERNEL),Linux) + ifneq ($(COMP),ndk) + ifneq ($(arch),armv8) + CXXFLAGS += -mfpu=neon + endif + endif + endif +endif + ### 3.7 pext ifeq ($(pext),yes) CXXFLAGS += -DUSE_PEXT ifeq ($(comp),$(filter $(comp),gcc clang mingw)) - CXXFLAGS += -msse4 -mbmi2 + CXXFLAGS += -mbmi2 endif endif -### 3.8 Link Time Optimization, it works since gcc 4.5 but not on mingw under Windows. +### 3.8 Link Time Optimization ### This is a mix of compile and link time options because the lto link phase ### needs access to the optimization flags. ifeq ($(optimize),yes) ifeq ($(debug), no) - ifeq ($(comp),$(filter $(comp),gcc clang)) + ifeq ($(comp),clang) + CXXFLAGS += -flto + ifneq ($(findstring MINGW,$(KERNEL)),) + CXXFLAGS += -fuse-ld=lld + else ifneq ($(findstring MSYS,$(KERNEL)),) + CXXFLAGS += -fuse-ld=lld + endif + LDFLAGS += $(CXXFLAGS) + +# GCC and CLANG use different methods for parallelizing LTO and CLANG pretends to be +# GCC on some systems. + else ifeq ($(comp),gcc) + ifeq ($(gccisclang),) + CXXFLAGS += -flto + LDFLAGS += $(CXXFLAGS) -flto=jobserver + ifneq ($(findstring MINGW,$(KERNEL)),) + LDFLAGS += -save-temps + else ifneq ($(findstring MSYS,$(KERNEL)),) + LDFLAGS += -save-temps + endif + else CXXFLAGS += -flto LDFLAGS += $(CXXFLAGS) endif - ifeq ($(comp),mingw) +# To use LTO and static linking on windows, the tool chain requires a recent gcc: +# gcc version 10.1 in msys2 or TDM-GCC version 9.2 are known to work, older might not. +# So, only enable it for a cross from Linux by default. + else ifeq ($(comp),mingw) ifeq ($(KERNEL),Linux) + ifneq ($(arch),i386) CXXFLAGS += -flto - LDFLAGS += $(CXXFLAGS) + LDFLAGS += $(CXXFLAGS) -flto=jobserver + endif endif endif endif @@ -368,9 +647,10 @@ ifneq (,$(findstring mpi, $(CXX))) endif ### ========================================================================== -### Section 4. Public targets +### Section 4. Public Targets ### ========================================================================== + help: @echo "" @echo "To compile stockfish, type: " @@ -379,22 +659,35 @@ help: @echo "" @echo "Supported targets:" @echo "" + @echo "help > Display architecture details" @echo "build > Standard build" - @echo "profile-build > PGO build" + @echo "net > Download the default nnue net" + @echo "profile-build > Faster build (with profile-guided optimization)" @echo "strip > Strip executable" @echo "install > Install executable" @echo "clean > Clean up" @echo "" @echo "Supported archs:" @echo "" - @echo "x86-64-bmi2 > x86 64-bit with pext support (also enables SSE4)" - @echo "x86-64-modern > x86 64-bit with popcnt support (also enables SSE3)" - @echo "x86-64 > x86 64-bit generic" - @echo "x86-32 > x86 32-bit (also enables SSE)" - @echo "x86-32-old > x86 32-bit fall back for old hardware" + @echo "x86-64-vnni512 > x86 64-bit with vnni support 512bit wide" + @echo "x86-64-vnni256 > x86 64-bit with vnni support 256bit wide" + @echo "x86-64-avx512 > x86 64-bit with avx512 support" + @echo "x86-64-bmi2 > x86 64-bit with bmi2 support" + @echo "x86-64-avx2 > x86 64-bit with avx2 support" + @echo "x86-64-sse41-popcnt > x86 64-bit with sse41 and popcnt support" + @echo "x86-64-modern > common modern CPU, currently x86-64-sse41-popcnt" + @echo "x86-64-ssse3 > x86 64-bit with ssse3 support" + @echo "x86-64-sse3-popcnt > x86 64-bit with sse3 and popcnt support" + @echo "x86-64 > x86 64-bit generic (with sse2 support)" + @echo "x86-32-sse41-popcnt > x86 32-bit with sse41 and popcnt support" + @echo "x86-32-sse2 > x86 32-bit with sse2 support" + @echo "x86-32 > x86 32-bit generic (with mmx and sse support)" @echo "ppc-64 > PPC 64-bit" @echo "ppc-32 > PPC 32-bit" @echo "armv7 > ARMv7 32-bit" + @echo "armv7-neon > ARMv7 32-bit with popcnt and neon" + @echo "armv8 > ARMv8 64-bit with popcnt and neon" + @echo "apple-silicon > Apple silicon ARM64" @echo "general-64 > unspecified 64-bit" @echo "general-32 > unspecified 32-bit" @echo "" @@ -404,27 +697,37 @@ help: @echo "mingw > Gnu compiler with MinGW under Windows" @echo "clang > LLVM Clang compiler" @echo "icc > Intel compiler" + @echo "ndk > Google NDK to cross-compile for Android" @echo "" @echo "Simple examples. If you don't know what to do, you likely want to run: " @echo "" - @echo "make build ARCH=x86-64 (This is for 64-bit systems)" - @echo "make build ARCH=x86-32 (This is for 32-bit systems)" + @echo "make -j build ARCH=x86-64 (A portable, slow compile for 64-bit systems)" + @echo "make -j build ARCH=x86-32 (A portable, slow compile for 32-bit systems)" @echo "" - @echo "Advanced examples, for experienced users: " + @echo "Advanced examples, for experienced users looking for performance: " @echo "" - @echo "make build ARCH=x86-64 COMP=clang" - @echo "make profile-build ARCH=x86-64-bmi2 COMP=gcc COMPCXX=g++-4.8" + @echo "make help ARCH=x86-64-bmi2" + @echo "make -j profile-build ARCH=x86-64-bmi2 COMP=gcc COMPCXX=g++-9.0" + @echo "make -j build ARCH=x86-64-ssse3 COMP=clang" @echo "" + @echo "-------------------------------" +ifeq ($(SUPPORTED_ARCH)$(help_skip_sanity), true) + @echo "The selected architecture $(ARCH) will enable the following configuration: " + @$(MAKE) ARCH=$(ARCH) COMP=$(COMP) config-sanity +else + @echo "Specify a supported architecture with the ARCH option for more details" + @echo "" +endif -.PHONY: help build profile-build strip install clean objclean profileclean help \ +.PHONY: help build profile-build strip install clean net objclean profileclean \ config-sanity icc-profile-use icc-profile-make gcc-profile-use gcc-profile-make \ clang-profile-use clang-profile-make -build: config-sanity +build: net config-sanity $(MAKE) ARCH=$(ARCH) COMP=$(COMP) all -profile-build: config-sanity objclean profileclean +profile-build: net config-sanity objclean profileclean @echo "" @echo "Step 1/4. Building instrumented executable ..." $(MAKE) ARCH=$(ARCH) COMP=$(COMP) $(profile_make) @@ -440,37 +743,61 @@ profile-build: config-sanity objclean profileclean $(MAKE) ARCH=$(ARCH) COMP=$(COMP) profileclean strip: - strip $(EXE) + $(STRIP) $(EXE) install: -mkdir -p -m 755 $(BINDIR) -cp $(EXE) $(BINDIR) -strip $(BINDIR)/$(EXE) -#clean all +# clean all clean: objclean profileclean @rm -f .depend *~ core +# evaluation network (nnue) +net: + $(eval nnuenet := $(shell grep EvalFileDefaultName evaluate.h | grep define | sed 's/.*\(nn-[a-z0-9]\{12\}.nnue\).*/\1/')) + @echo "Default net: $(nnuenet)" + $(eval nnuedownloadurl := https://tests.stockfishchess.org/api/nn/$(nnuenet)) + $(eval curl_or_wget := $(shell if hash curl 2>/dev/null; then echo "curl -skL"; elif hash wget 2>/dev/null; then echo "wget -qO-"; fi)) + @if test -f "$(nnuenet)"; then \ + echo "Already available."; \ + else \ + if [ "x$(curl_or_wget)" = "x" ]; then \ + echo "Automatic download failed: neither curl nor wget is installed. Install one of these tools or download the net manually"; exit 1; \ + else \ + echo "Downloading $(nnuedownloadurl)"; $(curl_or_wget) $(nnuedownloadurl) > $(nnuenet);\ + fi; \ + fi; + $(eval shasum_command := $(shell if hash shasum 2>/dev/null; then echo "shasum -a 256 "; elif hash sha256sum 2>/dev/null; then echo "sha256sum "; fi)) + @if [ "x$(shasum_command)" != "x" ]; then \ + if [ "$(nnuenet)" != "nn-"`$(shasum_command) $(nnuenet) | cut -c1-12`".nnue" ]; then \ + echo "Failed download or $(nnuenet) corrupted, please delete!"; exit 1; \ + fi \ + else \ + echo "shasum / sha256sum not found, skipping net validation"; \ + fi + # clean binaries and objects objclean: - @rm -f $(EXE) *.o ./syzygy/*.o + @rm -f $(EXE) *.o ./syzygy/*.o ./nnue/*.o ./nnue/features/*.o # clean auxiliary profiling files profileclean: @rm -rf profdir - @rm -f bench.txt *.gcda ./syzygy/*.gcda *.gcno ./syzygy/*.gcno + @rm -f bench.txt *.gcda *.gcno ./syzygy/*.gcda ./nnue/*.gcda ./nnue/features/*.gcda *.s @rm -f stockfish.profdata *.profraw default: help ### ========================================================================== -### Section 5. Private targets +### Section 5. Private Targets ### ========================================================================== all: $(EXE) .depend -config-sanity: +config-sanity: net @echo "" @echo "Config:" @echo "debug: '$(debug)'" @@ -482,8 +809,17 @@ config-sanity: @echo "os: '$(OS)'" @echo "prefetch: '$(prefetch)'" @echo "popcnt: '$(popcnt)'" - @echo "sse: '$(sse)'" @echo "pext: '$(pext)'" + @echo "sse: '$(sse)'" + @echo "mmx: '$(mmx)'" + @echo "sse2: '$(sse2)'" + @echo "ssse3: '$(ssse3)'" + @echo "sse41: '$(sse41)'" + @echo "avx2: '$(avx2)'" + @echo "avx512: '$(avx512)'" + @echo "vnni256: '$(vnni256)'" + @echo "vnni512: '$(vnni512)'" + @echo "neon: '$(neon)'" @echo "mpi: '$(mpi)'" @echo "" @echo "Flags:" @@ -496,17 +832,29 @@ config-sanity: @test "$(debug)" = "yes" || test "$(debug)" = "no" @test "$(sanitize)" = "undefined" || test "$(sanitize)" = "thread" || test "$(sanitize)" = "address" || test "$(sanitize)" = "no" @test "$(optimize)" = "yes" || test "$(optimize)" = "no" + @test "$(SUPPORTED_ARCH)" = "true" @test "$(arch)" = "any" || test "$(arch)" = "x86_64" || test "$(arch)" = "i386" || \ - test "$(arch)" = "ppc64" || test "$(arch)" = "ppc" || test "$(arch)" = "armv7" + test "$(arch)" = "ppc64" || test "$(arch)" = "ppc" || \ + test "$(arch)" = "armv7" || test "$(arch)" = "armv8" || test "$(arch)" = "arm64" @test "$(bits)" = "32" || test "$(bits)" = "64" @test "$(prefetch)" = "yes" || test "$(prefetch)" = "no" @test "$(popcnt)" = "yes" || test "$(popcnt)" = "no" - @test "$(sse)" = "yes" || test "$(sse)" = "no" @test "$(pext)" = "yes" || test "$(pext)" = "no" - @test "$(comp)" = "gcc" || test "$(comp)" = "icc" || test "$(comp)" = "mingw" || test "$(comp)" = "clang" + @test "$(sse)" = "yes" || test "$(sse)" = "no" + @test "$(mmx)" = "yes" || test "$(mmx)" = "no" + @test "$(sse2)" = "yes" || test "$(sse2)" = "no" + @test "$(ssse3)" = "yes" || test "$(ssse3)" = "no" + @test "$(sse41)" = "yes" || test "$(sse41)" = "no" + @test "$(avx2)" = "yes" || test "$(avx2)" = "no" + @test "$(avx512)" = "yes" || test "$(avx512)" = "no" + @test "$(vnni256)" = "yes" || test "$(vnni256)" = "no" + @test "$(vnni512)" = "yes" || test "$(vnni512)" = "no" + @test "$(neon)" = "yes" || test "$(neon)" = "no" + @test "$(comp)" = "gcc" || test "$(comp)" = "icc" || test "$(comp)" = "mingw" || test "$(comp)" = "clang" \ + || test "$(comp)" = "armv7a-linux-androideabi16-clang" || test "$(comp)" = "aarch64-linux-android21-clang" $(EXE): $(OBJS) - $(CXX) -o $@ $(OBJS) $(LDFLAGS) + +$(CXX) -o $@ $(OBJS) $(LDFLAGS) clang-profile-make: $(MAKE) ARCH=$(ARCH) COMP=$(COMP) \ @@ -515,7 +863,7 @@ clang-profile-make: all clang-profile-use: - llvm-profdata merge -output=stockfish.profdata *.profraw + $(XCRUN) llvm-profdata merge -output=stockfish.profdata *.profraw $(MAKE) ARCH=$(ARCH) COMP=$(COMP) \ EXTRACXXFLAGS='-fprofile-instr-use=stockfish.profdata' \ EXTRALDFLAGS='-fprofile-use ' \ @@ -545,7 +893,6 @@ icc-profile-use: all .depend: - -@$(CXX) $(DEPENDFLAGS) -MM $(OBJS:.o=.cpp) > $@ 2> /dev/null + -@$(CXX) $(DEPENDFLAGS) -MM $(SRCS) > $@ 2> /dev/null -include .depend - diff --git a/src/benchmark.cpp b/src/benchmark.cpp index 8f30bee4..7945a453 100644 --- a/src/benchmark.cpp +++ b/src/benchmark.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -65,6 +63,7 @@ const vector Defaults = { "4rrk1/1p1nq3/p7/2p1P1pp/3P2bp/3Q1Bn1/PPPB4/1K2R1NR w - - 40 21", "r3k2r/3nnpbp/q2pp1p1/p7/Pp1PPPP1/4BNN1/1P5P/R2Q1RK1 w kq - 0 16", "3Qb1k1/1r2ppb1/pN1n2q1/Pp1Pp1Pr/4P2p/4BP2/4B1R1/1R5K b - - 11 40", + "4k3/3q1r2/1N2r1b1/3ppN2/2nPP3/1B1R2n1/2R1Q3/3K4 w - - 5 1", // 5-man positions "8/8/8/8/5kp1/P7/8/1K1N4 w - - 0 1", // Kc2 - mate @@ -87,17 +86,20 @@ const vector Defaults = { // Chess 960 "setoption name UCI_Chess960 value true", - "bbqnnrkr/pppppppp/8/8/8/8/PPPPPPPP/BBQNNRKR w KQkq - 0 1 moves g2g3 d7d5 d2d4 c8h3 c1g5 e8d6 g5e7 f7f6", + "bbqnnrkr/pppppppp/8/8/8/8/PPPPPPPP/BBQNNRKR w HFhf - 0 1 moves g2g3 d7d5 d2d4 c8h3 c1g5 e8d6 g5e7 f7f6", "setoption name UCI_Chess960 value false" }; } // namespace +namespace Stockfish { + /// setup_bench() builds a list of UCI commands to be run by bench. There /// are five parameters: TT size in MB, number of search threads that /// should be used, the limit value spent for each position, a file name -/// where to look for positions in FEN format and the type of the limit: -/// depth, perft, nodes and movetime (in millisecs). +/// where to look for positions in FEN format, the type of the limit: +/// depth, perft, nodes and movetime (in millisecs), and evaluation type +/// mixed (default), classical, NNUE. /// /// bench -> search default positions up to depth 13 /// bench 64 1 15 -> search default positions up to depth 15 (TT = 64MB) @@ -116,8 +118,9 @@ vector setup_bench(const Position& current, istream& is) { string limit = (is >> token) ? token : "13"; string fenFile = (is >> token) ? token : "default"; string limitType = (is >> token) ? token : "depth"; + string evalType = (is >> token) ? token : "mixed"; - go = "go " + limitType + " " + limit; + go = limitType == "eval" ? "eval" : "go " + limitType + " " + limit; if (fenFile == "default") fens = Defaults; @@ -147,14 +150,25 @@ vector setup_bench(const Position& current, istream& is) { list.emplace_back("setoption name Hash value " + ttSize); list.emplace_back("ucinewgame"); + size_t posCounter = 0; + for (const string& fen : fens) if (fen.find("setoption") != string::npos) list.emplace_back(fen); else { + if (evalType == "classical" || (evalType == "mixed" && posCounter % 2 == 0)) + list.emplace_back("setoption name Use NNUE value false"); + else if (evalType == "NNUE" || (evalType == "mixed" && posCounter % 2 != 0)) + list.emplace_back("setoption name Use NNUE value true"); list.emplace_back("position fen " + fen); list.emplace_back(go); + ++posCounter; } + list.emplace_back("setoption name Use NNUE value true"); + return list; } + +} // namespace Stockfish diff --git a/src/bitbase.cpp b/src/bitbase.cpp index 2b1a5517..ece9ec72 100644 --- a/src/bitbase.cpp +++ b/src/bitbase.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -19,20 +17,21 @@ */ #include -#include #include +#include #include "bitboard.h" #include "types.h" +namespace Stockfish { + namespace { // There are 24 possible pawn squares: files A to D and ranks from 2 to 7. // Positions with the pawn on files E to H will be mirrored before probing. constexpr unsigned MAX_INDEX = 2*24*64*64; // stm * psq * wksq * bksq = 196608 - // Each uint32_t stores results of 32 positions, one per bit - uint32_t KPKBitbase[MAX_INDEX / 32]; + std::bitset KPKBitbase; // A KPK bitbase index is an integer in [0, IndexMax] range // @@ -43,8 +42,8 @@ namespace { // bit 12: side to move (WHITE or BLACK) // bit 13-14: white pawn file (from FILE_A to FILE_D) // bit 15-17: white pawn RANK_7 - rank (from RANK_7 - RANK_7 to RANK_7 - RANK_2) - unsigned index(Color us, Square bksq, Square wksq, Square psq) { - return wksq | (bksq << 6) | (us << 12) | (file_of(psq) << 13) | ((RANK_7 - rank_of(psq)) << 15); + unsigned index(Color stm, Square bksq, Square wksq, Square psq) { + return int(wksq) | (bksq << 6) | (stm << 12) | (file_of(psq) << 13) | ((RANK_7 - rank_of(psq)) << 15); } enum Result { @@ -60,25 +59,20 @@ namespace { KPKPosition() = default; explicit KPKPosition(unsigned idx); operator Result() const { return result; } - Result classify(const std::vector& db) - { return us == WHITE ? classify(db) : classify(db); } + Result classify(const std::vector& db); - template Result classify(const std::vector& db); - - Color us; + Color stm; Square ksq[COLOR_NB], psq; Result result; }; } // namespace - -bool Bitbases::probe(Square wksq, Square wpsq, Square bksq, Color us) { +bool Bitbases::probe(Square wksq, Square wpsq, Square bksq, Color stm) { assert(file_of(wpsq) <= FILE_D); - unsigned idx = index(us, bksq, wksq, wpsq); - return KPKBitbase[idx / 32] & (1 << (idx & 0x1F)); + return KPKBitbase[index(stm, bksq, wksq, wpsq)]; } @@ -97,41 +91,40 @@ void Bitbases::init() { for (repeat = idx = 0; idx < MAX_INDEX; ++idx) repeat |= (db[idx] == UNKNOWN && db[idx].classify(db) != UNKNOWN); - // Map 32 results into one KPKBitbase[] entry + // Fill the bitbase with the decisive results for (idx = 0; idx < MAX_INDEX; ++idx) if (db[idx] == WIN) - KPKBitbase[idx / 32] |= 1 << (idx & 0x1F); + KPKBitbase.set(idx); } - namespace { KPKPosition::KPKPosition(unsigned idx) { ksq[WHITE] = Square((idx >> 0) & 0x3F); ksq[BLACK] = Square((idx >> 6) & 0x3F); - us = Color ((idx >> 12) & 0x01); + stm = Color ((idx >> 12) & 0x01); psq = make_square(File((idx >> 13) & 0x3), Rank(RANK_7 - ((idx >> 15) & 0x7))); - // Check if two pieces are on the same square or if a king can be captured + // Invalid if two pieces are on the same square or if a king can be captured if ( distance(ksq[WHITE], ksq[BLACK]) <= 1 || ksq[WHITE] == psq || ksq[BLACK] == psq - || (us == WHITE && (PawnAttacks[WHITE][psq] & ksq[BLACK]))) + || (stm == WHITE && (pawn_attacks_bb(WHITE, psq) & ksq[BLACK]))) result = INVALID; - // Immediate win if a pawn can be promoted without getting captured - else if ( us == WHITE + // Win if the pawn can be promoted without getting captured + else if ( stm == WHITE && rank_of(psq) == RANK_7 - && ksq[us] != psq + NORTH - && ( distance(ksq[~us], psq + NORTH) > 1 - || (PseudoAttacks[KING][ksq[us]] & (psq + NORTH)))) + && ksq[WHITE] != psq + NORTH + && ( distance(ksq[BLACK], psq + NORTH) > 1 + || (distance(ksq[WHITE], psq + NORTH) == 1))) result = WIN; - // Immediate draw if it is a stalemate or a king captures undefended pawn - else if ( us == BLACK - && ( !(PseudoAttacks[KING][ksq[us]] & ~(PseudoAttacks[KING][ksq[~us]] | PawnAttacks[~us][psq])) - || (PseudoAttacks[KING][ksq[us]] & psq & ~PseudoAttacks[KING][ksq[~us]]))) + // Draw if it is stalemate or the black king can capture the pawn + else if ( stm == BLACK + && ( !(attacks_bb(ksq[BLACK]) & ~(attacks_bb(ksq[WHITE]) | pawn_attacks_bb(WHITE, psq))) + || (attacks_bb(ksq[BLACK]) & ~attacks_bb(ksq[WHITE]) & psq))) result = DRAW; // Position will be classified later @@ -139,7 +132,6 @@ namespace { result = UNKNOWN; } - template Result KPKPosition::classify(const std::vector& db) { // White to move: If one move leads to a position classified as WIN, the result @@ -151,30 +143,30 @@ namespace { // of the current position is DRAW. If all moves lead to positions classified // as WIN, the position is classified as WIN, otherwise the current position is // classified as UNKNOWN. - - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Result Good = (Us == WHITE ? WIN : DRAW); - constexpr Result Bad = (Us == WHITE ? DRAW : WIN); + const Result Good = (stm == WHITE ? WIN : DRAW); + const Result Bad = (stm == WHITE ? DRAW : WIN); Result r = INVALID; - Bitboard b = PseudoAttacks[KING][ksq[Us]]; + Bitboard b = attacks_bb(ksq[stm]); while (b) - r |= Us == WHITE ? db[index(Them, ksq[Them] , pop_lsb(&b), psq)] - : db[index(Them, pop_lsb(&b), ksq[Them] , psq)]; + r |= stm == WHITE ? db[index(BLACK, ksq[BLACK] , pop_lsb(&b), psq)] + : db[index(WHITE, pop_lsb(&b), ksq[WHITE], psq)]; - if (Us == WHITE) + if (stm == WHITE) { if (rank_of(psq) < RANK_7) // Single push - r |= db[index(Them, ksq[Them], ksq[Us], psq + NORTH)]; + r |= db[index(BLACK, ksq[BLACK], ksq[WHITE], psq + NORTH)]; if ( rank_of(psq) == RANK_2 // Double push - && psq + NORTH != ksq[Us] - && psq + NORTH != ksq[Them]) - r |= db[index(Them, ksq[Them], ksq[Us], psq + NORTH + NORTH)]; + && psq + NORTH != ksq[WHITE] + && psq + NORTH != ksq[BLACK]) + r |= db[index(BLACK, ksq[BLACK], ksq[WHITE], psq + NORTH + NORTH)]; } return result = r & Good ? Good : r & UNKNOWN ? UNKNOWN : Bad; } } // namespace + +} // namespace Stockfish diff --git a/src/bitboard.cpp b/src/bitboard.cpp index 2afd3766..a2021449 100644 --- a/src/bitboard.cpp +++ b/src/bitboard.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -24,6 +22,8 @@ #include "bitboard.h" #include "misc.h" +namespace Stockfish { + uint8_t PopCnt16[1 << 16]; uint8_t SquareDistance[SQUARE_NB][SQUARE_NB]; @@ -40,14 +40,23 @@ namespace { Bitboard RookTable[0x19000]; // To store rook attacks Bitboard BishopTable[0x1480]; // To store bishop attacks - void init_magics(Bitboard table[], Magic magics[], Direction directions[]); + void init_magics(PieceType pt, Bitboard table[], Magic magics[]); + +} + +/// safe_destination() returns the bitboard of target square for the given step +/// from the given square. If the step is off the board, returns empty bitboard. + +inline Bitboard safe_destination(Square s, int step) { + Square to = Square(s + step); + return is_ok(to) && distance(s, to) <= 2 ? square_bb(to) : Bitboard(0); } /// Bitboards::pretty() returns an ASCII representation of a bitboard suitable /// to be printed to standard output. Useful for debugging. -const std::string Bitboards::pretty(Bitboard b) { +std::string Bitboards::pretty(Bitboard b) { std::string s = "+---+---+---+---+---+---+---+---+\n"; @@ -56,8 +65,9 @@ const std::string Bitboards::pretty(Bitboard b) { for (File f = FILE_A; f <= FILE_H; ++f) s += b & make_square(f, r) ? "| X " : "| "; - s += "|\n+---+---+---+---+---+---+---+---+\n"; + s += "| " + std::to_string(1 + r) + "\n+---+---+---+---+---+---+---+---+\n"; } + s += " a b c d e f g h\n"; return s; } @@ -69,41 +79,29 @@ const std::string Bitboards::pretty(Bitboard b) { void Bitboards::init() { for (unsigned i = 0; i < (1 << 16); ++i) - PopCnt16[i] = std::bitset<16>(i).count(); + PopCnt16[i] = uint8_t(std::bitset<16>(i).count()); for (Square s = SQ_A1; s <= SQ_H8; ++s) SquareBB[s] = (1ULL << s); for (Square s1 = SQ_A1; s1 <= SQ_H8; ++s1) for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2) - SquareDistance[s1][s2] = std::max(distance(s1, s2), distance(s1, s2)); + SquareDistance[s1][s2] = std::max(distance(s1, s2), distance(s1, s2)); - int steps[][5] = { {}, { 7, 9 }, { 6, 10, 15, 17 }, {}, {}, {}, { 1, 7, 8, 9 } }; - - for (Color c : { WHITE, BLACK }) - for (PieceType pt : { PAWN, KNIGHT, KING }) - for (Square s = SQ_A1; s <= SQ_H8; ++s) - for (int i = 0; steps[pt][i]; ++i) - { - Square to = s + Direction(c == WHITE ? steps[pt][i] : -steps[pt][i]); - - if (is_ok(to) && distance(s, to) < 3) - { - if (pt == PAWN) - PawnAttacks[c][s] |= to; - else - PseudoAttacks[pt][s] |= to; - } - } - - Direction RookDirections[] = { NORTH, EAST, SOUTH, WEST }; - Direction BishopDirections[] = { NORTH_EAST, SOUTH_EAST, SOUTH_WEST, NORTH_WEST }; - - init_magics(RookTable, RookMagics, RookDirections); - init_magics(BishopTable, BishopMagics, BishopDirections); + init_magics(ROOK, RookTable, RookMagics); + init_magics(BISHOP, BishopTable, BishopMagics); for (Square s1 = SQ_A1; s1 <= SQ_H8; ++s1) { + PawnAttacks[WHITE][s1] = pawn_attacks_bb(square_bb(s1)); + PawnAttacks[BLACK][s1] = pawn_attacks_bb(square_bb(s1)); + + for (int step : {-9, -8, -7, -1, 1, 7, 8, 9} ) + PseudoAttacks[KING][s1] |= safe_destination(s1, step); + + for (int step : {-17, -15, -10, -6, 6, 10, 15, 17} ) + PseudoAttacks[KNIGHT][s1] |= safe_destination(s1, step); + PseudoAttacks[QUEEN][s1] = PseudoAttacks[BISHOP][s1] = attacks_bb(s1, 0); PseudoAttacks[QUEEN][s1] |= PseudoAttacks[ ROOK][s1] = attacks_bb< ROOK>(s1, 0); @@ -114,25 +112,22 @@ void Bitboards::init() { } } - namespace { - Bitboard sliding_attack(Direction directions[], Square sq, Bitboard occupied) { + Bitboard sliding_attack(PieceType pt, Square sq, Bitboard occupied) { - Bitboard attack = 0; + Bitboard attacks = 0; + Direction RookDirections[4] = {NORTH, SOUTH, EAST, WEST}; + Direction BishopDirections[4] = {NORTH_EAST, SOUTH_EAST, SOUTH_WEST, NORTH_WEST}; - for (int i = 0; i < 4; ++i) - for (Square s = sq + directions[i]; - is_ok(s) && distance(s, s - directions[i]) == 1; - s += directions[i]) - { - attack |= s; + for (Direction d : (pt == ROOK ? RookDirections : BishopDirections)) + { + Square s = sq; + while(safe_destination(s, d) && !(occupied & s)) + attacks |= (s += d); + } - if (occupied & s) - break; - } - - return attack; + return attacks; } @@ -141,7 +136,7 @@ namespace { // www.chessprogramming.org/Magic_Bitboards. In particular, here we use the so // called "fancy" approach. - void init_magics(Bitboard table[], Magic magics[], Direction directions[]) { + void init_magics(PieceType pt, Bitboard table[], Magic magics[]) { // Optimal PRNG seeds to pick the correct magics in the shortest time int seeds[][RANK_NB] = { { 8977, 44560, 54343, 38998, 5731, 95205, 104912, 17020 }, @@ -161,7 +156,7 @@ namespace { // the number of 1s of the mask. Hence we deduce the size of the shift to // apply to the 64 or 32 bits word to get the index. Magic& m = magics[s]; - m.mask = sliding_attack(directions, s, 0) & ~edges; + m.mask = sliding_attack(pt, s, 0) & ~edges; m.shift = (Is64Bit ? 64 : 32) - popcount(m.mask); // Set the offset for the attacks table of the square. We have individual @@ -173,7 +168,7 @@ namespace { b = size = 0; do { occupancy[size] = b; - reference[size] = sliding_attack(directions, s, b); + reference[size] = sliding_attack(pt, s, b); if (HasPext) m.attacks[pext(b, m.mask)] = reference[size]; @@ -216,3 +211,5 @@ namespace { } } } + +} // namespace Stockfish diff --git a/src/bitboard.h b/src/bitboard.h index 477b1655..e14fe0df 100644 --- a/src/bitboard.h +++ b/src/bitboard.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,19 +23,21 @@ #include "types.h" +namespace Stockfish { + namespace Bitbases { void init(); bool probe(Square wksq, Square wpsq, Square bksq, Color us); -} +} // namespace Stockfish::Bitbases namespace Bitboards { void init(); -const std::string pretty(Bitboard b); +std::string pretty(Bitboard b); -} +} // namespace Stockfish::Bitboards constexpr Bitboard AllSquares = ~Bitboard(0); constexpr Bitboard DarkSquares = 0xAA55AA55AA55AA55ULL; @@ -106,10 +106,11 @@ extern Magic RookMagics[SQUARE_NB]; extern Magic BishopMagics[SQUARE_NB]; inline Bitboard square_bb(Square s) { - assert(s >= SQ_A1 && s <= SQ_H8); + assert(is_ok(s)); return SquareBB[s]; } + /// Overloads of bitwise operators between a Bitboard and a Square for testing /// whether a given bit is set in a bitboard, and for setting and clearing bits. @@ -119,36 +120,43 @@ inline Bitboard operator^( Bitboard b, Square s) { return b ^ square_bb(s); } inline Bitboard& operator|=(Bitboard& b, Square s) { return b |= square_bb(s); } inline Bitboard& operator^=(Bitboard& b, Square s) { return b ^= square_bb(s); } +inline Bitboard operator&(Square s, Bitboard b) { return b & s; } +inline Bitboard operator|(Square s, Bitboard b) { return b | s; } +inline Bitboard operator^(Square s, Bitboard b) { return b ^ s; } + +inline Bitboard operator|(Square s1, Square s2) { return square_bb(s1) | s2; } + constexpr bool more_than_one(Bitboard b) { return b & (b - 1); } -inline bool opposite_colors(Square s1, Square s2) { - return bool(DarkSquares & s1) != bool(DarkSquares & s2); + +constexpr bool opposite_colors(Square s1, Square s2) { + return (s1 + rank_of(s1) + s2 + rank_of(s2)) & 1; } /// rank_bb() and file_bb() return a bitboard representing all the squares on /// the given file or rank. -inline Bitboard rank_bb(Rank r) { +constexpr Bitboard rank_bb(Rank r) { return Rank1BB << (8 * r); } -inline Bitboard rank_bb(Square s) { +constexpr Bitboard rank_bb(Square s) { return rank_bb(rank_of(s)); } -inline Bitboard file_bb(File f) { +constexpr Bitboard file_bb(File f) { return FileABB << f; } -inline Bitboard file_bb(Square s) { +constexpr Bitboard file_bb(Square s) { return file_bb(file_of(s)); } -/// shift() moves a bitboard one step along direction D +/// shift() moves a bitboard one or two steps as specified by the direction D template constexpr Bitboard shift(Bitboard b) { @@ -170,6 +178,12 @@ constexpr Bitboard pawn_attacks_bb(Bitboard b) { : shift(b) | shift(b); } +inline Bitboard pawn_attacks_bb(Color c, Square s) { + + assert(is_ok(s)); + return PawnAttacks[c][s]; +} + /// pawn_double_attacks_bb() returns the squares doubly attacked by pawns of the /// given color from the squares in the given bitboard. @@ -182,19 +196,33 @@ constexpr Bitboard pawn_double_attacks_bb(Bitboard b) { /// adjacent_files_bb() returns a bitboard representing all the squares on the -/// adjacent files of the given one. +/// adjacent files of a given square. -inline Bitboard adjacent_files_bb(Square s) { +constexpr Bitboard adjacent_files_bb(Square s) { return shift(file_bb(s)) | shift(file_bb(s)); } -/// between_bb() returns squares that are linearly between the given squares -/// If the given squares are not on a same file/rank/diagonal, return 0. +/// line_bb() returns a bitboard representing an entire line (from board edge +/// to board edge) that intersects the two given squares. If the given squares +/// are not on a same file/rank/diagonal, the function returns 0. For instance, +/// line_bb(SQ_C4, SQ_F7) will return a bitboard with the A2-G8 diagonal. + +inline Bitboard line_bb(Square s1, Square s2) { + + assert(is_ok(s1) && is_ok(s2)); + return LineBB[s1][s2]; +} + + +/// between_bb() returns a bitboard representing squares that are linearly +/// between the two given squares (excluding the given squares). If the given +/// squares are not on a same file/rank/diagonal, we return 0. For instance, +/// between_bb(SQ_C4, SQ_F7) will return a bitboard with squares D5 and E6. inline Bitboard between_bb(Square s1, Square s2) { - return LineBB[s1][s2] & ( (AllSquares << (s1 + (s1 < s2))) - ^(AllSquares << (s2 + !(s1 < s2)))); + Bitboard b = line_bb(s1, s2) & ((AllSquares << s1) ^ (AllSquares << s2)); + return b & (b - 1); //exclude lsb } @@ -202,25 +230,25 @@ inline Bitboard between_bb(Square s1, Square s2) { /// in front of the given one, from the point of view of the given color. For instance, /// forward_ranks_bb(BLACK, SQ_D3) will return the 16 squares on ranks 1 and 2. -inline Bitboard forward_ranks_bb(Color c, Square s) { - return c == WHITE ? ~Rank1BB << 8 * (rank_of(s) - RANK_1) - : ~Rank8BB >> 8 * (RANK_8 - rank_of(s)); +constexpr Bitboard forward_ranks_bb(Color c, Square s) { + return c == WHITE ? ~Rank1BB << 8 * relative_rank(WHITE, s) + : ~Rank8BB >> 8 * relative_rank(BLACK, s); } /// forward_file_bb() returns a bitboard representing all the squares along the /// line in front of the given one, from the point of view of the given color. -inline Bitboard forward_file_bb(Color c, Square s) { +constexpr Bitboard forward_file_bb(Color c, Square s) { return forward_ranks_bb(c, s) & file_bb(s); } /// pawn_attack_span() returns a bitboard representing all the squares that can -/// be attacked by a pawn of the given color when it moves along its file, -/// starting from the given square. +/// be attacked by a pawn of the given color when it moves along its file, starting +/// from the given square. -inline Bitboard pawn_attack_span(Color c, Square s) { +constexpr Bitboard pawn_attack_span(Color c, Square s) { return forward_ranks_bb(c, s) & adjacent_files_bb(s); } @@ -228,8 +256,8 @@ inline Bitboard pawn_attack_span(Color c, Square s) { /// passed_pawn_span() returns a bitboard which can be used to test if a pawn of /// the given color and on the given square is a passed pawn. -inline Bitboard passed_pawn_span(Color c, Square s) { - return forward_ranks_bb(c, s) & (adjacent_files_bb(s) | file_bb(s)); +constexpr Bitboard passed_pawn_span(Color c, Square s) { + return pawn_attack_span(c, s) | forward_file_bb(c, s); } @@ -237,7 +265,7 @@ inline Bitboard passed_pawn_span(Color c, Square s) { /// straight or on a diagonal line. inline bool aligned(Square s1, Square s2, Square s3) { - return LineBB[s1][s2] & s3; + return line_bb(s1, s2) & s3; } @@ -249,23 +277,43 @@ template<> inline int distance(Square x, Square y) { return std::abs(file_ template<> inline int distance(Square x, Square y) { return std::abs(rank_of(x) - rank_of(y)); } template<> inline int distance(Square x, Square y) { return SquareDistance[x][y]; } -template constexpr const T& clamp(const T& v, const T& lo, const T& hi) { - return v < lo ? lo : v > hi ? hi : v; +inline int edge_distance(File f) { return std::min(f, File(FILE_H - f)); } +inline int edge_distance(Rank r) { return std::min(r, Rank(RANK_8 - r)); } + + +/// attacks_bb(Square) returns the pseudo attacks of the give piece type +/// assuming an empty board. + +template +inline Bitboard attacks_bb(Square s) { + + assert((Pt != PAWN) && (is_ok(s))); + + return PseudoAttacks[Pt][s]; } -/// attacks_bb() returns a bitboard representing all the squares attacked by a -/// piece of type Pt (bishop or rook) placed on 's'. + +/// attacks_bb(Square, Bitboard) returns the attacks by the given piece +/// assuming the board is occupied according to the passed Bitboard. +/// Sliding piece attacks do not continue passed an occupied square. template inline Bitboard attacks_bb(Square s, Bitboard occupied) { - const Magic& m = Pt == ROOK ? RookMagics[s] : BishopMagics[s]; - return m.attacks[m.index(occupied)]; + assert((Pt != PAWN) && (is_ok(s))); + + switch (Pt) + { + case BISHOP: return BishopMagics[s].attacks[BishopMagics[s].index(occupied)]; + case ROOK : return RookMagics[s].attacks[ RookMagics[s].index(occupied)]; + case QUEEN : return attacks_bb(s, occupied) | attacks_bb(s, occupied); + default : return PseudoAttacks[Pt][s]; + } } inline Bitboard attacks_bb(PieceType pt, Square s, Bitboard occupied) { - assert(pt != PAWN); + assert((pt != PAWN) && (is_ok(s))); switch (pt) { @@ -370,15 +418,20 @@ inline Square msb(Bitboard b) { /// pop_lsb() finds and clears the least significant bit in a non-zero bitboard inline Square pop_lsb(Bitboard* b) { + assert(*b); const Square s = lsb(*b); *b &= *b - 1; return s; } -/// frontmost_sq() returns the most advanced square for the given color +/// frontmost_sq() returns the most advanced square for the given color, +/// requires a non-zero bitboard. inline Square frontmost_sq(Color c, Bitboard b) { + assert(b); return c == WHITE ? msb(b) : lsb(b); } +} // namespace Stockfish + #endif // #ifndef BITBOARD_H_INCLUDED diff --git a/src/cluster.cpp b/src/cluster.cpp index b56359e6..7d0cc0f8 100644 --- a/src/cluster.cpp +++ b/src/cluster.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -35,6 +33,7 @@ #include "tt.h" #include "timeman.h" +namespace Stockfish { namespace Cluster { // Total number of ranks and rank within the communicator @@ -452,6 +451,7 @@ uint64_t TT_saves() { } +} } #else @@ -459,6 +459,7 @@ uint64_t TT_saves() { #include "cluster.h" #include "thread.h" +namespace Stockfish { namespace Cluster { uint64_t nodes_searched() { @@ -476,6 +477,7 @@ uint64_t TT_saves() { return Threads.TT_saves(); } +} } #endif // USE_MPI diff --git a/src/cluster.h b/src/cluster.h index 0e074554..0d2fff57 100644 --- a/src/cluster.h +++ b/src/cluster.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -28,6 +26,7 @@ #include "tt.h" +namespace Stockfish { class Thread; /// The Cluster namespace contains functionality required to run on distributed @@ -121,6 +120,7 @@ inline void signals_sync() { } #endif /* USE_MPI */ +} } #endif // #ifndef CLUSTER_H_INCLUDED diff --git a/src/endgame.cpp b/src/endgame.cpp index e10f8d5d..a44d3a1c 100644 --- a/src/endgame.cpp +++ b/src/endgame.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -24,42 +22,27 @@ #include "endgame.h" #include "movegen.h" -using std::string; +namespace Stockfish { namespace { - // Table used to drive the king towards the edge of the board + // Used to drive the king towards the edge of the board // in KX vs K and KQ vs KR endgames. - constexpr int PushToEdges[SQUARE_NB] = { - 100, 90, 80, 70, 70, 80, 90, 100, - 90, 70, 60, 50, 50, 60, 70, 90, - 80, 60, 40, 30, 30, 40, 60, 80, - 70, 50, 30, 20, 20, 30, 50, 70, - 70, 50, 30, 20, 20, 30, 50, 70, - 80, 60, 40, 30, 30, 40, 60, 80, - 90, 70, 60, 50, 50, 60, 70, 90, - 100, 90, 80, 70, 70, 80, 90, 100 - }; + // Values range from 27 (center squares) to 90 (in the corners) + inline int push_to_edge(Square s) { + int rd = edge_distance(rank_of(s)), fd = edge_distance(file_of(s)); + return 90 - (7 * fd * fd / 2 + 7 * rd * rd / 2); + } - // Table used to drive the king towards a corner square of the - // right color in KBN vs K endgames. - constexpr int PushToCorners[SQUARE_NB] = { - 6400, 6080, 5760, 5440, 5120, 4800, 4480, 4160, - 6080, 5760, 5440, 5120, 4800, 4480, 4160, 4480, - 5760, 5440, 4960, 4480, 4480, 4000, 4480, 4800, - 5440, 5120, 4480, 3840, 3520, 4480, 4800, 5120, - 5120, 4800, 4480, 3520, 3840, 4480, 5120, 5440, - 4800, 4480, 4000, 4480, 4480, 4960, 5440, 5760, - 4480, 4160, 4480, 4800, 5120, 5440, 5760, 6080, - 4160, 4480, 4800, 5120, 5440, 5760, 6080, 6400 - }; + // Used to drive the king towards A1H8 corners in KBN vs K endgames. + // Values range from 0 on A8H1 diagonal to 7 in A1H8 corners + inline int push_to_corner(Square s) { + return abs(7 - rank_of(s) - file_of(s)); + } - // Tables used to drive a piece towards or away from another piece - constexpr int PushClose[8] = { 0, 0, 100, 80, 60, 40, 20, 10 }; - constexpr int PushAway [8] = { 0, 5, 20, 40, 60, 80, 90, 100 }; - - // Pawn Rank based scaling factors used in KRPPKRP endgame - constexpr int KRPPKRPScaleFactors[RANK_NB] = { 0, 9, 10, 14, 21, 44, 0, 0 }; + // Drive a piece close to or away from another piece + inline int push_close(Square s1, Square s2) { return 140 - 20 * distance(s1, s2); } + inline int push_away(Square s1, Square s2) { return 120 - push_close(s1, s2); } #ifndef NDEBUG bool verify_material(const Position& pos, Color c, Value npm, int pawnsCnt) { @@ -74,9 +57,9 @@ namespace { assert(pos.count(strongSide) == 1); if (file_of(pos.square(strongSide)) >= FILE_E) - sq = Square(sq ^ 7); // Mirror SQ_H1 -> SQ_A1 + sq = flip_file(sq); - return strongSide == WHITE ? sq : ~sq; + return strongSide == WHITE ? sq : flip_rank(sq); } } // namespace @@ -98,8 +81,6 @@ namespace Endgames { add("KQKR"); add("KNNKP"); - add("KNPK"); - add("KNPKB"); add("KRPKR"); add("KRPKB"); add("KBPKB"); @@ -124,20 +105,20 @@ Value Endgame::operator()(const Position& pos) const { if (pos.side_to_move() == weakSide && !MoveList(pos).size()) return VALUE_DRAW; - Square winnerKSq = pos.square(strongSide); - Square loserKSq = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + Square weakKing = pos.square(weakSide); Value result = pos.non_pawn_material(strongSide) + pos.count(strongSide) * PawnValueEg - + PushToEdges[loserKSq] - + PushClose[distance(winnerKSq, loserKSq)]; + + push_to_edge(weakKing) + + push_close(strongKing, weakKing); if ( pos.count(strongSide) || pos.count(strongSide) ||(pos.count(strongSide) && pos.count(strongSide)) || ( (pos.pieces(strongSide, BISHOP) & ~DarkSquares) && (pos.pieces(strongSide, BISHOP) & DarkSquares))) - result = std::min(result + VALUE_KNOWN_WIN, VALUE_MATE_IN_MAX_PLY - 1); + result = std::min(result + VALUE_KNOWN_WIN, VALUE_TB_WIN_IN_MAX_PLY - 1); return strongSide == pos.side_to_move() ? result : -result; } @@ -151,23 +132,23 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, KnightValueMg + BishopValueMg, 0)); assert(verify_material(pos, weakSide, VALUE_ZERO, 0)); - Square winnerKSq = pos.square(strongSide); - Square loserKSq = pos.square(weakSide); - Square bishopSq = pos.square(strongSide); + Square strongKing = pos.square(strongSide); + Square strongBishop = pos.square(strongSide); + Square weakKing = pos.square(weakSide); - // If our Bishop does not attack A1/H8, we flip the enemy king square + // If our bishop does not attack A1/H8, we flip the enemy king square // to drive to opposite corners (A8/H1). - Value result = VALUE_KNOWN_WIN - + PushClose[distance(winnerKSq, loserKSq)] - + PushToCorners[opposite_colors(bishopSq, SQ_A1) ? ~loserKSq : loserKSq]; + Value result = (VALUE_KNOWN_WIN + 3520) + + push_close(strongKing, weakKing) + + 420 * push_to_corner(opposite_colors(strongBishop, SQ_A1) ? flip_file(weakKing) : weakKing); - assert(abs(result) < VALUE_MATE_IN_MAX_PLY); + assert(abs(result) < VALUE_TB_WIN_IN_MAX_PLY); return strongSide == pos.side_to_move() ? result : -result; } -/// KP vs K. This endgame is evaluated with the help of a bitbase. +/// KP vs K. This endgame is evaluated with the help of a bitbase template<> Value Endgame::operator()(const Position& pos) const { @@ -175,16 +156,16 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, weakSide, VALUE_ZERO, 0)); // Assume strongSide is white and the pawn is on files A-D - Square wksq = normalize(pos, strongSide, pos.square(strongSide)); - Square bksq = normalize(pos, strongSide, pos.square(weakSide)); - Square psq = normalize(pos, strongSide, pos.square(strongSide)); + Square strongKing = normalize(pos, strongSide, pos.square(strongSide)); + Square strongPawn = normalize(pos, strongSide, pos.square(strongSide)); + Square weakKing = normalize(pos, strongSide, pos.square(weakSide)); Color us = strongSide == pos.side_to_move() ? WHITE : BLACK; - if (!Bitbases::probe(wksq, psq, bksq, us)) + if (!Bitbases::probe(strongKing, strongPawn, weakKing, us)) return VALUE_DRAW; - Value result = VALUE_KNOWN_WIN + PawnValueEg + Value(rank_of(psq)); + Value result = VALUE_KNOWN_WIN + PawnValueEg + Value(rank_of(strongPawn)); return strongSide == pos.side_to_move() ? result : -result; } @@ -200,36 +181,35 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, RookValueMg, 0)); assert(verify_material(pos, weakSide, VALUE_ZERO, 1)); - Square wksq = relative_square(strongSide, pos.square(strongSide)); - Square bksq = relative_square(strongSide, pos.square(weakSide)); - Square rsq = relative_square(strongSide, pos.square(strongSide)); - Square psq = relative_square(strongSide, pos.square(weakSide)); - - Square queeningSq = make_square(file_of(psq), RANK_1); + Square strongKing = pos.square(strongSide); + Square weakKing = pos.square(weakSide); + Square strongRook = pos.square(strongSide); + Square weakPawn = pos.square(weakSide); + Square queeningSquare = make_square(file_of(weakPawn), relative_rank(weakSide, RANK_8)); Value result; // If the stronger side's king is in front of the pawn, it's a win - if (forward_file_bb(WHITE, wksq) & psq) - result = RookValueEg - distance(wksq, psq); + if (forward_file_bb(strongSide, strongKing) & weakPawn) + result = RookValueEg - distance(strongKing, weakPawn); // If the weaker side's king is too far from the pawn and the rook, // it's a win. - else if ( distance(bksq, psq) >= 3 + (pos.side_to_move() == weakSide) - && distance(bksq, rsq) >= 3) - result = RookValueEg - distance(wksq, psq); + else if ( distance(weakKing, weakPawn) >= 3 + (pos.side_to_move() == weakSide) + && distance(weakKing, strongRook) >= 3) + result = RookValueEg - distance(strongKing, weakPawn); // If the pawn is far advanced and supported by the defending king, // the position is drawish - else if ( rank_of(bksq) <= RANK_3 - && distance(bksq, psq) == 1 - && rank_of(wksq) >= RANK_4 - && distance(wksq, psq) > 2 + (pos.side_to_move() == strongSide)) - result = Value(80) - 8 * distance(wksq, psq); + else if ( relative_rank(strongSide, weakKing) <= RANK_3 + && distance(weakKing, weakPawn) == 1 + && relative_rank(strongSide, strongKing) >= RANK_4 + && distance(strongKing, weakPawn) > 2 + (pos.side_to_move() == strongSide)) + result = Value(80) - 8 * distance(strongKing, weakPawn); else - result = Value(200) - 8 * ( distance(wksq, psq + SOUTH) - - distance(bksq, psq + SOUTH) - - distance(psq, queeningSq)); + result = Value(200) - 8 * ( distance(strongKing, weakPawn + pawn_push(weakSide)) + - distance(weakKing, weakPawn + pawn_push(weakSide)) + - distance(weakPawn, queeningSquare)); return strongSide == pos.side_to_move() ? result : -result; } @@ -243,7 +223,7 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, RookValueMg, 0)); assert(verify_material(pos, weakSide, BishopValueMg, 0)); - Value result = Value(PushToEdges[pos.square(weakSide)]); + Value result = Value(push_to_edge(pos.square(weakSide))); return strongSide == pos.side_to_move() ? result : -result; } @@ -256,9 +236,9 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, RookValueMg, 0)); assert(verify_material(pos, weakSide, KnightValueMg, 0)); - Square bksq = pos.square(weakSide); - Square bnsq = pos.square(weakSide); - Value result = Value(PushToEdges[bksq] + PushAway[distance(bksq, bnsq)]); + Square weakKing = pos.square(weakSide); + Square weakKnight = pos.square(weakSide); + Value result = Value(push_to_edge(weakKing) + push_away(weakKing, weakKnight)); return strongSide == pos.side_to_move() ? result : -result; } @@ -273,22 +253,22 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, QueenValueMg, 0)); assert(verify_material(pos, weakSide, VALUE_ZERO, 1)); - Square winnerKSq = pos.square(strongSide); - Square loserKSq = pos.square(weakSide); - Square pawnSq = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + Square weakKing = pos.square(weakSide); + Square weakPawn = pos.square(weakSide); - Value result = Value(PushClose[distance(winnerKSq, loserKSq)]); + Value result = Value(push_close(strongKing, weakKing)); - if ( relative_rank(weakSide, pawnSq) != RANK_7 - || distance(loserKSq, pawnSq) != 1 - || !((FileABB | FileCBB | FileFBB | FileHBB) & pawnSq)) + if ( relative_rank(weakSide, weakPawn) != RANK_7 + || distance(weakKing, weakPawn) != 1 + || ((FileBBB | FileDBB | FileEBB | FileGBB) & weakPawn)) result += QueenValueEg - PawnValueEg; return strongSide == pos.side_to_move() ? result : -result; } -/// KQ vs KR. This is almost identical to KX vs K: We give the attacking +/// KQ vs KR. This is almost identical to KX vs K: we give the attacking /// king a bonus for having the kings close together, and for forcing the /// defending king towards the edge. If we also take care to avoid null move for /// the defending side in the search, this is usually sufficient to win KQ vs KR. @@ -298,28 +278,32 @@ Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, QueenValueMg, 0)); assert(verify_material(pos, weakSide, RookValueMg, 0)); - Square winnerKSq = pos.square(strongSide); - Square loserKSq = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + Square weakKing = pos.square(weakSide); Value result = QueenValueEg - RookValueEg - + PushToEdges[loserKSq] - + PushClose[distance(winnerKSq, loserKSq)]; + + push_to_edge(weakKing) + + push_close(strongKing, weakKing); return strongSide == pos.side_to_move() ? result : -result; } -/// KNN vs KP. Simply push the opposing king to the corner +/// KNN vs KP. Very drawish, but there are some mate opportunities if we can +/// press the weakSide King to a corner before the pawn advances too much. template<> Value Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, 2 * KnightValueMg, 0)); assert(verify_material(pos, weakSide, VALUE_ZERO, 1)); - Value result = 2 * KnightValueEg - - PawnValueEg - + PushToEdges[pos.square(weakSide)]; + Square weakKing = pos.square(weakSide); + Square weakPawn = pos.square(weakSide); + + Value result = PawnValueEg + + 2 * push_to_edge(weakKing) + - 10 * relative_rank(weakSide, weakPawn); return strongSide == pos.side_to_move() ? result : -result; } @@ -342,51 +326,47 @@ ScaleFactor Endgame::operator()(const Position& pos) const { // No assertions about the material of weakSide, because we want draws to // be detected even when the weaker side has some pawns. - Bitboard pawns = pos.pieces(strongSide, PAWN); - File pawnsFile = file_of(lsb(pawns)); + Bitboard strongPawns = pos.pieces(strongSide, PAWN); + Bitboard allPawns = pos.pieces(PAWN); - // All pawns are on a single rook file? - if ( (pawnsFile == FILE_A || pawnsFile == FILE_H) - && !(pawns & ~file_bb(pawnsFile))) + Square strongBishop = pos.square(strongSide); + Square weakKing = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + + // All strongSide pawns are on a single rook file? + if (!(strongPawns & ~FileABB) || !(strongPawns & ~FileHBB)) { - Square bishopSq = pos.square(strongSide); - Square queeningSq = relative_square(strongSide, make_square(pawnsFile, RANK_8)); - Square kingSq = pos.square(weakSide); + Square queeningSquare = relative_square(strongSide, make_square(file_of(lsb(strongPawns)), RANK_8)); - if ( opposite_colors(queeningSq, bishopSq) - && distance(queeningSq, kingSq) <= 1) + if ( opposite_colors(queeningSquare, strongBishop) + && distance(queeningSquare, weakKing) <= 1) return SCALE_FACTOR_DRAW; } // If all the pawns are on the same B or G file, then it's potentially a draw - if ( (pawnsFile == FILE_B || pawnsFile == FILE_G) - && !(pos.pieces(PAWN) & ~file_bb(pawnsFile)) + if ((!(allPawns & ~FileBBB) || !(allPawns & ~FileGBB)) && pos.non_pawn_material(weakSide) == 0 && pos.count(weakSide) >= 1) { - // Get weakSide pawn that is closest to the home rank - Square weakPawnSq = frontmost_sq(strongSide, pos.pieces(weakSide, PAWN)); - - Square strongKingSq = pos.square(strongSide); - Square weakKingSq = pos.square(weakSide); - Square bishopSq = pos.square(strongSide); + // Get the least advanced weakSide pawn + Square weakPawn = frontmost_sq(strongSide, pos.pieces(weakSide, PAWN)); // There's potential for a draw if our pawn is blocked on the 7th rank, - // the bishop cannot attack it or they only have one pawn left - if ( relative_rank(strongSide, weakPawnSq) == RANK_7 - && (pos.pieces(strongSide, PAWN) & (weakPawnSq + pawn_push(weakSide))) - && (opposite_colors(bishopSq, weakPawnSq) || pos.count(strongSide) == 1)) + // the bishop cannot attack it or they only have one pawn left. + if ( relative_rank(strongSide, weakPawn) == RANK_7 + && (strongPawns & (weakPawn + pawn_push(weakSide))) + && (opposite_colors(strongBishop, weakPawn) || !more_than_one(strongPawns))) { - int strongKingDist = distance(weakPawnSq, strongKingSq); - int weakKingDist = distance(weakPawnSq, weakKingSq); + int strongKingDist = distance(weakPawn, strongKing); + int weakKingDist = distance(weakPawn, weakKing); // It's a draw if the weak king is on its back two ranks, within 2 // squares of the blocking pawn and the strong king is not // closer. (I think this rule only fails in practically // unreachable positions such as 5k1K/6p1/6P1/8/8/3B4/8/8 w // and positions where qsearch will immediately correct the - // problem such as 8/4k1p1/6P1/1K6/3B4/8/8/8 w) - if ( relative_rank(strongSide, weakKingSq) >= RANK_7 + // problem such as 8/4k1p1/6P1/1K6/3B4/8/8/8 w). + if ( relative_rank(strongSide, weakKing) >= RANK_7 && weakKingDist <= 2 && weakKingDist <= strongKingDist) return SCALE_FACTOR_DRAW; @@ -406,15 +386,16 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(pos.count(weakSide) == 1); assert(pos.count(weakSide) >= 1); - Square kingSq = pos.square(weakSide); - Square rsq = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + Square weakKing = pos.square(weakSide); + Square weakRook = pos.square(weakSide); - if ( relative_rank(weakSide, kingSq) <= RANK_2 - && relative_rank(weakSide, pos.square(strongSide)) >= RANK_4 - && relative_rank(weakSide, rsq) == RANK_3 + if ( relative_rank(weakSide, weakKing) <= RANK_2 + && relative_rank(weakSide, strongKing) >= RANK_4 + && relative_rank(weakSide, weakRook) == RANK_3 && ( pos.pieces(weakSide, PAWN) - & pos.attacks_from(kingSq) - & pos.attacks_from(rsq, strongSide))) + & attacks_bb(weakKing) + & pawn_attacks_bb(strongSide, weakRook))) return SCALE_FACTOR_DRAW; return SCALE_FACTOR_NONE; @@ -434,89 +415,89 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, weakSide, RookValueMg, 0)); // Assume strongSide is white and the pawn is on files A-D - Square wksq = normalize(pos, strongSide, pos.square(strongSide)); - Square bksq = normalize(pos, strongSide, pos.square(weakSide)); - Square wrsq = normalize(pos, strongSide, pos.square(strongSide)); - Square wpsq = normalize(pos, strongSide, pos.square(strongSide)); - Square brsq = normalize(pos, strongSide, pos.square(weakSide)); + Square strongKing = normalize(pos, strongSide, pos.square(strongSide)); + Square strongRook = normalize(pos, strongSide, pos.square(strongSide)); + Square strongPawn = normalize(pos, strongSide, pos.square(strongSide)); + Square weakKing = normalize(pos, strongSide, pos.square(weakSide)); + Square weakRook = normalize(pos, strongSide, pos.square(weakSide)); - File f = file_of(wpsq); - Rank r = rank_of(wpsq); - Square queeningSq = make_square(f, RANK_8); + File pawnFile = file_of(strongPawn); + Rank pawnRank = rank_of(strongPawn); + Square queeningSquare = make_square(pawnFile, RANK_8); int tempo = (pos.side_to_move() == strongSide); // If the pawn is not too far advanced and the defending king defends the // queening square, use the third-rank defence. - if ( r <= RANK_5 - && distance(bksq, queeningSq) <= 1 - && wksq <= SQ_H5 - && (rank_of(brsq) == RANK_6 || (r <= RANK_3 && rank_of(wrsq) != RANK_6))) + if ( pawnRank <= RANK_5 + && distance(weakKing, queeningSquare) <= 1 + && strongKing <= SQ_H5 + && (rank_of(weakRook) == RANK_6 || (pawnRank <= RANK_3 && rank_of(strongRook) != RANK_6))) return SCALE_FACTOR_DRAW; // The defending side saves a draw by checking from behind in case the pawn // has advanced to the 6th rank with the king behind. - if ( r == RANK_6 - && distance(bksq, queeningSq) <= 1 - && rank_of(wksq) + tempo <= RANK_6 - && (rank_of(brsq) == RANK_1 || (!tempo && distance(brsq, wpsq) >= 3))) + if ( pawnRank == RANK_6 + && distance(weakKing, queeningSquare) <= 1 + && rank_of(strongKing) + tempo <= RANK_6 + && (rank_of(weakRook) == RANK_1 || (!tempo && distance(weakRook, strongPawn) >= 3))) return SCALE_FACTOR_DRAW; - if ( r >= RANK_6 - && bksq == queeningSq - && rank_of(brsq) == RANK_1 - && (!tempo || distance(wksq, wpsq) >= 2)) + if ( pawnRank >= RANK_6 + && weakKing == queeningSquare + && rank_of(weakRook) == RANK_1 + && (!tempo || distance(strongKing, strongPawn) >= 2)) return SCALE_FACTOR_DRAW; // White pawn on a7 and rook on a8 is a draw if black's king is on g7 or h7 // and the black rook is behind the pawn. - if ( wpsq == SQ_A7 - && wrsq == SQ_A8 - && (bksq == SQ_H7 || bksq == SQ_G7) - && file_of(brsq) == FILE_A - && (rank_of(brsq) <= RANK_3 || file_of(wksq) >= FILE_D || rank_of(wksq) <= RANK_5)) + if ( strongPawn == SQ_A7 + && strongRook == SQ_A8 + && (weakKing == SQ_H7 || weakKing == SQ_G7) + && file_of(weakRook) == FILE_A + && (rank_of(weakRook) <= RANK_3 || file_of(strongKing) >= FILE_D || rank_of(strongKing) <= RANK_5)) return SCALE_FACTOR_DRAW; // If the defending king blocks the pawn and the attacking king is too far // away, it's a draw. - if ( r <= RANK_5 - && bksq == wpsq + NORTH - && distance(wksq, wpsq) - tempo >= 2 - && distance(wksq, brsq) - tempo >= 2) + if ( pawnRank <= RANK_5 + && weakKing == strongPawn + NORTH + && distance(strongKing, strongPawn) - tempo >= 2 + && distance(strongKing, weakRook) - tempo >= 2) return SCALE_FACTOR_DRAW; // Pawn on the 7th rank supported by the rook from behind usually wins if the // attacking king is closer to the queening square than the defending king, // and the defending king cannot gain tempi by threatening the attacking rook. - if ( r == RANK_7 - && f != FILE_A - && file_of(wrsq) == f - && wrsq != queeningSq - && (distance(wksq, queeningSq) < distance(bksq, queeningSq) - 2 + tempo) - && (distance(wksq, queeningSq) < distance(bksq, wrsq) + tempo)) - return ScaleFactor(SCALE_FACTOR_MAX - 2 * distance(wksq, queeningSq)); + if ( pawnRank == RANK_7 + && pawnFile != FILE_A + && file_of(strongRook) == pawnFile + && strongRook != queeningSquare + && (distance(strongKing, queeningSquare) < distance(weakKing, queeningSquare) - 2 + tempo) + && (distance(strongKing, queeningSquare) < distance(weakKing, strongRook) + tempo)) + return ScaleFactor(SCALE_FACTOR_MAX - 2 * distance(strongKing, queeningSquare)); // Similar to the above, but with the pawn further back - if ( f != FILE_A - && file_of(wrsq) == f - && wrsq < wpsq - && (distance(wksq, queeningSq) < distance(bksq, queeningSq) - 2 + tempo) - && (distance(wksq, wpsq + NORTH) < distance(bksq, wpsq + NORTH) - 2 + tempo) - && ( distance(bksq, wrsq) + tempo >= 3 - || ( distance(wksq, queeningSq) < distance(bksq, wrsq) + tempo - && (distance(wksq, wpsq + NORTH) < distance(bksq, wrsq) + tempo)))) + if ( pawnFile != FILE_A + && file_of(strongRook) == pawnFile + && strongRook < strongPawn + && (distance(strongKing, queeningSquare) < distance(weakKing, queeningSquare) - 2 + tempo) + && (distance(strongKing, strongPawn + NORTH) < distance(weakKing, strongPawn + NORTH) - 2 + tempo) + && ( distance(weakKing, strongRook) + tempo >= 3 + || ( distance(strongKing, queeningSquare) < distance(weakKing, strongRook) + tempo + && (distance(strongKing, strongPawn + NORTH) < distance(weakKing, strongPawn) + tempo)))) return ScaleFactor( SCALE_FACTOR_MAX - - 8 * distance(wpsq, queeningSq) - - 2 * distance(wksq, queeningSq)); + - 8 * distance(strongPawn, queeningSquare) + - 2 * distance(strongKing, queeningSquare)); // If the pawn is not far advanced and the defending king is somewhere in // the pawn's path, it's probably a draw. - if (r <= RANK_4 && bksq > wpsq) + if (pawnRank <= RANK_4 && weakKing > strongPawn) { - if (file_of(bksq) == file_of(wpsq)) + if (file_of(weakKing) == file_of(strongPawn)) return ScaleFactor(10); - if ( distance(bksq, wpsq) == 1 - && distance(wksq, bksq) > 2) - return ScaleFactor(24 - 2 * distance(wksq, bksq)); + if ( distance(weakKing, strongPawn) == 1 + && distance(strongKing, weakKing) > 2) + return ScaleFactor(24 - 2 * distance(strongKing, weakKing)); } return SCALE_FACTOR_NONE; } @@ -530,10 +511,11 @@ ScaleFactor Endgame::operator()(const Position& pos) const { // Test for a rook pawn if (pos.pieces(PAWN) & (FileABB | FileHBB)) { - Square ksq = pos.square(weakSide); - Square bsq = pos.square(weakSide); - Square psq = pos.square(strongSide); - Rank rk = relative_rank(strongSide, psq); + Square weakKing = pos.square(weakSide); + Square weakBishop = pos.square(weakSide); + Square strongKing = pos.square(strongSide); + Square strongPawn = pos.square(strongSide); + Rank pawnRank = relative_rank(strongSide, strongPawn); Direction push = pawn_push(strongSide); // If the pawn is on the 5th rank and the pawn (currently) is on @@ -541,11 +523,11 @@ ScaleFactor Endgame::operator()(const Position& pos) const { // a fortress. Depending on the king position give a moderate // reduction or a stronger one if the defending king is near the // corner but not trapped there. - if (rk == RANK_5 && !opposite_colors(bsq, psq)) + if (pawnRank == RANK_5 && !opposite_colors(weakBishop, strongPawn)) { - int d = distance(psq + 3 * push, ksq); + int d = distance(strongPawn + 3 * push, weakKing); - if (d <= 2 && !(d == 0 && ksq == pos.square(strongSide) + 2 * push)) + if (d <= 2 && !(d == 0 && weakKing == strongKing + 2 * push)) return ScaleFactor(24); else return ScaleFactor(48); @@ -555,10 +537,10 @@ ScaleFactor Endgame::operator()(const Position& pos) const { // it's drawn if the bishop attacks the square in front of the // pawn from a reasonable distance and the defending king is near // the corner - if ( rk == RANK_6 - && distance(psq + 2 * push, ksq) <= 1 - && (PseudoAttacks[BISHOP][bsq] & (psq + push)) - && distance(bsq, psq) >= 2) + if ( pawnRank == RANK_6 + && distance(strongPawn + 2 * push, weakKing) <= 1 + && (attacks_bb(weakBishop) & (strongPawn + push)) + && distance(weakBishop, strongPawn) >= 2) return ScaleFactor(8); } @@ -573,28 +555,28 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, RookValueMg, 2)); assert(verify_material(pos, weakSide, RookValueMg, 1)); - Square wpsq1 = pos.squares(strongSide)[0]; - Square wpsq2 = pos.squares(strongSide)[1]; - Square bksq = pos.square(weakSide); + Square strongPawn1 = lsb(pos.pieces(strongSide, PAWN)); + Square strongPawn2 = msb(pos.pieces(strongSide, PAWN)); + Square weakKing = pos.square(weakSide); // Does the stronger side have a passed pawn? - if (pos.pawn_passed(strongSide, wpsq1) || pos.pawn_passed(strongSide, wpsq2)) + if (pos.pawn_passed(strongSide, strongPawn1) || pos.pawn_passed(strongSide, strongPawn2)) return SCALE_FACTOR_NONE; - Rank r = std::max(relative_rank(strongSide, wpsq1), relative_rank(strongSide, wpsq2)); + Rank pawnRank = std::max(relative_rank(strongSide, strongPawn1), relative_rank(strongSide, strongPawn2)); - if ( distance(bksq, wpsq1) <= 1 - && distance(bksq, wpsq2) <= 1 - && relative_rank(strongSide, bksq) > r) + if ( distance(weakKing, strongPawn1) <= 1 + && distance(weakKing, strongPawn2) <= 1 + && relative_rank(strongSide, weakKing) > pawnRank) { - assert(r > RANK_1 && r < RANK_7); - return ScaleFactor(KRPPKRPScaleFactors[r]); + assert(pawnRank > RANK_1 && pawnRank < RANK_7); + return ScaleFactor(7 * pawnRank); } return SCALE_FACTOR_NONE; } -/// K and two or more pawns vs K. There is just a single rule here: If all pawns +/// K and two or more pawns vs K. There is just a single rule here: if all pawns /// are on the same rook file and are blocked by the defending king, it's a draw. template<> ScaleFactor Endgame::operator()(const Position& pos) const { @@ -603,14 +585,12 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(pos.count(strongSide) >= 2); assert(verify_material(pos, weakSide, VALUE_ZERO, 0)); - Square ksq = pos.square(weakSide); - Bitboard pawns = pos.pieces(strongSide, PAWN); + Square weakKing = pos.square(weakSide); + Bitboard strongPawns = pos.pieces(strongSide, PAWN); - // If all pawns are ahead of the king, on a single rook file and - // the king is within one file of the pawns, it's a draw. - if ( !(pawns & ~forward_ranks_bb(weakSide, ksq)) - && !((pawns & ~FileABB) && (pawns & ~FileHBB)) - && distance(ksq, lsb(pawns)) <= 1) + // If all pawns are ahead of the king on a single rook file, it's a draw. + if ( !(strongPawns & ~(FileABB | FileHBB)) + && !(strongPawns & ~passed_pawn_span(weakSide, weakKing))) return SCALE_FACTOR_DRAW; return SCALE_FACTOR_NONE; @@ -627,20 +607,19 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, BishopValueMg, 1)); assert(verify_material(pos, weakSide, BishopValueMg, 0)); - Square pawnSq = pos.square(strongSide); - Square strongBishopSq = pos.square(strongSide); - Square weakBishopSq = pos.square(weakSide); - Square weakKingSq = pos.square(weakSide); + Square strongPawn = pos.square(strongSide); + Square strongBishop = pos.square(strongSide); + Square weakBishop = pos.square(weakSide); + Square weakKing = pos.square(weakSide); // Case 1: Defending king blocks the pawn, and cannot be driven away - if ( file_of(weakKingSq) == file_of(pawnSq) - && relative_rank(strongSide, pawnSq) < relative_rank(strongSide, weakKingSq) - && ( opposite_colors(weakKingSq, strongBishopSq) - || relative_rank(strongSide, weakKingSq) <= RANK_6)) + if ( (forward_file_bb(strongSide, strongPawn) & weakKing) + && ( opposite_colors(weakKing, strongBishop) + || relative_rank(strongSide, weakKing) <= RANK_6)) return SCALE_FACTOR_DRAW; // Case 2: Opposite colored bishops - if (opposite_colors(strongBishopSq, weakBishopSq)) + if (opposite_colors(strongBishop, weakBishop)) return SCALE_FACTOR_DRAW; return SCALE_FACTOR_NONE; @@ -654,36 +633,36 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, BishopValueMg, 2)); assert(verify_material(pos, weakSide, BishopValueMg, 0)); - Square wbsq = pos.square(strongSide); - Square bbsq = pos.square(weakSide); + Square strongBishop = pos.square(strongSide); + Square weakBishop = pos.square(weakSide); - if (!opposite_colors(wbsq, bbsq)) + if (!opposite_colors(strongBishop, weakBishop)) return SCALE_FACTOR_NONE; - Square ksq = pos.square(weakSide); - Square psq1 = pos.squares(strongSide)[0]; - Square psq2 = pos.squares(strongSide)[1]; + Square weakKing = pos.square(weakSide); + Square strongPawn1 = lsb(pos.pieces(strongSide, PAWN)); + Square strongPawn2 = msb(pos.pieces(strongSide, PAWN)); Square blockSq1, blockSq2; - if (relative_rank(strongSide, psq1) > relative_rank(strongSide, psq2)) + if (relative_rank(strongSide, strongPawn1) > relative_rank(strongSide, strongPawn2)) { - blockSq1 = psq1 + pawn_push(strongSide); - blockSq2 = make_square(file_of(psq2), rank_of(psq1)); + blockSq1 = strongPawn1 + pawn_push(strongSide); + blockSq2 = make_square(file_of(strongPawn2), rank_of(strongPawn1)); } else { - blockSq1 = psq2 + pawn_push(strongSide); - blockSq2 = make_square(file_of(psq1), rank_of(psq2)); + blockSq1 = strongPawn2 + pawn_push(strongSide); + blockSq2 = make_square(file_of(strongPawn1), rank_of(strongPawn2)); } - switch (distance(psq1, psq2)) + switch (distance(strongPawn1, strongPawn2)) { case 0: // Both pawns are on the same file. It's an easy draw if the defender firmly // controls some square in the frontmost pawn's path. - if ( file_of(ksq) == file_of(blockSq1) - && relative_rank(strongSide, ksq) >= relative_rank(strongSide, blockSq1) - && opposite_colors(ksq, wbsq)) + if ( file_of(weakKing) == file_of(blockSq1) + && relative_rank(strongSide, weakKing) >= relative_rank(strongSide, blockSq1) + && opposite_colors(weakKing, strongBishop)) return SCALE_FACTOR_DRAW; else return SCALE_FACTOR_NONE; @@ -692,17 +671,17 @@ ScaleFactor Endgame::operator()(const Position& pos) const { // Pawns on adjacent files. It's a draw if the defender firmly controls the // square in front of the frontmost pawn's path, and the square diagonally // behind this square on the file of the other pawn. - if ( ksq == blockSq1 - && opposite_colors(ksq, wbsq) - && ( bbsq == blockSq2 - || (pos.attacks_from(blockSq2) & pos.pieces(weakSide, BISHOP)) - || distance(psq1, psq2) >= 2)) + if ( weakKing == blockSq1 + && opposite_colors(weakKing, strongBishop) + && ( weakBishop == blockSq2 + || (attacks_bb(blockSq2, pos.pieces()) & pos.pieces(weakSide, BISHOP)) + || distance(strongPawn1, strongPawn2) >= 2)) return SCALE_FACTOR_DRAW; - else if ( ksq == blockSq2 - && opposite_colors(ksq, wbsq) - && ( bbsq == blockSq1 - || (pos.attacks_from(blockSq1) & pos.pieces(weakSide, BISHOP)))) + else if ( weakKing == blockSq2 + && opposite_colors(weakKing, strongBishop) + && ( weakBishop == blockSq1 + || (attacks_bb(blockSq1, pos.pieces()) & pos.pieces(weakSide, BISHOP)))) return SCALE_FACTOR_DRAW; else return SCALE_FACTOR_NONE; @@ -714,7 +693,7 @@ ScaleFactor Endgame::operator()(const Position& pos) const { } -/// KBP vs KN. There is a single rule: If the defending king is somewhere along +/// KBP vs KN. There is a single rule: if the defending king is somewhere along /// the path of the pawn, and the square of the king is not of the same color as /// the stronger side's bishop, it's a draw. template<> @@ -723,62 +702,22 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, strongSide, BishopValueMg, 1)); assert(verify_material(pos, weakSide, KnightValueMg, 0)); - Square pawnSq = pos.square(strongSide); - Square strongBishopSq = pos.square(strongSide); - Square weakKingSq = pos.square(weakSide); + Square strongPawn = pos.square(strongSide); + Square strongBishop = pos.square(strongSide); + Square weakKing = pos.square(weakSide); - if ( file_of(weakKingSq) == file_of(pawnSq) - && relative_rank(strongSide, pawnSq) < relative_rank(strongSide, weakKingSq) - && ( opposite_colors(weakKingSq, strongBishopSq) - || relative_rank(strongSide, weakKingSq) <= RANK_6)) + if ( file_of(weakKing) == file_of(strongPawn) + && relative_rank(strongSide, strongPawn) < relative_rank(strongSide, weakKing) + && ( opposite_colors(weakKing, strongBishop) + || relative_rank(strongSide, weakKing) <= RANK_6)) return SCALE_FACTOR_DRAW; return SCALE_FACTOR_NONE; } -/// KNP vs K. There is a single rule: if the pawn is a rook pawn on the 7th rank -/// and the defending king prevents the pawn from advancing, the position is drawn. -template<> -ScaleFactor Endgame::operator()(const Position& pos) const { - - assert(verify_material(pos, strongSide, KnightValueMg, 1)); - assert(verify_material(pos, weakSide, VALUE_ZERO, 0)); - - // Assume strongSide is white and the pawn is on files A-D - Square pawnSq = normalize(pos, strongSide, pos.square(strongSide)); - Square weakKingSq = normalize(pos, strongSide, pos.square(weakSide)); - - if (pawnSq == SQ_A7 && distance(SQ_A8, weakKingSq) <= 1) - return SCALE_FACTOR_DRAW; - - return SCALE_FACTOR_NONE; -} - - -/// KNP vs KB. If knight can block bishop from taking pawn, it's a win. -/// Otherwise the position is drawn. -template<> -ScaleFactor Endgame::operator()(const Position& pos) const { - - assert(verify_material(pos, strongSide, KnightValueMg, 1)); - assert(verify_material(pos, weakSide, BishopValueMg, 0)); - - Square pawnSq = pos.square(strongSide); - Square bishopSq = pos.square(weakSide); - Square weakKingSq = pos.square(weakSide); - - // King needs to get close to promoting pawn to prevent knight from blocking. - // Rules for this are very tricky, so just approximate. - if (forward_file_bb(strongSide, pawnSq) & pos.attacks_from(bishopSq)) - return ScaleFactor(distance(weakKingSq, pawnSq)); - - return SCALE_FACTOR_NONE; -} - - /// KP vs KP. This is done by removing the weakest side's pawn and probing the -/// KP vs K bitbase: If the weakest side has a draw without the pawn, it probably +/// KP vs K bitbase: if the weakest side has a draw without the pawn, it probably /// has at least a draw with the pawn as well. The exception is when the stronger /// side's pawn is far advanced and not on a rook file; in this case it is often /// possible to win (e.g. 8/4k3/3p4/3P4/6K1/8/8/8 w - - 0 1). @@ -789,18 +728,20 @@ ScaleFactor Endgame::operator()(const Position& pos) const { assert(verify_material(pos, weakSide, VALUE_ZERO, 1)); // Assume strongSide is white and the pawn is on files A-D - Square wksq = normalize(pos, strongSide, pos.square(strongSide)); - Square bksq = normalize(pos, strongSide, pos.square(weakSide)); - Square psq = normalize(pos, strongSide, pos.square(strongSide)); + Square strongKing = normalize(pos, strongSide, pos.square(strongSide)); + Square weakKing = normalize(pos, strongSide, pos.square(weakSide)); + Square strongPawn = normalize(pos, strongSide, pos.square(strongSide)); Color us = strongSide == pos.side_to_move() ? WHITE : BLACK; // If the pawn has advanced to the fifth rank or further, and is not a // rook pawn, it's too dangerous to assume that it's at least a draw. - if (rank_of(psq) >= RANK_5 && file_of(psq) != FILE_A) + if (rank_of(strongPawn) >= RANK_5 && file_of(strongPawn) != FILE_A) return SCALE_FACTOR_NONE; // Probe the KPK bitbase with the weakest side's pawn removed. If it's a draw, // it's probably at least a draw even with the pawn. - return Bitbases::probe(wksq, psq, bksq, us) ? SCALE_FACTOR_NONE : SCALE_FACTOR_DRAW; + return Bitbases::probe(strongKing, strongPawn, weakKing, us) ? SCALE_FACTOR_NONE : SCALE_FACTOR_DRAW; } + +} // namespace Stockfish diff --git a/src/endgame.h b/src/endgame.h index e29f8777..146111b9 100644 --- a/src/endgame.h +++ b/src/endgame.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -21,15 +19,16 @@ #ifndef ENDGAME_H_INCLUDED #define ENDGAME_H_INCLUDED -#include #include #include #include +#include #include #include "position.h" #include "types.h" +namespace Stockfish { /// EndgameCode lists all supported endgame functions by corresponding codes @@ -57,8 +56,6 @@ enum EndgameCode { KBPKB, // KBP vs KB KBPPKB, // KBPP vs KB KBPKN, // KBP vs KN - KNPK, // KNP vs K - KNPKB, // KNP vs KB KPKP // KP vs KP }; @@ -124,4 +121,6 @@ namespace Endgames { } } +} // namespace Stockfish + #endif // #ifndef ENDGAME_H_INCLUDED diff --git a/src/evaluate.cpp b/src/evaluate.cpp index a0ad09f0..a89ea533 100644 --- a/src/evaluate.cpp +++ b/src/evaluate.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -20,22 +18,144 @@ #include #include +#include #include // For std::memset +#include #include #include +#include +#include +#include #include "bitboard.h" +#include "cluster.h" #include "evaluate.h" #include "material.h" +#include "misc.h" #include "pawns.h" #include "thread.h" +#include "uci.h" +#include "incbin/incbin.h" + + +// Macro to embed the default efficiently updatable neural network (NNUE) file +// data in the engine binary (using incbin.h, by Dale Weiler). +// This macro invocation will declare the following three variables +// const unsigned char gEmbeddedNNUEData[]; // a pointer to the embedded data +// const unsigned char *const gEmbeddedNNUEEnd; // a marker to the end +// const unsigned int gEmbeddedNNUESize; // the size of the embedded file +// Note that this does not work in Microsoft Visual Studio. +#if !defined(_MSC_VER) && !defined(NNUE_EMBEDDING_OFF) + INCBIN(EmbeddedNNUE, EvalFileDefaultName); +#else + const unsigned char gEmbeddedNNUEData[1] = {0x0}; + const unsigned char *const gEmbeddedNNUEEnd = &gEmbeddedNNUEData[1]; + const unsigned int gEmbeddedNNUESize = 1; +#endif + + +using namespace std; +using namespace Stockfish::Eval::NNUE; + +namespace Stockfish { + +namespace Eval { + + bool useNNUE; + string eval_file_loaded = "None"; + + /// NNUE::init() tries to load a NNUE network at startup time, or when the engine + /// receives a UCI command "setoption name EvalFile value nn-[a-z0-9]{12}.nnue" + /// The name of the NNUE network is always retrieved from the EvalFile option. + /// We search the given network in three locations: internally (the default + /// network may be embedded in the binary), in the active working directory and + /// in the engine directory. Distro packagers may define the DEFAULT_NNUE_DIRECTORY + /// variable to have the engine search in a special directory in their distro. + + void NNUE::init() { + + useNNUE = Options["Use NNUE"]; + if (!useNNUE) + return; + + string eval_file = string(Options["EvalFile"]); + + #if defined(DEFAULT_NNUE_DIRECTORY) + #define stringify2(x) #x + #define stringify(x) stringify2(x) + vector dirs = { "" , "" , CommandLine::binaryDirectory , stringify(DEFAULT_NNUE_DIRECTORY) }; + #else + vector dirs = { "" , "" , CommandLine::binaryDirectory }; + #endif + + for (string directory : dirs) + if (eval_file_loaded != eval_file) + { + if (directory != "") + { + ifstream stream(directory + eval_file, ios::binary); + if (load_eval(eval_file, stream)) + eval_file_loaded = eval_file; + } + + if (directory == "" && eval_file == EvalFileDefaultName) + { + // C++ way to prepare a buffer for a memory stream + class MemoryBuffer : public basic_streambuf { + public: MemoryBuffer(char* p, size_t n) { setg(p, p, p + n); setp(p, p + n); } + }; + + MemoryBuffer buffer(const_cast(reinterpret_cast(gEmbeddedNNUEData)), + size_t(gEmbeddedNNUESize)); + + istream stream(&buffer); + if (load_eval(eval_file, stream)) + eval_file_loaded = eval_file; + } + } + } + + /// NNUE::verify() verifies that the last net used was loaded successfully + void NNUE::verify() { + + string eval_file = string(Options["EvalFile"]); + + if (useNNUE && eval_file_loaded != eval_file) + { + UCI::OptionsMap defaults; + UCI::init(defaults); + + string msg1 = "If the UCI option \"Use NNUE\" is set to true, network evaluation parameters compatible with the engine must be available."; + string msg2 = "The option is set to true, but the network file " + eval_file + " was not loaded successfully."; + string msg3 = "The UCI option EvalFile might need to specify the full path, including the directory name, to the network file."; + string msg4 = "The default net can be downloaded from: https://tests.stockfishchess.org/api/nn/" + string(defaults["EvalFile"]); + string msg5 = "The engine will be terminated now."; + + sync_cout << "info string ERROR: " << msg1 << sync_endl; + sync_cout << "info string ERROR: " << msg2 << sync_endl; + sync_cout << "info string ERROR: " << msg3 << sync_endl; + sync_cout << "info string ERROR: " << msg4 << sync_endl; + sync_cout << "info string ERROR: " << msg5 << sync_endl; + + exit(EXIT_FAILURE); + } + + if (Cluster::is_root()) + { + if (useNNUE) + sync_cout << "info string NNUE evaluation using " << eval_file << " enabled" << sync_endl; + else + sync_cout << "info string classical evaluation enabled" << sync_endl; + } + } +} namespace Trace { enum Tracing { NO_TRACE, TRACE }; enum Term { // The first 8 entries are reserved for PieceType - MATERIAL = 8, IMBALANCE, MOBILITY, THREAT, PASSED, SPACE, INITIATIVE, TOTAL, TERM_NB + MATERIAL = 8, IMBALANCE, MOBILITY, THREAT, PASSED, SPACE, WINNABLE, TOTAL, TERM_NB }; Score scores[TERM_NB][COLOR_NB]; @@ -59,7 +179,7 @@ namespace Trace { std::ostream& operator<<(std::ostream& os, Term t) { - if (t == MATERIAL || t == IMBALANCE || t == INITIATIVE || t == TOTAL) + if (t == MATERIAL || t == IMBALANCE || t == WINNABLE || t == TOTAL) os << " ---- ----" << " | " << " ---- ----"; else os << scores[t][WHITE] << " | " << scores[t][BLACK]; @@ -74,78 +194,96 @@ using namespace Trace; namespace { // Threshold for lazy and space evaluation - constexpr Value LazyThreshold = Value(1400); - constexpr Value SpaceThreshold = Value(12222); + constexpr Value LazyThreshold1 = Value(1565); + constexpr Value LazyThreshold2 = Value(1102); + constexpr Value SpaceThreshold = Value(11551); + constexpr Value NNUEThreshold1 = Value(682); + constexpr Value NNUEThreshold2 = Value(176); // KingAttackWeights[PieceType] contains king attack weights by piece type constexpr int KingAttackWeights[PIECE_TYPE_NB] = { 0, 0, 81, 52, 44, 10 }; - // Penalties for enemy's safe checks - constexpr int QueenSafeCheck = 780; - constexpr int RookSafeCheck = 1080; - constexpr int BishopSafeCheck = 635; - constexpr int KnightSafeCheck = 790; + // SafeCheck[PieceType][single/multiple] contains safe check bonus by piece type, + // higher if multiple safe checks are possible for that piece type. + constexpr int SafeCheck[][2] = { + {}, {}, {803, 1292}, {639, 974}, {1087, 1878}, {759, 1132} + }; #define S(mg, eg) make_score(mg, eg) // MobilityBonus[PieceType-2][attacked] contains bonuses for middle and end game, // indexed by piece type and number of attacked squares in the mobility area. constexpr Score MobilityBonus[][32] = { - { S(-62,-81), S(-53,-56), S(-12,-30), S( -4,-14), S( 3, 8), S( 13, 15), // Knights - S( 22, 23), S( 28, 27), S( 33, 33) }, - { S(-48,-59), S(-20,-23), S( 16, -3), S( 26, 13), S( 38, 24), S( 51, 42), // Bishops - S( 55, 54), S( 63, 57), S( 63, 65), S( 68, 73), S( 81, 78), S( 81, 86), - S( 91, 88), S( 98, 97) }, - { S(-58,-76), S(-27,-18), S(-15, 28), S(-10, 55), S( -5, 69), S( -2, 82), // Rooks - S( 9,112), S( 16,118), S( 30,132), S( 29,142), S( 32,155), S( 38,165), - S( 46,166), S( 48,169), S( 58,171) }, - { S(-39,-36), S(-21,-15), S( 3, 8), S( 3, 18), S( 14, 34), S( 22, 54), // Queens - S( 28, 61), S( 41, 73), S( 43, 79), S( 48, 92), S( 56, 94), S( 60,104), - S( 60,113), S( 66,120), S( 67,123), S( 70,126), S( 71,133), S( 73,136), - S( 79,140), S( 88,143), S( 88,148), S( 99,166), S(102,170), S(102,175), - S(106,184), S(109,191), S(113,206), S(116,212) } + { S(-62,-79), S(-53,-57), S(-12,-31), S( -3,-17), S( 3, 7), S( 12, 13), // Knight + S( 21, 16), S( 28, 21), S( 37, 26) }, + { S(-47,-59), S(-20,-25), S( 14, -8), S( 29, 12), S( 39, 21), S( 53, 40), // Bishop + S( 53, 56), S( 60, 58), S( 62, 65), S( 69, 72), S( 78, 78), S( 83, 87), + S( 91, 88), S( 96, 98) }, + { S(-60,-82), S(-24,-15), S( 0, 17) ,S( 3, 43), S( 4, 72), S( 14,100), // Rook + S( 20,102), S( 30,122), S( 41,133), S(41 ,139), S( 41,153), S( 45,160), + S( 57,165), S( 58,170), S( 67,175) }, + { S(-29,-49), S(-16,-29), S( -8, -8), S( -8, 17), S( 18, 39), S( 25, 54), // Queen + S( 23, 59), S( 37, 73), S( 41, 76), S( 54, 95), S( 65, 95) ,S( 68,101), + S( 69,124), S( 70,128), S( 70,132), S( 70,133) ,S( 71,136), S( 72,140), + S( 74,147), S( 76,149), S( 90,153), S(104,169), S(105,171), S(106,171), + S(112,178), S(114,185), S(114,187), S(119,221) } }; - // RookOnFile[semiopen/open] contains bonuses for each rook when there is - // no (friendly) pawn on the rook file. - constexpr Score RookOnFile[] = { S(21, 4), S(47, 25) }; + // BishopPawns[distance from edge] contains a file-dependent penalty for pawns on + // squares of the same color as our bishop. + constexpr Score BishopPawns[int(FILE_NB) / 2] = { + S(3, 8), S(3, 9), S(2, 8), S(3, 8) + }; + + // KingProtector[knight/bishop] contains penalty for each distance unit to own king + constexpr Score KingProtector[] = { S(8, 9), S(6, 9) }; + + // Outpost[knight/bishop] contains bonuses for each knight or bishop occupying a + // pawn protected square on rank 4 to 6 which is also safe from a pawn attack. + constexpr Score Outpost[] = { S(57, 38), S(31, 24) }; + + // PassedRank[Rank] contains a bonus according to the rank of a passed pawn + constexpr Score PassedRank[RANK_NB] = { + S(0, 0), S(7, 27), S(16, 32), S(17, 40), S(64, 71), S(170, 174), S(278, 262) + }; + + constexpr Score RookOnClosedFile = S(10, 5); + constexpr Score RookOnOpenFile[] = { S(19, 6), S(47, 26) }; // ThreatByMinor/ByRook[attacked PieceType] contains bonuses according to // which piece type attacks which one. Attacks on lesser pieces which are // pawn-defended are not considered. constexpr Score ThreatByMinor[PIECE_TYPE_NB] = { - S(0, 0), S(6, 32), S(59, 41), S(79, 56), S(90, 119), S(79, 161) + S(0, 0), S(5, 32), S(55, 41), S(77, 56), S(89, 119), S(79, 162) }; constexpr Score ThreatByRook[PIECE_TYPE_NB] = { - S(0, 0), S(3, 44), S(38, 71), S(38, 61), S(0, 38), S(51, 38) - }; - - // PassedRank[Rank] contains a bonus according to the rank of a passed pawn - constexpr Score PassedRank[RANK_NB] = { - S(0, 0), S(10, 28), S(17, 33), S(15, 41), S(62, 72), S(168, 177), S(276, 260) + S(0, 0), S(3, 44), S(37, 68), S(42, 60), S(0, 39), S(58, 43) }; // Assorted bonuses and penalties - constexpr Score BishopPawns = S( 3, 7); - constexpr Score CorneredBishop = S( 50, 50); - constexpr Score FlankAttacks = S( 8, 0); - constexpr Score Hanging = S( 69, 36); - constexpr Score KingProtector = S( 7, 8); - constexpr Score KnightOnQueen = S( 16, 12); - constexpr Score LongDiagonalBishop = S( 45, 0); - constexpr Score MinorBehindPawn = S( 18, 3); - constexpr Score Outpost = S( 32, 10); - constexpr Score PassedFile = S( 11, 8); - constexpr Score PawnlessFlank = S( 17, 95); - constexpr Score RestrictedPiece = S( 7, 7); - constexpr Score RookOnQueenFile = S( 7, 6); - constexpr Score SliderOnQueen = S( 59, 18); - constexpr Score ThreatByKing = S( 24, 89); - constexpr Score ThreatByPawnPush = S( 48, 39); - constexpr Score ThreatBySafePawn = S(173, 94); - constexpr Score TrappedRook = S( 47, 4); - constexpr Score WeakQueen = S( 49, 15); + constexpr Score UncontestedOutpost = S( 1, 10); + constexpr Score BishopOnKingRing = S( 24, 0); + constexpr Score BishopXRayPawns = S( 4, 5); + constexpr Score CorneredBishop = S( 50, 50); + constexpr Score FlankAttacks = S( 8, 0); + constexpr Score Hanging = S( 69, 36); + constexpr Score KnightOnQueen = S( 16, 11); + constexpr Score LongDiagonalBishop = S( 45, 0); + constexpr Score MinorBehindPawn = S( 18, 3); + constexpr Score PassedFile = S( 11, 8); + constexpr Score PawnlessFlank = S( 17, 95); + constexpr Score ReachableOutpost = S( 31, 22); + constexpr Score RestrictedPiece = S( 7, 7); + constexpr Score RookOnKingRing = S( 16, 0); + constexpr Score SliderOnQueen = S( 60, 18); + constexpr Score ThreatByKing = S( 24, 89); + constexpr Score ThreatByPawnPush = S( 48, 39); + constexpr Score ThreatBySafePawn = S(173, 94); + constexpr Score TrappedRook = S( 55, 13); + constexpr Score WeakQueenProtection = S( 14, 0); + constexpr Score WeakQueen = S( 56, 15); + #undef S @@ -166,8 +304,7 @@ namespace { template Score threats() const; template Score passed() const; template Score space() const; - ScaleFactor scale_factor(Value eg) const; - Score initiative(Score score) const; + Value winnable(Score score) const; const Position& pos; Material::Entry* me; @@ -209,12 +346,13 @@ namespace { // Evaluation::initialize() computes king and pawn attacks, and the king ring // bitboard for a given color. This is done at the beginning of the evaluation. + template template void Evaluation::initialize() { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Up = (Us == WHITE ? NORTH : SOUTH); - constexpr Direction Down = (Us == WHITE ? SOUTH : NORTH); + constexpr Color Them = ~Us; + constexpr Direction Up = pawn_push(Us); + constexpr Direction Down = -Up; constexpr Bitboard LowRanks = (Us == WHITE ? Rank2BB | Rank3BB : Rank7BB | Rank6BB); const Square ksq = pos.square(Us); @@ -224,26 +362,20 @@ namespace { // Find our pawns that are blocked or on the first two ranks Bitboard b = pos.pieces(Us, PAWN) & (shift(pos.pieces()) | LowRanks); - // Squares occupied by those pawns, by our king or queen or controlled by - // enemy pawns are excluded from the mobility area. - mobilityArea[Us] = ~(b | pos.pieces(Us, KING, QUEEN) | pe->pawn_attacks(Them)); + // Squares occupied by those pawns, by our king or queen, by blockers to attacks on our king + // or controlled by enemy pawns are excluded from the mobility area. + mobilityArea[Us] = ~(b | pos.pieces(Us, KING, QUEEN) | pos.blockers_for_king(Us) | pe->pawn_attacks(Them)); // Initialize attackedBy[] for king and pawns - attackedBy[Us][KING] = pos.attacks_from(ksq); + attackedBy[Us][KING] = attacks_bb(ksq); attackedBy[Us][PAWN] = pe->pawn_attacks(Us); attackedBy[Us][ALL_PIECES] = attackedBy[Us][KING] | attackedBy[Us][PAWN]; attackedBy2[Us] = dblAttackByPawn | (attackedBy[Us][KING] & attackedBy[Us][PAWN]); // Init our king safety tables - kingRing[Us] = attackedBy[Us][KING]; - if (relative_rank(Us, ksq) == RANK_1) - kingRing[Us] |= shift(kingRing[Us]); - - if (file_of(ksq) == FILE_H) - kingRing[Us] |= shift(kingRing[Us]); - - else if (file_of(ksq) == FILE_A) - kingRing[Us] |= shift(kingRing[Us]); + Square s = make_square(std::clamp(file_of(ksq), FILE_B, FILE_G), + std::clamp(rank_of(ksq), RANK_2, RANK_7)); + kingRing[Us] = attacks_bb(s) | s; kingAttackersCount[Them] = popcount(kingRing[Us] & pe->pawn_attacks(Them)); kingAttacksCount[Them] = kingAttackersWeight[Them] = 0; @@ -254,29 +386,30 @@ namespace { // Evaluation::pieces() scores pieces of a given color and type + template template Score Evaluation::pieces() { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Down = (Us == WHITE ? SOUTH : NORTH); + constexpr Color Them = ~Us; + constexpr Direction Down = -pawn_push(Us); constexpr Bitboard OutpostRanks = (Us == WHITE ? Rank4BB | Rank5BB | Rank6BB : Rank5BB | Rank4BB | Rank3BB); - const Square* pl = pos.squares(Us); - + Bitboard b1 = pos.pieces(Us, Pt); Bitboard b, bb; Score score = SCORE_ZERO; attackedBy[Us][Pt] = 0; - for (Square s = *pl; s != SQ_NONE; s = *++pl) - { + while (b1) { + Square s = pop_lsb(&b1); + // Find attacked squares, including x-ray attacks for bishops and rooks b = Pt == BISHOP ? attacks_bb(s, pos.pieces() ^ pos.pieces(QUEEN)) : Pt == ROOK ? attacks_bb< ROOK>(s, pos.pieces() ^ pos.pieces(QUEEN) ^ pos.pieces(Us, ROOK)) - : pos.attacks_from(s); + : attacks_bb(s, pos.pieces()); if (pos.blockers_for_king(Us) & s) - b &= LineBB[pos.square(Us)][s]; + b &= line_bb(pos.square(Us), s); attackedBy2[Us] |= attackedBy[Us][ALL_PIECES] & b; attackedBy[Us][Pt] |= b; @@ -289,76 +422,100 @@ namespace { kingAttacksCount[Us] += popcount(b & attackedBy[Them][KING]); } - int mob = popcount(b & mobilityArea[Us]); + else if (Pt == ROOK && (file_bb(s) & kingRing[Them])) + score += RookOnKingRing; + else if (Pt == BISHOP && (attacks_bb(s, pos.pieces(PAWN)) & kingRing[Them])) + score += BishopOnKingRing; + + int mob = popcount(b & mobilityArea[Us]); mobility[Us] += MobilityBonus[Pt - 2][mob]; if (Pt == BISHOP || Pt == KNIGHT) { - // Bonus if piece is on an outpost square or can reach one - bb = OutpostRanks & attackedBy[Us][PAWN] & ~pe->pawn_attacks_span(Them); - if (bb & s) - score += Outpost * (Pt == KNIGHT ? 2 : 1); + // Bonus if the piece is on an outpost square or can reach one + // Bonus for knights (UncontestedOutpost) if few relevant targets + bb = OutpostRanks & (attackedBy[Us][PAWN] | shift(pos.pieces(PAWN))) + & ~pe->pawn_attacks_span(Them); + Bitboard targets = pos.pieces(Them) & ~pos.pieces(PAWN); + if ( Pt == KNIGHT + && bb & s & ~CenterFiles // on a side outpost + && !(b & targets) // no relevant attacks + && (!more_than_one(targets & (s & QueenSide ? QueenSide : KingSide)))) + score += UncontestedOutpost * popcount(pos.pieces(PAWN) & (s & QueenSide ? QueenSide : KingSide)); + else if (bb & s) + score += Outpost[Pt == BISHOP]; else if (Pt == KNIGHT && bb & b & ~pos.pieces(Us)) - score += Outpost; + score += ReachableOutpost; - // Knight and Bishop bonus for being right behind a pawn + // Bonus for a knight or bishop shielded by pawn if (shift(pos.pieces(PAWN)) & s) score += MinorBehindPawn; // Penalty if the piece is far from the king - score -= KingProtector * distance(s, pos.square(Us)); + score -= KingProtector[Pt == BISHOP] * distance(pos.square(Us), s); - if (Pt == BISHOP) + if constexpr (Pt == BISHOP) { - // Penalty according to number of pawns on the same color square as the - // bishop, bigger when the center files are blocked with pawns. + // Penalty according to the number of our pawns on the same color square as the + // bishop, bigger when the center files are blocked with pawns and smaller + // when the bishop is outside the pawn chain. Bitboard blocked = pos.pieces(Us, PAWN) & shift(pos.pieces()); - score -= BishopPawns * pos.pawns_on_same_color_squares(Us, s) - * (1 + popcount(blocked & CenterFiles)); + score -= BishopPawns[edge_distance(file_of(s))] * pos.pawns_on_same_color_squares(Us, s) + * (!(attackedBy[Us][PAWN] & s) + popcount(blocked & CenterFiles)); + + // Penalty for all enemy pawns x-rayed + score -= BishopXRayPawns * popcount(attacks_bb(s) & pos.pieces(Them, PAWN)); // Bonus for bishop on a long diagonal which can "see" both center squares if (more_than_one(attacks_bb(s, pos.pieces(PAWN)) & Center)) score += LongDiagonalBishop; - } - // An important Chess960 pattern: A cornered bishop blocked by a friendly - // pawn diagonally in front of it is a very serious problem, especially - // when that pawn is also blocked. - if ( Pt == BISHOP - && pos.is_chess960() - && (s == relative_square(Us, SQ_A1) || s == relative_square(Us, SQ_H1))) - { - Direction d = pawn_push(Us) + (file_of(s) == FILE_A ? EAST : WEST); - if (pos.piece_on(s + d) == make_piece(Us, PAWN)) - score -= !pos.empty(s + d + pawn_push(Us)) ? CorneredBishop * 4 - : pos.piece_on(s + d + d) == make_piece(Us, PAWN) ? CorneredBishop * 2 - : CorneredBishop; + // An important Chess960 pattern: a cornered bishop blocked by a friendly + // pawn diagonally in front of it is a very serious problem, especially + // when that pawn is also blocked. + if ( pos.is_chess960() + && (s == relative_square(Us, SQ_A1) || s == relative_square(Us, SQ_H1))) + { + Direction d = pawn_push(Us) + (file_of(s) == FILE_A ? EAST : WEST); + if (pos.piece_on(s + d) == make_piece(Us, PAWN)) + score -= !pos.empty(s + d + pawn_push(Us)) ? CorneredBishop * 4 + : pos.piece_on(s + d + d) == make_piece(Us, PAWN) ? CorneredBishop * 2 + : CorneredBishop; + } } } - if (Pt == ROOK) + if constexpr (Pt == ROOK) { - // Bonus for rook on the same file as a queen - if (file_bb(s) & pos.pieces(QUEEN)) - score += RookOnQueenFile; - - // Bonus for rook on an open or semi-open file + // Bonuses for rook on a (semi-)open or closed file if (pos.is_on_semiopen_file(Us, s)) - score += RookOnFile[pos.is_on_semiopen_file(Them, s)]; - - // Penalty when trapped by the king, even more if the king cannot castle - else if (mob <= 3) { - File kf = file_of(pos.square(Us)); - if ((kf < FILE_E) == (file_of(s) < kf)) - score -= TrappedRook * (1 + !pos.castling_rights(Us)); + score += RookOnOpenFile[pos.is_on_semiopen_file(Them, s)]; + } + else + { + // If our pawn on this file is blocked, increase penalty + if ( pos.pieces(Us, PAWN) + & shift(pos.pieces()) + & file_bb(s)) + { + score -= RookOnClosedFile; + } + + // Penalty when trapped by the king, even more if the king cannot castle + if (mob <= 3) + { + File kf = file_of(pos.square(Us)); + if ((kf < FILE_E) == (file_of(s) < kf)) + score -= TrappedRook * (1 + !pos.castling_rights(Us)); + } } } - if (Pt == QUEEN) + if constexpr (Pt == QUEEN) { // Penalty if any relative pin or discovered attack against the queen Bitboard queenPinners; @@ -366,7 +523,7 @@ namespace { score -= WeakQueen; } } - if (T) + if constexpr (T) Trace::add(Pt, Us, score); return score; @@ -374,14 +531,15 @@ namespace { // Evaluation::king() assigns bonuses and penalties to a king of a given color + template template Score Evaluation::king() const { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); + constexpr Color Them = ~Us; constexpr Bitboard Camp = (Us == WHITE ? AllSquares ^ Rank6BB ^ Rank7BB ^ Rank8BB : AllSquares ^ Rank1BB ^ Rank2BB ^ Rank3BB); - Bitboard weak, b1, b2, safe, unsafeChecks = 0; + Bitboard weak, b1, b2, b3, safe, unsafeChecks = 0; Bitboard rookChecks, queenChecks, bishopChecks, knightChecks; int kingDanger = 0; const Square ksq = pos.square(Us); @@ -402,63 +560,57 @@ namespace { b2 = attacks_bb(ksq, pos.pieces() ^ pos.pieces(Us, QUEEN)); // Enemy rooks checks - rookChecks = b1 & safe & attackedBy[Them][ROOK]; - + rookChecks = b1 & attackedBy[Them][ROOK] & safe; if (rookChecks) - kingDanger += RookSafeCheck; + kingDanger += SafeCheck[ROOK][more_than_one(rookChecks)]; else unsafeChecks |= b1 & attackedBy[Them][ROOK]; - // Enemy queen safe checks: we count them only if they are from squares from - // which we can't give a rook check, because rook checks are more valuable. - queenChecks = (b1 | b2) - & attackedBy[Them][QUEEN] - & safe - & ~attackedBy[Us][QUEEN] - & ~rookChecks; - + // Enemy queen safe checks: count them only if the checks are from squares from + // which opponent cannot give a rook check, because rook checks are more valuable. + queenChecks = (b1 | b2) & attackedBy[Them][QUEEN] & safe + & ~(attackedBy[Us][QUEEN] | rookChecks); if (queenChecks) - kingDanger += QueenSafeCheck; + kingDanger += SafeCheck[QUEEN][more_than_one(queenChecks)]; - // Enemy bishops checks: we count them only if they are from squares from - // which we can't give a queen check, because queen checks are more valuable. - bishopChecks = b2 - & attackedBy[Them][BISHOP] - & safe + // Enemy bishops checks: count them only if they are from squares from which + // opponent cannot give a queen check, because queen checks are more valuable. + bishopChecks = b2 & attackedBy[Them][BISHOP] & safe & ~queenChecks; - if (bishopChecks) - kingDanger += BishopSafeCheck; + kingDanger += SafeCheck[BISHOP][more_than_one(bishopChecks)]; + else unsafeChecks |= b2 & attackedBy[Them][BISHOP]; // Enemy knights checks - knightChecks = pos.attacks_from(ksq) & attackedBy[Them][KNIGHT]; - + knightChecks = attacks_bb(ksq) & attackedBy[Them][KNIGHT]; if (knightChecks & safe) - kingDanger += KnightSafeCheck; + kingDanger += SafeCheck[KNIGHT][more_than_one(knightChecks & safe)]; else unsafeChecks |= knightChecks; - // Find the squares that opponent attacks in our king flank, and the squares - // which are attacked twice in that flank. + // Find the squares that opponent attacks in our king flank, the squares + // which they attack twice in that flank, and the squares that we defend. b1 = attackedBy[Them][ALL_PIECES] & KingFlank[file_of(ksq)] & Camp; b2 = b1 & attackedBy2[Them]; + b3 = attackedBy[Us][ALL_PIECES] & KingFlank[file_of(ksq)] & Camp; - int kingFlankAttacks = popcount(b1) + popcount(b2); + int kingFlankAttack = popcount(b1) + popcount(b2); + int kingFlankDefense = popcount(b3); - kingDanger += kingAttackersCount[Them] * kingAttackersWeight[Them] - + 69 * kingAttacksCount[Them] - + 185 * popcount(kingRing[Us] & weak) - - 100 * bool(attackedBy[Us][KNIGHT] & attackedBy[Us][KING]) - - 35 * bool(attackedBy[Us][BISHOP] & attackedBy[Us][KING]) - + 148 * popcount(unsafeChecks) - + 98 * popcount(pos.blockers_for_king(Us)) - - 873 * !pos.count(Them) - - 6 * mg_value(score) / 8 - + mg_value(mobility[Them] - mobility[Us]) - + 3 * kingFlankAttacks * kingFlankAttacks / 8 - - 7; + kingDanger += kingAttackersCount[Them] * kingAttackersWeight[Them] // (~10 Elo) + + 183 * popcount(kingRing[Us] & weak) // (~15 Elo) + + 148 * popcount(unsafeChecks) // (~4 Elo) + + 98 * popcount(pos.blockers_for_king(Us)) // (~2 Elo) + + 69 * kingAttacksCount[Them] // (~0.5 Elo) + + 3 * kingFlankAttack * kingFlankAttack / 8 // (~0.5 Elo) + + mg_value(mobility[Them] - mobility[Us]) // (~0.5 Elo) + - 873 * !pos.count(Them) // (~24 Elo) + - 100 * bool(attackedBy[Us][KNIGHT] & attackedBy[Us][KING]) // (~5 Elo) + - 6 * mg_value(score) / 8 // (~8 Elo) + - 4 * kingFlankDefense // (~5 Elo) + + 37; // (~0.5 Elo) // Transform the kingDanger units into a Score, and subtract it from the evaluation if (kingDanger > 100) @@ -469,9 +621,9 @@ namespace { score -= PawnlessFlank; // Penalty if king flank is under attack, potentially moving toward the king - score -= FlankAttacks * kingFlankAttacks; + score -= FlankAttacks * kingFlankAttack; - if (T) + if constexpr (T) Trace::add(KING, Us, score); return score; @@ -480,11 +632,12 @@ namespace { // Evaluation::threats() assigns bonuses according to the types of the // attacking and the attacked pieces. + template template Score Evaluation::threats() const { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Up = (Us == WHITE ? NORTH : SOUTH); + constexpr Color Them = ~Us; + constexpr Direction Up = pawn_push(Us); constexpr Bitboard TRank3BB = (Us == WHITE ? Rank3BB : Rank6BB); Bitboard b, weak, defended, nonPawnEnemies, stronglyProtected, safe; @@ -521,13 +674,15 @@ namespace { b = ~attackedBy[Them][ALL_PIECES] | (nonPawnEnemies & attackedBy2[Us]); score += Hanging * popcount(weak & b); + + // Additional bonus if weak piece is only protected by a queen + score += WeakQueenProtection * popcount(weak & attackedBy[Them][QUEEN]); } // Bonus for restricting their piece moves b = attackedBy[Them][ALL_PIECES] & ~stronglyProtected & attackedBy[Us][ALL_PIECES]; - score += RestrictedPiece * popcount(b); // Protected or unattacked squares @@ -552,20 +707,24 @@ namespace { // Bonus for threats on the next moves against enemy queen if (pos.count(Them) == 1) { + bool queenImbalance = pos.count() == 1; + Square s = pos.square(Them); - safe = mobilityArea[Us] & ~stronglyProtected; + safe = mobilityArea[Us] + & ~pos.pieces(Us, PAWN) + & ~stronglyProtected; - b = attackedBy[Us][KNIGHT] & pos.attacks_from(s); + b = attackedBy[Us][KNIGHT] & attacks_bb(s); - score += KnightOnQueen * popcount(b & safe); + score += KnightOnQueen * popcount(b & safe) * (1 + queenImbalance); - b = (attackedBy[Us][BISHOP] & pos.attacks_from(s)) - | (attackedBy[Us][ROOK ] & pos.attacks_from(s)); + b = (attackedBy[Us][BISHOP] & attacks_bb(s, pos.pieces())) + | (attackedBy[Us][ROOK ] & attacks_bb(s, pos.pieces())); - score += SliderOnQueen * popcount(b & safe & attackedBy2[Us]); + score += SliderOnQueen * popcount(b & safe & attackedBy2[Us]) * (1 + queenImbalance); } - if (T) + if constexpr (T) Trace::add(THREAT, Us, score); return score; @@ -577,18 +736,32 @@ namespace { template template Score Evaluation::passed() const { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Up = (Us == WHITE ? NORTH : SOUTH); + constexpr Color Them = ~Us; + constexpr Direction Up = pawn_push(Us); + constexpr Direction Down = -Up; auto king_proximity = [&](Color c, Square s) { return std::min(distance(pos.square(c), s), 5); }; - Bitboard b, bb, squaresToQueen, unsafeSquares; + Bitboard b, bb, squaresToQueen, unsafeSquares, blockedPassers, helpers; Score score = SCORE_ZERO; b = pe->passed_pawns(Us); + blockedPassers = b & shift(pos.pieces(Them, PAWN)); + if (blockedPassers) + { + helpers = shift(pos.pieces(Us, PAWN)) + & ~pos.pieces(Them) + & (~attackedBy2[Them] | attackedBy[Us][ALL_PIECES]); + + // Remove blocked candidate passers that don't have help to pass + b &= ~blockedPassers + | shift(helpers) + | shift(helpers); + } + while (b) { Square s = pop_lsb(&b); @@ -596,7 +769,6 @@ namespace { assert(!(pos.pieces(Them, PAWN) & forward_file_bb(Us, s + Up))); int r = relative_rank(Us, s); - File f = file_of(s); Score bonus = PassedRank[r]; @@ -606,8 +778,8 @@ namespace { Square blockSq = s + Up; // Adjust bonus based on the king's proximity - bonus += make_score(0, ( king_proximity(Them, blockSq) * 5 - - king_proximity(Us, blockSq) * 2) * w); + bonus += make_score(0, ( king_proximity(Them, blockSq) * 19 / 4 + - king_proximity(Us, blockSq) * 2) * w); // If blockSq is not the queening square then consider also a second push if (r != RANK_7) @@ -622,14 +794,16 @@ namespace { bb = forward_file_bb(Them, s) & pos.pieces(ROOK, QUEEN); if (!(pos.pieces(Them) & bb)) - unsafeSquares &= attackedBy[Them][ALL_PIECES]; + unsafeSquares &= attackedBy[Them][ALL_PIECES] | pos.pieces(Them); - // If there are no enemy attacks on passed pawn span, assign a big bonus. + // If there are no enemy pieces or attacks on passed pawn span, assign a big bonus. + // Or if there is some, but they are all attacked by our pawns, assign a bit smaller bonus. // Otherwise assign a smaller bonus if the path to queen is not attacked // and even smaller bonus if it is attacked but block square is not. - int k = !unsafeSquares ? 35 : - !(unsafeSquares & squaresToQueen) ? 20 : - !(unsafeSquares & blockSq) ? 9 : + int k = !unsafeSquares ? 36 : + !(unsafeSquares & ~attackedBy[Us][PAWN]) ? 30 : + !(unsafeSquares & squaresToQueen) ? 17 : + !(unsafeSquares & blockSq) ? 7 : 0 ; // Assign a larger bonus if the block square is defended @@ -640,37 +814,30 @@ namespace { } } // r > RANK_3 - // Scale down bonus for candidate passers which need more than one - // pawn push to become passed, or have a pawn in front of them. - if ( !pos.pawn_passed(Us, s + Up) - || (pos.pieces(PAWN) & (s + Up))) - bonus = bonus / 2; - - score += bonus - PassedFile * map_to_queenside(f); + score += bonus - PassedFile * edge_distance(file_of(s)); } - if (T) + if constexpr (T) Trace::add(PASSED, Us, score); return score; } - // Evaluation::space() computes the space evaluation for a given side. The - // space evaluation is a simple bonus based on the number of safe squares - // available for minor pieces on the central four files on ranks 2--4. Safe - // squares one, two or three squares behind a friendly pawn are counted - // twice. Finally, the space bonus is multiplied by a weight. The aim is to - // improve play on game opening. + // Evaluation::space() computes a space evaluation for a given side, aiming to improve game + // play in the opening. It is based on the number of safe squares on the four central files + // on ranks 2 to 4. Completely safe squares behind a friendly pawn are counted twice. + // Finally, the space bonus is multiplied by a weight which decreases according to occupancy. template template Score Evaluation::space() const { + // Early exit if, for example, both queens or 6 minor pieces have been exchanged if (pos.non_pawn_material() < SpaceThreshold) return SCORE_ZERO; - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Down = (Us == WHITE ? SOUTH : NORTH); + constexpr Color Them = ~Us; + constexpr Direction Down = -pawn_push(Us); constexpr Bitboard SpaceMask = Us == WHITE ? CenterFiles & (Rank2BB | Rank3BB | Rank4BB) : CenterFiles & (Rank7BB | Rank6BB | Rank5BB); @@ -685,80 +852,114 @@ namespace { behind |= shift(behind); behind |= shift(behind); + // Compute space score based on the number of safe squares and number of our pieces + // increased with number of total blocked pawns in position. int bonus = popcount(safe) + popcount(behind & safe & ~attackedBy[Them][ALL_PIECES]); - int weight = pos.count(Us) - 1; + int weight = pos.count(Us) - 3 + std::min(pe->blocked_count(), 9); Score score = make_score(bonus * weight * weight / 16, 0); - if (T) + if constexpr (T) Trace::add(SPACE, Us, score); return score; } - // Evaluation::initiative() computes the initiative correction value - // for the position. It is a second order bonus/malus based on the - // known attacking/defending status of the players. + // Evaluation::winnable() adjusts the midgame and endgame score components, based on + // the known attacking/defending status of the players. The final value is derived + // by interpolation from the midgame and endgame values. template - Score Evaluation::initiative(Score score) const { - - Value mg = mg_value(score); - Value eg = eg_value(score); + Value Evaluation::winnable(Score score) const { int outflanking = distance(pos.square(WHITE), pos.square(BLACK)) - - distance(pos.square(WHITE), pos.square(BLACK)); + + int(rank_of(pos.square(WHITE)) - rank_of(pos.square(BLACK))); bool pawnsOnBothFlanks = (pos.pieces(PAWN) & QueenSide) && (pos.pieces(PAWN) & KingSide); - bool almostUnwinnable = !pe->passed_count() - && outflanking < 0 + bool almostUnwinnable = outflanking < 0 && !pawnsOnBothFlanks; + bool infiltration = rank_of(pos.square(WHITE)) > RANK_4 + || rank_of(pos.square(BLACK)) < RANK_5; + // Compute the initiative bonus for the attacking side int complexity = 9 * pe->passed_count() - + 11 * pos.count() + + 12 * pos.count() + 9 * outflanking - + 18 * pawnsOnBothFlanks - + 49 * !pos.non_pawn_material() - - 36 * almostUnwinnable - -103 ; + + 21 * pawnsOnBothFlanks + + 24 * infiltration + + 51 * !pos.non_pawn_material() + - 43 * almostUnwinnable + -110 ; + + Value mg = mg_value(score); + Value eg = eg_value(score); // Now apply the bonus: note that we find the attacking side by extracting the // sign of the midgame or endgame values, and that we carefully cap the bonus // so that the midgame and endgame scores do not change sign after the bonus. - int u = ((mg > 0) - (mg < 0)) * std::max(std::min(complexity + 50, 0), -abs(mg)); + int u = ((mg > 0) - (mg < 0)) * std::clamp(complexity + 50, -abs(mg), 0); int v = ((eg > 0) - (eg < 0)) * std::max(complexity, -abs(eg)); - if (T) - Trace::add(INITIATIVE, make_score(u, v)); - - return make_score(u, v); - } - - - // Evaluation::scale_factor() computes the scale factor for the winning side - - template - ScaleFactor Evaluation::scale_factor(Value eg) const { + mg += u; + eg += v; + // Compute the scale factor for the winning side Color strongSide = eg > VALUE_DRAW ? WHITE : BLACK; int sf = me->scale_factor(pos, strongSide); - // If scale is not already specific, scale down the endgame via general heuristics + // If scale factor is not already specific, scale down via general heuristics if (sf == SCALE_FACTOR_NORMAL) { - if ( pos.opposite_bishops() - && pos.non_pawn_material() == 2 * BishopValueMg) - sf = 16 + 4 * pe->passed_count(); + if (pos.opposite_bishops()) + { + // For pure opposite colored bishops endgames use scale factor + // based on the number of passed pawns of the strong side. + if ( pos.non_pawn_material(WHITE) == BishopValueMg + && pos.non_pawn_material(BLACK) == BishopValueMg) + sf = 18 + 4 * popcount(pe->passed_pawns(strongSide)); + // For every other opposite colored bishops endgames use scale factor + // based on the number of all pieces of the strong side. + else + sf = 22 + 3 * pos.count(strongSide); + } + // For rook endgames with strong side not having overwhelming pawn number advantage + // and its pawns being on one flank and weak side protecting its pieces with a king + // use lower scale factor. + else if ( pos.non_pawn_material(WHITE) == RookValueMg + && pos.non_pawn_material(BLACK) == RookValueMg + && pos.count(strongSide) - pos.count(~strongSide) <= 1 + && bool(KingSide & pos.pieces(strongSide, PAWN)) != bool(QueenSide & pos.pieces(strongSide, PAWN)) + && (attacks_bb(pos.square(~strongSide)) & pos.pieces(~strongSide, PAWN))) + sf = 36; + // For queen vs no queen endgames use scale factor + // based on number of minors of side that doesn't have queen. + else if (pos.count() == 1) + sf = 37 + 3 * (pos.count(WHITE) == 1 ? pos.count(BLACK) + pos.count(BLACK) + : pos.count(WHITE) + pos.count(WHITE)); + // In every other case use scale factor based on + // the number of pawns of the strong side reduced if pawns are on a single flank. else - sf = std::min(sf, 36 + (pos.opposite_bishops() ? 2 : 7) * pos.count(strongSide)); + sf = std::min(sf, 36 + 7 * pos.count(strongSide)) - 4 * !pawnsOnBothFlanks; - sf = std::max(0, sf - (pos.rule50_count() - 12) / 4 ); + // Reduce scale factor in case of pawns being on a single flank + sf -= 4 * !pawnsOnBothFlanks; } - return ScaleFactor(sf); + // Interpolate between the middlegame and (scaled by 'sf') endgame score + v = mg * int(me->game_phase()) + + eg * int(PHASE_MIDGAME - me->game_phase()) * ScaleFactor(sf) / SCALE_FACTOR_NORMAL; + v /= PHASE_MIDGAME; + + if constexpr (T) + { + Trace::add(WINNABLE, make_score(u, eg * ScaleFactor(sf) / SCALE_FACTOR_NORMAL - eg_value(score))); + Trace::add(TOTAL, make_score(mg, eg * ScaleFactor(sf) / SCALE_FACTOR_NORMAL)); + } + + return Value(v); } @@ -789,16 +990,19 @@ namespace { score += pe->pawn_score(WHITE) - pe->pawn_score(BLACK); // Early exit if score is high - Value v = (mg_value(score) + eg_value(score)) / 2; - if (abs(v) > LazyThreshold + pos.non_pawn_material() / 64) - return pos.side_to_move() == WHITE ? v : -v; + auto lazy_skip = [&](Value lazyThreshold) { + return abs(mg_value(score) + eg_value(score)) / 2 > lazyThreshold + pos.non_pawn_material() / 64; + }; + + if (lazy_skip(LazyThreshold1)) + goto make_v; // Main evaluation begins here - initialize(); initialize(); - // Pieces should be evaluated first (populate attack tables) + // Pieces evaluated first (also populates attackedBy, attackedBy2). + // Note that the order of evaluation of the terms is left unspecified. score += pieces() - pieces() + pieces() - pieces() + pieces() - pieces() @@ -806,32 +1010,36 @@ namespace { score += mobility[WHITE] - mobility[BLACK]; + // More complex interactions that require fully populated attack bitboards score += king< WHITE>() - king< BLACK>() - + threats() - threats() - + passed< WHITE>() - passed< BLACK>() + + passed< WHITE>() - passed< BLACK>(); + + if (lazy_skip(LazyThreshold2)) + goto make_v; + + score += threats() - threats() + space< WHITE>() - space< BLACK>(); - score += initiative(score); - - // Interpolate between a middlegame and a (scaled by 'sf') endgame score - ScaleFactor sf = scale_factor(eg_value(score)); - v = mg_value(score) * int(me->game_phase()) - + eg_value(score) * int(PHASE_MIDGAME - me->game_phase()) * sf / SCALE_FACTOR_NORMAL; - - v /= PHASE_MIDGAME; +make_v: + // Derive single value from mg and eg parts of score + Value v = winnable(score); // In case of tracing add all remaining individual evaluation terms - if (T) + if constexpr (T) { Trace::add(MATERIAL, pos.psq_score()); Trace::add(IMBALANCE, me->imbalance()); Trace::add(PAWN, pe->pawn_score(WHITE), pe->pawn_score(BLACK)); Trace::add(MOBILITY, mobility[WHITE], mobility[BLACK]); - Trace::add(TOTAL, score); } - return (pos.side_to_move() == WHITE ? v : -v) // Side to move point of view - + Eval::Tempo; + // Evaluation grain + v = (v / 16) * 16; + + // Side to move point of view + v = (pos.side_to_move() == WHITE ? v : -v) + Tempo; + + return v; } } // namespace @@ -841,25 +1049,72 @@ namespace { /// evaluation of the position from the point of view of the side to move. Value Eval::evaluate(const Position& pos) { - return Evaluation(pos).value(); -} + Value v; + + if (!Eval::useNNUE) + v = Evaluation(pos).value(); + else + { + // Scale and shift NNUE for compatibility with search and classical evaluation + auto adjusted_NNUE = [&](){ + int mat = pos.non_pawn_material() + 2 * PawnValueMg * pos.count(); + return NNUE::evaluate(pos) * (641 + mat / 32 - 4 * pos.rule50_count()) / 1024 + Tempo; + }; + + // If there is PSQ imbalance use classical eval, with small probability if it is small + Value psq = Value(abs(eg_value(pos.psq_score()))); + int r50 = 16 + pos.rule50_count(); + bool largePsq = psq * 16 > (NNUEThreshold1 + pos.non_pawn_material() / 64) * r50; + bool classical = largePsq || (psq > PawnValueMg / 4 && !(pos.this_thread()->nodes & 0xB)); + + // Use classical evaluation for really low piece endgames. + // The most critical case is a bishop + A/H file pawn vs naked king draw. + bool strongClassical = pos.non_pawn_material() < 2 * RookValueMg && pos.count() < 2; + + v = classical || strongClassical ? Evaluation(pos).value() : adjusted_NNUE(); + + // If the classical eval is small and imbalance large, use NNUE nevertheless. + // For the case of opposite colored bishops, switch to NNUE eval with + // small probability if the classical eval is less than the threshold. + if ( largePsq && !strongClassical + && ( abs(v) * 16 < NNUEThreshold2 * r50 + || ( pos.opposite_bishops() + && abs(v) * 16 < (NNUEThreshold1 + pos.non_pawn_material() / 64) * r50 + && !(pos.this_thread()->nodes & 0xB)))) + v = adjusted_NNUE(); + } + + // Damp down the evaluation linearly when shuffling + v = v * (100 - pos.rule50_count()) / 100; + + // Guarantee evaluation does not hit the tablebase range + v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1); + + return v; +} /// trace() is like evaluate(), but instead of returning a value, it returns /// a string (suitable for outputting to stdout) that contains the detailed /// descriptions and values of each evaluation term. Useful for debugging. +/// Trace scores are from white's point of view std::string Eval::trace(const Position& pos) { + if (pos.checkers()) + return "Final evaluation: none (in check)"; + + std::stringstream ss; + ss << std::showpoint << std::noshowpos << std::fixed << std::setprecision(2); + + Value v; + std::memset(scores, 0, sizeof(scores)); pos.this_thread()->contempt = SCORE_ZERO; // Reset any dynamic contempt - Value v = Evaluation(pos).value(); + v = Evaluation(pos).value(); - v = pos.side_to_move() == WHITE ? v : -v; // Trace scores are from white's point of view - - std::stringstream ss; ss << std::showpoint << std::noshowpos << std::fixed << std::setprecision(2) << " Term | White | Black | Total \n" << " | MG EG | MG EG | MG EG \n" @@ -876,11 +1131,26 @@ std::string Eval::trace(const Position& pos) { << " Threats | " << Term(THREAT) << " Passed | " << Term(PASSED) << " Space | " << Term(SPACE) - << " Initiative | " << Term(INITIATIVE) + << " Winnable | " << Term(WINNABLE) << " ------------+-------------+-------------+------------\n" << " Total | " << Term(TOTAL); - ss << "\nTotal evaluation: " << to_cp(v) << " (white side)\n"; + v = pos.side_to_move() == WHITE ? v : -v; + + ss << "\nClassical evaluation: " << to_cp(v) << " (white side)\n"; + + if (Eval::useNNUE) + { + v = NNUE::evaluate(pos); + v = pos.side_to_move() == WHITE ? v : -v; + ss << "\nNNUE evaluation: " << to_cp(v) << " (white side)\n"; + } + + v = evaluate(pos); + v = pos.side_to_move() == WHITE ? v : -v; + ss << "\nFinal evaluation: " << to_cp(v) << " (white side)\n"; return ss.str(); } + +} // namespace Stockfish diff --git a/src/evaluate.h b/src/evaluate.h index cccdd25d..6210bd58 100644 --- a/src/evaluate.h +++ b/src/evaluate.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,15 +23,34 @@ #include "types.h" +namespace Stockfish { + class Position; namespace Eval { -constexpr Value Tempo = Value(28); // Must be visible to search + std::string trace(const Position& pos); + Value evaluate(const Position& pos); -std::string trace(const Position& pos); + extern bool useNNUE; + extern std::string eval_file_loaded; -Value evaluate(const Position& pos); -} + // 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-62ef826d1a6d.nnue" + + namespace NNUE { + + Value evaluate(const Position& pos); + bool load_eval(std::string name, std::istream& stream); + void init(); + void verify(); + + } // namespace NNUE + +} // namespace Eval + +} // namespace Stockfish #endif // #ifndef EVALUATE_H_INCLUDED diff --git a/src/incbin/UNLICENCE b/src/incbin/UNLICENCE new file mode 100644 index 00000000..32484ab5 --- /dev/null +++ b/src/incbin/UNLICENCE @@ -0,0 +1,26 @@ +The file "incbin.h" is free and unencumbered software released into +the public domain by Dale Weiler, see: + + +Anyone is free to copy, modify, publish, use, compile, sell, or +distribute this software, either in source code form or as a compiled +binary, for any purpose, commercial or non-commercial, and by any +means. + +In jurisdictions that recognize copyright laws, the author or authors +of this software dedicate any and all copyright interest in the +software to the public domain. We make this dedication for the benefit +of the public at large and to the detriment of our heirs and +successors. We intend this dedication to be an overt act of +relinquishment in perpetuity of all present and future rights to this +software under copyright law. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, +EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF +MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. +IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR +OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR +OTHER DEALINGS IN THE SOFTWARE. + +For more information, please refer to diff --git a/src/incbin/incbin.h b/src/incbin/incbin.h new file mode 100755 index 00000000..c19684d7 --- /dev/null +++ b/src/incbin/incbin.h @@ -0,0 +1,368 @@ +/** + * @file incbin.h + * @author Dale Weiler + * @brief Utility for including binary files + * + * Facilities for including binary files into the current translation unit and + * making use from them externally in other translation units. + */ +#ifndef INCBIN_HDR +#define INCBIN_HDR +#include +#if defined(__AVX512BW__) || \ + defined(__AVX512CD__) || \ + defined(__AVX512DQ__) || \ + defined(__AVX512ER__) || \ + defined(__AVX512PF__) || \ + defined(__AVX512VL__) || \ + defined(__AVX512F__) +# define INCBIN_ALIGNMENT_INDEX 6 +#elif defined(__AVX__) || \ + defined(__AVX2__) +# define INCBIN_ALIGNMENT_INDEX 5 +#elif defined(__SSE__) || \ + defined(__SSE2__) || \ + defined(__SSE3__) || \ + defined(__SSSE3__) || \ + defined(__SSE4_1__) || \ + defined(__SSE4_2__) || \ + defined(__neon__) +# define INCBIN_ALIGNMENT_INDEX 4 +#elif ULONG_MAX != 0xffffffffu +# define INCBIN_ALIGNMENT_INDEX 3 +# else +# define INCBIN_ALIGNMENT_INDEX 2 +#endif + +/* Lookup table of (1 << n) where `n' is `INCBIN_ALIGNMENT_INDEX' */ +#define INCBIN_ALIGN_SHIFT_0 1 +#define INCBIN_ALIGN_SHIFT_1 2 +#define INCBIN_ALIGN_SHIFT_2 4 +#define INCBIN_ALIGN_SHIFT_3 8 +#define INCBIN_ALIGN_SHIFT_4 16 +#define INCBIN_ALIGN_SHIFT_5 32 +#define INCBIN_ALIGN_SHIFT_6 64 + +/* Actual alignment value */ +#define INCBIN_ALIGNMENT \ + INCBIN_CONCATENATE( \ + INCBIN_CONCATENATE(INCBIN_ALIGN_SHIFT, _), \ + INCBIN_ALIGNMENT_INDEX) + +/* Stringize */ +#define INCBIN_STR(X) \ + #X +#define INCBIN_STRINGIZE(X) \ + INCBIN_STR(X) +/* Concatenate */ +#define INCBIN_CAT(X, Y) \ + X ## Y +#define INCBIN_CONCATENATE(X, Y) \ + INCBIN_CAT(X, Y) +/* Deferred macro expansion */ +#define INCBIN_EVAL(X) \ + X +#define INCBIN_INVOKE(N, ...) \ + INCBIN_EVAL(N(__VA_ARGS__)) + +/* Green Hills uses a different directive for including binary data */ +#if defined(__ghs__) +# if (__ghs_asm == 2) +# define INCBIN_MACRO ".file" +/* Or consider the ".myrawdata" entry in the ld file */ +# else +# define INCBIN_MACRO "\tINCBIN" +# endif +#else +# define INCBIN_MACRO ".incbin" +#endif + +#ifndef _MSC_VER +# define INCBIN_ALIGN \ + __attribute__((aligned(INCBIN_ALIGNMENT))) +#else +# define INCBIN_ALIGN __declspec(align(INCBIN_ALIGNMENT)) +#endif + +#if defined(__arm__) || /* GNU C and RealView */ \ + defined(__arm) || /* Diab */ \ + defined(_ARM) /* ImageCraft */ +# define INCBIN_ARM +#endif + +#ifdef __GNUC__ +/* Utilize .balign where supported */ +# define INCBIN_ALIGN_HOST ".balign " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n" +# define INCBIN_ALIGN_BYTE ".balign 1\n" +#elif defined(INCBIN_ARM) +/* + * On arm assemblers, the alignment value is calculated as (1 << n) where `n' is + * the shift count. This is the value passed to `.align' + */ +# define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT_INDEX) "\n" +# define INCBIN_ALIGN_BYTE ".align 0\n" +#else +/* We assume other inline assembler's treat `.align' as `.balign' */ +# define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n" +# define INCBIN_ALIGN_BYTE ".align 1\n" +#endif + +/* INCBIN_CONST is used by incbin.c generated files */ +#if defined(__cplusplus) +# define INCBIN_EXTERNAL extern "C" +# define INCBIN_CONST extern const +#else +# define INCBIN_EXTERNAL extern +# define INCBIN_CONST const +#endif + +/** + * @brief Optionally override the linker section into which data is emitted. + * + * @warning If you use this facility, you'll have to deal with platform-specific linker output + * section naming on your own + * + * Overriding the default linker output section, e.g for esp8266/Arduino: + * @code + * #define INCBIN_OUTPUT_SECTION ".irom.text" + * #include "incbin.h" + * INCBIN(Foo, "foo.txt"); + * // Data is emitted into program memory that never gets copied to RAM + * @endcode + */ +#if !defined(INCBIN_OUTPUT_SECTION) +# if defined(__APPLE__) +# define INCBIN_OUTPUT_SECTION ".const_data" +# else +# define INCBIN_OUTPUT_SECTION ".rodata" +# endif +#endif + +#if defined(__APPLE__) +/* The directives are different for Apple branded compilers */ +# define INCBIN_SECTION INCBIN_OUTPUT_SECTION "\n" +# define INCBIN_GLOBAL(NAME) ".globl " INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n" +# define INCBIN_INT ".long " +# define INCBIN_MANGLE "_" +# define INCBIN_BYTE ".byte " +# define INCBIN_TYPE(...) +#else +# define INCBIN_SECTION ".section " INCBIN_OUTPUT_SECTION "\n" +# define INCBIN_GLOBAL(NAME) ".global " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n" +# if defined(__ghs__) +# define INCBIN_INT ".word " +# else +# define INCBIN_INT ".int " +# endif +# if defined(__USER_LABEL_PREFIX__) +# define INCBIN_MANGLE INCBIN_STRINGIZE(__USER_LABEL_PREFIX__) +# else +# define INCBIN_MANGLE "" +# endif +# if defined(INCBIN_ARM) +/* On arm assemblers, `@' is used as a line comment token */ +# define INCBIN_TYPE(NAME) ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", %object\n" +# elif defined(__MINGW32__) || defined(__MINGW64__) +/* Mingw doesn't support this directive either */ +# define INCBIN_TYPE(NAME) +# else +/* It's safe to use `@' on other architectures */ +# define INCBIN_TYPE(NAME) ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", @object\n" +# endif +# define INCBIN_BYTE ".byte " +#endif + +/* List of style types used for symbol names */ +#define INCBIN_STYLE_CAMEL 0 +#define INCBIN_STYLE_SNAKE 1 + +/** + * @brief Specify the prefix to use for symbol names. + * + * By default this is `g', producing symbols of the form: + * @code + * #include "incbin.h" + * INCBIN(Foo, "foo.txt"); + * + * // Now you have the following symbols: + * // const unsigned char gFooData[]; + * // const unsigned char *const gFooEnd; + * // const unsigned int gFooSize; + * @endcode + * + * If however you specify a prefix before including: e.g: + * @code + * #define INCBIN_PREFIX incbin + * #include "incbin.h" + * INCBIN(Foo, "foo.txt"); + * + * // Now you have the following symbols instead: + * // const unsigned char incbinFooData[]; + * // const unsigned char *const incbinFooEnd; + * // const unsigned int incbinFooSize; + * @endcode + */ +#if !defined(INCBIN_PREFIX) +# define INCBIN_PREFIX g +#endif + +/** + * @brief Specify the style used for symbol names. + * + * Possible options are + * - INCBIN_STYLE_CAMEL "CamelCase" + * - INCBIN_STYLE_SNAKE "snake_case" + * + * Default option is *INCBIN_STYLE_CAMEL* producing symbols of the form: + * @code + * #include "incbin.h" + * INCBIN(Foo, "foo.txt"); + * + * // Now you have the following symbols: + * // const unsigned char FooData[]; + * // const unsigned char *const FooEnd; + * // const unsigned int FooSize; + * @endcode + * + * If however you specify a style before including: e.g: + * @code + * #define INCBIN_STYLE INCBIN_STYLE_SNAKE + * #include "incbin.h" + * INCBIN(foo, "foo.txt"); + * + * // Now you have the following symbols: + * // const unsigned char foo_data[]; + * // const unsigned char *const foo_end; + * // const unsigned int foo_size; + * @endcode + */ +#if !defined(INCBIN_STYLE) +# define INCBIN_STYLE INCBIN_STYLE_CAMEL +#endif + +/* Style lookup tables */ +#define INCBIN_STYLE_0_DATA Data +#define INCBIN_STYLE_0_END End +#define INCBIN_STYLE_0_SIZE Size +#define INCBIN_STYLE_1_DATA _data +#define INCBIN_STYLE_1_END _end +#define INCBIN_STYLE_1_SIZE _size + +/* Style lookup: returning identifier */ +#define INCBIN_STYLE_IDENT(TYPE) \ + INCBIN_CONCATENATE( \ + INCBIN_STYLE_, \ + INCBIN_CONCATENATE( \ + INCBIN_EVAL(INCBIN_STYLE), \ + INCBIN_CONCATENATE(_, TYPE))) + +/* Style lookup: returning string literal */ +#define INCBIN_STYLE_STRING(TYPE) \ + INCBIN_STRINGIZE( \ + INCBIN_STYLE_IDENT(TYPE)) \ + +/* Generate the global labels by indirectly invoking the macro with our style + * type and concatenating the name against them. */ +#define INCBIN_GLOBAL_LABELS(NAME, TYPE) \ + INCBIN_INVOKE( \ + INCBIN_GLOBAL, \ + INCBIN_CONCATENATE( \ + NAME, \ + INCBIN_INVOKE( \ + INCBIN_STYLE_IDENT, \ + TYPE))) \ + INCBIN_INVOKE( \ + INCBIN_TYPE, \ + INCBIN_CONCATENATE( \ + NAME, \ + INCBIN_INVOKE( \ + INCBIN_STYLE_IDENT, \ + TYPE))) + +/** + * @brief Externally reference binary data included in another translation unit. + * + * Produces three external symbols that reference the binary data included in + * another translation unit. + * + * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with + * "Data", as well as "End" and "Size" after. An example is provided below. + * + * @param NAME The name given for the binary data + * + * @code + * INCBIN_EXTERN(Foo); + * + * // Now you have the following symbols: + * // extern const unsigned char FooData[]; + * // extern const unsigned char *const FooEnd; + * // extern const unsigned int FooSize; + * @endcode + */ +#define INCBIN_EXTERN(NAME) \ + INCBIN_EXTERNAL const INCBIN_ALIGN unsigned char \ + INCBIN_CONCATENATE( \ + INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \ + INCBIN_STYLE_IDENT(DATA))[]; \ + INCBIN_EXTERNAL const INCBIN_ALIGN unsigned char *const \ + INCBIN_CONCATENATE( \ + INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \ + INCBIN_STYLE_IDENT(END)); \ + INCBIN_EXTERNAL const unsigned int \ + INCBIN_CONCATENATE( \ + INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \ + INCBIN_STYLE_IDENT(SIZE)) + +/** + * @brief Include a binary file into the current translation unit. + * + * Includes a binary file into the current translation unit, producing three symbols + * for objects that encode the data and size respectively. + * + * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with + * "Data", as well as "End" and "Size" after. An example is provided below. + * + * @param NAME The name to associate with this binary data (as an identifier.) + * @param FILENAME The file to include (as a string literal.) + * + * @code + * INCBIN(Icon, "icon.png"); + * + * // Now you have the following symbols: + * // const unsigned char IconData[]; + * // const unsigned char *const IconEnd; + * // const unsigned int IconSize; + * @endcode + * + * @warning This must be used in global scope + * @warning The identifiers may be different if INCBIN_STYLE is not default + * + * To externally reference the data included by this in another translation unit + * please @see INCBIN_EXTERN. + */ +#ifdef _MSC_VER +#define INCBIN(NAME, FILENAME) \ + INCBIN_EXTERN(NAME) +#else +#define INCBIN(NAME, FILENAME) \ + __asm__(INCBIN_SECTION \ + INCBIN_GLOBAL_LABELS(NAME, DATA) \ + INCBIN_ALIGN_HOST \ + INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) ":\n" \ + INCBIN_MACRO " \"" FILENAME "\"\n" \ + INCBIN_GLOBAL_LABELS(NAME, END) \ + INCBIN_ALIGN_BYTE \ + INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) ":\n" \ + INCBIN_BYTE "1\n" \ + INCBIN_GLOBAL_LABELS(NAME, SIZE) \ + INCBIN_ALIGN_HOST \ + INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(SIZE) ":\n" \ + INCBIN_INT INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) " - " \ + INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) "\n" \ + INCBIN_ALIGN_HOST \ + ".text\n" \ + ); \ + INCBIN_EXTERN(NAME) + +#endif +#endif diff --git a/src/main.cpp b/src/main.cpp index e0b54ec7..8e4cdacc 100644 --- a/src/main.cpp +++ b/src/main.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -21,17 +19,16 @@ #include #include "bitboard.h" +#include "endgame.h" #include "position.h" +#include "psqt.h" #include "search.h" +#include "syzygy/tbprobe.h" #include "thread.h" #include "tt.h" #include "uci.h" -#include "endgame.h" -#include "syzygy/tbprobe.h" -namespace PSQT { - void init(); -} +using namespace Stockfish; int main(int argc, char* argv[]) { @@ -39,14 +36,17 @@ int main(int argc, char* argv[]) { if (Cluster::is_root()) std::cout << engine_info() << std::endl; + CommandLine::init(argc, argv); UCI::init(Options); + Tune::init(); PSQT::init(); Bitboards::init(); Position::init(); Bitbases::init(); Endgames::init(); - Threads.set(Options["Threads"]); + Threads.set(size_t(Options["Threads"])); Search::clear(); // After threads are up + Eval::NNUE::init(); UCI::loop(argc, argv); diff --git a/src/material.cpp b/src/material.cpp index 11d4c687..84d7a4bd 100644 --- a/src/material.cpp +++ b/src/material.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,32 +24,39 @@ using namespace std; +namespace Stockfish { + namespace { + #define S(mg, eg) make_score(mg, eg) // Polynomial material imbalance parameters - constexpr int QuadraticOurs[][PIECE_TYPE_NB] = { - // OUR PIECES - // pair pawn knight bishop rook queen - {1438 }, // Bishop pair - { 40, 38 }, // Pawn - { 32, 255, -62 }, // Knight OUR PIECES - { 0, 104, 4, 0 }, // Bishop - { -26, -2, 47, 105, -208 }, // Rook - {-189, 24, 117, 133, -134, -6 } // Queen + // One Score parameter for each pair (our piece, another of our pieces) + constexpr Score QuadraticOurs[][PIECE_TYPE_NB] = { + // OUR PIECE 2 + // bishop pair pawn knight bishop rook queen + {S(1419, 1455) }, // Bishop pair + {S( 101, 28), S( 37, 39) }, // Pawn + {S( 57, 64), S(249, 187), S(-49, -62) }, // Knight OUR PIECE 1 + {S( 0, 0), S(118, 137), S( 10, 27), S( 0, 0) }, // Bishop + {S( -63, -68), S( -5, 3), S(100, 81), S(132, 118), S(-246, -244) }, // Rook + {S(-210, -211), S( 37, 14), S(147, 141), S(161, 105), S(-158, -174), S(-9,-31) } // Queen }; - constexpr int QuadraticTheirs[][PIECE_TYPE_NB] = { - // THEIR PIECES - // pair pawn knight bishop rook queen - { 0 }, // Bishop pair - { 36, 0 }, // Pawn - { 9, 63, 0 }, // Knight OUR PIECES - { 59, 65, 42, 0 }, // Bishop - { 46, 39, 24, -24, 0 }, // Rook - { 97, 100, -42, 137, 268, 0 } // Queen + // One Score parameter for each pair (our piece, their piece) + constexpr Score QuadraticTheirs[][PIECE_TYPE_NB] = { + // THEIR PIECE + // bishop pair pawn knight bishop rook queen + { }, // Bishop pair + {S( 33, 30) }, // Pawn + {S( 46, 18), S(106, 84) }, // Knight OUR PIECE + {S( 75, 35), S( 59, 44), S( 60, 15) }, // Bishop + {S( 26, 35), S( 6, 22), S( 38, 39), S(-12, -2) }, // Rook + {S( 97, 93), S(100, 163), S(-58, -91), S(112, 192), S(276, 225) } // Queen }; + #undef S + // Endgame evaluation and scaling functions are accessed directly and not through // the function maps because they correspond to more than one material hash key. Endgame EvaluateKXK[] = { Endgame(WHITE), Endgame(BLACK) }; @@ -79,14 +84,16 @@ namespace { && pos.count(~us) >= 1; } + /// imbalance() calculates the imbalance by comparing the piece count of each /// piece type for both colors. + template - int imbalance(const int pieceCount[][PIECE_TYPE_NB]) { + Score imbalance(const int pieceCount[][PIECE_TYPE_NB]) { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); + constexpr Color Them = ~Us; - int bonus = 0; + Score bonus = SCORE_ZERO; // Second-degree polynomial material imbalance, by Tord Romstad for (int pt1 = NO_PIECE_TYPE; pt1 <= QUEEN; ++pt1) @@ -94,9 +101,9 @@ namespace { if (!pieceCount[Us][pt1]) continue; - int v = 0; + int v = QuadraticOurs[pt1][pt1] * pieceCount[Us][pt1]; - for (int pt2 = NO_PIECE_TYPE; pt2 <= pt1; ++pt2) + for (int pt2 = NO_PIECE_TYPE; pt2 < pt1; ++pt2) v += QuadraticOurs[pt1][pt2] * pieceCount[Us][pt2] + QuadraticTheirs[pt1][pt2] * pieceCount[Them][pt2]; @@ -110,6 +117,7 @@ namespace { namespace Material { + /// Material::probe() looks up the current position's material configuration in /// the material hash table. It returns a pointer to the Entry if the position /// is found. Otherwise a new Entry is computed and stored there, so we don't @@ -129,7 +137,7 @@ Entry* probe(const Position& pos) { Value npm_w = pos.non_pawn_material(WHITE); Value npm_b = pos.non_pawn_material(BLACK); - Value npm = clamp(npm_w + npm_b, EndgameLimit, MidgameLimit); + Value npm = std::clamp(npm_w + npm_b, EndgameLimit, MidgameLimit); // Map total non-pawn material into [PHASE_ENDGAME, PHASE_MIDGAME] e->gamePhase = Phase(((npm - EndgameLimit) * PHASE_MIDGAME) / (MidgameLimit - EndgameLimit)); @@ -212,8 +220,10 @@ Entry* probe(const Position& pos) { { pos.count(BLACK) > 1, pos.count(BLACK), pos.count(BLACK), pos.count(BLACK) , pos.count(BLACK), pos.count(BLACK) } }; - e->value = int16_t((imbalance(pieceCount) - imbalance(pieceCount)) / 16); + e->score = (imbalance(pieceCount) - imbalance(pieceCount)) / 16; return e; } } // namespace Material + +} // namespace Stockfish diff --git a/src/material.h b/src/material.h index b472c3fd..26535a53 100644 --- a/src/material.h +++ b/src/material.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,7 +24,7 @@ #include "position.h" #include "types.h" -namespace Material { +namespace Stockfish::Material { /// Material::Entry contains various information about a material configuration. /// It contains a material imbalance evaluation, a function pointer to a special @@ -39,12 +37,12 @@ namespace Material { struct Entry { - Score imbalance() const { return make_score(value, value); } - Phase game_phase() const { return gamePhase; } + Score imbalance() const { return score; } + Phase game_phase() const { return (Phase)gamePhase; } bool specialized_eval_exists() const { return evaluationFunction != nullptr; } Value evaluate(const Position& pos) const { return (*evaluationFunction)(pos); } - // scale_factor takes a position and a color as input and returns a scale factor + // scale_factor() takes a position and a color as input and returns a scale factor // for the given color. We have to provide the position in addition to the color // because the scale factor may also be a function which should be applied to // the position. For instance, in KBP vs K endgames, the scaling function looks @@ -59,15 +57,15 @@ struct Entry { const EndgameBase* evaluationFunction; const EndgameBase* scalingFunction[COLOR_NB]; // Could be one for each // side (e.g. KPKP, KBPsK) - int16_t value; + Score score; + int16_t gamePhase; uint8_t factor[COLOR_NB]; - Phase gamePhase; }; typedef HashTable Table; Entry* probe(const Position& pos); -} // namespace Material +} // namespace Stockfish::Material #endif // #ifndef MATERIAL_H_INCLUDED diff --git a/src/misc.cpp b/src/misc.cpp index 17644eed..7600fc11 100644 --- a/src/misc.cpp +++ b/src/misc.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -46,12 +44,25 @@ typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY); #include #include #include +#include + +#if defined(__linux__) && !defined(__ANDROID__) +#include +#include +#endif + +#if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) +#define POSIXALIGNEDALLOC +#include +#endif #include "misc.h" #include "thread.h" using namespace std; +namespace Stockfish { + namespace { /// Version number. If Version is left empty, then compile date in the format @@ -102,6 +113,13 @@ public: if (!fname.empty() && !l.file.is_open()) { l.file.open(fname, ifstream::out); + + if (!l.file.is_open()) + { + cerr << "Unable to open debug log file " << fname << endl; + exit(EXIT_FAILURE); + } + cin.rdbuf(&l.in); cout.rdbuf(&l.out); } @@ -116,12 +134,13 @@ public: } // namespace + /// engine_info() returns the full name of the current Stockfish version. This /// will be either "Stockfish DD-MM-YY" (where DD-MM-YY is the date when /// the program was compiled) or "Stockfish ", depending on whether /// Version is empty. -const string engine_info(bool to_uci) { +string engine_info(bool to_uci) { const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec"); string month, day, year; @@ -135,15 +154,117 @@ const string engine_info(bool to_uci) { ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2); } - ss << (Is64Bit ? " 64" : "") - << (HasPext ? " BMI2" : (HasPopCnt ? " POPCNT" : "")) - << (to_uci ? "\nid author ": " by ") - << "T. Romstad, M. Costalba, J. Kiiski, G. Linscott"; + ss << (to_uci ? "\nid author ": " by ") + << "the Stockfish developers (see AUTHORS file)"; return ss.str(); } +/// compiler_info() returns a string trying to describe the compiler we use + +std::string compiler_info() { + + #define stringify2(x) #x + #define stringify(x) stringify2(x) + #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch) + +/// Predefined macros hell: +/// +/// __GNUC__ Compiler is gcc, Clang or Intel on Linux +/// __INTEL_COMPILER Compiler is Intel +/// _MSC_VER Compiler is MSVC or Intel on Windows +/// _WIN32 Building on Windows (any) +/// _WIN64 Building on Windows 64 bit + + std::string compiler = "\nCompiled by "; + + #ifdef __clang__ + compiler += "clang++ "; + compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__); + #elif __INTEL_COMPILER + compiler += "Intel compiler "; + compiler += "(version "; + compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE); + compiler += ")"; + #elif _MSC_VER + compiler += "MSVC "; + compiler += "(version "; + compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD); + compiler += ")"; + #elif __GNUC__ + compiler += "g++ (GNUC) "; + compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__); + #else + compiler += "Unknown compiler "; + compiler += "(unknown version)"; + #endif + + #if defined(__APPLE__) + compiler += " on Apple"; + #elif defined(__CYGWIN__) + compiler += " on Cygwin"; + #elif defined(__MINGW64__) + compiler += " on MinGW64"; + #elif defined(__MINGW32__) + compiler += " on MinGW32"; + #elif defined(__ANDROID__) + compiler += " on Android"; + #elif defined(__linux__) + compiler += " on Linux"; + #elif defined(_WIN64) + compiler += " on Microsoft Windows 64-bit"; + #elif defined(_WIN32) + compiler += " on Microsoft Windows 32-bit"; + #else + compiler += " on unknown system"; + #endif + + compiler += "\nCompilation settings include: "; + compiler += (Is64Bit ? " 64bit" : " 32bit"); + #if defined(USE_VNNI) + compiler += " VNNI"; + #endif + #if defined(USE_AVX512) + compiler += " AVX512"; + #endif + compiler += (HasPext ? " BMI2" : ""); + #if defined(USE_AVX2) + compiler += " AVX2"; + #endif + #if defined(USE_SSE41) + compiler += " SSE41"; + #endif + #if defined(USE_SSSE3) + compiler += " SSSE3"; + #endif + #if defined(USE_SSE2) + compiler += " SSE2"; + #endif + compiler += (HasPopCnt ? " POPCNT" : ""); + #if defined(USE_MMX) + compiler += " MMX"; + #endif + #if defined(USE_NEON) + compiler += " NEON"; + #endif + + #if !defined(NDEBUG) + compiler += " DEBUG"; + #endif + + compiler += "\n__VERSION__ macro expands to: "; + #ifdef __VERSION__ + compiler += __VERSION__; + #else + compiler += "(undefined macro)"; + #endif + compiler += "\n"; + + return compiler; +} + + /// Debug functions used mainly to collect run-time statistics static std::atomic hits[2], means[2]; @@ -210,6 +331,146 @@ void prefetch(void* addr) { #endif + +/// std_aligned_alloc() is our wrapper for systems where the c++17 implementation +/// does not guarantee the availability of aligned_alloc(). Memory allocated with +/// std_aligned_alloc() must be freed with std_aligned_free(). + +void* std_aligned_alloc(size_t alignment, size_t size) { + +#if defined(POSIXALIGNEDALLOC) + void *mem; + return posix_memalign(&mem, alignment, size) ? nullptr : mem; +#elif defined(_WIN32) + return _mm_malloc(size, alignment); +#else + return std::aligned_alloc(alignment, size); +#endif +} + +void std_aligned_free(void* ptr) { + +#if defined(POSIXALIGNEDALLOC) + free(ptr); +#elif defined(_WIN32) + _mm_free(ptr); +#else + free(ptr); +#endif +} + +/// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages. + +#if defined(_WIN32) +#if defined(_WIN64) +static void* aligned_large_pages_alloc_win(size_t allocSize) { + + HANDLE hProcessToken { }; + LUID luid { }; + void* mem = nullptr; + + const size_t largePageSize = GetLargePageMinimum(); + if (!largePageSize) + return nullptr; + + // We need SeLockMemoryPrivilege, so try to enable it for the process + if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken)) + return nullptr; + + if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid)) + { + TOKEN_PRIVILEGES tp { }; + TOKEN_PRIVILEGES prevTp { }; + DWORD prevTpLen = 0; + + tp.PrivilegeCount = 1; + tp.Privileges[0].Luid = luid; + tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED; + + // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds, + // we still need to query GetLastError() to ensure that the privileges were actually obtained. + if (AdjustTokenPrivileges( + hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) && + GetLastError() == ERROR_SUCCESS) + { + // Round up size to full pages and allocate + allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1); + mem = VirtualAlloc( + NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE); + + // Privilege no longer needed, restore previous state + AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL); + } + } + + CloseHandle(hProcessToken); + + return mem; +} +#endif + +void* aligned_large_pages_alloc(size_t allocSize) { + +#if defined(_WIN64) + // Try to allocate large pages + void* mem = aligned_large_pages_alloc_win(allocSize); + + // Fall back to regular, page aligned, allocation if necessary + if (!mem) + mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE); +#else + void* mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE); +#endif + + return mem; +} + +#else + +void* aligned_large_pages_alloc(size_t allocSize) { + +#if defined(__linux__) + constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size +#else + constexpr size_t alignment = 4096; // assumed small page size +#endif + + // round up to multiples of alignment + size_t size = ((allocSize + alignment - 1) / alignment) * alignment; + void *mem = std_aligned_alloc(alignment, size); +#if defined(MADV_HUGEPAGE) + madvise(mem, size, MADV_HUGEPAGE); +#endif + return mem; +} + +#endif + + +/// aligned_large_pages_free() will free the previously allocated ttmem + +#if defined(_WIN32) + +void aligned_large_pages_free(void* mem) { + + if (mem && !VirtualFree(mem, 0, MEM_RELEASE)) + { + DWORD err = GetLastError(); + std::cerr << "Failed to free transposition table. Error code: 0x" << + std::hex << err << std::dec << std::endl; + exit(EXIT_FAILURE); + } +} + +#else + +void aligned_large_pages_free(void *mem) { + std_aligned_free(mem); +} + +#endif + + namespace WinProcGroup { #ifndef _WIN32 @@ -315,3 +576,62 @@ void bindThisThread(size_t idx) { #endif } // namespace WinProcGroup + +#ifdef _WIN32 +#include +#define GETCWD _getcwd +#else +#include +#define GETCWD getcwd +#endif + +namespace CommandLine { + +string argv0; // path+name of the executable binary, as given by argv[0] +string binaryDirectory; // path of the executable directory +string workingDirectory; // path of the working directory + +void init(int argc, char* argv[]) { + (void)argc; + string pathSeparator; + + // extract the path+name of the executable binary + argv0 = argv[0]; + +#ifdef _WIN32 + pathSeparator = "\\"; + #ifdef _MSC_VER + // Under windows argv[0] may not have the extension. Also _get_pgmptr() had + // issues in some windows 10 versions, so check returned values carefully. + char* pgmptr = nullptr; + if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr) + argv0 = pgmptr; + #endif +#else + pathSeparator = "/"; +#endif + + // extract the working directory + workingDirectory = ""; + char buff[40000]; + char* cwd = GETCWD(buff, 40000); + if (cwd) + workingDirectory = cwd; + + // extract the binary directory path from argv0 + binaryDirectory = argv0; + size_t pos = binaryDirectory.find_last_of("\\/"); + if (pos == std::string::npos) + binaryDirectory = "." + pathSeparator; + else + binaryDirectory.resize(pos + 1); + + // pattern replacement: "./" at the start of path is replaced by the working directory + if (binaryDirectory.find("." + pathSeparator) == 0) + binaryDirectory.replace(0, 1, workingDirectory); +} + + +} // namespace CommandLine + +} // namespace Stockfish diff --git a/src/misc.h b/src/misc.h index ddd05e4e..f834e470 100644 --- a/src/misc.h +++ b/src/misc.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,12 +24,20 @@ #include #include #include +#include #include "types.h" -const std::string engine_info(bool to_uci = false); +namespace Stockfish { + +std::string engine_info(bool to_uci = false); +std::string compiler_info(); void prefetch(void* addr); void start_logger(const std::string& fname); +void* std_aligned_alloc(size_t alignment, size_t size); +void std_aligned_free(void* ptr); +void* aligned_large_pages_alloc(size_t size); // memory aligned by page size, min alignment: 4096 bytes +void aligned_large_pages_free(void* mem); // nop if mem == nullptr void dbg_hit_on(bool b); void dbg_hit_on(bool c, bool b); @@ -39,9 +45,7 @@ void dbg_mean_of(int v); void dbg_print(); typedef std::chrono::milliseconds::rep TimePoint; // A value in milliseconds - static_assert(sizeof(TimePoint) == sizeof(int64_t), "TimePoint should be 64 bits"); - inline TimePoint now() { return std::chrono::duration_cast (std::chrono::steady_clock::now().time_since_epoch()).count(); @@ -62,6 +66,17 @@ std::ostream& operator<<(std::ostream&, SyncCout); #define sync_cout std::cout << IO_LOCK #define sync_endl std::endl << IO_UNLOCK +// `ptr` must point to an array of size at least +// `sizeof(T) * N + alignment` bytes, where `N` is the +// number of elements in the array. +template +T* align_ptr_up(T* ptr) +{ + static_assert(alignof(T) < Alignment); + + const uintptr_t ptrint = reinterpret_cast(reinterpret_cast(ptr)); + return reinterpret_cast(reinterpret_cast((ptrint + (Alignment - 1)) / Alignment * Alignment)); +} /// xorshift64star Pseudo-Random Number Generator /// This class is based on original code written and dedicated @@ -99,6 +114,19 @@ public: { return T(rand64() & rand64() & rand64()); } }; +inline uint64_t mul_hi64(uint64_t a, uint64_t b) { +#if defined(__GNUC__) && defined(IS_64BIT) + __extension__ typedef unsigned __int128 uint128; + return ((uint128)a * (uint128)b) >> 64; +#else + uint64_t aL = (uint32_t)a, aH = a >> 32; + uint64_t bL = (uint32_t)b, bH = b >> 32; + uint64_t c1 = (aL * bL) >> 32; + uint64_t c2 = aH * bL + c1; + uint64_t c3 = aL * bH + (uint32_t)c2; + return aH * bH + (c2 >> 32) + (c3 >> 32); +#endif +} /// 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 @@ -110,4 +138,13 @@ namespace WinProcGroup { void bindThisThread(size_t idx); } +namespace CommandLine { + void init(int argc, char* argv[]); + + extern std::string binaryDirectory; // path of the executable directory + extern std::string workingDirectory; // path of the working directory +} + +} // namespace Stockfish + #endif // #ifndef MISC_H_INCLUDED diff --git a/src/movegen.cpp b/src/movegen.cpp index fc99ec26..c5d76afa 100644 --- a/src/movegen.cpp +++ b/src/movegen.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -23,28 +21,28 @@ #include "movegen.h" #include "position.h" +namespace Stockfish { + namespace { template ExtMove* make_promotions(ExtMove* moveList, Square to, Square ksq) { if (Type == CAPTURES || Type == EVASIONS || Type == NON_EVASIONS) + { *moveList++ = make(to - D, to, QUEEN); + if (attacks_bb(to) & ksq) + *moveList++ = make(to - D, to, KNIGHT); + } if (Type == QUIETS || Type == EVASIONS || Type == NON_EVASIONS) { *moveList++ = make(to - D, to, ROOK); *moveList++ = make(to - D, to, BISHOP); - *moveList++ = make(to - D, to, KNIGHT); + if (!(attacks_bb(to) & ksq)) + *moveList++ = make(to - D, to, KNIGHT); } - // Knight promotion is the only promotion that can give a direct check - // that's not already included in the queen promotion. - if (Type == QUIET_CHECKS && (PseudoAttacks[KNIGHT][to] & ksq)) - *moveList++ = make(to - D, to, KNIGHT); - else - (void)ksq; // Silence a warning under MSVC - return moveList; } @@ -52,14 +50,14 @@ namespace { template ExtMove* generate_pawn_moves(const Position& pos, ExtMove* moveList, Bitboard target) { - // Compute some compile time parameters relative to the white side - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); + constexpr Color Them = ~Us; constexpr Bitboard TRank7BB = (Us == WHITE ? Rank7BB : Rank2BB); constexpr Bitboard TRank3BB = (Us == WHITE ? Rank3BB : Rank6BB); - constexpr Direction Up = (Us == WHITE ? NORTH : SOUTH); + constexpr Direction Up = pawn_push(Us); constexpr Direction UpRight = (Us == WHITE ? NORTH_EAST : SOUTH_WEST); constexpr Direction UpLeft = (Us == WHITE ? NORTH_WEST : SOUTH_EAST); + const Square ksq = pos.square(Them); Bitboard emptySquares; Bitboard pawnsOn7 = pos.pieces(Us, PAWN) & TRank7BB; @@ -84,14 +82,12 @@ namespace { if (Type == QUIET_CHECKS) { - Square ksq = pos.square(Them); - - b1 &= pos.attacks_from(ksq, Them); - b2 &= pos.attacks_from(ksq, Them); + b1 &= pawn_attacks_bb(Them, ksq); + b2 &= pawn_attacks_bb(Them, ksq); // Add pawn pushes which give discovered check. This is possible only // if the pawn is not on the same file as the enemy king, because we - // don't generate captures. Note that a possible discovery check + // don't generate captures. Note that a possible discovered check // promotion has been already generated amongst the captures. Bitboard dcCandidateQuiets = pos.blockers_for_king(Them) & pawnsNotOn7; if (dcCandidateQuiets) @@ -130,8 +126,6 @@ namespace { Bitboard b2 = shift(pawnsOn7) & enemies; Bitboard b3 = shift(pawnsOn7) & emptySquares; - Square ksq = pos.square(Them); - while (b1) moveList = make_promotions(moveList, pop_lsb(&b1), ksq); @@ -142,7 +136,7 @@ namespace { moveList = make_promotions(moveList, pop_lsb(&b3), ksq); } - // Standard and en-passant captures + // Standard and en passant captures if (Type == CAPTURES || Type == EVASIONS || Type == NON_EVASIONS) { Bitboard b1 = shift(pawnsNotOn7) & enemies; @@ -164,18 +158,16 @@ namespace { { assert(rank_of(pos.ep_square()) == relative_rank(Us, RANK_6)); - // An en passant capture can be an evasion only if the checking piece - // is the double pushed pawn and so is in the target. Otherwise this - // is a discovery check and we are forced to do otherwise. - if (Type == EVASIONS && !(target & (pos.ep_square() - Up))) + // An en passant capture cannot resolve a discovered check. + if (Type == EVASIONS && (target & (pos.ep_square() + Up))) return moveList; - b1 = pawnsNotOn7 & pos.attacks_from(pos.ep_square(), Them); + b1 = pawnsNotOn7 & pawn_attacks_bb(Them, pos.ep_square()); assert(b1); while (b1) - *moveList++ = make(pop_lsb(&b1), pos.ep_square()); + *moveList++ = make(pop_lsb(&b1), pos.ep_square()); } } @@ -184,29 +176,23 @@ namespace { template - ExtMove* generate_moves(const Position& pos, ExtMove* moveList, Color us, - Bitboard target) { + ExtMove* generate_moves(const Position& pos, ExtMove* moveList, Bitboard piecesToMove, Bitboard target) { - assert(Pt != KING && Pt != PAWN); + static_assert(Pt != KING && Pt != PAWN, "Unsupported piece type in generate_moves()"); - const Square* pl = pos.squares(us); + Bitboard bb = piecesToMove & pos.pieces(Pt); - for (Square from = *pl; from != SQ_NONE; from = *++pl) - { - if (Checks) - { - if ( (Pt == BISHOP || Pt == ROOK || Pt == QUEEN) - && !(PseudoAttacks[Pt][from] & target & pos.check_squares(Pt))) - continue; + if (!bb) + return moveList; - if (pos.blockers_for_king(~us) & from) - continue; - } + [[maybe_unused]] const Bitboard checkSquares = pos.check_squares(Pt); - Bitboard b = pos.attacks_from(from) & target; + while (bb) { + Square from = pop_lsb(&bb); - if (Checks) - b &= pos.check_squares(Pt); + Bitboard b = attacks_bb(from, pos.pieces()) & target; + if constexpr (Checks) + b &= checkSquares; while (b) *moveList++ = make_move(from, pop_lsb(&b)); @@ -217,33 +203,50 @@ namespace { template - ExtMove* generate_all(const Position& pos, ExtMove* moveList, Bitboard target) { + ExtMove* generate_all(const Position& pos, ExtMove* moveList) { - constexpr CastlingRights OO = Us & KING_SIDE; - constexpr CastlingRights OOO = Us & QUEEN_SIDE; - constexpr bool Checks = Type == QUIET_CHECKS; // Reduce template instantations + static_assert(Type != LEGAL, "Unsupported type in generate_all()"); + + constexpr bool Checks = Type == QUIET_CHECKS; // Reduce template instantiations + Bitboard target, piecesToMove = pos.pieces(Us); + + if(Type == QUIET_CHECKS) + piecesToMove &= ~pos.blockers_for_king(~Us); + + switch (Type) + { + case CAPTURES: + target = pos.pieces(~Us); + break; + case QUIETS: + case QUIET_CHECKS: + target = ~pos.pieces(); + break; + case EVASIONS: + target = between_bb(pos.square(Us), lsb(pos.checkers())) | pos.checkers(); + break; + case NON_EVASIONS: + target = ~pos.pieces(Us); + break; + } moveList = generate_pawn_moves(pos, moveList, target); - moveList = generate_moves(pos, moveList, Us, target); - moveList = generate_moves(pos, moveList, Us, target); - moveList = generate_moves< ROOK, Checks>(pos, moveList, Us, target); - moveList = generate_moves< QUEEN, Checks>(pos, moveList, Us, target); + moveList = generate_moves(pos, moveList, piecesToMove, target); + moveList = generate_moves(pos, moveList, piecesToMove, target); + moveList = generate_moves< ROOK, Checks>(pos, moveList, piecesToMove, target); + moveList = generate_moves< QUEEN, Checks>(pos, moveList, piecesToMove, target); if (Type != QUIET_CHECKS && Type != EVASIONS) { Square ksq = pos.square(Us); - Bitboard b = pos.attacks_from(ksq) & target; + Bitboard b = attacks_bb(ksq) & target; while (b) *moveList++ = make_move(ksq, pop_lsb(&b)); - if (Type != CAPTURES && pos.can_castle(CastlingRights(OO | OOO))) - { - if (!pos.castling_impeded(OO) && pos.can_castle(OO)) - *moveList++ = make(ksq, pos.castling_rook_square(OO)); - - if (!pos.castling_impeded(OOO) && pos.can_castle(OOO)) - *moveList++ = make(ksq, pos.castling_rook_square(OOO)); - } + if ((Type != CAPTURES) && pos.can_castle(Us & ANY_CASTLING)) + for (CastlingRights cr : { Us & KING_SIDE, Us & QUEEN_SIDE } ) + if (!pos.castling_impeded(cr) && pos.can_castle(cr)) + *moveList++ = make(ksq, pos.castling_rook_square(cr)); } return moveList; @@ -252,8 +255,8 @@ namespace { } // namespace -/// Generates all pseudo-legal captures and queen promotions -/// Generates all pseudo-legal non-captures and underpromotions +/// Generates all pseudo-legal captures plus queen and checking knight promotions +/// Generates all pseudo-legal non-captures and underpromotions (except checking knight) /// Generates all pseudo-legal captures and non-captures /// /// Returns a pointer to the end of the move list. @@ -261,17 +264,13 @@ namespace { template ExtMove* generate(const Position& pos, ExtMove* moveList) { - assert(Type == CAPTURES || Type == QUIETS || Type == NON_EVASIONS); + static_assert(Type == CAPTURES || Type == QUIETS || Type == NON_EVASIONS, "Unsupported type in generate()"); assert(!pos.checkers()); Color us = pos.side_to_move(); - Bitboard target = Type == CAPTURES ? pos.pieces(~us) - : Type == QUIETS ? ~pos.pieces() - : Type == NON_EVASIONS ? ~pos.pieces(us) : 0; - - return us == WHITE ? generate_all(pos, moveList, target) - : generate_all(pos, moveList, target); + return us == WHITE ? generate_all(pos, moveList) + : generate_all(pos, moveList); } // Explicit template instantiations @@ -280,35 +279,32 @@ template ExtMove* generate(const Position&, ExtMove*); template ExtMove* generate(const Position&, ExtMove*); -/// generate generates all pseudo-legal non-captures and knight -/// underpromotions that give check. Returns a pointer to the end of the move list. +/// generate generates all pseudo-legal non-captures giving check, +/// except castling. Returns a pointer to the end of the move list. template<> ExtMove* generate(const Position& pos, ExtMove* moveList) { assert(!pos.checkers()); Color us = pos.side_to_move(); - Bitboard dc = pos.blockers_for_king(~us) & pos.pieces(us); + Bitboard dc = pos.blockers_for_king(~us) & pos.pieces(us) & ~pos.pieces(PAWN); while (dc) { Square from = pop_lsb(&dc); PieceType pt = type_of(pos.piece_on(from)); - if (pt == PAWN) - continue; // Will be generated together with direct checks - - Bitboard b = pos.attacks_from(pt, from) & ~pos.pieces(); + Bitboard b = attacks_bb(pt, from, pos.pieces()) & ~pos.pieces(); if (pt == KING) - b &= ~PseudoAttacks[QUEEN][pos.square(~us)]; + b &= ~attacks_bb(pos.square(~us)); while (b) *moveList++ = make_move(from, pop_lsb(&b)); } - return us == WHITE ? generate_all(pos, moveList, ~pos.pieces()) - : generate_all(pos, moveList, ~pos.pieces()); + return us == WHITE ? generate_all(pos, moveList) + : generate_all(pos, moveList); } @@ -328,13 +324,10 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) { // the king evasions in order to skip known illegal moves, which avoids any // useless legality checks later on. while (sliders) - { - Square checksq = pop_lsb(&sliders); - sliderAttacks |= LineBB[checksq][ksq] ^ checksq; - } + sliderAttacks |= line_bb(ksq, pop_lsb(&sliders)) & ~pos.checkers(); // Generate evasions for king, capture and non capture moves - Bitboard b = pos.attacks_from(ksq) & ~pos.pieces(us) & ~sliderAttacks; + Bitboard b = attacks_bb(ksq) & ~pos.pieces(us) & ~sliderAttacks; while (b) *moveList++ = make_move(ksq, pop_lsb(&b)); @@ -342,11 +335,8 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) { return moveList; // Double check, only a king move can save the day // Generate blocking evasions or captures of the checking piece - Square checksq = lsb(pos.checkers()); - Bitboard target = between_bb(checksq, ksq) | checksq; - - return us == WHITE ? generate_all(pos, moveList, target) - : generate_all(pos, moveList, target); + return us == WHITE ? generate_all(pos, moveList) + : generate_all(pos, moveList); } @@ -363,7 +353,7 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) { moveList = pos.checkers() ? generate(pos, moveList) : generate(pos, moveList); while (cur != moveList) - if ( (pinned || from_sq(*cur) == ksq || type_of(*cur) == ENPASSANT) + if ( ((pinned && pinned & from_sq(*cur)) || from_sq(*cur) == ksq || type_of(*cur) == EN_PASSANT) && !pos.legal(*cur)) *cur = (--moveList)->move; else @@ -371,3 +361,5 @@ ExtMove* generate(const Position& pos, ExtMove* moveList) { return moveList; } + +} // namespace Stockfish diff --git a/src/movegen.h b/src/movegen.h index aeba93ad..3f895f05 100644 --- a/src/movegen.h +++ b/src/movegen.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,6 +23,8 @@ #include "types.h" +namespace Stockfish { + class Position; enum GenType { @@ -72,4 +72,6 @@ private: ExtMove moveList[MAX_MOVES], *last; }; +} // namespace Stockfish + #endif // #ifndef MOVEGEN_H_INCLUDED diff --git a/src/movepick.cpp b/src/movepick.cpp index e39f2afa..4ff4cff4 100644 --- a/src/movepick.cpp +++ b/src/movepick.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -22,6 +20,8 @@ #include "movepick.h" +namespace Stockfish { + namespace { enum Stages { @@ -56,45 +56,40 @@ namespace { /// ordering is at the current node. /// MovePicker constructor for the main search -MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh, - const CapturePieceToHistory* cph, const PieceToHistory** ch, Move cm, Move* killers) - : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch), - refutations{{killers[0], 0}, {killers[1], 0}, {cm, 0}}, depth(d) { +MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh, const LowPlyHistory* lp, + const CapturePieceToHistory* cph, const PieceToHistory** ch, Move cm, const Move* killers, int pl) + : pos(p), mainHistory(mh), lowPlyHistory(lp), captureHistory(cph), continuationHistory(ch), + ttMove(ttm), refutations{{killers[0], 0}, {killers[1], 0}, {cm, 0}}, depth(d), ply(pl) { assert(d > 0); - stage = pos.checkers() ? EVASION_TT : MAIN_TT; - ttMove = ttm && pos.pseudo_legal(ttm) ? ttm : MOVE_NONE; - stage += (ttMove == MOVE_NONE); + stage = (pos.checkers() ? EVASION_TT : MAIN_TT) + + !(ttm && pos.pseudo_legal(ttm)); } /// MovePicker constructor for quiescence search MovePicker::MovePicker(const Position& p, Move ttm, Depth d, const ButterflyHistory* mh, const CapturePieceToHistory* cph, const PieceToHistory** ch, Square rs) - : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch), recaptureSquare(rs), depth(d) { + : pos(p), mainHistory(mh), captureHistory(cph), continuationHistory(ch), ttMove(ttm), recaptureSquare(rs), depth(d) { assert(d <= 0); - stage = pos.checkers() ? EVASION_TT : QSEARCH_TT; - ttMove = ttm - && (depth > DEPTH_QS_RECAPTURES || to_sq(ttm) == recaptureSquare) - && pos.pseudo_legal(ttm) ? ttm : MOVE_NONE; - stage += (ttMove == MOVE_NONE); + stage = (pos.checkers() ? EVASION_TT : QSEARCH_TT) + + !( ttm + && (pos.checkers() || depth > DEPTH_QS_RECAPTURES || to_sq(ttm) == recaptureSquare) + && pos.pseudo_legal(ttm)); } /// MovePicker constructor for ProbCut: we generate captures with SEE greater /// than or equal to the given threshold. MovePicker::MovePicker(const Position& p, Move ttm, Value th, const CapturePieceToHistory* cph) - : pos(p), captureHistory(cph), threshold(th) { + : pos(p), captureHistory(cph), ttMove(ttm), threshold(th) { assert(!pos.checkers()); - stage = PROBCUT_TT; - ttMove = ttm - && pos.capture(ttm) - && pos.pseudo_legal(ttm) - && pos.see_ge(ttm, threshold) ? ttm : MOVE_NONE; - stage += (ttMove == MOVE_NONE); + stage = PROBCUT_TT + !(ttm && pos.capture(ttm) + && pos.pseudo_legal(ttm) + && pos.see_ge(ttm, threshold)); } /// MovePicker::score() assigns a numerical value to each move in a list, used @@ -106,16 +101,17 @@ void MovePicker::score() { static_assert(Type == CAPTURES || Type == QUIETS || Type == EVASIONS, "Wrong type"); for (auto& m : *this) - if (Type == CAPTURES) + if constexpr (Type == CAPTURES) m.value = int(PieceValue[MG][pos.piece_on(to_sq(m))]) * 6 + (*captureHistory)[pos.moved_piece(m)][to_sq(m)][type_of(pos.piece_on(to_sq(m)))]; - else if (Type == QUIETS) + else if constexpr (Type == QUIETS) m.value = (*mainHistory)[pos.side_to_move()][from_to(m)] + 2 * (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)] - + 2 * (*continuationHistory[1])[pos.moved_piece(m)][to_sq(m)] - + 2 * (*continuationHistory[3])[pos.moved_piece(m)][to_sq(m)] - + (*continuationHistory[5])[pos.moved_piece(m)][to_sq(m)]; + + (*continuationHistory[1])[pos.moved_piece(m)][to_sq(m)] + + (*continuationHistory[3])[pos.moved_piece(m)][to_sq(m)] + + (*continuationHistory[5])[pos.moved_piece(m)][to_sq(m)] + + (ply < MAX_LPH ? std::min(4, depth / 3) * (*lowPlyHistory)[ply][from_to(m)] : 0); else // Type == EVASIONS { @@ -123,8 +119,8 @@ void MovePicker::score() { m.value = PieceValue[MG][pos.piece_on(to_sq(m))] - Value(type_of(pos.moved_piece(m))); else - m.value = (*mainHistory)[pos.side_to_move()][from_to(m)] - + (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)] + m.value = (*mainHistory)[pos.side_to_move()][from_to(m)] + + 2 * (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)] - (1 << 28); } } @@ -148,7 +144,7 @@ Move MovePicker::select(Pred filter) { } /// MovePicker::next_move() is the most important method of the MovePicker class. It -/// returns a new pseudo legal move every time it is called until there are no more +/// returns a new pseudo-legal move every time it is called until there are no more /// moves left, picking the move with the highest score from a list of generated moves. Move MovePicker::next_move(bool skipQuiets) { @@ -174,7 +170,7 @@ top: case GOOD_CAPTURE: if (select([&](){ - return pos.see_ge(*cur, Value(-55 * cur->value / 1024)) ? + return pos.see_ge(*cur, Value(-69 * cur->value / 1024)) ? // Move losing capture to endBadCaptures to be tried later true : (*endBadCaptures++ = *cur, false); })) return *(cur - 1); @@ -189,7 +185,7 @@ top: --endMoves; ++stage; - /* fallthrough */ + [[fallthrough]]; case REFUTATION: if (select([&](){ return *cur != MOVE_NONE @@ -197,7 +193,7 @@ top: && pos.pseudo_legal(*cur); })) return *(cur - 1); ++stage; - /* fallthrough */ + [[fallthrough]]; case QUIET_INIT: if (!skipQuiets) @@ -210,7 +206,7 @@ top: } ++stage; - /* fallthrough */ + [[fallthrough]]; case QUIET: if ( !skipQuiets @@ -224,7 +220,7 @@ top: endMoves = endBadCaptures; ++stage; - /* fallthrough */ + [[fallthrough]]; case BAD_CAPTURE: return select([](){ return true; }); @@ -235,7 +231,7 @@ top: score(); ++stage; - /* fallthrough */ + [[fallthrough]]; case EVASION: return select([](){ return true; }); @@ -253,14 +249,14 @@ top: return MOVE_NONE; ++stage; - /* fallthrough */ + [[fallthrough]]; case QCHECK_INIT: cur = moves; endMoves = generate(pos, cur); ++stage; - /* fallthrough */ + [[fallthrough]]; case QCHECK: return select([](){ return true; }); @@ -269,3 +265,5 @@ top: assert(false); return MOVE_NONE; // Silence warning } + +} // namespace Stockfish diff --git a/src/movepick.h b/src/movepick.h index 105c95d7..c76d4957 100644 --- a/src/movepick.h +++ b/src/movepick.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -29,6 +27,8 @@ #include "position.h" #include "types.h" +namespace Stockfish { + /// StatsEntry stores the stat table value. It is usually a number but could /// be a move or even a nested history. We use a class instead of naked value /// to directly call history update operator<<() on the entry so to use stats @@ -86,7 +86,13 @@ enum StatsType { NoCaptures, Captures }; /// unsuccessful during the current search, and is used for reduction and move /// ordering decisions. It uses 2 tables (one for each color) indexed by /// the move's from and to squares, see www.chessprogramming.org/Butterfly_Boards -typedef Stats ButterflyHistory; +typedef Stats ButterflyHistory; + +/// At higher depths LowPlyHistory records successful quiet moves near the root +/// and quiet moves which are/were in the PV (ttPv). It is cleared with each new +/// search and filled during iterative deepening. +constexpr int MAX_LPH = 4; +typedef Stats LowPlyHistory; /// CounterMoveHistory stores counter moves indexed by [piece][to] of the previous /// move, see www.chessprogramming.org/Countermove_Heuristic @@ -104,12 +110,12 @@ typedef Stats PieceToHistory; typedef Stats ContinuationHistory; -/// MovePicker class is used to pick one pseudo legal move at a time from the +/// MovePicker class is used to pick one pseudo-legal move at a time from the /// current position. The most important method is next_move(), which returns a -/// new pseudo legal move each time it is called, until there are no moves left, -/// when MOVE_NONE is returned. In order to improve the efficiency of the alpha -/// beta algorithm, MovePicker attempts to return the moves which are most likely -/// to get a cut-off first. +/// new pseudo-legal move each time it is called, until there are no moves left, +/// when MOVE_NONE is returned. In order to improve the efficiency of the +/// alpha-beta algorithm, MovePicker attempts to return the moves which are most +/// likely to get a cut-off first. class MovePicker { enum PickType { Next, Best }; @@ -123,10 +129,12 @@ public: const PieceToHistory**, Square); MovePicker(const Position&, Move, Depth, const ButterflyHistory*, + const LowPlyHistory*, const CapturePieceToHistory*, const PieceToHistory**, Move, - Move*); + const Move*, + int); Move next_move(bool skipQuiets = false); private: @@ -137,6 +145,7 @@ private: const Position& pos; const ButterflyHistory* mainHistory; + const LowPlyHistory* lowPlyHistory; const CapturePieceToHistory* captureHistory; const PieceToHistory** continuationHistory; Move ttMove; @@ -145,7 +154,10 @@ private: Square recaptureSquare; Value threshold; Depth depth; + int ply; ExtMove moves[MAX_MOVES]; }; +} // namespace Stockfish + #endif // #ifndef MOVEPICK_H_INCLUDED diff --git a/src/nnue/architectures/halfkp_256x2-32-32.h b/src/nnue/architectures/halfkp_256x2-32-32.h new file mode 100644 index 00000000..a6768204 --- /dev/null +++ b/src/nnue/architectures/halfkp_256x2-32-32.h @@ -0,0 +1,54 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Definition of input features and network structure used in NNUE evaluation function + +#ifndef NNUE_HALFKP_256X2_32_32_H_INCLUDED +#define NNUE_HALFKP_256X2_32_32_H_INCLUDED + +#include "../features/feature_set.h" +#include "../features/half_kp.h" + +#include "../layers/input_slice.h" +#include "../layers/affine_transform.h" +#include "../layers/clipped_relu.h" + +namespace Stockfish::Eval::NNUE { + +// Input features used in evaluation function +using RawFeatures = Features::FeatureSet< + Features::HalfKP>; + +// Number of input feature dimensions after conversion +constexpr IndexType kTransformedFeatureDimensions = 256; + +namespace Layers { + +// Define network structure +using InputLayer = InputSlice; +using HiddenLayer1 = ClippedReLU>; +using HiddenLayer2 = ClippedReLU>; +using OutputLayer = AffineTransform; + +} // namespace Layers + +using Network = Layers::OutputLayer; + +} // namespace Stockfish::Eval::NNUE + +#endif // #ifndef NNUE_HALFKP_256X2_32_32_H_INCLUDED diff --git a/src/nnue/evaluate_nnue.cpp b/src/nnue/evaluate_nnue.cpp new file mode 100644 index 00000000..5416f13e --- /dev/null +++ b/src/nnue/evaluate_nnue.cpp @@ -0,0 +1,144 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Code for calculating NNUE evaluation function + +#include +#include + +#include "../evaluate.h" +#include "../position.h" +#include "../misc.h" +#include "../uci.h" +#include "../types.h" + +#include "evaluate_nnue.h" + +namespace Stockfish::Eval::NNUE { + + // Input feature converter + LargePagePtr feature_transformer; + + // Evaluation function + AlignedPtr network; + + // Evaluation function file name + std::string fileName; + + namespace Detail { + + // Initialize the evaluation function parameters + template + void Initialize(AlignedPtr& pointer) { + + pointer.reset(reinterpret_cast(std_aligned_alloc(alignof(T), sizeof(T)))); + std::memset(pointer.get(), 0, sizeof(T)); + } + + template + void Initialize(LargePagePtr& pointer) { + + static_assert(alignof(T) <= 4096, "aligned_large_pages_alloc() may fail for such a big alignment requirement of T"); + pointer.reset(reinterpret_cast(aligned_large_pages_alloc(sizeof(T)))); + std::memset(pointer.get(), 0, sizeof(T)); + } + + // Read evaluation function parameters + template + bool ReadParameters(std::istream& stream, T& reference) { + + std::uint32_t header; + header = read_little_endian(stream); + if (!stream || header != T::GetHashValue()) return false; + return reference.ReadParameters(stream); + } + + } // namespace Detail + + // Initialize the evaluation function parameters + void Initialize() { + + Detail::Initialize(feature_transformer); + Detail::Initialize(network); + } + + // Read network header + bool ReadHeader(std::istream& stream, std::uint32_t* hash_value, std::string* architecture) + { + std::uint32_t version, size; + + version = read_little_endian(stream); + *hash_value = read_little_endian(stream); + size = read_little_endian(stream); + if (!stream || version != kVersion) return false; + architecture->resize(size); + stream.read(&(*architecture)[0], size); + return !stream.fail(); + } + + // Read network parameters + bool ReadParameters(std::istream& stream) { + + std::uint32_t hash_value; + std::string architecture; + if (!ReadHeader(stream, &hash_value, &architecture)) return false; + if (hash_value != kHashValue) return false; + if (!Detail::ReadParameters(stream, *feature_transformer)) return false; + if (!Detail::ReadParameters(stream, *network)) return false; + return stream && stream.peek() == std::ios::traits_type::eof(); + } + + // Evaluation function. Perform differential calculation. + Value evaluate(const Position& pos) { + + // We manually align the arrays on the stack because with gcc < 9.3 + // overaligning stack variables with alignas() doesn't work correctly. + + constexpr uint64_t alignment = kCacheLineSize; + +#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN) + TransformedFeatureType transformed_features_unaligned[ + FeatureTransformer::kBufferSize + alignment / sizeof(TransformedFeatureType)]; + char buffer_unaligned[Network::kBufferSize + alignment]; + + auto* transformed_features = align_ptr_up(&transformed_features_unaligned[0]); + auto* buffer = align_ptr_up(&buffer_unaligned[0]); +#else + alignas(alignment) + TransformedFeatureType transformed_features[FeatureTransformer::kBufferSize]; + alignas(alignment) char buffer[Network::kBufferSize]; +#endif + + ASSERT_ALIGNED(transformed_features, alignment); + ASSERT_ALIGNED(buffer, alignment); + + feature_transformer->Transform(pos, transformed_features); + const auto output = network->Propagate(transformed_features, buffer); + + return static_cast(output[0] / FV_SCALE); + } + + // Load eval, from a file stream or a memory stream + bool load_eval(std::string name, std::istream& stream) { + + Initialize(); + fileName = name; + return ReadParameters(stream); + } + +} // namespace Stockfish::Eval::NNUE diff --git a/src/nnue/evaluate_nnue.h b/src/nnue/evaluate_nnue.h new file mode 100644 index 00000000..24aa6cc0 --- /dev/null +++ b/src/nnue/evaluate_nnue.h @@ -0,0 +1,59 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// header used in NNUE evaluation function + +#ifndef NNUE_EVALUATE_NNUE_H_INCLUDED +#define NNUE_EVALUATE_NNUE_H_INCLUDED + +#include "nnue_feature_transformer.h" + +#include + +namespace Stockfish::Eval::NNUE { + + // Hash value of evaluation function structure + constexpr std::uint32_t kHashValue = + FeatureTransformer::GetHashValue() ^ Network::GetHashValue(); + + // Deleter for automating release of memory area + template + struct AlignedDeleter { + void operator()(T* ptr) const { + ptr->~T(); + std_aligned_free(ptr); + } + }; + + template + struct LargePageDeleter { + void operator()(T* ptr) const { + ptr->~T(); + aligned_large_pages_free(ptr); + } + }; + + template + using AlignedPtr = std::unique_ptr>; + + template + using LargePagePtr = std::unique_ptr>; + +} // namespace Stockfish::Eval::NNUE + +#endif // #ifndef NNUE_EVALUATE_NNUE_H_INCLUDED diff --git a/src/nnue/features/feature_set.h b/src/nnue/features/feature_set.h new file mode 100644 index 00000000..a3fea9c0 --- /dev/null +++ b/src/nnue/features/feature_set.h @@ -0,0 +1,69 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// A class template that represents the input feature set of the NNUE evaluation function + +#ifndef NNUE_FEATURE_SET_H_INCLUDED +#define NNUE_FEATURE_SET_H_INCLUDED + +#include "features_common.h" +#include + +namespace Stockfish::Eval::NNUE::Features { + + // Class template that represents a list of values + template + struct CompileTimeList; + + template + struct CompileTimeList { + static constexpr bool Contains(T value) { + return value == First || CompileTimeList::Contains(value); + } + static constexpr std::array + kValues = {{First, Remaining...}}; + }; + + // Base class of feature set + template + class FeatureSetBase { + + }; + + // Class template that represents the feature set + template + class FeatureSet : public FeatureSetBase> { + + public: + // Hash value embedded in the evaluation file + static constexpr std::uint32_t kHashValue = FeatureType::kHashValue; + // Number of feature dimensions + static constexpr IndexType kDimensions = FeatureType::kDimensions; + // Maximum number of simultaneously active features + static constexpr IndexType kMaxActiveDimensions = + FeatureType::kMaxActiveDimensions; + // Trigger for full calculation instead of difference calculation + using SortedTriggerSet = + CompileTimeList; + static constexpr auto kRefreshTriggers = SortedTriggerSet::kValues; + + }; + +} // namespace Stockfish::Eval::NNUE::Features + +#endif // #ifndef NNUE_FEATURE_SET_H_INCLUDED diff --git a/src/nnue/features/features_common.h b/src/nnue/features/features_common.h new file mode 100644 index 00000000..118ec953 --- /dev/null +++ b/src/nnue/features/features_common.h @@ -0,0 +1,45 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +//Common header of input features of NNUE evaluation function + +#ifndef NNUE_FEATURES_COMMON_H_INCLUDED +#define NNUE_FEATURES_COMMON_H_INCLUDED + +#include "../../evaluate.h" +#include "../nnue_common.h" + +namespace Stockfish::Eval::NNUE::Features { + + class IndexList; + + template + class FeatureSet; + + // Trigger to perform full calculations instead of difference only + enum class TriggerEvent { + kFriendKingMoved // calculate full evaluation when own king moves + }; + + enum class Side { + kFriend // side to move + }; + +} // namespace Stockfish::Eval::NNUE::Features + +#endif // #ifndef NNUE_FEATURES_COMMON_H_INCLUDED diff --git a/src/nnue/features/half_kp.cpp b/src/nnue/features/half_kp.cpp new file mode 100644 index 00000000..ac6317e7 --- /dev/null +++ b/src/nnue/features/half_kp.cpp @@ -0,0 +1,68 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +//Definition of input features HalfKP of NNUE evaluation function + +#include "half_kp.h" +#include "index_list.h" + +namespace Stockfish::Eval::NNUE::Features { + + // Orient a square according to perspective (rotates by 180 for black) + inline Square orient(Color perspective, Square s) { + return Square(int(s) ^ (bool(perspective) * 63)); + } + + // Index of a feature for a given king position and another piece on some square + inline IndexType make_index(Color perspective, Square s, Piece pc, Square ksq) { + return IndexType(orient(perspective, s) + kpp_board_index[perspective][pc] + PS_END * ksq); + } + + // Get a list of indices for active features + template + void HalfKP::AppendActiveIndices( + const Position& pos, Color perspective, IndexList* active) { + + Square ksq = orient(perspective, pos.square(perspective)); + Bitboard bb = pos.pieces() & ~pos.pieces(KING); + while (bb) { + Square s = pop_lsb(&bb); + active->push_back(make_index(perspective, s, pos.piece_on(s), ksq)); + } + } + + // Get a list of indices for recently changed features + template + void HalfKP::AppendChangedIndices( + const Position& pos, const DirtyPiece& dp, Color perspective, + IndexList* removed, IndexList* added) { + + Square ksq = orient(perspective, pos.square(perspective)); + for (int i = 0; i < dp.dirty_num; ++i) { + Piece pc = dp.piece[i]; + if (type_of(pc) == KING) continue; + if (dp.from[i] != SQ_NONE) + removed->push_back(make_index(perspective, dp.from[i], pc, ksq)); + if (dp.to[i] != SQ_NONE) + added->push_back(make_index(perspective, dp.to[i], pc, ksq)); + } + } + + template class HalfKP; + +} // namespace Stockfish::Eval::NNUE::Features diff --git a/src/nnue/features/half_kp.h b/src/nnue/features/half_kp.h new file mode 100644 index 00000000..2461acb7 --- /dev/null +++ b/src/nnue/features/half_kp.h @@ -0,0 +1,59 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +//Definition of input features HalfKP of NNUE evaluation function + +#ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED +#define NNUE_FEATURES_HALF_KP_H_INCLUDED + +#include "../../evaluate.h" +#include "features_common.h" + +namespace Stockfish::Eval::NNUE::Features { + + // Feature HalfKP: Combination of the position of own king + // and the position of pieces other than kings + template + class HalfKP { + + public: + // Feature name + static constexpr const char* kName = "HalfKP(Friend)"; + // Hash value embedded in the evaluation file + static constexpr std::uint32_t kHashValue = + 0x5D69D5B9u ^ (AssociatedKing == Side::kFriend); + // Number of feature dimensions + static constexpr IndexType kDimensions = + static_cast(SQUARE_NB) * static_cast(PS_END); + // Maximum number of simultaneously active features + static constexpr IndexType kMaxActiveDimensions = 30; // Kings don't count + // Trigger for full calculation instead of difference calculation + static constexpr TriggerEvent kRefreshTrigger = TriggerEvent::kFriendKingMoved; + + // Get a list of indices for active features + static void AppendActiveIndices(const Position& pos, Color perspective, + IndexList* active); + + // Get a list of indices for recently changed features + static void AppendChangedIndices(const Position& pos, const DirtyPiece& dp, Color perspective, + IndexList* removed, IndexList* added); + }; + +} // namespace Stockfish::Eval::NNUE::Features + +#endif // #ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED diff --git a/src/nnue/features/index_list.h b/src/nnue/features/index_list.h new file mode 100644 index 00000000..9f03993b --- /dev/null +++ b/src/nnue/features/index_list.h @@ -0,0 +1,64 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Definition of index list of input features + +#ifndef NNUE_FEATURES_INDEX_LIST_H_INCLUDED +#define NNUE_FEATURES_INDEX_LIST_H_INCLUDED + +#include "../../position.h" +#include "../nnue_architecture.h" + +namespace Stockfish::Eval::NNUE::Features { + + // Class template used for feature index list + template + class ValueList { + + public: + std::size_t size() const { return size_; } + void resize(std::size_t size) { size_ = size; } + void push_back(const T& value) { values_[size_++] = value; } + T& operator[](std::size_t index) { return values_[index]; } + T* begin() { return values_; } + T* end() { return values_ + size_; } + const T& operator[](std::size_t index) const { return values_[index]; } + const T* begin() const { return values_; } + const T* end() const { return values_ + size_; } + + void swap(ValueList& other) { + const std::size_t max_size = std::max(size_, other.size_); + for (std::size_t i = 0; i < max_size; ++i) { + std::swap(values_[i], other.values_[i]); + } + std::swap(size_, other.size_); + } + + private: + T values_[MaxSize]; + std::size_t size_ = 0; + }; + + //Type of feature index list + class IndexList + : public ValueList { + }; + +} // namespace Stockfish::Eval::NNUE::Features + +#endif // NNUE_FEATURES_INDEX_LIST_H_INCLUDED diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h new file mode 100644 index 00000000..d2713c5a --- /dev/null +++ b/src/nnue/layers/affine_transform.h @@ -0,0 +1,464 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Definition of layer AffineTransform of NNUE evaluation function + +#ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED +#define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED + +#include +#include "../nnue_common.h" + +namespace Stockfish::Eval::NNUE::Layers { + + // Affine transformation layer + template + class AffineTransform { + public: + // Input/output type + using InputType = typename PreviousLayer::OutputType; + using OutputType = std::int32_t; + static_assert(std::is_same::value, ""); + + // Number of input/output dimensions + static constexpr IndexType kInputDimensions = + PreviousLayer::kOutputDimensions; + static constexpr IndexType kOutputDimensions = OutputDimensions; + static constexpr IndexType kPaddedInputDimensions = + CeilToMultiple(kInputDimensions, kMaxSimdWidth); +#if defined (USE_AVX512) + static constexpr const IndexType kOutputSimdWidth = kSimdWidth / 2; +#elif defined (USE_SSSE3) + static constexpr const IndexType kOutputSimdWidth = kSimdWidth / 4; +#endif + + // Size of forward propagation buffer used in this layer + static constexpr std::size_t kSelfBufferSize = + CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize); + + // Size of the forward propagation buffer used from the input layer to this layer + static constexpr std::size_t kBufferSize = + PreviousLayer::kBufferSize + kSelfBufferSize; + + // Hash value embedded in the evaluation file + static constexpr std::uint32_t GetHashValue() { + std::uint32_t hash_value = 0xCC03DAE4u; + hash_value += kOutputDimensions; + hash_value ^= PreviousLayer::GetHashValue() >> 1; + hash_value ^= PreviousLayer::GetHashValue() << 31; + return hash_value; + } + + // Read network parameters + bool ReadParameters(std::istream& stream) { + if (!previous_layer_.ReadParameters(stream)) return false; + for (std::size_t i = 0; i < kOutputDimensions; ++i) + biases_[i] = read_little_endian(stream); + for (std::size_t i = 0; i < kOutputDimensions * kPaddedInputDimensions; ++i) +#if !defined (USE_SSSE3) + weights_[i] = read_little_endian(stream); +#else + weights_[ + (i / 4) % (kPaddedInputDimensions / 4) * kOutputDimensions * 4 + + i / kPaddedInputDimensions * 4 + + i % 4 + ] = read_little_endian(stream); + + // Determine if eights of weight and input products can be summed using 16bits + // without saturation. We assume worst case combinations of 0 and 127 for all inputs. + if (kOutputDimensions > 1 && !stream.fail()) + { + canSaturate16.count = 0; +#if !defined(USE_VNNI) + for (IndexType i = 0; i < kPaddedInputDimensions; i += 16) + for (IndexType j = 0; j < kOutputDimensions; ++j) + for (int x = 0; x < 2; ++x) + { + WeightType* w = &weights_[i * kOutputDimensions + j * 4 + x * 2]; + int sum[2] = {0, 0}; + for (int k = 0; k < 8; ++k) + { + IndexType idx = k / 2 * kOutputDimensions * 4 + k % 2; + sum[w[idx] < 0] += w[idx]; + } + for (int sign : {-1, 1}) + while (sign * sum[sign == -1] > 258) + { + int maxK = 0, maxW = 0; + for (int k = 0; k < 8; ++k) + { + IndexType idx = k / 2 * kOutputDimensions * 4 + k % 2; + if (maxW < sign * w[idx]) + maxK = k, maxW = sign * w[idx]; + } + + IndexType idx = maxK / 2 * kOutputDimensions * 4 + maxK % 2; + sum[sign == -1] -= w[idx]; + canSaturate16.add(j, i + maxK / 2 * 4 + maxK % 2 + x * 2, w[idx]); + w[idx] = 0; + } + } + + // Non functional optimization for faster more linear access + std::sort(canSaturate16.ids, canSaturate16.ids + canSaturate16.count, + [](const typename CanSaturate::Entry& e1, const typename CanSaturate::Entry& e2) + { return e1.in == e2.in ? e1.out < e2.out : e1.in < e2.in; }); +#endif + } +#endif + + return !stream.fail(); + } + + // Forward propagation + const OutputType* Propagate( + const TransformedFeatureType* transformed_features, char* buffer) const { + const auto input = previous_layer_.Propagate( + transformed_features, buffer + kSelfBufferSize); + +#if defined (USE_AVX512) + + [[maybe_unused]] const __m512i kOnes512 = _mm512_set1_epi16(1); + + [[maybe_unused]] auto m512_hadd = [](__m512i sum, int bias) -> int { + return _mm512_reduce_add_epi32(sum) + bias; + }; + + [[maybe_unused]] auto m512_add_dpbusd_epi32 = [=](__m512i& acc, __m512i a, __m512i b) { +#if defined (USE_VNNI) + acc = _mm512_dpbusd_epi32(acc, a, b); +#else + __m512i product0 = _mm512_maddubs_epi16(a, b); + product0 = _mm512_madd_epi16(product0, kOnes512); + acc = _mm512_add_epi32(acc, product0); +#endif + }; + + [[maybe_unused]] auto m512_add_dpbusd_epi32x4 = [=](__m512i& acc, __m512i a0, __m512i b0, __m512i a1, __m512i b1, + __m512i a2, __m512i b2, __m512i a3, __m512i b3) { +#if defined (USE_VNNI) + acc = _mm512_dpbusd_epi32(acc, a0, b0); + acc = _mm512_dpbusd_epi32(acc, a1, b1); + acc = _mm512_dpbusd_epi32(acc, a2, b2); + acc = _mm512_dpbusd_epi32(acc, a3, b3); +#else + __m512i product0 = _mm512_maddubs_epi16(a0, b0); + __m512i product1 = _mm512_maddubs_epi16(a1, b1); + __m512i product2 = _mm512_maddubs_epi16(a2, b2); + __m512i product3 = _mm512_maddubs_epi16(a3, b3); + product0 = _mm512_add_epi16(product0, product1); + product2 = _mm512_add_epi16(product2, product3); + product0 = _mm512_add_epi16(product0, product2); + product0 = _mm512_madd_epi16(product0, kOnes512); + acc = _mm512_add_epi32(acc, product0); +#endif + }; + +#endif +#if defined (USE_AVX2) + + [[maybe_unused]] const __m256i kOnes256 = _mm256_set1_epi16(1); + + [[maybe_unused]] auto m256_hadd = [](__m256i sum, int bias) -> int { + __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1)); + sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC)); + sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB)); + return _mm_cvtsi128_si32(sum128) + bias; + }; + + [[maybe_unused]] auto m256_add_dpbusd_epi32 = [=](__m256i& acc, __m256i a, __m256i b) { +#if defined (USE_VNNI) + acc = _mm256_dpbusd_epi32(acc, a, b); +#else + __m256i product0 = _mm256_maddubs_epi16(a, b); + product0 = _mm256_madd_epi16(product0, kOnes256); + acc = _mm256_add_epi32(acc, product0); +#endif + }; + + [[maybe_unused]] auto m256_add_dpbusd_epi32x4 = [=](__m256i& acc, __m256i a0, __m256i b0, __m256i a1, __m256i b1, + __m256i a2, __m256i b2, __m256i a3, __m256i b3) { +#if defined (USE_VNNI) + acc = _mm256_dpbusd_epi32(acc, a0, b0); + acc = _mm256_dpbusd_epi32(acc, a1, b1); + acc = _mm256_dpbusd_epi32(acc, a2, b2); + acc = _mm256_dpbusd_epi32(acc, a3, b3); +#else + __m256i product0 = _mm256_maddubs_epi16(a0, b0); + __m256i product1 = _mm256_maddubs_epi16(a1, b1); + __m256i product2 = _mm256_maddubs_epi16(a2, b2); + __m256i product3 = _mm256_maddubs_epi16(a3, b3); + product0 = _mm256_add_epi16(product0, product1); + product2 = _mm256_add_epi16(product2, product3); + product0 = _mm256_add_epi16(product0, product2); + product0 = _mm256_madd_epi16(product0, kOnes256); + acc = _mm256_add_epi32(acc, product0); +#endif + }; + +#endif +#if defined (USE_SSSE3) + + [[maybe_unused]] const __m128i kOnes128 = _mm_set1_epi16(1); + + [[maybe_unused]] auto m128_hadd = [](__m128i sum, int bias) -> int { + sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC + sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB + return _mm_cvtsi128_si32(sum) + bias; + }; + + [[maybe_unused]] auto m128_add_dpbusd_epi32 = [=](__m128i& acc, __m128i a, __m128i b) { + __m128i product0 = _mm_maddubs_epi16(a, b); + product0 = _mm_madd_epi16(product0, kOnes128); + acc = _mm_add_epi32(acc, product0); + }; + + [[maybe_unused]] auto m128_add_dpbusd_epi32x4 = [=](__m128i& acc, __m128i a0, __m128i b0, __m128i a1, __m128i b1, + __m128i a2, __m128i b2, __m128i a3, __m128i b3) { + __m128i product0 = _mm_maddubs_epi16(a0, b0); + __m128i product1 = _mm_maddubs_epi16(a1, b1); + __m128i product2 = _mm_maddubs_epi16(a2, b2); + __m128i product3 = _mm_maddubs_epi16(a3, b3); + product0 = _mm_adds_epi16(product0, product1); + product2 = _mm_adds_epi16(product2, product3); + product0 = _mm_adds_epi16(product0, product2); + product0 = _mm_madd_epi16(product0, kOnes128); + acc = _mm_add_epi32(acc, product0); + }; + +#endif + +#if defined (USE_AVX512) + using vec_t = __m512i; + #define vec_setzero _mm512_setzero_si512 + #define vec_set_32 _mm512_set1_epi32 + auto& vec_add_dpbusd_32 = m512_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m512_add_dpbusd_epi32x4; + auto& vec_hadd = m512_hadd; +#elif defined (USE_AVX2) + using vec_t = __m256i; + #define vec_setzero _mm256_setzero_si256 + #define vec_set_32 _mm256_set1_epi32 + auto& vec_add_dpbusd_32 = m256_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m256_add_dpbusd_epi32x4; + auto& vec_hadd = m256_hadd; +#elif defined (USE_SSSE3) + using vec_t = __m128i; + #define vec_setzero _mm_setzero_si128 + #define vec_set_32 _mm_set1_epi32 + auto& vec_add_dpbusd_32 = m128_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m128_add_dpbusd_epi32x4; + auto& vec_hadd = m128_hadd; +#endif + +#if defined (USE_SSSE3) + + const auto output = reinterpret_cast(buffer); + const auto input_vector = reinterpret_cast(input); + + static_assert(kOutputDimensions % kOutputSimdWidth == 0 || kOutputDimensions == 1); + + // kOutputDimensions is either 1 or a multiple of kSimdWidth + // because then it is also an input dimension. + if constexpr (kOutputDimensions % kOutputSimdWidth == 0) + { + constexpr IndexType kNumChunks = kPaddedInputDimensions / 4; + + const auto input32 = reinterpret_cast(input); + vec_t* outptr = reinterpret_cast(output); + std::memcpy(output, biases_, kOutputDimensions * sizeof(OutputType)); + + for (int i = 0; i < (int)kNumChunks - 3; i += 4) + { + const vec_t in0 = vec_set_32(input32[i + 0]); + const vec_t in1 = vec_set_32(input32[i + 1]); + const vec_t in2 = vec_set_32(input32[i + 2]); + const vec_t in3 = vec_set_32(input32[i + 3]); + const auto col0 = reinterpret_cast(&weights_[(i + 0) * kOutputDimensions * 4]); + const auto col1 = reinterpret_cast(&weights_[(i + 1) * kOutputDimensions * 4]); + const auto col2 = reinterpret_cast(&weights_[(i + 2) * kOutputDimensions * 4]); + const auto col3 = reinterpret_cast(&weights_[(i + 3) * kOutputDimensions * 4]); + for (int j = 0; j * kOutputSimdWidth < kOutputDimensions; ++j) + vec_add_dpbusd_32x4(outptr[j], in0, col0[j], in1, col1[j], in2, col2[j], in3, col3[j]); + } + for (int i = 0; i < canSaturate16.count; ++i) + output[canSaturate16.ids[i].out] += input[canSaturate16.ids[i].in] * canSaturate16.ids[i].w; + } + else if constexpr (kOutputDimensions == 1) + { +#if defined (USE_AVX512) + if constexpr (kPaddedInputDimensions % (kSimdWidth * 2) != 0) + { + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const auto input_vector256 = reinterpret_cast(input); + + __m256i sum0 = _mm256_setzero_si256(); + const auto row0 = reinterpret_cast(&weights_[0]); + + for (int j = 0; j < (int)kNumChunks; ++j) + { + const __m256i in = input_vector256[j]; + m256_add_dpbusd_epi32(sum0, in, row0[j]); + } + output[0] = m256_hadd(sum0, biases_[0]); + } + else +#endif + { +#if defined (USE_AVX512) + constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2); +#else + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; +#endif + vec_t sum0 = vec_setzero(); + const auto row0 = reinterpret_cast(&weights_[0]); + + for (int j = 0; j < (int)kNumChunks; ++j) + { + const vec_t in = input_vector[j]; + vec_add_dpbusd_32(sum0, in, row0[j]); + } + output[0] = vec_hadd(sum0, biases_[0]); + } + } + +#else + +// Use old implementation for the other architectures. + + auto output = reinterpret_cast(buffer); + +#if defined(USE_SSE2) + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const __m128i kZeros = _mm_setzero_si128(); + const auto input_vector = reinterpret_cast(input); + +#elif defined(USE_MMX) + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const __m64 kZeros = _mm_setzero_si64(); + const auto input_vector = reinterpret_cast(input); + +#elif defined(USE_NEON) + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const auto input_vector = reinterpret_cast(input); +#endif + + for (IndexType i = 0; i < kOutputDimensions; ++i) { + const IndexType offset = i * kPaddedInputDimensions; + +#if defined(USE_SSE2) + __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]); + __m128i sum_hi = kZeros; + const auto row = reinterpret_cast(&weights_[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m128i row_j = _mm_load_si128(&row[j]); + __m128i input_j = _mm_load_si128(&input_vector[j]); + __m128i extended_row_lo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8); + __m128i extended_row_hi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8); + __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros); + __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros); + __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo); + __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi); + sum_lo = _mm_add_epi32(sum_lo, product_lo); + sum_hi = _mm_add_epi32(sum_hi, product_hi); + } + __m128i sum = _mm_add_epi32(sum_lo, sum_hi); + __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sum_high_64); + __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sum_second_32); + output[i] = _mm_cvtsi128_si32(sum); + +#elif defined(USE_MMX) + __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]); + __m64 sum_hi = kZeros; + const auto row = reinterpret_cast(&weights_[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m64 row_j = row[j]; + __m64 input_j = input_vector[j]; + __m64 extended_row_lo = _mm_srai_pi16(_mm_unpacklo_pi8(row_j, row_j), 8); + __m64 extended_row_hi = _mm_srai_pi16(_mm_unpackhi_pi8(row_j, row_j), 8); + __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros); + __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros); + __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo); + __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi); + sum_lo = _mm_add_pi32(sum_lo, product_lo); + sum_hi = _mm_add_pi32(sum_hi, product_hi); + } + __m64 sum = _mm_add_pi32(sum_lo, sum_hi); + sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum)); + output[i] = _mm_cvtsi64_si32(sum); + +#elif defined(USE_NEON) + int32x4_t sum = {biases_[i]}; + const auto row = reinterpret_cast(&weights_[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]); + product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]); + sum = vpadalq_s16(sum, product); + } + output[i] = sum[0] + sum[1] + sum[2] + sum[3]; + +#else + OutputType sum = biases_[i]; + for (IndexType j = 0; j < kInputDimensions; ++j) { + sum += weights_[offset + j] * input[j]; + } + output[i] = sum; +#endif + + } +#if defined(USE_MMX) + _mm_empty(); +#endif + +#endif + + return output; + } + + private: + using BiasType = OutputType; + using WeightType = std::int8_t; + + PreviousLayer previous_layer_; + + alignas(kCacheLineSize) BiasType biases_[kOutputDimensions]; + alignas(kCacheLineSize) WeightType weights_[kOutputDimensions * kPaddedInputDimensions]; +#if defined (USE_SSSE3) + struct CanSaturate { + int count; + struct Entry { + uint16_t out; + uint16_t in; + int8_t w; + } ids[kPaddedInputDimensions * kOutputDimensions * 3 / 4]; + + void add(int i, int j, int8_t w) { + ids[count].out = i; + ids[count].in = j; + ids[count].w = w; + ++count; + } + } canSaturate16; +#endif + }; + +} // namespace Stockfish::Eval::NNUE::Layers + +#endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED diff --git a/src/nnue/layers/clipped_relu.h b/src/nnue/layers/clipped_relu.h new file mode 100644 index 00000000..a10e3e48 --- /dev/null +++ b/src/nnue/layers/clipped_relu.h @@ -0,0 +1,166 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Definition of layer ClippedReLU of NNUE evaluation function + +#ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED +#define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED + +#include "../nnue_common.h" + +namespace Stockfish::Eval::NNUE::Layers { + + // Clipped ReLU + template + class ClippedReLU { + public: + // Input/output type + using InputType = typename PreviousLayer::OutputType; + using OutputType = std::uint8_t; + static_assert(std::is_same::value, ""); + + // Number of input/output dimensions + static constexpr IndexType kInputDimensions = + PreviousLayer::kOutputDimensions; + static constexpr IndexType kOutputDimensions = kInputDimensions; + + // Size of forward propagation buffer used in this layer + static constexpr std::size_t kSelfBufferSize = + CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize); + + // Size of the forward propagation buffer used from the input layer to this layer + static constexpr std::size_t kBufferSize = + PreviousLayer::kBufferSize + kSelfBufferSize; + + // Hash value embedded in the evaluation file + static constexpr std::uint32_t GetHashValue() { + std::uint32_t hash_value = 0x538D24C7u; + hash_value += PreviousLayer::GetHashValue(); + return hash_value; + } + + // Read network parameters + bool ReadParameters(std::istream& stream) { + return previous_layer_.ReadParameters(stream); + } + + // Forward propagation + const OutputType* Propagate( + const TransformedFeatureType* transformed_features, char* buffer) const { + const auto input = previous_layer_.Propagate( + transformed_features, buffer + kSelfBufferSize); + const auto output = reinterpret_cast(buffer); + + #if defined(USE_AVX2) + constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth; + const __m256i kZero = _mm256_setzero_si256(); + const __m256i kOffsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m256i*>(output); + for (IndexType i = 0; i < kNumChunks; ++i) { + const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32( + _mm256_load_si256(&in[i * 4 + 0]), + _mm256_load_si256(&in[i * 4 + 1])), kWeightScaleBits); + const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32( + _mm256_load_si256(&in[i * 4 + 2]), + _mm256_load_si256(&in[i * 4 + 3])), kWeightScaleBits); + _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8( + _mm256_packs_epi16(words0, words1), kZero), kOffsets)); + } + constexpr IndexType kStart = kNumChunks * kSimdWidth; + + #elif defined(USE_SSE2) + constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth; + + #ifdef USE_SSE41 + const __m128i kZero = _mm_setzero_si128(); + #else + const __m128i k0x80s = _mm_set1_epi8(-128); + #endif + + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m128i*>(output); + for (IndexType i = 0; i < kNumChunks; ++i) { + const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32( + _mm_load_si128(&in[i * 4 + 0]), + _mm_load_si128(&in[i * 4 + 1])), kWeightScaleBits); + const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32( + _mm_load_si128(&in[i * 4 + 2]), + _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits); + const __m128i packedbytes = _mm_packs_epi16(words0, words1); + _mm_store_si128(&out[i], + + #ifdef USE_SSE41 + _mm_max_epi8(packedbytes, kZero) + #else + _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) + #endif + + ); + } + constexpr IndexType kStart = kNumChunks * kSimdWidth; + + #elif defined(USE_MMX) + constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth; + const __m64 k0x80s = _mm_set1_pi8(-128); + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m64*>(output); + for (IndexType i = 0; i < kNumChunks; ++i) { + const __m64 words0 = _mm_srai_pi16( + _mm_packs_pi32(in[i * 4 + 0], in[i * 4 + 1]), + kWeightScaleBits); + const __m64 words1 = _mm_srai_pi16( + _mm_packs_pi32(in[i * 4 + 2], in[i * 4 + 3]), + kWeightScaleBits); + const __m64 packedbytes = _mm_packs_pi16(words0, words1); + out[i] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s); + } + _mm_empty(); + constexpr IndexType kStart = kNumChunks * kSimdWidth; + + #elif defined(USE_NEON) + constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2); + const int8x8_t kZero = {0}; + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast(output); + for (IndexType i = 0; i < kNumChunks; ++i) { + int16x8_t shifted; + const auto pack = reinterpret_cast(&shifted); + pack[0] = vqshrn_n_s32(in[i * 2 + 0], kWeightScaleBits); + pack[1] = vqshrn_n_s32(in[i * 2 + 1], kWeightScaleBits); + out[i] = vmax_s8(vqmovn_s16(shifted), kZero); + } + constexpr IndexType kStart = kNumChunks * (kSimdWidth / 2); + #else + constexpr IndexType kStart = 0; + #endif + + for (IndexType i = kStart; i < kInputDimensions; ++i) { + output[i] = static_cast( + std::max(0, std::min(127, input[i] >> kWeightScaleBits))); + } + return output; + } + + private: + PreviousLayer previous_layer_; + }; + +} // namespace Stockfish::Eval::NNUE::Layers + +#endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED diff --git a/src/nnue/layers/input_slice.h b/src/nnue/layers/input_slice.h new file mode 100644 index 00000000..43b06eec --- /dev/null +++ b/src/nnue/layers/input_slice.h @@ -0,0 +1,68 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// NNUE evaluation function layer InputSlice definition + +#ifndef NNUE_LAYERS_INPUT_SLICE_H_INCLUDED +#define NNUE_LAYERS_INPUT_SLICE_H_INCLUDED + +#include "../nnue_common.h" + +namespace Stockfish::Eval::NNUE::Layers { + +// Input layer +template +class InputSlice { + public: + // Need to maintain alignment + static_assert(Offset % kMaxSimdWidth == 0, ""); + + // Output type + using OutputType = TransformedFeatureType; + + // Output dimensionality + static constexpr IndexType kOutputDimensions = OutputDimensions; + + // Size of forward propagation buffer used from the input layer to this layer + static constexpr std::size_t kBufferSize = 0; + + // Hash value embedded in the evaluation file + static constexpr std::uint32_t GetHashValue() { + std::uint32_t hash_value = 0xEC42E90Du; + hash_value ^= kOutputDimensions ^ (Offset << 10); + return hash_value; + } + + // Read network parameters + bool ReadParameters(std::istream& /*stream*/) { + return true; + } + + // Forward propagation + const OutputType* Propagate( + const TransformedFeatureType* transformed_features, + char* /*buffer*/) const { + return transformed_features + Offset; + } + + private: +}; + +} // namespace Stockfish::Eval::NNUE::Layers + +#endif // #ifndef NNUE_LAYERS_INPUT_SLICE_H_INCLUDED diff --git a/src/nnue/nnue_accumulator.h b/src/nnue/nnue_accumulator.h new file mode 100644 index 00000000..55fafa13 --- /dev/null +++ b/src/nnue/nnue_accumulator.h @@ -0,0 +1,40 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Class for difference calculation of NNUE evaluation function + +#ifndef NNUE_ACCUMULATOR_H_INCLUDED +#define NNUE_ACCUMULATOR_H_INCLUDED + +#include "nnue_architecture.h" + +namespace Stockfish::Eval::NNUE { + + // The accumulator of a StateInfo without parent is set to the INIT state + enum AccumulatorState { EMPTY, COMPUTED, INIT }; + + // Class that holds the result of affine transformation of input features + struct alignas(kCacheLineSize) Accumulator { + std::int16_t + accumulation[2][kRefreshTriggers.size()][kTransformedFeatureDimensions]; + AccumulatorState state[2]; + }; + +} // namespace Stockfish::Eval::NNUE + +#endif // NNUE_ACCUMULATOR_H_INCLUDED diff --git a/src/nnue/nnue_architecture.h b/src/nnue/nnue_architecture.h new file mode 100644 index 00000000..1680368e --- /dev/null +++ b/src/nnue/nnue_architecture.h @@ -0,0 +1,38 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Input features and network structure used in NNUE evaluation function + +#ifndef NNUE_ARCHITECTURE_H_INCLUDED +#define NNUE_ARCHITECTURE_H_INCLUDED + +// Defines the network structure +#include "architectures/halfkp_256x2-32-32.h" + +namespace Stockfish::Eval::NNUE { + + static_assert(kTransformedFeatureDimensions % kMaxSimdWidth == 0, ""); + static_assert(Network::kOutputDimensions == 1, ""); + static_assert(std::is_same::value, ""); + + // Trigger for full calculation instead of difference calculation + constexpr auto kRefreshTriggers = RawFeatures::kRefreshTriggers; + +} // namespace Stockfish::Eval::NNUE + +#endif // #ifndef NNUE_ARCHITECTURE_H_INCLUDED diff --git a/src/nnue/nnue_common.h b/src/nnue/nnue_common.h new file mode 100644 index 00000000..09a152a5 --- /dev/null +++ b/src/nnue/nnue_common.h @@ -0,0 +1,132 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// Constants used in NNUE evaluation function + +#ifndef NNUE_COMMON_H_INCLUDED +#define NNUE_COMMON_H_INCLUDED + +#include +#include + +#if defined(USE_AVX2) +#include + +#elif defined(USE_SSE41) +#include + +#elif defined(USE_SSSE3) +#include + +#elif defined(USE_SSE2) +#include + +#elif defined(USE_MMX) +#include + +#elif defined(USE_NEON) +#include +#endif + +namespace Stockfish::Eval::NNUE { + + // Version of the evaluation file + constexpr std::uint32_t kVersion = 0x7AF32F16u; + + // Constant used in evaluation value calculation + constexpr int FV_SCALE = 16; + constexpr int kWeightScaleBits = 6; + + // Size of cache line (in bytes) + constexpr std::size_t kCacheLineSize = 64; + + // SIMD width (in bytes) + #if defined(USE_AVX2) + constexpr std::size_t kSimdWidth = 32; + + #elif defined(USE_SSE2) + constexpr std::size_t kSimdWidth = 16; + + #elif defined(USE_MMX) + constexpr std::size_t kSimdWidth = 8; + + #elif defined(USE_NEON) + constexpr std::size_t kSimdWidth = 16; + #endif + + constexpr std::size_t kMaxSimdWidth = 32; + + // unique number for each piece type on each square + enum { + PS_NONE = 0, + PS_W_PAWN = 1, + PS_B_PAWN = 1 * SQUARE_NB + 1, + PS_W_KNIGHT = 2 * SQUARE_NB + 1, + PS_B_KNIGHT = 3 * SQUARE_NB + 1, + PS_W_BISHOP = 4 * SQUARE_NB + 1, + PS_B_BISHOP = 5 * SQUARE_NB + 1, + PS_W_ROOK = 6 * SQUARE_NB + 1, + PS_B_ROOK = 7 * SQUARE_NB + 1, + PS_W_QUEEN = 8 * SQUARE_NB + 1, + PS_B_QUEEN = 9 * SQUARE_NB + 1, + PS_W_KING = 10 * SQUARE_NB + 1, + PS_END = PS_W_KING, // pieces without kings (pawns included) + PS_B_KING = 11 * SQUARE_NB + 1, + PS_END2 = 12 * SQUARE_NB + 1 + }; + + constexpr uint32_t kpp_board_index[COLOR_NB][PIECE_NB] = { + // convention: W - us, B - them + // viewed from other side, W and B are reversed + { PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_W_KING, PS_NONE, + PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_B_KING, PS_NONE }, + { PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_B_KING, PS_NONE, + PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_W_KING, PS_NONE } + }; + + // Type of input feature after conversion + using TransformedFeatureType = std::uint8_t; + using IndexType = std::uint32_t; + + // Round n up to be a multiple of base + template + constexpr IntType CeilToMultiple(IntType n, IntType base) { + return (n + base - 1) / base * base; + } + + // read_little_endian() is our utility to read an integer (signed or unsigned, any size) + // from a stream in little-endian order. We swap the byte order after the read if + // necessary to return a result with the byte ordering of the compiling machine. + template + inline IntType read_little_endian(std::istream& stream) { + + IntType result; + std::uint8_t u[sizeof(IntType)]; + typename std::make_unsigned::type v = 0; + + stream.read(reinterpret_cast(u), sizeof(IntType)); + for (std::size_t i = 0; i < sizeof(IntType); ++i) + v = (v << 8) | u[sizeof(IntType) - i - 1]; + + std::memcpy(&result, &v, sizeof(IntType)); + return result; + } + +} // namespace Stockfish::Eval::NNUE + +#endif // #ifndef NNUE_COMMON_H_INCLUDED diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h new file mode 100644 index 00000000..1e0b0e6d --- /dev/null +++ b/src/nnue/nnue_feature_transformer.h @@ -0,0 +1,417 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +// A class that converts the input features of the NNUE evaluation function + +#ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED +#define NNUE_FEATURE_TRANSFORMER_H_INCLUDED + +#include "nnue_common.h" +#include "nnue_architecture.h" +#include "features/index_list.h" + +#include // std::memset() + +namespace Stockfish::Eval::NNUE { + + // If vector instructions are enabled, we update and refresh the + // accumulator tile by tile such that each tile fits in the CPU's + // vector registers. + #define VECTOR + + #ifdef USE_AVX512 + typedef __m512i vec_t; + #define vec_load(a) _mm512_load_si512(a) + #define vec_store(a,b) _mm512_store_si512(a,b) + #define vec_add_16(a,b) _mm512_add_epi16(a,b) + #define vec_sub_16(a,b) _mm512_sub_epi16(a,b) + static constexpr IndexType kNumRegs = 8; // only 8 are needed + + #elif USE_AVX2 + typedef __m256i vec_t; + #define vec_load(a) _mm256_load_si256(a) + #define vec_store(a,b) _mm256_store_si256(a,b) + #define vec_add_16(a,b) _mm256_add_epi16(a,b) + #define vec_sub_16(a,b) _mm256_sub_epi16(a,b) + static constexpr IndexType kNumRegs = 16; + + #elif USE_SSE2 + typedef __m128i vec_t; + #define vec_load(a) (*(a)) + #define vec_store(a,b) *(a)=(b) + #define vec_add_16(a,b) _mm_add_epi16(a,b) + #define vec_sub_16(a,b) _mm_sub_epi16(a,b) + static constexpr IndexType kNumRegs = Is64Bit ? 16 : 8; + + #elif USE_MMX + typedef __m64 vec_t; + #define vec_load(a) (*(a)) + #define vec_store(a,b) *(a)=(b) + #define vec_add_16(a,b) _mm_add_pi16(a,b) + #define vec_sub_16(a,b) _mm_sub_pi16(a,b) + static constexpr IndexType kNumRegs = 8; + + #elif USE_NEON + typedef int16x8_t vec_t; + #define vec_load(a) (*(a)) + #define vec_store(a,b) *(a)=(b) + #define vec_add_16(a,b) vaddq_s16(a,b) + #define vec_sub_16(a,b) vsubq_s16(a,b) + static constexpr IndexType kNumRegs = 16; + + #else + #undef VECTOR + + #endif + + // Input feature converter + class FeatureTransformer { + + private: + // Number of output dimensions for one side + static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions; + + #ifdef VECTOR + static constexpr IndexType kTileHeight = kNumRegs * sizeof(vec_t) / 2; + static_assert(kHalfDimensions % kTileHeight == 0, "kTileHeight must divide kHalfDimensions"); + #endif + + public: + // Output type + using OutputType = TransformedFeatureType; + + // Number of input/output dimensions + static constexpr IndexType kInputDimensions = RawFeatures::kDimensions; + static constexpr IndexType kOutputDimensions = kHalfDimensions * 2; + + // Size of forward propagation buffer + static constexpr std::size_t kBufferSize = + kOutputDimensions * sizeof(OutputType); + + // Hash value embedded in the evaluation file + static constexpr std::uint32_t GetHashValue() { + + return RawFeatures::kHashValue ^ kOutputDimensions; + } + + // Read network parameters + bool ReadParameters(std::istream& stream) { + + for (std::size_t i = 0; i < kHalfDimensions; ++i) + biases_[i] = read_little_endian(stream); + for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i) + weights_[i] = read_little_endian(stream); + return !stream.fail(); + } + + // Convert input features + void Transform(const Position& pos, OutputType* output) const { + + UpdateAccumulator(pos, WHITE); + UpdateAccumulator(pos, BLACK); + + const auto& accumulation = pos.state()->accumulator.accumulation; + + #if defined(USE_AVX512) + constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth * 2); + static_assert(kHalfDimensions % (kSimdWidth * 2) == 0); + const __m512i kControl = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7); + const __m512i kZero = _mm512_setzero_si512(); + + #elif defined(USE_AVX2) + constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + constexpr int kControl = 0b11011000; + const __m256i kZero = _mm256_setzero_si256(); + + #elif defined(USE_SSE2) + constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + + #ifdef USE_SSE41 + const __m128i kZero = _mm_setzero_si128(); + #else + const __m128i k0x80s = _mm_set1_epi8(-128); + #endif + + #elif defined(USE_MMX) + constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + const __m64 k0x80s = _mm_set1_pi8(-128); + + #elif defined(USE_NEON) + constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); + const int8x8_t kZero = {0}; + #endif + + const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()}; + for (IndexType p = 0; p < 2; ++p) { + const IndexType offset = kHalfDimensions * p; + + #if defined(USE_AVX512) + auto out = reinterpret_cast<__m512i*>(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m512i sum0 = _mm512_load_si512( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); + __m512i sum1 = _mm512_load_si512( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); + _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(kControl, + _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), kZero))); + } + + #elif defined(USE_AVX2) + auto out = reinterpret_cast<__m256i*>(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m256i sum0 = _mm256_load_si256( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); + __m256i sum1 = _mm256_load_si256( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); + _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8( + _mm256_packs_epi16(sum0, sum1), kZero), kControl)); + } + + #elif defined(USE_SSE2) + auto out = reinterpret_cast<__m128i*>(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m128i sum0 = _mm_load_si128(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 0]); + __m128i sum1 = _mm_load_si128(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 1]); + const __m128i packedbytes = _mm_packs_epi16(sum0, sum1); + + _mm_store_si128(&out[j], + + #ifdef USE_SSE41 + _mm_max_epi8(packedbytes, kZero) + #else + _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) + #endif + + ); + } + + #elif defined(USE_MMX) + auto out = reinterpret_cast<__m64*>(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m64 sum0 = *(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 0]); + __m64 sum1 = *(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 1]); + const __m64 packedbytes = _mm_packs_pi16(sum0, sum1); + out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s); + } + + #elif defined(USE_NEON) + const auto out = reinterpret_cast(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + int16x8_t sum = reinterpret_cast( + accumulation[perspectives[p]][0])[j]; + out[j] = vmax_s8(vqmovn_s16(sum), kZero); + } + + #else + for (IndexType j = 0; j < kHalfDimensions; ++j) { + BiasType sum = accumulation[static_cast(perspectives[p])][0][j]; + output[offset + j] = static_cast( + std::max(0, std::min(127, sum))); + } + #endif + + } + #if defined(USE_MMX) + _mm_empty(); + #endif + } + + private: + void UpdateAccumulator(const Position& pos, const Color c) const { + + #ifdef VECTOR + // Gcc-10.2 unnecessarily spills AVX2 registers if this array + // is defined in the VECTOR code below, once in each branch + vec_t acc[kNumRegs]; + #endif + + // Look for a usable accumulator of an earlier position. We keep track + // of the estimated gain in terms of features to be added/subtracted. + StateInfo *st = pos.state(), *next = nullptr; + int gain = pos.count() - 2; + while (st->accumulator.state[c] == EMPTY) + { + auto& dp = st->dirtyPiece; + // The first condition tests whether an incremental update is + // possible at all: if this side's king has moved, it is not possible. + static_assert(std::is_same_v>, + "Current code assumes that only kFriendlyKingMoved refresh trigger is being used."); + if ( dp.piece[0] == make_piece(c, KING) + || (gain -= dp.dirty_num + 1) < 0) + break; + next = st; + st = st->previous; + } + + if (st->accumulator.state[c] == COMPUTED) + { + if (next == nullptr) + return; + + // Update incrementally in two steps. First, we update the "next" + // accumulator. Then, we update the current accumulator (pos.state()). + + // Gather all features to be updated. This code assumes HalfKP features + // only and doesn't support refresh triggers. + static_assert(std::is_same_v>, + RawFeatures>); + Features::IndexList removed[2], added[2]; + Features::HalfKP::AppendChangedIndices(pos, + next->dirtyPiece, c, &removed[0], &added[0]); + for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous) + Features::HalfKP::AppendChangedIndices(pos, + st2->dirtyPiece, c, &removed[1], &added[1]); + + // Mark the accumulators as computed. + next->accumulator.state[c] = COMPUTED; + pos.state()->accumulator.state[c] = COMPUTED; + + // Now update the accumulators listed in info[], where the last element is a sentinel. + StateInfo *info[3] = + { next, next == pos.state() ? nullptr : pos.state(), nullptr }; + #ifdef VECTOR + for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j) + { + // Load accumulator + auto accTile = reinterpret_cast( + &st->accumulator.accumulation[c][0][j * kTileHeight]); + for (IndexType k = 0; k < kNumRegs; ++k) + acc[k] = vec_load(&accTile[k]); + + for (IndexType i = 0; info[i]; ++i) + { + // Difference calculation for the deactivated features + for (const auto index : removed[i]) + { + const IndexType offset = kHalfDimensions * index + j * kTileHeight; + auto column = reinterpret_cast(&weights_[offset]); + for (IndexType k = 0; k < kNumRegs; ++k) + acc[k] = vec_sub_16(acc[k], column[k]); + } + + // Difference calculation for the activated features + for (const auto index : added[i]) + { + const IndexType offset = kHalfDimensions * index + j * kTileHeight; + auto column = reinterpret_cast(&weights_[offset]); + for (IndexType k = 0; k < kNumRegs; ++k) + acc[k] = vec_add_16(acc[k], column[k]); + } + + // Store accumulator + accTile = reinterpret_cast( + &info[i]->accumulator.accumulation[c][0][j * kTileHeight]); + for (IndexType k = 0; k < kNumRegs; ++k) + vec_store(&accTile[k], acc[k]); + } + } + + #else + for (IndexType i = 0; info[i]; ++i) + { + std::memcpy(info[i]->accumulator.accumulation[c][0], + st->accumulator.accumulation[c][0], + kHalfDimensions * sizeof(BiasType)); + st = info[i]; + + // Difference calculation for the deactivated features + for (const auto index : removed[i]) + { + const IndexType offset = kHalfDimensions * index; + + for (IndexType j = 0; j < kHalfDimensions; ++j) + st->accumulator.accumulation[c][0][j] -= weights_[offset + j]; + } + + // Difference calculation for the activated features + for (const auto index : added[i]) + { + const IndexType offset = kHalfDimensions * index; + + for (IndexType j = 0; j < kHalfDimensions; ++j) + st->accumulator.accumulation[c][0][j] += weights_[offset + j]; + } + } + #endif + } + else + { + // Refresh the accumulator + auto& accumulator = pos.state()->accumulator; + accumulator.state[c] = COMPUTED; + Features::IndexList active; + Features::HalfKP::AppendActiveIndices(pos, c, &active); + + #ifdef VECTOR + for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j) + { + auto biasesTile = reinterpret_cast( + &biases_[j * kTileHeight]); + for (IndexType k = 0; k < kNumRegs; ++k) + acc[k] = biasesTile[k]; + + for (const auto index : active) + { + const IndexType offset = kHalfDimensions * index + j * kTileHeight; + auto column = reinterpret_cast(&weights_[offset]); + + for (unsigned k = 0; k < kNumRegs; ++k) + acc[k] = vec_add_16(acc[k], column[k]); + } + + auto accTile = reinterpret_cast( + &accumulator.accumulation[c][0][j * kTileHeight]); + for (unsigned k = 0; k < kNumRegs; k++) + vec_store(&accTile[k], acc[k]); + } + + #else + std::memcpy(accumulator.accumulation[c][0], biases_, + kHalfDimensions * sizeof(BiasType)); + + for (const auto index : active) + { + const IndexType offset = kHalfDimensions * index; + + for (IndexType j = 0; j < kHalfDimensions; ++j) + accumulator.accumulation[c][0][j] += weights_[offset + j]; + } + #endif + } + + #if defined(USE_MMX) + _mm_empty(); + #endif + } + + using BiasType = std::int16_t; + using WeightType = std::int16_t; + + alignas(kCacheLineSize) BiasType biases_[kHalfDimensions]; + alignas(kCacheLineSize) + WeightType weights_[kHalfDimensions * kInputDimensions]; + }; + +} // namespace Stockfish::Eval::NNUE + +#endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED diff --git a/src/pawns.cpp b/src/pawns.cpp index 1825b6e2..9a0610a0 100644 --- a/src/pawns.cpp +++ b/src/pawns.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,29 +24,38 @@ #include "position.h" #include "thread.h" +namespace Stockfish { + namespace { #define V Value #define S(mg, eg) make_score(mg, eg) // Pawn penalties - constexpr Score Backward = S( 9, 24); - constexpr Score BlockedStorm = S(82, 82); - constexpr Score Doubled = S(11, 56); - constexpr Score Isolated = S( 5, 15); - constexpr Score WeakLever = S( 0, 56); - constexpr Score WeakUnopposed = S(13, 27); + constexpr Score Backward = S( 9, 22); + constexpr Score Doubled = S(13, 51); + constexpr Score DoubledEarly = S(20, 7); + constexpr Score Isolated = S( 3, 15); + constexpr Score WeakLever = S( 4, 58); + constexpr Score WeakUnopposed = S(13, 24); + + // Bonus for blocked pawns at 5th or 6th rank + constexpr Score BlockedPawn[2] = { S(-17, -6), S(-9, 2) }; + + constexpr Score BlockedStorm[RANK_NB] = { + S(0, 0), S(0, 0), S(75, 78), S(-8, 16), S(-6, 10), S(-6, 6), S(0, 2) + }; // Connected pawn bonus - constexpr int Connected[RANK_NB] = { 0, 7, 8, 12, 29, 48, 86 }; + constexpr int Connected[RANK_NB] = { 0, 5, 7, 11, 23, 48, 87 }; // Strength of pawn shelter for our king by [distance from edge][rank]. // RANK_1 = 0 is used for files where we have no pawn, or pawn is behind our king. constexpr Value ShelterStrength[int(FILE_NB) / 2][RANK_NB] = { - { V( -6), V( 81), V( 93), V( 58), V( 39), V( 18), V( 25) }, - { V(-43), V( 61), V( 35), V(-49), V(-29), V(-11), V( -63) }, - { V(-10), V( 75), V( 23), V( -2), V( 32), V( 3), V( -45) }, - { V(-39), V(-13), V(-29), V(-52), V(-48), V(-67), V(-166) } + { V( -5), V( 82), V( 92), V( 54), V( 36), V( 22), V( 28) }, + { V(-44), V( 63), V( 33), V(-50), V(-30), V(-12), V( -62) }, + { V(-11), V( 77), V( 22), V( -6), V( 31), V( 8), V( -45) }, + { V(-39), V(-12), V(-29), V(-50), V(-43), V(-68), V(-164) } }; // Danger of enemy pawns moving toward our king by [distance from edge][rank]. @@ -56,27 +63,40 @@ namespace { // is behind our king. Note that UnblockedStorm[0][1-2] accommodate opponent pawn // on edge, likely blocked by our king. constexpr Value UnblockedStorm[int(FILE_NB) / 2][RANK_NB] = { - { V( 89), V(-285), V(-185), V(93), V(57), V( 45), V( 51) }, - { V( 44), V( -18), V( 123), V(46), V(39), V( -7), V( 23) }, - { V( 4), V( 52), V( 162), V(37), V( 7), V(-14), V( -2) }, - { V(-10), V( -14), V( 90), V(15), V( 2), V( -7), V(-16) } + { V( 87), V(-288), V(-168), V( 96), V( 47), V( 44), V( 46) }, + { V( 42), V( -25), V( 120), V( 45), V( 34), V( -9), V( 24) }, + { V( -8), V( 51), V( 167), V( 35), V( -4), V(-16), V(-12) }, + { V(-17), V( -13), V( 100), V( 4), V( 9), V(-16), V(-31) } }; + + // KingOnFile[semi-open Us][semi-open Them] contains bonuses/penalties + // for king when the king is on a semi-open or open file. + constexpr Score KingOnFile[2][2] = {{ S(-21,10), S(-7, 1) }, + { S( 0,-3), S( 9,-4) }}; + #undef S #undef V + + /// evaluate() calculates a score for the static pawn structure of the given position. + /// We cannot use the location of pieces or king in this function, as the evaluation + /// of the pawn structure will be stored in a small cache for speed reasons, and will + /// be re-used even when the pieces have moved. + template Score evaluate(const Position& pos, Pawns::Entry* e) { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); - constexpr Direction Up = (Us == WHITE ? NORTH : SOUTH); + constexpr Color Them = ~Us; + constexpr Direction Up = pawn_push(Us); + constexpr Direction Down = -Up; Bitboard neighbours, stoppers, support, phalanx, opposed; Bitboard lever, leverPush, blocked; Square s; bool backward, passed, doubled; Score score = SCORE_ZERO; - const Square* pl = pos.squares(Us); + Bitboard b = pos.pieces(Us, PAWN); Bitboard ourPawns = pos.pieces( Us, PAWN); Bitboard theirPawns = pos.pieces(Them, PAWN); @@ -86,10 +106,12 @@ namespace { e->passedPawns[Us] = 0; e->kingSquares[Us] = SQ_NONE; e->pawnAttacks[Us] = e->pawnAttacksSpan[Us] = pawn_attacks_bb(ourPawns); + e->blockedCount += popcount(shift(ourPawns) & (theirPawns | doubleAttackThem)); // Loop through all pawns of the current color and score each pawn - while ((s = *pl++) != SQ_NONE) - { + while (b) { + s = pop_lsb(&b); + assert(pos.piece_on(s) == make_piece(Us, PAWN)); Rank r = relative_rank(Us, s); @@ -98,17 +120,24 @@ namespace { opposed = theirPawns & forward_file_bb(Us, s); blocked = theirPawns & (s + Up); stoppers = theirPawns & passed_pawn_span(Us, s); - lever = theirPawns & PawnAttacks[Us][s]; - leverPush = theirPawns & PawnAttacks[Us][s + Up]; + lever = theirPawns & pawn_attacks_bb(Us, s); + leverPush = theirPawns & pawn_attacks_bb(Us, s + Up); doubled = ourPawns & (s - Up); neighbours = ourPawns & adjacent_files_bb(s); phalanx = neighbours & rank_bb(s); support = neighbours & rank_bb(s - Up); + if (doubled) + { + // Additional doubled penalty if none of their pawns is fixed + if (!(ourPawns & shift(theirPawns | pawn_attacks_bb(theirPawns)))) + score -= DoubledEarly; + } + // A pawn is backward when it is behind all pawns of the same color on // the adjacent files and cannot safely advance. backward = !(neighbours & forward_ranks_bb(Them, s + Up)) - && (stoppers & (leverPush | blocked)); + && (leverPush | blocked); // Compute additional span if pawn is not backward nor blocked if (!backward && !blocked) @@ -118,12 +147,15 @@ namespace { // (a) there is no stoppers except some levers // (b) the only stoppers are the leverPush, but we outnumber them // (c) there is only one front stopper which can be levered. + // (Refined in Evaluation::passed) passed = !(stoppers ^ lever) || ( !(stoppers ^ leverPush) && popcount(phalanx) >= popcount(leverPush)) || ( stoppers == blocked && r >= RANK_5 && (shift(support) & ~(theirPawns | doubleAttackThem))); + passed &= !(forward_file_bb(Us, s) & ourPawns); + // Passed pawns will be properly scored later in evaluation when we have // full attack info. if (passed) @@ -133,22 +165,32 @@ namespace { if (support | phalanx) { int v = Connected[r] * (2 + bool(phalanx) - bool(opposed)) - + 21 * popcount(support); + + 22 * popcount(support); score += make_score(v, v * (r - 2) / 4); } else if (!neighbours) - score -= Isolated - + WeakUnopposed * !opposed; + { + if ( opposed + && (ourPawns & forward_file_bb(Them, s)) + && !(theirPawns & adjacent_files_bb(s))) + score -= Doubled; + else + score -= Isolated + + WeakUnopposed * !opposed; + } else if (backward) - score -= Backward - + WeakUnopposed * !opposed; + score -= Backward + + WeakUnopposed * !opposed * bool(~(FileABB | FileHBB) & s); if (!support) - score -= Doubled * doubled - + WeakLever * more_than_one(lever); + score -= Doubled * doubled + + WeakLever * more_than_one(lever); + + if (blocked && r >= RANK_5) + score += BlockedPawn[r - RANK_5]; } return score; @@ -158,6 +200,7 @@ namespace { namespace Pawns { + /// Pawns::probe() looks up the current position's pawns configuration in /// the pawns hash table. It returns a pointer to the Entry if the position /// is found. Otherwise a new Entry is computed and stored there, so we don't @@ -172,6 +215,7 @@ Entry* probe(const Position& pos) { return e; e->key = key; + e->blockedCount = 0; e->scores[WHITE] = evaluate(pos, e); e->scores[BLACK] = evaluate(pos, e); @@ -183,17 +227,17 @@ Entry* probe(const Position& pos) { /// penalty for a king, looking at the king file and the two closest files. template -Score Entry::evaluate_shelter(const Position& pos, Square ksq) { +Score Entry::evaluate_shelter(const Position& pos, Square ksq) const { - constexpr Color Them = (Us == WHITE ? BLACK : WHITE); + constexpr Color Them = ~Us; Bitboard b = pos.pieces(PAWN) & ~forward_ranks_bb(Them, ksq); - Bitboard ourPawns = b & pos.pieces(Us); + Bitboard ourPawns = b & pos.pieces(Us) & ~pawnAttacks[Them]; Bitboard theirPawns = b & pos.pieces(Them); Score bonus = make_score(5, 5); - File center = clamp(file_of(ksq), FILE_B, FILE_G); + File center = std::clamp(file_of(ksq), FILE_B, FILE_G); for (File f = File(center - 1); f <= File(center + 1); ++f) { b = ourPawns & file_bb(f); @@ -202,15 +246,18 @@ Score Entry::evaluate_shelter(const Position& pos, Square ksq) { b = theirPawns & file_bb(f); int theirRank = b ? relative_rank(Us, frontmost_sq(Them, b)) : 0; - File d = map_to_queenside(f); + int d = edge_distance(f); bonus += make_score(ShelterStrength[d][ourRank], 0); if (ourRank && (ourRank == theirRank - 1)) - bonus -= BlockedStorm * int(theirRank == RANK_3); + bonus -= BlockedStorm[theirRank]; else bonus -= make_score(UnblockedStorm[d][theirRank], 0); } + // King On File + bonus -= KingOnFile[pos.is_on_semiopen_file(Us, ksq)][pos.is_on_semiopen_file(Them, ksq)]; + return bonus; } @@ -238,9 +285,9 @@ Score Entry::do_king_safety(const Position& pos) { // In endgame we like to bring our king near our closest pawn Bitboard pawns = pos.pieces(Us, PAWN); - int minPawnDist = pawns ? 8 : 0; + int minPawnDist = 6; - if (pawns & PseudoAttacks[KING][ksq]) + if (pawns & attacks_bb(ksq)) minPawnDist = 1; else while (pawns) minPawnDist = std::min(minPawnDist, distance(ksq, pop_lsb(&pawns))); @@ -253,3 +300,5 @@ template Score Entry::do_king_safety(const Position& pos); template Score Entry::do_king_safety(const Position& pos); } // namespace Pawns + +} // namespace Stockfish diff --git a/src/pawns.h b/src/pawns.h index 4c041716..124619d6 100644 --- a/src/pawns.h +++ b/src/pawns.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,7 +23,7 @@ #include "position.h" #include "types.h" -namespace Pawns { +namespace Stockfish::Pawns { /// Pawns::Entry contains various information about a pawn structure. A lookup /// to the pawn hash table (performed by calling the probe function) returns a @@ -38,6 +36,7 @@ struct Entry { Bitboard passed_pawns(Color c) const { return passedPawns[c]; } Bitboard pawn_attacks_span(Color c) const { return pawnAttacksSpan[c]; } int passed_count() const { return popcount(passedPawns[WHITE] | passedPawns[BLACK]); } + int blocked_count() const { return blockedCount; } template Score king_safety(const Position& pos) { @@ -49,7 +48,7 @@ struct Entry { Score do_king_safety(const Position& pos); template - Score evaluate_shelter(const Position& pos, Square ksq); + Score evaluate_shelter(const Position& pos, Square ksq) const; Key key; Score scores[COLOR_NB]; @@ -59,12 +58,13 @@ struct Entry { Square kingSquares[COLOR_NB]; Score kingSafety[COLOR_NB]; int castlingRights[COLOR_NB]; + int blockedCount; }; typedef HashTable Table; Entry* probe(const Position& pos); -} // namespace Pawns +} // namespace Stockfish::Pawns #endif // #ifndef PAWNS_H_INCLUDED diff --git a/src/position.cpp b/src/position.cpp index 6336a5ed..17b165b9 100644 --- a/src/position.cpp +++ b/src/position.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -36,6 +34,8 @@ using std::string; +namespace Stockfish { + namespace Zobrist { Key psq[PIECE_NB][SQUARE_NB]; @@ -50,41 +50,6 @@ const string PieceToChar(" PNBRQK pnbrqk"); constexpr Piece Pieces[] = { W_PAWN, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING, B_PAWN, B_KNIGHT, B_BISHOP, B_ROOK, B_QUEEN, B_KING }; - -// min_attacker() is a helper function used by see_ge() to locate the least -// valuable attacker for the side to move, remove the attacker we just found -// from the bitboards and scan for new X-ray attacks behind it. - -template -PieceType min_attacker(const Bitboard* byTypeBB, Square to, Bitboard stmAttackers, - Bitboard& occupied, Bitboard& attackers) { - - Bitboard b = stmAttackers & byTypeBB[Pt]; - if (!b) - return min_attacker(byTypeBB, to, stmAttackers, occupied, attackers); - - occupied ^= lsb(b); // Remove the attacker from occupied - - // Add any X-ray attack behind the just removed piece. For instance with - // rooks in a8 and a7 attacking a1, after removing a7 we add rook in a8. - // Note that new added attackers can be of any color. - if (Pt == PAWN || Pt == BISHOP || Pt == QUEEN) - attackers |= attacks_bb(to, occupied) & (byTypeBB[BISHOP] | byTypeBB[QUEEN]); - - if (Pt == ROOK || Pt == QUEEN) - attackers |= attacks_bb(to, occupied) & (byTypeBB[ROOK] | byTypeBB[QUEEN]); - - // X-ray may add already processed pieces because byTypeBB[] is constant: in - // the rook example, now attackers contains _again_ rook in a7, so remove it. - attackers &= occupied; - return Pt; -} - -template<> -PieceType min_attacker(const Bitboard*, Square, Bitboard, Bitboard&, Bitboard&) { - return KING; // No need to update bitboards: it is the last cycle -} - } // namespace @@ -99,10 +64,11 @@ std::ostream& operator<<(std::ostream& os, const Position& pos) { for (File f = FILE_A; f <= FILE_H; ++f) os << " | " << PieceToChar[pos.piece_on(make_square(f, r))]; - os << " |\n +---+---+---+---+---+---+---+---+\n"; + os << " | " << (1 + r) << "\n +---+---+---+---+---+---+---+---+\n"; } - os << "\nFen: " << pos.fen() << "\nKey: " << std::hex << std::uppercase + os << " a b c d e f g h\n" + << "\nFen: " << pos.fen() << "\nKey: " << std::hex << std::uppercase << std::setfill('0') << std::setw(16) << pos.key() << std::setfill(' ') << std::dec << "\nCheckers: "; @@ -113,6 +79,8 @@ std::ostream& operator<<(std::ostream& os, const Position& pos) { && !pos.can_castle(ANY_CASTLING)) { StateInfo st; + ASSERT_ALIGNED(&st, Eval::NNUE::kCacheLineSize); + Position p; p.set(pos.fen(), pos.is_chess960(), &st, pos.this_thread()); Tablebases::ProbeState s1, s2; @@ -139,8 +107,7 @@ Key cuckoo[8192]; Move cuckooMove[8192]; -/// Position::init() initializes at startup the various arrays used to compute -/// hash keys. +/// Position::init() initializes at startup the various arrays used to compute hash keys void Position::init() { @@ -154,15 +121,7 @@ void Position::init() { Zobrist::enpassant[f] = rng.rand(); for (int cr = NO_CASTLING; cr <= ANY_CASTLING; ++cr) - { - Zobrist::castling[cr] = 0; - Bitboard b = cr; - while (b) - { - Key k = Zobrist::castling[1ULL << pop_lsb(&b)]; - Zobrist::castling[cr] ^= k ? k : rng.rand(); - } - } + Zobrist::castling[cr] = rng.rand(); Zobrist::side = rng.rand(); Zobrist::noPawns = rng.rand(); @@ -174,7 +133,7 @@ void Position::init() { for (Piece pc : Pieces) for (Square s1 = SQ_A1; s1 <= SQ_H8; ++s1) for (Square s2 = Square(s1 + 1); s2 <= SQ_H8; ++s2) - if (PseudoAttacks[type_of(pc)][s1] & s2) + if ((type_of(pc) != PAWN) && (attacks_bb(type_of(pc), s1, 0) & s2)) { Move move = make_move(s1, s2); Key key = Zobrist::psq[pc][s1] ^ Zobrist::psq[pc][s2] ^ Zobrist::side; @@ -221,9 +180,9 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th 4) En passant target square (in algebraic notation). If there's no en passant target square, this is "-". If a pawn has just made a 2-square move, this - is the position "behind" the pawn. This is recorded only if there is a pawn - in position to make an en passant capture, and if there really is a pawn - that might have advanced two squares. + is the position "behind" the pawn. Following X-FEN standard, this is recorded only + if there is a pawn in position to make an en passant capture, and if there really + is a pawn that might have advanced two squares. 5) Halfmove clock. This is the number of halfmoves since the last pawn advance or capture. This is used to determine if a draw can be claimed under the @@ -240,7 +199,6 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th std::memset(this, 0, sizeof(Position)); std::memset(si, 0, sizeof(StateInfo)); - std::fill_n(&pieceList[0][0], sizeof(pieceList) / sizeof(Square), SQ_NONE); st = si; ss >> std::noskipws; @@ -254,8 +212,7 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th else if (token == '/') sq += 2 * SOUTH; - else if ((idx = PieceToChar.find(token)) != string::npos) - { + else if ((idx = PieceToChar.find(token)) != string::npos) { put_piece(Piece(idx), sq); ++sq; } @@ -294,15 +251,37 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th set_castling_right(c, rsq); } - // 4. En passant square. Ignore if no pawn capture is possible + set_state(st); + + // 4. En passant square. + // Ignore if square is invalid or not on side to move relative rank 6. + bool enpassant = false; + if ( ((ss >> col) && (col >= 'a' && col <= 'h')) - && ((ss >> row) && (row == '3' || row == '6'))) + && ((ss >> row) && (row == (sideToMove == WHITE ? '6' : '3')))) { st->epSquare = make_square(File(col - 'a'), Rank(row - '1')); - if ( !(attackers_to(st->epSquare) & pieces(sideToMove, PAWN)) - || !(pieces(~sideToMove, PAWN) & (st->epSquare + pawn_push(~sideToMove)))) - st->epSquare = SQ_NONE; + // En passant square will be considered only if + // a) side to move have a pawn threatening epSquare + // b) there is an enemy pawn in front of epSquare + // c) there is no piece on epSquare or behind epSquare + // d) enemy pawn didn't block a check of its own color by moving forward + enpassant = pawn_attacks_bb(~sideToMove, st->epSquare) & pieces(sideToMove, PAWN) + && (pieces(~sideToMove, PAWN) & (st->epSquare + pawn_push(~sideToMove))) + && !(pieces() & (st->epSquare | (st->epSquare + pawn_push(sideToMove)))) + && ( file_of(square(sideToMove)) == file_of(st->epSquare) + || !(blockers_for_king(sideToMove) & (st->epSquare + pawn_push(~sideToMove)))); + } + + // It's necessary for st->previous to be intialized in this way because legality check relies on its existence + if (enpassant) { + st->previous = new StateInfo(); + remove_piece(st->epSquare - pawn_push(sideToMove)); + st->previous->checkersBB = attackers_to(square(~sideToMove)) & pieces(sideToMove); + st->previous->blockersForKing[WHITE] = slider_blockers(pieces(BLACK), square(WHITE), st->previous->pinners[BLACK]); + st->previous->blockersForKing[BLACK] = slider_blockers(pieces(WHITE), square(BLACK), st->previous->pinners[WHITE]); + put_piece(make_piece(~sideToMove, PAWN), st->epSquare - pawn_push(sideToMove)); } else st->epSquare = SQ_NONE; @@ -316,7 +295,8 @@ Position& Position::set(const string& fenStr, bool isChess960, StateInfo* si, Th chess960 = isChess960; thisThread = th; - set_state(st); + st->accumulator.state[WHITE] = Eval::NNUE::INIT; + st->accumulator.state[BLACK] = Eval::NNUE::INIT; assert(pos_is_ok()); @@ -341,7 +321,7 @@ void Position::set_castling_right(Color c, Square rfrom) { Square rto = relative_square(c, cr & KING_SIDE ? SQ_F1 : SQ_D1); castlingPath[cr] = (between_bb(rfrom, rto) | between_bb(kfrom, kto) | rto | kto) - & ~(square_bb(kfrom) | rfrom); + & ~(kfrom | rfrom); } @@ -354,10 +334,10 @@ void Position::set_check_info(StateInfo* si) const { Square ksq = square(~sideToMove); - si->checkSquares[PAWN] = attacks_from(ksq, ~sideToMove); - si->checkSquares[KNIGHT] = attacks_from(ksq); - si->checkSquares[BISHOP] = attacks_from(ksq); - si->checkSquares[ROOK] = attacks_from(ksq); + si->checkSquares[PAWN] = pawn_attacks_bb(~sideToMove, ksq); + si->checkSquares[KNIGHT] = attacks_bb(ksq); + si->checkSquares[BISHOP] = attacks_bb(ksq, pieces()); + si->checkSquares[ROOK] = attacks_bb(ksq, pieces()); si->checkSquares[QUEEN] = si->checkSquares[BISHOP] | si->checkSquares[ROOK]; si->checkSquares[KING] = 0; } @@ -410,11 +390,13 @@ void Position::set_state(StateInfo* si) const { Position& Position::set(const string& code, Color c, StateInfo* si) { - assert(code.length() > 0 && code.length() < 8); assert(code[0] == 'K'); string sides[] = { code.substr(code.find('K', 1)), // Weak - code.substr(0, code.find('K', 1)) }; // Strong + code.substr(0, std::min(code.find('v'), code.find('K', 1))) }; // Strong + + assert(sides[0].length() > 0 && sides[0].length() < 8); + assert(sides[1].length() > 0 && sides[1].length() < 8); std::transform(sides[c].begin(), sides[c].end(), sides[c].begin(), tolower); @@ -428,7 +410,7 @@ Position& Position::set(const string& code, Color c, StateInfo* si) { /// Position::fen() returns a FEN representation of the position. In case of /// Chess960 the Shredder-FEN notation is used. This is mainly a debugging function. -const string Position::fen() const { +string Position::fen() const { int emptyCnt; std::ostringstream ss; @@ -488,8 +470,8 @@ Bitboard Position::slider_blockers(Bitboard sliders, Square s, Bitboard& pinners pinners = 0; // Snipers are sliders that attack 's' when a piece and other snipers are removed - Bitboard snipers = ( (PseudoAttacks[ ROOK][s] & pieces(QUEEN, ROOK)) - | (PseudoAttacks[BISHOP][s] & pieces(QUEEN, BISHOP))) & sliders; + Bitboard snipers = ( (attacks_bb< ROOK>(s) & pieces(QUEEN, ROOK)) + | (attacks_bb(s) & pieces(QUEEN, BISHOP))) & sliders; Bitboard occupancy = pieces() ^ snipers; while (snipers) @@ -513,12 +495,12 @@ Bitboard Position::slider_blockers(Bitboard sliders, Square s, Bitboard& pinners Bitboard Position::attackers_to(Square s, Bitboard occupied) const { - return (attacks_from(s, BLACK) & pieces(WHITE, PAWN)) - | (attacks_from(s, WHITE) & pieces(BLACK, PAWN)) - | (attacks_from(s) & pieces(KNIGHT)) + return (pawn_attacks_bb(BLACK, s) & pieces(WHITE, PAWN)) + | (pawn_attacks_bb(WHITE, s) & pieces(BLACK, PAWN)) + | (attacks_bb(s) & pieces(KNIGHT)) | (attacks_bb< ROOK>(s, occupied) & pieces( ROOK, QUEEN)) | (attacks_bb(s, occupied) & pieces(BISHOP, QUEEN)) - | (attacks_from(s) & pieces(KING)); + | (attacks_bb(s) & pieces(KING)); } @@ -535,23 +517,11 @@ bool Position::legal(Move m) const { assert(color_of(moved_piece(m)) == us); assert(piece_on(square(us)) == make_piece(us, KING)); - // En passant captures are a tricky special case. Because they are rather - // uncommon, we do it simply by testing whether the king is attacked after - // the move is made. - if (type_of(m) == ENPASSANT) - { - Square ksq = square(us); - Square capsq = to - pawn_push(us); - Bitboard occupied = (pieces() ^ from ^ capsq) | to; - - assert(to == ep_square()); - assert(moved_piece(m) == make_piece(us, PAWN)); - assert(piece_on(capsq) == make_piece(~us, PAWN)); - assert(piece_on(to) == NO_PIECE); - - return !(attacks_bb< ROOK>(ksq, occupied) & pieces(~us, QUEEN, ROOK)) - && !(attacks_bb(ksq, occupied) & pieces(~us, QUEEN, BISHOP)); - } + // st->previous->blockersForKing consider capsq as empty. + // If pinned, it has to move along the king ray. + if (type_of(m) == EN_PASSANT) + return !(st->previous->blockersForKing[sideToMove] & from) + || aligned(from, to, square(us)); // Castling moves generation does not check if the castling path is clear of // enemy attacks, it is delayed at a later time: now! @@ -566,11 +536,9 @@ bool Position::legal(Move m) const { if (attackers_to(s) & pieces(~us)) return false; - // In case of Chess960, verify that when moving the castling rook we do - // not discover some hidden checker. + // In case of Chess960, verify if the Rook blocks some checks // For instance an enemy queen in SQ_A1 when castling rook is in SQ_B1. - return !chess960 - || !(attacks_bb(to, pieces() ^ to_sq(m)) & pieces(~us, ROOK, QUEEN)); + return !chess960 || !(blockers_for_king(us) & to_sq(m)); } // If the moving piece is a king, check whether the destination square is @@ -580,8 +548,8 @@ bool Position::legal(Move m) const { // A non-king move is legal if and only if it is not pinned or it // is moving along the ray towards or away from the king. - return !(blockers_for_king(us) & from) - || aligned(from, to, square(us)); + return !(blockers_for_king(us) & from) + || aligned(from, to, square(us)); } @@ -597,8 +565,10 @@ bool Position::pseudo_legal(const Move m) const { Piece pc = moved_piece(m); // Use a slower but simpler function for uncommon cases + // yet we skip the legality check of MoveList(). if (type_of(m) != NORMAL) - return MoveList(*this).contains(m); + return checkers() ? MoveList< EVASIONS>(*this).contains(m) + : MoveList(*this).contains(m); // Is not a promotion, so promotion piece must be empty if (promotion_type(m) - KNIGHT != NO_PIECE_TYPE) @@ -621,15 +591,15 @@ bool Position::pseudo_legal(const Move m) const { if ((Rank8BB | Rank1BB) & to) return false; - if ( !(attacks_from(from, us) & pieces(~us) & to) // Not a capture + if ( !(pawn_attacks_bb(us, from) & pieces(~us) & to) // Not a capture && !((from + pawn_push(us) == to) && empty(to)) // Not a single push && !( (from + 2 * pawn_push(us) == to) // Not a double push - && (rank_of(from) == relative_rank(us, RANK_2)) + && (relative_rank(us, from) == RANK_2) && empty(to) && empty(to - pawn_push(us)))) return false; } - else if (!(attacks_from(type_of(pc), from) & to)) + else if (!(attacks_bb(type_of(pc), from, pieces()) & to)) return false; // Evasions generator already takes care to avoid some kind of illegal moves @@ -668,11 +638,11 @@ bool Position::gives_check(Move m) const { Square to = to_sq(m); // Is there a direct check? - if (st->checkSquares[type_of(piece_on(from))] & to) + if (check_squares(type_of(piece_on(from))) & to) return true; // Is there a discovered check? - if ( (st->blockersForKing[~sideToMove] & from) + if ( (blockers_for_king(~sideToMove) & from) && !aligned(from, to, square(~sideToMove))) return true; @@ -684,31 +654,24 @@ bool Position::gives_check(Move m) const { case PROMOTION: return attacks_bb(promotion_type(m), to, pieces() ^ from) & square(~sideToMove); - // En passant capture with check? We have already handled the case - // of direct checks and ordinary discovered check, so the only case we - // need to handle is the unusual case of a discovered check through - // the captured pawn. - case ENPASSANT: - { - Square capsq = make_square(file_of(to), rank_of(from)); - Bitboard b = (pieces() ^ from ^ capsq) | to; + // The double-pushed pawn blocked a check? En Passant will remove the blocker. + // The only discovery check that wasn't handle is through capsq and fromsq + // So the King must be in the same rank as fromsq to consider this possibility. + // st->previous->blockersForKing consider capsq as empty. + case EN_PASSANT: + return st->previous->checkersBB + || ( rank_of(square(~sideToMove)) == rank_of(from) + && st->previous->blockersForKing[~sideToMove] & from); - return (attacks_bb< ROOK>(square(~sideToMove), b) & pieces(sideToMove, QUEEN, ROOK)) - | (attacks_bb(square(~sideToMove), b) & pieces(sideToMove, QUEEN, BISHOP)); - } - case CASTLING: + default: //CASTLING { - Square kfrom = from; - Square rfrom = to; // Castling is encoded as 'King captures the rook' - Square kto = relative_square(sideToMove, rfrom > kfrom ? SQ_G1 : SQ_C1); - Square rto = relative_square(sideToMove, rfrom > kfrom ? SQ_F1 : SQ_D1); + // Castling is encoded as 'king captures the rook' + Square ksq = square(~sideToMove); + Square rto = relative_square(sideToMove, to > from ? SQ_F1 : SQ_D1); - return (PseudoAttacks[ROOK][rto] & square(~sideToMove)) - && (attacks_bb(rto, (pieces() ^ kfrom ^ rfrom) | rto | kto) & square(~sideToMove)); + return (attacks_bb(rto) & ksq) + && (attacks_bb(rto, pieces() ^ from ^ to) & ksq); } - default: - assert(false); - return false; } } @@ -738,12 +701,18 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { ++st->rule50; ++st->pliesFromNull; + // Used by NNUE + st->accumulator.state[WHITE] = Eval::NNUE::EMPTY; + st->accumulator.state[BLACK] = Eval::NNUE::EMPTY; + auto& dp = st->dirtyPiece; + dp.dirty_num = 1; + Color us = sideToMove; Color them = ~us; Square from = from_sq(m); Square to = to_sq(m); Piece pc = piece_on(from); - Piece captured = type_of(m) == ENPASSANT ? make_piece(them, PAWN) : piece_on(to); + Piece captured = type_of(m) == EN_PASSANT ? make_piece(them, PAWN) : piece_on(to); assert(color_of(pc) == us); assert(captured == NO_PIECE || color_of(captured) == (type_of(m) != CASTLING ? them : us)); @@ -769,7 +738,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { // update non-pawn material. if (type_of(captured) == PAWN) { - if (type_of(m) == ENPASSANT) + if (type_of(m) == EN_PASSANT) { capsq -= pawn_push(us); @@ -778,8 +747,6 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { assert(relative_rank(us, to) == RANK_6); assert(piece_on(to) == NO_PIECE); assert(piece_on(capsq) == make_piece(them, PAWN)); - - board[capsq] = NO_PIECE; // Not done by remove_piece() } st->pawnKey ^= Zobrist::psq[captured][capsq]; @@ -787,8 +754,19 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { else st->nonPawnMaterial[them] -= PieceValue[MG][captured]; + if (Eval::useNNUE) + { + dp.dirty_num = 2; // 1 piece moved, 1 piece captured + dp.piece[1] = captured; + dp.from[1] = capsq; + dp.to[1] = SQ_NONE; + } + // Update board and piece lists - remove_piece(captured, capsq); + remove_piece(capsq); + + if (type_of(m) == EN_PASSANT) + board[capsq] = NO_PIECE; // Update material hash key and prefetch access to materialTable k ^= Zobrist::psq[captured][capsq]; @@ -812,21 +790,30 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { // Update castling rights if needed if (st->castlingRights && (castlingRightsMask[from] | castlingRightsMask[to])) { - int cr = castlingRightsMask[from] | castlingRightsMask[to]; - k ^= Zobrist::castling[st->castlingRights & cr]; - st->castlingRights &= ~cr; + k ^= Zobrist::castling[st->castlingRights]; + st->castlingRights &= ~(castlingRightsMask[from] | castlingRightsMask[to]); + k ^= Zobrist::castling[st->castlingRights]; } // Move the piece. The tricky Chess960 castling is handled earlier if (type_of(m) != CASTLING) - move_piece(pc, from, to); + { + if (Eval::useNNUE) + { + dp.piece[0] = pc; + dp.from[0] = from; + dp.to[0] = to; + } + + move_piece(from, to); + } // If the moving piece is a pawn do some special extra work if (type_of(pc) == PAWN) { - // Set en-passant square if the moved pawn can be captured + // Set en passant square if the moved pawn can be captured if ( (int(to) ^ int(from)) == 16 - && (attacks_from(to - pawn_push(us), us) & pieces(them, PAWN))) + && (pawn_attacks_bb(us, to - pawn_push(us)) & pieces(them, PAWN))) { st->epSquare = to - pawn_push(us); k ^= Zobrist::enpassant[file_of(st->epSquare)]; @@ -839,9 +826,19 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { assert(relative_rank(us, to) == RANK_8); assert(type_of(promotion) >= KNIGHT && type_of(promotion) <= QUEEN); - remove_piece(pc, to); + remove_piece(to); put_piece(promotion, to); + if (Eval::useNNUE) + { + // Promoting pawn to SQ_NONE, promoted piece from SQ_NONE + dp.to[0] = SQ_NONE; + dp.piece[dp.dirty_num] = promotion; + dp.from[dp.dirty_num] = SQ_NONE; + dp.to[dp.dirty_num] = to; + dp.dirty_num++; + } + // Update hash keys k ^= Zobrist::psq[pc][to] ^ Zobrist::psq[promotion][to]; st->pawnKey ^= Zobrist::psq[pc][to]; @@ -852,7 +849,7 @@ void Position::do_move(Move m, StateInfo& newSt, bool givesCheck) { st->nonPawnMaterial[us] += PieceValue[MG][promotion]; } - // Update pawn hash key and prefetch access to pawnsTable + // Update pawn hash key st->pawnKey ^= Zobrist::psq[pc][from] ^ Zobrist::psq[pc][to]; // Reset rule 50 draw counter @@ -919,7 +916,7 @@ void Position::undo_move(Move m) { assert(type_of(pc) == promotion_type(m)); assert(type_of(pc) >= KNIGHT && type_of(pc) <= QUEEN); - remove_piece(pc, to); + remove_piece(to); pc = make_piece(us, PAWN); put_piece(pc, to); } @@ -931,13 +928,13 @@ void Position::undo_move(Move m) { } else { - move_piece(pc, to, from); // Put the piece back at the source square + move_piece(to, from); // Put the piece back at the source square if (st->capturedPiece) { Square capsq = to; - if (type_of(m) == ENPASSANT) + if (type_of(m) == EN_PASSANT) { capsq -= pawn_push(us); @@ -970,16 +967,28 @@ 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) + { + auto& dp = st->dirtyPiece; + dp.piece[0] = make_piece(us, KING); + dp.from[0] = from; + dp.to[0] = to; + dp.piece[1] = make_piece(us, ROOK); + dp.from[1] = rfrom; + dp.to[1] = rto; + dp.dirty_num = 2; + } + // Remove both pieces first since squares could overlap in Chess960 - remove_piece(make_piece(us, KING), Do ? from : to); - remove_piece(make_piece(us, ROOK), Do ? rfrom : rto); - board[Do ? from : to] = board[Do ? rfrom : rto] = NO_PIECE; // Since remove_piece doesn't do it for us + remove_piece(Do ? from : to); + remove_piece(Do ? rfrom : rto); + board[Do ? from : to] = board[Do ? rfrom : rto] = NO_PIECE; // Since remove_piece doesn't do this for us put_piece(make_piece(us, KING), Do ? to : from); put_piece(make_piece(us, ROOK), Do ? rto : rfrom); } -/// Position::do(undo)_null_move() is used to do(undo) a "null move": It flips +/// Position::do(undo)_null_move() is used to do(undo) a "null move": it flips /// the side to move without executing any move on the board. void Position::do_null_move(StateInfo& newSt) { @@ -987,10 +996,16 @@ void Position::do_null_move(StateInfo& newSt) { assert(!checkers()); assert(&newSt != st); - std::memcpy(&newSt, st, sizeof(StateInfo)); + std::memcpy(&newSt, st, offsetof(StateInfo, accumulator)); + newSt.previous = st; st = &newSt; + st->dirtyPiece.dirty_num = 0; + st->dirtyPiece.piece[0] = NO_PIECE; // Avoid checks in UpdateAccumulator() + st->accumulator.state[WHITE] = Eval::NNUE::EMPTY; + st->accumulator.state[BLACK] = Eval::NNUE::EMPTY; + if (st->epSquare != SQ_NONE) { st->key ^= Zobrist::enpassant[file_of(st->epSquare)]; @@ -998,7 +1013,7 @@ void Position::do_null_move(StateInfo& newSt) { } st->key ^= Zobrist::side; - prefetch(TT.first_entry(st->key)); + prefetch(TT.first_entry(key())); ++st->rule50; st->pliesFromNull = 0; @@ -1023,7 +1038,7 @@ void Position::undo_null_move() { /// Position::key_after() computes the new hash key after the given move. Needed /// for speculative prefetch. It doesn't recognize special moves like castling, -/// en-passant and promotions. +/// en passant and promotions. Key Position::key_after(Move m) const { @@ -1048,79 +1063,99 @@ bool Position::see_ge(Move m, Value threshold) const { assert(is_ok(m)); - // Only deal with normal moves, assume others pass a simple see + // Only deal with normal moves, assume others pass a simple SEE if (type_of(m) != NORMAL) return VALUE_ZERO >= threshold; - Bitboard stmAttackers; Square from = from_sq(m), to = to_sq(m); - PieceType nextVictim = type_of(piece_on(from)); - Color us = color_of(piece_on(from)); - Color stm = ~us; // First consider opponent's move - Value balance; // Values of the pieces taken by us minus opponent's ones - // The opponent may be able to recapture so this is the best result - // we can hope for. - balance = PieceValue[MG][piece_on(to)] - threshold; - - if (balance < VALUE_ZERO) + int swap = PieceValue[MG][piece_on(to)] - threshold; + if (swap < 0) return false; - // Now assume the worst possible result: that the opponent can - // capture our piece for free. - balance -= PieceValue[MG][nextVictim]; - - // If it is enough (like in PxQ) then return immediately. Note that - // in case nextVictim == KING we always return here, this is ok - // if the given move is legal. - if (balance >= VALUE_ZERO) + swap = PieceValue[MG][piece_on(from)] - swap; + if (swap <= 0) return true; - // Find all attackers to the destination square, with the moving piece - // removed, but possibly an X-ray attacker added behind it. Bitboard occupied = pieces() ^ from ^ to; - Bitboard attackers = attackers_to(to, occupied) & occupied; + Color stm = color_of(piece_on(from)); + Bitboard attackers = attackers_to(to, occupied); + Bitboard stmAttackers, bb; + int res = 1; while (true) { - stmAttackers = attackers & pieces(stm); - - // Don't allow pinned pieces to attack (except the king) as long as - // any pinners are on their original square. - if (st->pinners[~stm] & occupied) - stmAttackers &= ~st->blockersForKing[stm]; + stm = ~stm; + attackers &= occupied; // If stm has no more attackers then give up: stm loses + if (!(stmAttackers = attackers & pieces(stm))) + break; + + // Don't allow pinned pieces to attack (except the king) as long as + // there are pinners on their original square. + if (pinners(~stm) & occupied) + stmAttackers &= ~blockers_for_king(stm); + if (!stmAttackers) break; + res ^= 1; + // Locate and remove the next least valuable attacker, and add to - // the bitboard 'attackers' the possibly X-ray attackers behind it. - nextVictim = min_attacker(byTypeBB, to, stmAttackers, occupied, attackers); - - stm = ~stm; // Switch side to move - - // Negamax the balance with alpha = balance, beta = balance+1 and - // add nextVictim's value. - // - // (balance, balance+1) -> (-balance-1, -balance) - // - assert(balance < VALUE_ZERO); - - balance = -balance - 1 - PieceValue[MG][nextVictim]; - - // If balance is still non-negative after giving away nextVictim then we - // win. The only thing to be careful about it is that we should revert - // stm if we captured with the king when the opponent still has attackers. - if (balance >= VALUE_ZERO) + // the bitboard 'attackers' any X-ray attackers behind it. + if ((bb = stmAttackers & pieces(PAWN))) { - if (nextVictim == KING && (attackers & pieces(stm))) - stm = ~stm; - break; + if ((swap = PawnValueMg - swap) < res) + break; + + occupied ^= lsb(bb); + attackers |= attacks_bb(to, occupied) & pieces(BISHOP, QUEEN); } - assert(nextVictim != KING); + + else if ((bb = stmAttackers & pieces(KNIGHT))) + { + if ((swap = KnightValueMg - swap) < res) + break; + + occupied ^= lsb(bb); + } + + else if ((bb = stmAttackers & pieces(BISHOP))) + { + if ((swap = BishopValueMg - swap) < res) + break; + + occupied ^= lsb(bb); + attackers |= attacks_bb(to, occupied) & pieces(BISHOP, QUEEN); + } + + else if ((bb = stmAttackers & pieces(ROOK))) + { + if ((swap = RookValueMg - swap) < res) + break; + + occupied ^= lsb(bb); + attackers |= attacks_bb(to, occupied) & pieces(ROOK, QUEEN); + } + + else if ((bb = stmAttackers & pieces(QUEEN))) + { + if ((swap = QueenValueMg - swap) < res) + break; + + occupied ^= lsb(bb); + attackers |= (attacks_bb(to, occupied) & pieces(BISHOP, QUEEN)) + | (attacks_bb(to, occupied) & pieces(ROOK , QUEEN)); + } + + else // KING + // If we "capture" with the king but opponent still has attackers, + // reverse the result. + return (attackers & ~pieces(stm)) ? res ^ 1 : res; } - return us != stm; // We break the above loop when stm loses + + return bool(res); } @@ -1134,10 +1169,7 @@ bool Position::is_draw(int ply) const { // Return a draw score if a position repeats once earlier but strictly // after the root, or repeats twice before or at the root. - if (st->repetition && st->repetition < ply) - return true; - - return false; + return st->repetition && st->repetition < ply; } @@ -1283,21 +1315,17 @@ bool Position::pos_is_ok() const { assert(0 && "pos_is_ok: Bitboards"); StateInfo si = *st; + ASSERT_ALIGNED(&si, Eval::NNUE::kCacheLineSize); + set_state(&si); if (std::memcmp(&si, st, sizeof(StateInfo))) assert(0 && "pos_is_ok: State"); for (Piece pc : Pieces) - { if ( pieceCount[pc] != popcount(pieces(color_of(pc), type_of(pc))) || pieceCount[pc] != std::count(board, board + SQUARE_NB, pc)) assert(0 && "pos_is_ok: Pieces"); - for (int i = 0; i < pieceCount[pc]; ++i) - if (board[pieceList[pc][i]] != pc || index[pieceList[pc][i]] != i) - assert(0 && "pos_is_ok: Index"); - } - for (Color c : { WHITE, BLACK }) for (CastlingRights cr : {c & KING_SIDE, c & QUEEN_SIDE}) { @@ -1312,3 +1340,5 @@ bool Position::pos_is_ok() const { return true; } + +} // namespace Stockfish diff --git a/src/position.h b/src/position.h index a0a9a306..a7654aa1 100644 --- a/src/position.h +++ b/src/position.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -27,8 +25,13 @@ #include #include "bitboard.h" +#include "evaluate.h" +#include "psqt.h" #include "types.h" +#include "nnue/nnue_accumulator.h" + +namespace Stockfish { /// StateInfo struct stores information needed to restore a Position object to /// its previous state when we retract a move. Whenever a move is made on the @@ -46,7 +49,6 @@ struct StateInfo { Square epSquare; // Not copied when making a move (will be recomputed anyhow) - int repetition; Key key; Bitboard checkersBB; Piece capturedPiece; @@ -54,8 +56,14 @@ struct StateInfo { Bitboard blockersForKing[COLOR_NB]; Bitboard pinners[COLOR_NB]; Bitboard checkSquares[PIECE_TYPE_NB]; + int repetition; + + // Used by NNUE + Eval::NNUE::Accumulator accumulator; + DirtyPiece dirtyPiece; }; + /// A list to keep track of the position states along the setup moves (from the /// start position to the position just before the search starts). Needed by /// 'draw by repetition' detection. Use a std::deque because pointers to @@ -80,10 +88,9 @@ public: // FEN string input/output Position& set(const std::string& fenStr, bool isChess960, StateInfo* si, Thread* th); Position& set(const std::string& code, Color c, StateInfo* si); - const std::string fen() const; + std::string fen() const; // Position representation - Bitboard pieces() const; Bitboard pieces(PieceType pt) const; Bitboard pieces(PieceType pt1, PieceType pt2) const; Bitboard pieces(Color c) const; @@ -94,12 +101,11 @@ public: bool empty(Square s) const; template int count(Color c) const; template int count() const; - template const Square* squares(Color c) const; template Square square(Color c) const; bool is_on_semiopen_file(Color c, Square s) const; // Castling - int castling_rights(Color c) const; + CastlingRights castling_rights(Color c) const; bool can_castle(CastlingRights cr) const; bool castling_impeded(CastlingRights cr) const; Square castling_rook_square(CastlingRights cr) const; @@ -108,14 +114,12 @@ public: Bitboard checkers() const; Bitboard blockers_for_king(Color c) const; Bitboard check_squares(PieceType pt) const; - bool is_discovery_check_on_king(Color c, Move m) const; + Bitboard pinners(Color c) const; + bool is_discovered_check_on_king(Color c, Move m) const; // Attacks to/from a given square Bitboard attackers_to(Square s) const; Bitboard attackers_to(Square s, Bitboard occupied) const; - Bitboard attacks_from(PieceType pt, Square s) const; - template Bitboard attacks_from(Square s) const; - template Bitboard attacks_from(Square s, Color c) const; Bitboard slider_blockers(Bitboard sliders, Square s, Bitboard& pinners) const; // Properties of moves @@ -166,6 +170,9 @@ public: bool pos_is_ok() const; void flip(); + // Used by NNUE + StateInfo* state() const; + private: // Initialization helpers (used while setting up a position) void set_castling_right(Color c, Square rfrom); @@ -174,8 +181,8 @@ private: // Other helpers void put_piece(Piece pc, Square s); - void remove_piece(Piece pc, Square s); - void move_piece(Piece pc, Square from, Square to); + void remove_piece(Square s); + void move_piece(Square from, Square to); template void do_castling(Color us, Square from, Square& to, Square& rfrom, Square& rto); @@ -184,8 +191,6 @@ private: Bitboard byTypeBB[PIECE_TYPE_NB]; Bitboard byColorBB[COLOR_NB]; int pieceCount[PIECE_NB]; - Square pieceList[PIECE_NB][16]; - int index[SQUARE_NB]; int castlingRightsMask[SQUARE_NB]; Square castlingRookSquare[CASTLING_RIGHT_NB]; Bitboard castlingPath[CASTLING_RIGHT_NB]; @@ -197,38 +202,31 @@ private: bool chess960; }; -namespace PSQT { - extern Score psq[PIECE_NB][SQUARE_NB]; -} - extern std::ostream& operator<<(std::ostream& os, const Position& pos); inline Color Position::side_to_move() const { return sideToMove; } -inline bool Position::empty(Square s) const { - return board[s] == NO_PIECE; -} - inline Piece Position::piece_on(Square s) const { + assert(is_ok(s)); return board[s]; } +inline bool Position::empty(Square s) const { + return piece_on(s) == NO_PIECE; +} + inline Piece Position::moved_piece(Move m) const { - return board[from_sq(m)]; + return piece_on(from_sq(m)); } -inline Bitboard Position::pieces() const { - return byTypeBB[ALL_PIECES]; -} - -inline Bitboard Position::pieces(PieceType pt) const { +inline Bitboard Position::pieces(PieceType pt = ALL_PIECES) const { return byTypeBB[pt]; } inline Bitboard Position::pieces(PieceType pt1, PieceType pt2) const { - return byTypeBB[pt1] | byTypeBB[pt2]; + return pieces(pt1) | pieces(pt2); } inline Bitboard Position::pieces(Color c) const { @@ -236,11 +234,11 @@ inline Bitboard Position::pieces(Color c) const { } inline Bitboard Position::pieces(Color c, PieceType pt) const { - return byColorBB[c] & byTypeBB[pt]; + return pieces(c) & pieces(pt); } inline Bitboard Position::pieces(Color c, PieceType pt1, PieceType pt2) const { - return byColorBB[c] & (byTypeBB[pt1] | byTypeBB[pt2]); + return pieces(c) & (pieces(pt1) | pieces(pt2)); } template inline int Position::count(Color c) const { @@ -248,16 +246,12 @@ template inline int Position::count(Color c) const { } template inline int Position::count() const { - return pieceCount[make_piece(WHITE, Pt)] + pieceCount[make_piece(BLACK, Pt)]; -} - -template inline const Square* Position::squares(Color c) const { - return pieceList[make_piece(c, Pt)]; + return count(WHITE) + count(BLACK); } template inline Square Position::square(Color c) const { - assert(pieceCount[make_piece(c, Pt)] == 1); - return pieceList[make_piece(c, Pt)][0]; + assert(count(c) == 1); + return lsb(pieces(c, Pt)); } inline Square Position::ep_square() const { @@ -272,37 +266,24 @@ inline bool Position::can_castle(CastlingRights cr) const { return st->castlingRights & cr; } -inline int Position::castling_rights(Color c) const { - return st->castlingRights & (c == WHITE ? WHITE_CASTLING : BLACK_CASTLING); +inline CastlingRights Position::castling_rights(Color c) const { + return c & CastlingRights(st->castlingRights); } inline bool Position::castling_impeded(CastlingRights cr) const { - return byTypeBB[ALL_PIECES] & castlingPath[cr]; + assert(cr == WHITE_OO || cr == WHITE_OOO || cr == BLACK_OO || cr == BLACK_OOO); + + return pieces() & castlingPath[cr]; } inline Square Position::castling_rook_square(CastlingRights cr) const { + assert(cr == WHITE_OO || cr == WHITE_OOO || cr == BLACK_OO || cr == BLACK_OOO); + return castlingRookSquare[cr]; } -template -inline Bitboard Position::attacks_from(Square s) const { - assert(Pt != PAWN); - return Pt == BISHOP || Pt == ROOK ? attacks_bb(s, byTypeBB[ALL_PIECES]) - : Pt == QUEEN ? attacks_from(s) | attacks_from(s) - : PseudoAttacks[Pt][s]; -} - -template<> -inline Bitboard Position::attacks_from(Square s, Color c) const { - return PawnAttacks[c][s]; -} - -inline Bitboard Position::attacks_from(PieceType pt, Square s) const { - return attacks_bb(pt, s, byTypeBB[ALL_PIECES]); -} - inline Bitboard Position::attackers_to(Square s) const { - return attackers_to(s, byTypeBB[ALL_PIECES]); + return attackers_to(s, pieces()); } inline Bitboard Position::checkers() const { @@ -313,11 +294,15 @@ inline Bitboard Position::blockers_for_king(Color c) const { return st->blockersForKing[c]; } +inline Bitboard Position::pinners(Color c) const { + return st->pinners[c]; +} + inline Bitboard Position::check_squares(PieceType pt) const { return st->checkSquares[pt]; } -inline bool Position::is_discovery_check_on_king(Color c, Move m) const { +inline bool Position::is_discovered_check_on_king(Color c, Move m) const { return st->blockersForKing[c] & from_sq(m); } @@ -327,7 +312,7 @@ inline bool Position::pawn_passed(Color c, Square s) const { inline bool Position::advanced_pawn_push(Move m) const { return type_of(moved_piece(m)) == PAWN - && relative_rank(sideToMove, to_sq(m)) > RANK_5; + && relative_rank(sideToMove, to_sq(m)) > RANK_6; } inline int Position::pawns_on_same_color_squares(Color c, Square s) const { @@ -335,7 +320,8 @@ inline int Position::pawns_on_same_color_squares(Color c, Square s) const { } inline Key Position::key() const { - return st->key; + return st->rule50 < 14 ? st->key + : st->key ^ make_key((st->rule50 - 14) / 8); } inline Key Position::pawn_key() const { @@ -355,7 +341,7 @@ inline Value Position::non_pawn_material(Color c) const { } inline Value Position::non_pawn_material() const { - return st->nonPawnMaterial[WHITE] + st->nonPawnMaterial[BLACK]; + return non_pawn_material(WHITE) + non_pawn_material(BLACK); } inline int Position::game_ply() const { @@ -367,8 +353,8 @@ inline int Position::rule50_count() const { } inline bool Position::opposite_bishops() const { - return pieceCount[W_BISHOP] == 1 - && pieceCount[B_BISHOP] == 1 + return count(WHITE) == 1 + && count(BLACK) == 1 && opposite_colors(square(WHITE), square(BLACK)); } @@ -384,7 +370,7 @@ inline bool Position::capture_or_promotion(Move m) const { inline bool Position::capture(Move m) const { assert(is_ok(m)); // Castling is encoded as "king captures rook" - return (!empty(to_sq(m)) && type_of(m) != CASTLING) || type_of(m) == ENPASSANT; + return (!empty(to_sq(m)) && type_of(m) != CASTLING) || type_of(m) == EN_PASSANT; } inline Piece Position::captured_piece() const { @@ -398,45 +384,34 @@ inline Thread* Position::this_thread() const { inline void Position::put_piece(Piece pc, Square s) { board[s] = pc; - byTypeBB[ALL_PIECES] |= s; - byTypeBB[type_of(pc)] |= s; + byTypeBB[ALL_PIECES] |= byTypeBB[type_of(pc)] |= s; byColorBB[color_of(pc)] |= s; - index[s] = pieceCount[pc]++; - pieceList[pc][index[s]] = s; + pieceCount[pc]++; pieceCount[make_piece(color_of(pc), ALL_PIECES)]++; psq += PSQT::psq[pc][s]; } -inline void Position::remove_piece(Piece pc, Square s) { +inline void Position::remove_piece(Square s) { - // WARNING: This is not a reversible operation. If we remove a piece in - // do_move() and then replace it in undo_move() we will put it at the end of - // the list and not in its original place, it means index[] and pieceList[] - // are not invariant to a do_move() + undo_move() sequence. + Piece pc = board[s]; byTypeBB[ALL_PIECES] ^= s; byTypeBB[type_of(pc)] ^= s; byColorBB[color_of(pc)] ^= s; /* board[s] = NO_PIECE; Not needed, overwritten by the capturing one */ - Square lastSquare = pieceList[pc][--pieceCount[pc]]; - index[lastSquare] = index[s]; - pieceList[pc][index[lastSquare]] = lastSquare; - pieceList[pc][pieceCount[pc]] = SQ_NONE; + pieceCount[pc]--; pieceCount[make_piece(color_of(pc), ALL_PIECES)]--; psq -= PSQT::psq[pc][s]; } -inline void Position::move_piece(Piece pc, Square from, Square to) { +inline void Position::move_piece(Square from, Square to) { - // index[from] is not updated and becomes stale. This works as long as index[] - // is accessed just by known occupied squares. - Bitboard fromTo = square_bb(from) | square_bb(to); + Piece pc = board[from]; + Bitboard fromTo = from | to; byTypeBB[ALL_PIECES] ^= fromTo; byTypeBB[type_of(pc)] ^= fromTo; byColorBB[color_of(pc)] ^= fromTo; board[from] = NO_PIECE; board[to] = pc; - index[to] = index[from]; - pieceList[pc][index[to]] = to; psq += PSQT::psq[pc][to] - PSQT::psq[pc][from]; } @@ -444,4 +419,11 @@ inline void Position::do_move(Move m, StateInfo& newSt) { do_move(m, newSt, gives_check(m)); } +inline StateInfo* Position::state() const { + + return st; +} + +} // namespace Stockfish + #endif // #ifndef POSITION_H_INCLUDED diff --git a/src/psqt.cpp b/src/psqt.cpp index 655eb993..33a3e00c 100644 --- a/src/psqt.cpp +++ b/src/psqt.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -18,45 +16,45 @@ along with this program. If not, see . */ + +#include "psqt.h" + #include +#include "bitboard.h" #include "types.h" -Value PieceValue[PHASE_NB][PIECE_NB] = { - { VALUE_ZERO, PawnValueMg, KnightValueMg, BishopValueMg, RookValueMg, QueenValueMg }, - { VALUE_ZERO, PawnValueEg, KnightValueEg, BishopValueEg, RookValueEg, QueenValueEg } -}; +namespace Stockfish { -namespace PSQT { +namespace +{ -#define S(mg, eg) make_score(mg, eg) +auto constexpr S = make_score; -// Bonus[PieceType][Square / 2] contains Piece-Square scores. For each piece -// type on a given square a (middlegame, endgame) score pair is assigned. Table -// is defined for files A..D and white side: it is symmetric for black side and -// second half of the files. +// 'Bonus' contains Piece-Square parameters. +// Scores are explicit for files A to D, implicitly mirrored for E to H. constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = { { }, { }, { // Knight - { S(-169,-105), S(-96,-74), S(-80,-46), S(-79,-18) }, - { S( -79, -70), S(-39,-56), S(-24,-15), S( -9, 6) }, - { S( -64, -38), S(-20,-33), S( 4, -5), S( 19, 27) }, - { S( -28, -36), S( 5, 0), S( 41, 13), S( 47, 34) }, - { S( -29, -41), S( 13,-20), S( 42, 4), S( 52, 35) }, - { S( -11, -51), S( 28,-38), S( 63,-17), S( 55, 19) }, - { S( -67, -64), S(-21,-45), S( 6,-37), S( 37, 16) }, - { S(-200, -98), S(-80,-89), S(-53,-53), S(-32,-16) } + { S(-175, -96), S(-92,-65), S(-74,-49), S(-73,-21) }, + { S( -77, -67), S(-41,-54), S(-27,-18), S(-15, 8) }, + { S( -61, -40), S(-17,-27), S( 6, -8), S( 12, 29) }, + { S( -35, -35), S( 8, -2), S( 40, 13), S( 49, 28) }, + { S( -34, -45), S( 13,-16), S( 44, 9), S( 51, 39) }, + { S( -9, -51), S( 22,-44), S( 58,-16), S( 53, 17) }, + { S( -67, -69), S(-27,-50), S( 4,-51), S( 37, 12) }, + { S(-201,-100), S(-83,-88), S(-56,-56), S(-26,-17) } }, { // Bishop - { S(-44,-63), S( -4,-30), S(-11,-35), S(-28, -8) }, - { S(-18,-38), S( 7,-13), S( 14,-14), S( 3, 0) }, - { S( -8,-18), S( 24, 0), S( -3, -7), S( 15, 13) }, - { S( 1,-26), S( 8, -3), S( 26, 1), S( 37, 16) }, - { S( -7,-24), S( 30, -6), S( 23,-10), S( 28, 17) }, - { S(-17,-26), S( 4, 2), S( -1, 1), S( 8, 16) }, - { S(-21,-34), S(-19,-18), S( 10, -7), S( -6, 9) }, - { S(-48,-51), S( -3,-40), S(-12,-39), S(-25,-20) } + { S(-37,-40), S(-4 ,-21), S( -6,-26), S(-16, -8) }, + { S(-11,-26), S( 6, -9), S( 13,-12), S( 3, 1) }, + { S(-5 ,-11), S( 15, -1), S( -4, -1), S( 12, 7) }, + { S(-4 ,-14), S( 8, -4), S( 18, 0), S( 27, 12) }, + { S(-8 ,-12), S( 20, -1), S( 15,-10), S( 22, 11) }, + { S(-11,-21), S( 4, 4), S( 1, 3), S( 8, 4) }, + { S(-12,-22), S(-10,-14), S( 4, -1), S( 0, 1) }, + { S(-34,-32), S( 1,-29), S(-10,-26), S(-16,-17) } }, { // Rook { S(-31, -9), S(-20,-13), S(-14,-10), S(-5, -9) }, @@ -70,61 +68,64 @@ constexpr Score Bonus[][RANK_NB][int(FILE_NB) / 2] = { }, { // Queen { S( 3,-69), S(-5,-57), S(-5,-47), S( 4,-26) }, - { S(-3,-55), S( 5,-31), S( 8,-22), S(12, -4) }, + { S(-3,-54), S( 5,-31), S( 8,-22), S(12, -4) }, { S(-3,-39), S( 6,-18), S(13, -9), S( 7, 3) }, { S( 4,-23), S( 5, -3), S( 9, 13), S( 8, 24) }, { S( 0,-29), S(14, -6), S(12, 9), S( 5, 21) }, - { S(-4,-38), S(10,-18), S( 6,-12), S( 8, 1) }, + { S(-4,-38), S(10,-18), S( 6,-11), S( 8, 1) }, { S(-5,-50), S( 6,-27), S(10,-24), S( 8, -8) }, - { S(-2,-75), S(-2,-52), S( 1,-43), S(-2,-36) } + { S(-2,-74), S(-2,-52), S( 1,-43), S(-2,-34) } }, { // King - { S(272, 0), S(325, 41), S(273, 80), S(190, 93) }, - { S(277, 57), S(305, 98), S(241,138), S(183,131) }, - { S(198, 86), S(253,138), S(168,165), S(120,173) }, - { S(169,103), S(191,152), S(136,168), S(108,169) }, - { S(145, 98), S(176,166), S(112,197), S( 69,194) }, - { S(122, 87), S(159,164), S( 85,174), S( 36,189) }, - { S( 87, 40), S(120, 99), S( 64,128), S( 25,141) }, - { S( 64, 5), S( 87, 60), S( 49, 75), S( 0, 75) } + { S(271, 1), S(327, 45), S(271, 85), S(198, 76) }, + { S(278, 53), S(303,100), S(234,133), S(179,135) }, + { S(195, 88), S(258,130), S(169,169), S(120,175) }, + { S(164,103), S(190,156), S(138,172), S( 98,172) }, + { S(154, 96), S(179,166), S(105,199), S( 70,199) }, + { S(123, 92), S(145,172), S( 81,184), S( 31,191) }, + { S( 88, 47), S(120,121), S( 65,116), S( 33,131) }, + { S( 59, 11), S( 89, 59), S( 45, 73), S( -1, 78) } } }; constexpr Score PBonus[RANK_NB][FILE_NB] = { // Pawn (asymmetric distribution) { }, - { S( 3,-10), S( 3, -6), S( 10, 10), S( 19, 0), S( 16, 14), S( 19, 7), S( 7, -5), S( -5,-19) }, - { S( -9,-10), S(-15,-10), S( 11,-10), S( 15, 4), S( 32, 4), S( 22, 3), S( 5, -6), S(-22, -4) }, - { S( -8, 6), S(-23, -2), S( 6, -8), S( 20, -4), S( 40,-13), S( 17,-12), S( 4,-10), S(-12, -9) }, - { S( 13, 9), S( 0, 4), S(-13, 3), S( 1,-12), S( 11,-12), S( -2, -6), S(-13, 13), S( 5, 8) }, - { S( -5, 28), S(-12, 20), S( -7, 21), S( 22, 28), S( -8, 30), S( -5, 7), S(-15, 6), S(-18, 13) }, - { S( -7, 0), S( 7,-11), S( -3, 12), S(-13, 21), S( 5, 25), S(-16, 19), S( 10, 4), S( -8, 7) } + { S( 2, -8), S( 4, -6), S( 11, 9), S( 18, 5), S( 16, 16), S( 21, 6), S( 9, -6), S( -3,-18) }, + { S( -9, -9), S(-15, -7), S( 11,-10), S( 15, 5), S( 31, 2), S( 23, 3), S( 6, -8), S(-20, -5) }, + { S( -3, 7), S(-20, 1), S( 8, -8), S( 19, -2), S( 39,-14), S( 17,-13), S( 2,-11), S( -5, -6) }, + { S( 11, 12), S( -4, 6), S(-11, 2), S( 2, -6), S( 11, -5), S( 0, -4), S(-12, 14), S( 5, 9) }, + { S( 3, 27), S(-11, 18), S( -6, 19), S( 22, 29), S( -8, 30), S( -5, 9), S(-14, 8), S(-11, 14) }, + { S( -7, -1), S( 6,-14), S( -2, 13), S(-11, 22), S( 4, 24), S(-14, 17), S( 10, 7), S( -9, 7) } }; -#undef S +} // namespace + + +namespace PSQT +{ Score psq[PIECE_NB][SQUARE_NB]; -// init() initializes piece-square tables: the white halves of the tables are -// copied from Bonus[] adding the piece value, then the black halves of the -// tables are initialized by flipping and changing the sign of the white scores. +// PSQT::init() initializes piece-square tables: the white halves of the tables are +// copied from Bonus[] and PBonus[], adding the piece value, then the black halves of +// the tables are initialized by flipping and changing the sign of the white scores. void init() { - for (Piece pc = W_PAWN; pc <= W_KING; ++pc) + for (Piece pc : {W_PAWN, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING}) { - PieceValue[MG][~pc] = PieceValue[MG][pc]; - PieceValue[EG][~pc] = PieceValue[EG][pc]; + Score score = make_score(PieceValue[MG][pc], PieceValue[EG][pc]); - Score score = make_score(PieceValue[MG][pc], PieceValue[EG][pc]); - - for (Square s = SQ_A1; s <= SQ_H8; ++s) - { - File f = map_to_queenside(file_of(s)); - psq[ pc][ s] = score + (type_of(pc) == PAWN ? PBonus[rank_of(s)][file_of(s)] - : Bonus[pc][rank_of(s)][f]); - psq[~pc][~s] = -psq[pc][s]; - } + for (Square s = SQ_A1; s <= SQ_H8; ++s) + { + File f = File(edge_distance(file_of(s))); + psq[ pc][s] = score + (type_of(pc) == PAWN ? PBonus[rank_of(s)][file_of(s)] + : Bonus[pc][rank_of(s)][f]); + psq[~pc][flip_rank(s)] = -psq[pc][s]; + } } } } // namespace PSQT + +} // namespace Stockfish diff --git a/src/psqt.h b/src/psqt.h new file mode 100644 index 00000000..7abb1483 --- /dev/null +++ b/src/psqt.h @@ -0,0 +1,38 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + + +#ifndef PSQT_H_INCLUDED +#define PSQT_H_INCLUDED + + +#include "types.h" + + +namespace Stockfish::PSQT +{ + +extern Score psq[PIECE_NB][SQUARE_NB]; + +// Fill psqt array from a set of internally linked parameters +extern void init(); + +} // namespace Stockfish::PSQT + + +#endif // PSQT_H_INCLUDED diff --git a/src/search.cpp b/src/search.cpp index 35ec26d4..d55fe59a 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -38,6 +36,8 @@ #include "uci.h" #include "syzygy/tbprobe.h" +namespace Stockfish { + namespace Search { LimitsType Limits; @@ -62,10 +62,12 @@ namespace { // Different node types, used as a template parameter enum NodeType { NonPV, PV }; - // Razor and futility margins - constexpr int RazorMargin = 661; + constexpr uint64_t TtHitAverageWindow = 4096; + constexpr uint64_t TtHitAverageResolution = 1024; + + // Futility margin Value futility_margin(Depth d, bool improving) { - return Value(198 * (d - improving)); + return Value(234 * (d - improving)); } // Reductions lookup table, initialized at startup @@ -73,19 +75,19 @@ namespace { Depth reduction(bool i, Depth d, int mn) { int r = Reductions[d] * Reductions[mn]; - return (r + 520) / 1024 + (!i && r > 999); + return (r + 503) / 1024 + (!i && r > 915); } - constexpr int futility_move_count(bool improving, int depth) { - return (5 + depth * depth) * (1 + improving) / 2; + constexpr int futility_move_count(bool improving, Depth depth) { + return (3 + depth * depth) / (2 - improving); } // History and stats update bonus, based on depth int stat_bonus(Depth d) { - return d > 17 ? -8 : 22 * d * d + 151 * d - 140; + return d > 14 ? 66 : 6 * d * d + 231 * d - 206; } - // Add a small random component to draw evaluations to avoid 3fold-blindness + // Add a small random component to draw evaluations to avoid 3-fold blindness Value value_draw(Thread* thisThread) { return VALUE_DRAW + Value(2 * (thisThread->nodes & 1) - 1); } @@ -151,11 +153,12 @@ namespace { Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth = 0); Value value_to_tt(Value v, int ply); - Value value_from_tt(Value v, int ply); + Value value_from_tt(Value v, int ply, int r50c); void update_pv(Move* pv, Move move, Move* childPv); void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus); - void update_quiet_stats(const Position& pos, Stack* ss, Move move, Move* quiets, int quietCount, int bonus); - void update_capture_stats(const Position& pos, Move move, Move* captures, int captureCount, int bonus); + void update_quiet_stats(const Position& pos, Stack* ss, Move move, int bonus, int depth); + void update_all_stats(const Position& pos, Stack* ss, Move bestMove, Value bestValue, Value beta, Square prevSq, + Move* quietsSearched, int quietCount, Move* capturesSearched, int captureCount, Depth depth); // perft() is our utility to verify move generation. All the leaf nodes up // to the given depth are generated and counted, and the sum is returned. @@ -163,6 +166,8 @@ namespace { uint64_t perft(Position& pos, Depth depth) { StateInfo st; + ASSERT_ALIGNED(&st, Eval::NNUE::kCacheLineSize); + uint64_t cnt, nodes = 0; const bool leaf = (depth == 2); @@ -191,7 +196,7 @@ namespace { void Search::init() { for (int i = 1; i < MAX_MOVES; ++i) - Reductions[i] = int((23.4 + std::log(Threads.size()) / 2) * std::log(i)); + Reductions[i] = int((21.3 + 2 * std::log(Threads.size())) * std::log(i + 0.25 * std::log(i))); } @@ -226,6 +231,8 @@ void MainThread::search() { Time.init(Limits, us, rootPos.game_ply()); TT.new_search(); + Eval::NNUE::verify(); + if (rootMoves.empty()) { rootMoves.emplace_back(MOVE_NONE); @@ -236,14 +243,8 @@ void MainThread::search() { } else { - for (Thread* th : Threads) - { - th->bestMoveChanges = 0; - if (th != this) - th->start_searching(); - } - - Thread::search(); // Let's start searching! + Threads.start_searching(); // start non-main threads + Thread::search(); // main thread start searching } // When we reach the maximum depth, we can arrive here without a raise of @@ -263,9 +264,7 @@ void MainThread::search() { Cluster::signals_sync(); // Wait until all threads have finished - for (Thread* th : Threads) - if (th != this) - th->wait_for_search_finished(); + Threads.wait_for_search_finished(); // When playing in 'nodes as time' mode, subtract the searched nodes from // the available ones before exiting. @@ -274,37 +273,11 @@ void MainThread::search() { Thread* bestThread = this; - // Check if there are threads with a better score than main thread - if ( Options["MultiPV"] == 1 + if ( int(Options["MultiPV"]) == 1 && !Limits.depth - && !(Skill(Options["Skill Level"]).enabled() || Options["UCI_LimitStrength"]) - && rootMoves[0].pv[0] != MOVE_NONE) - { - std::map votes; - Value minScore = this->rootMoves[0].score; - - // Find out minimum score - for (Thread* th: Threads) - minScore = std::min(minScore, th->rootMoves[0].score); - - // Vote according to score and depth, and select the best thread - for (Thread* th : Threads) - { - votes[th->rootMoves[0].pv[0]] += - (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth); - - if (bestThread->rootMoves[0].score >= VALUE_MATE_IN_MAX_PLY) - { - // Make sure we pick the shortest mate - if (th->rootMoves[0].score > bestThread->rootMoves[0].score) - bestThread = th; - } - else if ( th->rootMoves[0].score >= VALUE_MATE_IN_MAX_PLY - || votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]]) - bestThread = th; - } - } - + && !(Skill(Options["Skill Level"]).enabled() || int(Options["UCI_LimitStrength"])) + && rootMoves[0].pv[0] != MOVE_NONE) + bestThread = Threads.get_best_thread(); // Prepare PVLine and ponder move std::string PVLine = UCI::pv(bestThread->rootPos, bestThread->completedDepth, -VALUE_INFINITE, VALUE_INFINITE); @@ -322,7 +295,7 @@ void MainThread::search() { Cluster::rank()}; Cluster::pick_moves(mi, PVLine); - previousScore = static_cast(mi.score); + bestPreviousScore = static_cast(mi.score); if (Cluster::is_root()) { @@ -361,6 +334,7 @@ void Thread::search() { MainThread* mainThread = (this == Threads.main() ? Threads.main() : nullptr); double timeReduction = 1, totBestMoveChanges = 0; Color us = rootPos.side_to_move(); + int iterIdx = 0; std::memset(ss-7, 0, 10 * sizeof(Stack)); for (int i = 7; i > 0; i--) @@ -371,7 +345,20 @@ void Thread::search() { bestValue = delta = alpha = -VALUE_INFINITE; beta = VALUE_INFINITE; - size_t multiPV = Options["MultiPV"]; + if (mainThread) + { + if (mainThread->bestPreviousScore == VALUE_INFINITE) + for (int i = 0; i < 4; ++i) + mainThread->iterValue[i] = VALUE_ZERO; + else + for (int i = 0; i < 4; ++i) + mainThread->iterValue[i] = mainThread->bestPreviousScore; + } + + std::copy(&lowPlyHistory[2][0], &lowPlyHistory.back().back() + 1, &lowPlyHistory[0][0]); + std::fill(&lowPlyHistory[MAX_LPH - 2][0], &lowPlyHistory.back().back() + 1, 0); + + size_t multiPV = size_t(Options["MultiPV"]); // Pick integer skill levels, but non-deterministically round up or down // such that the average integer skill corresponds to the input floating point one. @@ -380,7 +367,7 @@ void Thread::search() { // for match (TC 60+0.6) results spanning a wide range of k values. PRNG rng(now()); double floatLevel = Options["UCI_LimitStrength"] ? - clamp(std::pow((Options["UCI_Elo"] - 1346.6) / 143.4, 1 / 0.806), 0.0, 20.0) : + std::clamp(std::pow((Options["UCI_Elo"] - 1346.6) / 143.4, 1 / 0.806), 0.0, 20.0) : double(Options["Skill Level"]); int intLevel = int(floatLevel) + ((floatLevel - int(floatLevel)) * 1024 > rng.rand() % 1024 ? 1 : 0); @@ -392,6 +379,7 @@ void Thread::search() { multiPV = std::max(multiPV, (size_t)4); multiPV = std::min(multiPV, rootMoves.size()); + ttHitAverage = TtHitAverageWindow * TtHitAverageResolution / 2; int ct = int(Options["Contempt"]) * PawnValueEg / 100; // From centipawns @@ -407,6 +395,8 @@ void Thread::search() { contempt = (us == WHITE ? make_score(ct, ct / 2) : -make_score(ct, ct / 2)); + int searchAgainCounter = 0; + // Iterative deepening loop until requested to stop or the target depth is reached while ( ++rootDepth < MAX_PLY && !Threads.stop @@ -425,6 +415,9 @@ void Thread::search() { size_t pvFirst = 0; pvLast = 0; + if (!Threads.increaseDepth) + searchAgainCounter++; + // MultiPV loop. We perform a full root search for each PV line for (pvIdx = 0; pvIdx < multiPV && !Threads.stop; ++pvIdx) { @@ -442,13 +435,13 @@ void Thread::search() { // Reset aspiration window starting size if (rootDepth >= 4) { - Value previousScore = rootMoves[pvIdx].previousScore; - delta = Value(21 + abs(previousScore) / 128); - alpha = std::max(previousScore - delta,-VALUE_INFINITE); - beta = std::min(previousScore + delta, VALUE_INFINITE); + Value prev = rootMoves[pvIdx].previousScore; + delta = Value(17); + alpha = std::max(prev - delta,-VALUE_INFINITE); + beta = std::min(prev + delta, VALUE_INFINITE); // Adjust contempt based on root move's previousScore (dynamic contempt) - int dct = ct + 86 * previousScore / (abs(previousScore) + 176); + int dct = ct + (113 - ct / 2) * prev / (abs(prev) + 147); contempt = (us == WHITE ? make_score(dct, dct / 2) : -make_score(dct, dct / 2)); @@ -457,11 +450,11 @@ void Thread::search() { // Start with a small aspiration window and, in the case of a fail // high/low, re-search with a bigger window until we don't fail // high/low anymore. - int failedHighCnt = 0; + failedHighCnt = 0; while (true) { - Depth adjustedDepth = std::max(1, rootDepth - failedHighCnt); - bestValue = ::search(rootPos, ss, alpha, beta, adjustedDepth, false); + Depth adjustedDepth = std::max(1, rootDepth - failedHighCnt - searchAgainCounter); + bestValue = Stockfish::search(rootPos, ss, alpha, beta, adjustedDepth, false); // Bring the best move to the front. It is critical that sorting // is done with a stable algorithm because all the values but the @@ -506,10 +499,7 @@ void Thread::search() { ++failedHighCnt; } else - { - ++rootMoves[pvIdx].bestMoveCount; break; - } delta += delta / 4 + 5; @@ -553,12 +543,13 @@ void Thread::search() { && !Threads.stop && !mainThread->stopOnPonderhit) { - double fallingEval = (354 + 10 * (mainThread->previousScore - bestValue)) / 692.0; - fallingEval = clamp(fallingEval, 0.5, 1.5); + double fallingEval = (318 + 6 * (mainThread->bestPreviousScore - bestValue) + + 6 * (mainThread->iterValue[iterIdx] - bestValue)) / 825.0; + fallingEval = std::clamp(fallingEval, 0.5, 1.5); // If the bestMove is stable over several iterations, reduce time accordingly - timeReduction = lastBestMoveDepth + 9 < completedDepth ? 1.97 : 0.98; - double reduction = (1.36 + mainThread->previousTimeReduction) / (2.29 * timeReduction); + timeReduction = lastBestMoveDepth + 9 < completedDepth ? 1.92 : 0.95; + double reduction = (1.47 + mainThread->previousTimeReduction) / (2.32 * timeReduction); // Use part of the gained time from a previous stable move for the current move for (Thread* th : Threads) @@ -566,11 +557,17 @@ void Thread::search() { totBestMoveChanges += th->bestMoveChanges; th->bestMoveChanges = 0; } - double bestMoveInstability = 1 + totBestMoveChanges / Threads.size(); + double bestMoveInstability = 1 + 2 * totBestMoveChanges / Threads.size(); - // Stop the search if we have only one legal move, or if available time elapsed - if ( rootMoves.size() == 1 - || Time.elapsed() > Time.optimum() * fallingEval * reduction * bestMoveInstability) + double totalTime = Time.optimum() * fallingEval * reduction * bestMoveInstability; + + // Cap used time in case of a single legal move for a better viewer experience in tournaments + // yielding correct scores and sufficiently fast moves. + if (rootMoves.size() == 1) + totalTime = std::min(500.0, totalTime); + + // Stop the search if we have exceeded the totalTime + if (Time.elapsed() > totalTime) { // If we are allowed to ponder do not stop the search now but // keep pondering until the GUI sends "ponderhit" or "stop". @@ -579,7 +576,16 @@ void Thread::search() { else Threads.stop = true; } + else if ( Threads.increaseDepth + && !mainThread->ponder + && Time.elapsed() > totalTime * 0.58) + Threads.increaseDepth = false; + else + Threads.increaseDepth = true; } + + mainThread->iterValue[iterIdx] = bestValue; + iterIdx = (iterIdx + 1) & 3; } if (!mainThread) @@ -603,6 +609,7 @@ namespace { constexpr bool PvNode = NT == PV; const bool rootNode = PvNode && ss->ply == 0; + const Depth maxNextDepth = rootNode ? depth : depth + 1; // Check if we have an upcoming move which draws by repetition, or // if the opponent had an alternative move earlier to this position. @@ -627,24 +634,28 @@ namespace { Move pv[MAX_PLY+1], capturesSearched[32], quietsSearched[64]; StateInfo st; + ASSERT_ALIGNED(&st, Eval::NNUE::kCacheLineSize); + TTEntry* tte; Key posKey; Move ttMove, move, excludedMove, bestMove; Depth extension, newDepth; - Value bestValue, value, ttValue, eval, maxValue; - bool ttHit, ttPv, inCheck, givesCheck, improving, doLMR, priorCapture; - bool captureOrPromotion, doFullDepthSearch, moveCountPruning, ttCapture; + Value bestValue, value, ttValue, eval, maxValue, probCutBeta; + bool formerPv, givesCheck, improving, didLMR, priorCapture; + bool captureOrPromotion, doFullDepthSearch, moveCountPruning, + ttCapture, singularQuietLMR; Piece movedPiece; - int moveCount, captureCount, quietCount, singularLMR; + int moveCount, captureCount, quietCount; // Step 1. Initialize node Thread* thisThread = pos.this_thread(); - inCheck = pos.checkers(); + ss->inCheck = pos.checkers(); priorCapture = pos.captured_piece(); Color us = pos.side_to_move(); - moveCount = captureCount = quietCount = singularLMR = ss->moveCount = 0; + moveCount = captureCount = quietCount = ss->moveCount = 0; bestValue = -VALUE_INFINITE; maxValue = VALUE_INFINITE; + ss->distanceFromPv = (PvNode ? 0 : ss->distanceFromPv); // Check for the available remaining time if (thisThread == Threads.main()) @@ -660,8 +671,8 @@ namespace { if ( Threads.stop.load(std::memory_order_relaxed) || pos.is_draw(ss->ply) || ss->ply >= MAX_PLY) - return (ss->ply >= MAX_PLY && !inCheck) ? evaluate(pos) - : value_draw(pos.this_thread()); + return (ss->ply >= MAX_PLY && !ss->inCheck) ? evaluate(pos) + : value_draw(pos.this_thread()); // Step 3. Mate distance pruning. Even if we mate at the next move our score // would be at best mate_in(ss->ply+1), but if alpha is already bigger because @@ -678,6 +689,7 @@ namespace { assert(0 <= ss->ply && ss->ply < MAX_PLY); (ss+1)->ply = ss->ply + 1; + (ss+1)->ttPv = false; (ss+1)->excludedMove = bestMove = MOVE_NONE; (ss+2)->killers[0] = (ss+2)->killers[1] = MOVE_NONE; Square prevSq = to_sq((ss-1)->currentMove); @@ -687,25 +699,37 @@ namespace { // starts with statScore = 0. Later grandchildren start with the last calculated // statScore of the previous grandchild. This influences the reduction rules in // LMR which are based on the statScore of parent position. - if (rootNode) - (ss+4)->statScore = 0; - else + if (!rootNode) (ss+2)->statScore = 0; // Step 4. Transposition table lookup. We don't want the score of a partial // search to overwrite a previous full search TT value, so we use a different // position key in case of an excluded move. excludedMove = ss->excludedMove; - posKey = pos.key() ^ Key(excludedMove << 16); // Isn't a very good hash - tte = TT.probe(posKey, ttHit); - ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE; + posKey = excludedMove == MOVE_NONE ? pos.key() : pos.key() ^ make_key(excludedMove); + tte = TT.probe(posKey, ss->ttHit); + ttValue = ss->ttHit ? value_from_tt(tte->value(), ss->ply, pos.rule50_count()) : VALUE_NONE; ttMove = rootNode ? thisThread->rootMoves[thisThread->pvIdx].pv[0] - : ttHit ? tte->move() : MOVE_NONE; - ttPv = PvNode || (ttHit && tte->is_pv()); + : ss->ttHit ? tte->move() : MOVE_NONE; + if (!excludedMove) + ss->ttPv = PvNode || (ss->ttHit && tte->is_pv()); + formerPv = ss->ttPv && !PvNode; + + // Update low ply history for previous move if we are near root and position is or has been in PV + if ( ss->ttPv + && depth > 12 + && ss->ply - 1 < MAX_LPH + && !priorCapture + && is_ok((ss-1)->currentMove)) + thisThread->lowPlyHistory[ss->ply - 1][from_to((ss-1)->currentMove)] << stat_bonus(depth - 5); + + // thisThread->ttHitAverage can be used to approximate the running average of ttHit + thisThread->ttHitAverage = (TtHitAverageWindow - 1) * thisThread->ttHitAverage / TtHitAverageWindow + + TtHitAverageResolution * ss->ttHit; // At non-PV nodes we check for an early TT cutoff if ( !PvNode - && ttHit + && ss->ttHit && tte->depth() >= depth && ttValue != VALUE_NONE // Possible in case of TT access race && (ttValue >= beta ? (tte->bound() & BOUND_LOWER) @@ -716,8 +740,9 @@ namespace { { if (ttValue >= beta) { + // Bonus for a quiet ttMove that fails high if (!pos.capture_or_promotion(ttMove)) - update_quiet_stats(pos, ss, ttMove, nullptr, 0, stat_bonus(depth)); + update_quiet_stats(pos, ss, ttMove, stat_bonus(depth), depth); // Extra penalty for early quiet moves of the previous ply if ((ss-1)->moveCount <= 2 && !priorCapture) @@ -731,7 +756,11 @@ namespace { update_continuation_histories(ss, pos.moved_piece(ttMove), to_sq(ttMove), penalty); } } - return ttValue; + + // Partial workaround for the graph history interaction problem + // For high rule50 counts don't produce transposition table cutoffs. + if (pos.rule50_count() < 90) + return ttValue; } // Step 5. Tablebases probe @@ -757,9 +786,10 @@ namespace { int drawScore = TB::UseRule50 ? 1 : 0; - value = wdl < -drawScore ? -VALUE_MATE + MAX_PLY + ss->ply + 1 - : wdl > drawScore ? VALUE_MATE - MAX_PLY - ss->ply - 1 - : VALUE_DRAW + 2 * wdl * drawScore; + // use the range VALUE_MATE_IN_MAX_PLY to VALUE_TB_WIN_IN_MAX_PLY to score + value = wdl < -drawScore ? VALUE_MATED_IN_MAX_PLY + ss->ply + 1 + : wdl > drawScore ? VALUE_MATE_IN_MAX_PLY - ss->ply - 1 + : VALUE_DRAW + 2 * wdl * drawScore; Bound b = wdl < -drawScore ? BOUND_UPPER : wdl > drawScore ? BOUND_LOWER : BOUND_EXACT; @@ -768,7 +798,7 @@ namespace { || (b == BOUND_LOWER ? value >= beta : value <= alpha)) { Cluster::save(thisThread, tte, - posKey, value_to_tt(value, ss->ply), ttPv, b, + posKey, value_to_tt(value, ss->ply), ss->ttPv, b, std::min(MAX_PLY - 1, depth + 6), MOVE_NONE, VALUE_NONE); @@ -786,20 +816,24 @@ namespace { } } + CapturePieceToHistory& captureHistory = thisThread->captureHistory; + // Step 6. Static evaluation of the position - if (inCheck) + if (ss->inCheck) { + // Skip early pruning when in check ss->staticEval = eval = VALUE_NONE; improving = false; - goto moves_loop; // Skip early pruning when in check + goto moves_loop; } - else if (ttHit) + else if (ss->ttHit) { // Never assume anything about values stored in TT ss->staticEval = eval = tte->eval(); if (eval == VALUE_NONE) ss->staticEval = eval = evaluate(pos); + // Randomize draw evaluation if (eval == VALUE_DRAW) eval = value_draw(thisThread); @@ -810,43 +844,48 @@ namespace { } else { + // In case of null move search use previous static eval with a different sign + // and addition of two tempos if ((ss-1)->currentMove != MOVE_NULL) - { - int bonus = -(ss-1)->statScore / 512; - - ss->staticEval = eval = evaluate(pos) + bonus; - } + ss->staticEval = eval = evaluate(pos); else - ss->staticEval = eval = -(ss-1)->staticEval + 2 * Eval::Tempo; + ss->staticEval = eval = -(ss-1)->staticEval + 2 * Tempo; + // Save static evaluation into transposition table Cluster::save(thisThread, tte, - posKey, VALUE_NONE, ttPv, BOUND_NONE, DEPTH_NONE, MOVE_NONE, + posKey, VALUE_NONE, ss->ttPv, BOUND_NONE, DEPTH_NONE, MOVE_NONE, eval); } - // Step 7. Razoring (~2 Elo) - if ( !rootNode // The required rootNode PV handling is not available in qsearch - && depth < 2 - && eval <= alpha - RazorMargin) - return qsearch(pos, ss, alpha, beta); + // Use static evaluation difference to improve quiet move ordering + if (is_ok((ss-1)->currentMove) && !(ss-1)->inCheck && !priorCapture) + { + int bonus = std::clamp(-depth * 4 * int((ss-1)->staticEval + ss->staticEval - 2 * Tempo), -1000, 1000); + thisThread->mainHistory[~us][from_to((ss-1)->currentMove)] << bonus; + } - improving = ss->staticEval >= (ss-2)->staticEval - || (ss-2)->staticEval == VALUE_NONE; + // Set up improving flag that is used in various pruning heuristics + // We define position as improving if static evaluation of position is better + // Than the previous static evaluation at our turn + // In case of us being in check at our previous move we look at move prior to it + improving = (ss-2)->staticEval == VALUE_NONE + ? ss->staticEval > (ss-4)->staticEval || (ss-4)->staticEval == VALUE_NONE + : ss->staticEval > (ss-2)->staticEval; - // Step 8. Futility pruning: child node (~30 Elo) + // Step 7. Futility pruning: child node (~50 Elo) if ( !PvNode - && depth < 7 + && depth < 9 && eval - futility_margin(depth, improving) >= beta && eval < VALUE_KNOWN_WIN) // Do not return unproven wins return eval; - // Step 9. Null move search with verification search (~40 Elo) + // Step 8. Null move search with verification search (~40 Elo) if ( !PvNode && (ss-1)->currentMove != MOVE_NULL - && (ss-1)->statScore < 22661 + && (ss-1)->statScore < 24185 && eval >= beta && eval >= ss->staticEval - && ss->staticEval >= beta - 33 * depth + 299 - improving * 30 + && ss->staticEval >= beta - 24 * depth - 34 * improving + 162 * ss->ttPv + 159 && !excludedMove && pos.non_pawn_material(us) && (ss->ply >= thisThread->nmpMinPly || us != thisThread->nmpColor)) @@ -854,7 +893,7 @@ namespace { assert(eval - beta >= 0); // Null move dynamic reduction based on depth and value - Depth R = (835 + 70 * depth) / 256 + std::min(int(eval - beta) / 185, 3); + Depth R = (1062 + 68 * depth) / 256 + std::min(int(eval - beta) / 190, 3); ss->currentMove = MOVE_NULL; ss->continuationHistory = &thisThread->continuationHistory[0][0][NO_PIECE][0]; @@ -867,11 +906,11 @@ namespace { if (nullValue >= beta) { - // Do not return unproven mate scores - if (nullValue >= VALUE_MATE_IN_MAX_PLY) + // Do not return unproven mate or TB scores + if (nullValue >= VALUE_TB_WIN_IN_MAX_PLY) nullValue = beta; - if (thisThread->nmpMinPly || (abs(beta) < VALUE_KNOWN_WIN && depth < 13)) + if (thisThread->nmpMinPly || (abs(beta) < VALUE_KNOWN_WIN && depth < 14)) return nullValue; assert(!thisThread->nmpMinPly); // Recursive verification is not allowed @@ -890,18 +929,40 @@ namespace { } } - // Step 10. ProbCut (~10 Elo) + probCutBeta = beta + 209 - 44 * improving; + + // Step 9. ProbCut (~10 Elo) // If we have a good enough capture and a reduced search returns a value // much above beta, we can (almost) safely prune the previous move. if ( !PvNode - && depth >= 5 - && abs(beta) < VALUE_MATE_IN_MAX_PLY) + && depth > 4 + && abs(beta) < VALUE_TB_WIN_IN_MAX_PLY + // if value from transposition table is lower than probCutBeta, don't attempt probCut + // there and in further interactions with transposition table cutoff depth is set to depth - 3 + // because probCut search has depth set to depth - 4 but we also do a move before it + // so effective depth is equal to depth - 3 + && !( ss->ttHit + && tte->depth() >= depth - 3 + && ttValue != VALUE_NONE + && ttValue < probCutBeta)) { - Value raisedBeta = std::min(beta + 191 - 46 * improving, VALUE_INFINITE); - MovePicker mp(pos, ttMove, raisedBeta - ss->staticEval, &thisThread->captureHistory); - int probCutCount = 0; + // if ttMove is a capture and value from transposition table is good enough produce probCut + // cutoff without digging into actual probCut search + if ( ss->ttHit + && tte->depth() >= depth - 3 + && ttValue != VALUE_NONE + && ttValue >= probCutBeta + && ttMove + && pos.capture_or_promotion(ttMove)) + return probCutBeta; - while ( (move = mp.next_move()) != MOVE_NONE + assert(probCutBeta < VALUE_INFINITE); + MovePicker mp(pos, ttMove, probCutBeta - ss->staticEval, &captureHistory); + int probCutCount = 0; + bool ttPv = ss->ttPv; + ss->ttPv = false; + + while ( (move = mp.next_move()) != MOVE_NONE && probCutCount < 2 + 2 * cutNode) if (move != excludedMove && pos.legal(move)) { @@ -912,7 +973,7 @@ namespace { probCutCount++; ss->currentMove = move; - ss->continuationHistory = &thisThread->continuationHistory[inCheck] + ss->continuationHistory = &thisThread->continuationHistory[ss->inCheck] [captureOrPromotion] [pos.moved_piece(move)] [to_sq(move)]; @@ -920,31 +981,54 @@ namespace { pos.do_move(move, st); // Perform a preliminary qsearch to verify that the move holds - value = -qsearch(pos, ss+1, -raisedBeta, -raisedBeta+1); + value = -qsearch(pos, ss+1, -probCutBeta, -probCutBeta+1); // If the qsearch held, perform the regular search - if (value >= raisedBeta) - value = -search(pos, ss+1, -raisedBeta, -raisedBeta+1, depth - 4, !cutNode); + if (value >= probCutBeta) + value = -search(pos, ss+1, -probCutBeta, -probCutBeta+1, depth - 4, !cutNode); pos.undo_move(move); - if (value >= raisedBeta) + if (value >= probCutBeta) + { + // if transposition table doesn't have equal or more deep info write probCut data into it + if ( !(ss->ttHit + && tte->depth() >= depth - 3 + && ttValue != VALUE_NONE)) + Cluster::save(thisThread, tte, posKey, value_to_tt(value, ss->ply), ttPv, + BOUND_LOWER, + depth - 3, move, ss->staticEval); return value; + } } + ss->ttPv = ttPv; } - // Step 11. Internal iterative deepening (~2 Elo) - if (depth >= 7 && !ttMove) - { - search(pos, ss, alpha, beta, depth - 7, cutNode); - - tte = TT.probe(posKey, ttHit); - ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE; - ttMove = ttHit ? tte->move() : MOVE_NONE; - } + // Step 10. If the position is not in TT, decrease depth by 2 + if ( PvNode + && depth >= 6 + && !ttMove) + depth -= 2; moves_loop: // When in check, search starts from here + ttCapture = ttMove && pos.capture_or_promotion(ttMove); + + // Step 11. A small Probcut idea, when we are in check + probCutBeta = beta + 400; + if ( ss->inCheck + && !PvNode + && depth >= 4 + && ttCapture + && (tte->bound() & BOUND_LOWER) + && tte->depth() >= depth - 3 + && ttValue >= probCutBeta + && abs(ttValue) <= VALUE_KNOWN_WIN + && abs(beta) <= VALUE_KNOWN_WIN + ) + return probCutBeta; + + const PieceToHistory* contHist[] = { (ss-1)->continuationHistory, (ss-2)->continuationHistory, nullptr , (ss-4)->continuationHistory, nullptr , (ss-6)->continuationHistory }; @@ -952,14 +1036,15 @@ moves_loop: // When in check, search starts from here Move countermove = thisThread->counterMoves[pos.piece_on(prevSq)][prevSq]; MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, - &thisThread->captureHistory, + &thisThread->lowPlyHistory, + &captureHistory, contHist, countermove, - ss->killers); + ss->killers, + ss->ply); value = bestValue; - moveCountPruning = false; - ttCapture = ttMove && pos.capture_or_promotion(ttMove); + singularQuietLMR = moveCountPruning = false; // Mark this node as being searched ThreadHolding th(thisThread, posKey, ss->ply); @@ -981,6 +1066,10 @@ moves_loop: // When in check, search starts from here thisThread->rootMoves.begin() + thisThread->pvLast, move)) continue; + // Check for legality + if (!rootNode && !pos.legal(move)) + continue; + ss->moveCount = ++moveCount; if (rootNode && Cluster::is_root() && thisThread == Threads.main() && Time.elapsed() > 3000) @@ -995,36 +1084,91 @@ moves_loop: // When in check, search starts from here movedPiece = pos.moved_piece(move); givesCheck = pos.gives_check(move); - // Step 13. Extensions (~70 Elo) + // Indicate PvNodes that will probably fail low if node was searched with non-PV search + // at depth equal or greater to current depth and result of this search was far below alpha + bool likelyFailLow = PvNode + && ttMove + && (tte->bound() & BOUND_UPPER) + && ttValue < alpha + 200 + 100 * depth + && tte->depth() >= depth; - // Singular extension search (~60 Elo). If all moves but one fail low on a + // Calculate new depth for this move + newDepth = depth - 1; + + // Step 13. Pruning at shallow depth (~200 Elo) + if ( !rootNode + && pos.non_pawn_material(us) + && bestValue > VALUE_TB_LOSS_IN_MAX_PLY) + { + // Skip quiet moves if movecount exceeds our FutilityMoveCount threshold + moveCountPruning = moveCount >= futility_move_count(improving, depth); + + // Reduced depth of the next LMR search + int lmrDepth = std::max(newDepth - reduction(improving, depth, moveCount), 0); + + if ( captureOrPromotion + || givesCheck) + { + // Capture history based pruning when the move doesn't give check + if ( !givesCheck + && lmrDepth < 1 + && captureHistory[movedPiece][to_sq(move)][type_of(pos.piece_on(to_sq(move)))] < 0) + continue; + + // SEE based pruning + if (!pos.see_ge(move, Value(-218) * depth)) // (~25 Elo) + continue; + } + else + { + // Countermoves based pruning (~20 Elo) + if ( lmrDepth < 4 + ((ss-1)->statScore > 0 || (ss-1)->moveCount == 1) + && (*contHist[0])[movedPiece][to_sq(move)] < CounterMovePruneThreshold + && (*contHist[1])[movedPiece][to_sq(move)] < CounterMovePruneThreshold) + continue; + + // Futility pruning: parent node (~5 Elo) + if ( lmrDepth < 7 + && !ss->inCheck + && ss->staticEval + 174 + 157 * lmrDepth <= alpha + && (*contHist[0])[movedPiece][to_sq(move)] + + (*contHist[1])[movedPiece][to_sq(move)] + + (*contHist[3])[movedPiece][to_sq(move)] + + (*contHist[5])[movedPiece][to_sq(move)] / 3 < 28255) + continue; + + // Prune moves with negative SEE (~20 Elo) + if (!pos.see_ge(move, Value(-(30 - std::min(lmrDepth, 18)) * lmrDepth * lmrDepth))) + continue; + } + } + + // Step 14. Extensions (~75 Elo) + + // Singular extension search (~70 Elo). If all moves but one fail low on a // search of (alpha-s, beta-s), and just one fails high on (alpha, beta), // then that move is singular and should be extended. To verify this we do // a reduced search on all the other moves but the ttMove and if the - // result is lower than ttValue minus a margin then we will extend the ttMove. - if ( depth >= 6 + // result is lower than ttValue minus a margin, then we will extend the ttMove. + if ( depth >= 7 && move == ttMove && !rootNode && !excludedMove // Avoid recursive singular search /* && ttValue != VALUE_NONE Already implicit in the next condition */ && abs(ttValue) < VALUE_KNOWN_WIN && (tte->bound() & BOUND_LOWER) - && tte->depth() >= depth - 3 - && pos.legal(move)) + && tte->depth() >= depth - 3) { - Value singularBeta = ttValue - 2 * depth; - Depth halfDepth = depth / 2; + Value singularBeta = ttValue - ((formerPv + 4) * depth) / 2; + Depth singularDepth = (depth - 1 + 3 * formerPv) / 2; ss->excludedMove = move; - value = search(pos, ss, singularBeta - 1, singularBeta, halfDepth, cutNode); + value = search(pos, ss, singularBeta - 1, singularBeta, singularDepth, cutNode); ss->excludedMove = MOVE_NONE; if (value < singularBeta) { extension = 1; - singularLMR++; - - if (value < singularBeta - std::min(4 * depth, 36)) - singularLMR++; + singularQuietLMR = !ttCapture; } // Multi-cut pruning @@ -1032,89 +1176,41 @@ moves_loop: // When in check, search starts from here // search without the ttMove. So we assume this expected Cut-node is not singular, // that multiple moves fail high, and we can prune the whole subtree by returning // a soft bound. - else if ( eval >= beta - && singularBeta >= beta) + else if (singularBeta >= beta) return singularBeta; + + // If the eval of ttMove is greater than beta we try also if there is another + // move that pushes it over beta, if so also produce a cutoff. + else if (ttValue >= beta) + { + ss->excludedMove = move; + value = search(pos, ss, beta - 1, beta, (depth + 3) / 2, cutNode); + ss->excludedMove = MOVE_NONE; + + if (value >= beta) + return beta; + } } // Check extension (~2 Elo) else if ( givesCheck - && (pos.is_discovery_check_on_king(~us, move) || pos.see_ge(move))) + && (pos.is_discovered_check_on_king(~us, move) || pos.see_ge(move))) extension = 1; - // Shuffle extension - else if ( PvNode - && pos.rule50_count() > 18 - && depth < 3 - && ++thisThread->shuffleExts < thisThread->nodes.load(std::memory_order_relaxed) / 4) // To avoid too many extensions + // Last captures extension + else if ( PieceValue[EG][pos.captured_piece()] > PawnValueEg + && pos.non_pawn_material() <= 2 * RookValueMg) extension = 1; - // Passed pawn extension - else if ( move == ss->killers[0] - && pos.advanced_pawn_push(move) - && pos.pawn_passed(us, to_sq(move))) - extension = 1; - - // Castling extension - if (type_of(move) == CASTLING) - extension = 1; - - // Calculate new depth for this move - newDepth = depth - 1 + extension; - - // Step 14. Pruning at shallow depth (~170 Elo) - if ( !rootNode - && pos.non_pawn_material(us) - && bestValue > VALUE_MATED_IN_MAX_PLY) - { - // Skip quiet moves if movecount exceeds our FutilityMoveCount threshold - moveCountPruning = moveCount >= futility_move_count(improving, depth); - - if ( !captureOrPromotion - && !givesCheck - && (!pos.advanced_pawn_push(move) || pos.non_pawn_material(~us) > BishopValueMg)) - { - // Move count based pruning - if (moveCountPruning) - continue; - - // Reduced depth of the next LMR search - int lmrDepth = std::max(newDepth - reduction(improving, depth, moveCount), 0); - - // Countermoves based pruning (~20 Elo) - if ( lmrDepth < 4 + ((ss-1)->statScore > 0 || (ss-1)->moveCount == 1) - && (*contHist[0])[movedPiece][to_sq(move)] < CounterMovePruneThreshold - && (*contHist[1])[movedPiece][to_sq(move)] < CounterMovePruneThreshold) - continue; - - // Futility pruning: parent node (~2 Elo) - if ( lmrDepth < 6 - && !inCheck - && ss->staticEval + 250 + 211 * lmrDepth <= alpha) - continue; - - // Prune moves with negative SEE (~10 Elo) - if (!pos.see_ge(move, Value(-(31 - std::min(lmrDepth, 18)) * lmrDepth * lmrDepth))) - continue; - } - else if ( !(givesCheck && extension) - && !pos.see_ge(move, Value(-199) * depth)) // (~20 Elo) - continue; - } + // Add extension to new depth + newDepth += extension; // Speculative prefetch as early as possible prefetch(TT.first_entry(pos.key_after(move))); - // Check for legality just before making the move - if (!rootNode && !pos.legal(move)) - { - ss->moveCount = --moveCount; - continue; - } - // Update the current move (this must be done after singular extension search) ss->currentMove = move; - ss->continuationHistory = &thisThread->continuationHistory[inCheck] + ss->continuationHistory = &thisThread->continuationHistory[ss->inCheck] [captureOrPromotion] [movedPiece] [to_sq(move)]; @@ -1122,96 +1218,130 @@ moves_loop: // When in check, search starts from here // Step 15. Make the move pos.do_move(move, st, givesCheck); - // Step 16. Reduced depth search (LMR). If the move fails high it will be - // re-searched at full depth. + (ss+1)->distanceFromPv = ss->distanceFromPv + moveCount - 1; + + // Step 16. Late moves reduction / extension (LMR, ~200 Elo) + // We use various heuristics for the sons of a node after the first son has + // been searched. In general we would like to reduce them, but there are many + // cases where we extend a son if it has good chances to be "interesting". if ( depth >= 3 && moveCount > 1 + 2 * rootNode - && (!rootNode || thisThread->best_move_count(move) == 0) && ( !captureOrPromotion || moveCountPruning || ss->staticEval + PieceValue[EG][pos.captured_piece()] <= alpha - || cutNode)) + || cutNode + || (!PvNode && !formerPv && captureHistory[movedPiece][to_sq(move)][type_of(pos.captured_piece())] < 3678) + || thisThread->ttHitAverage < 432 * TtHitAverageResolution * TtHitAverageWindow / 1024)) { Depth r = reduction(improving, depth, moveCount); - // Reduction if other threads are searching this position. + // Decrease reduction if the ttHit running average is large + if (thisThread->ttHitAverage > 537 * TtHitAverageResolution * TtHitAverageWindow / 1024) + r--; + + // Increase reduction if other threads are searching this position if (th.marked()) r++; // Decrease reduction if position is or has been on the PV - if (ttPv) + // and node is not likely to fail low. (~10 Elo) + if (ss->ttPv && !likelyFailLow) r -= 2; - // Decrease reduction if opponent's move count is high (~10 Elo) - if ((ss-1)->moveCount > 15) + // Increase reduction at root and non-PV nodes when the best move does not change frequently + if ((rootNode || !PvNode) && thisThread->rootDepth > 10 && thisThread->bestMoveChanges <= 2) + r++; + + // More reductions for late moves if position was not in previous PV + if (moveCountPruning && !formerPv) + r++; + + // Decrease reduction if opponent's move count is high (~5 Elo) + if ((ss-1)->moveCount > 13) r--; - // Decrease reduction if ttMove has been singularly extended - r -= singularLMR; + // Decrease reduction if ttMove has been singularly extended (~3 Elo) + if (singularQuietLMR) + r--; - if (!captureOrPromotion) + if (captureOrPromotion) { - // Increase reduction if ttMove is a capture (~0 Elo) + // Unless giving check, this capture is likely bad + if ( !givesCheck + && ss->staticEval + PieceValue[EG][pos.captured_piece()] + 210 * depth <= alpha) + r++; + } + else + { + // Increase reduction if ttMove is a capture (~5 Elo) if (ttCapture) r++; - // Increase reduction for cut nodes (~5 Elo) + // Increase reduction at root if failing high + r += rootNode ? thisThread->failedHighCnt * thisThread->failedHighCnt * moveCount / 512 : 0; + + // Increase reduction for cut nodes (~10 Elo) if (cutNode) r += 2; // Decrease reduction for moves that escape a capture. Filter out // castling moves, because they are coded as "king captures rook" and - // hence break make_move(). (~5 Elo) + // hence break make_move(). (~2 Elo) else if ( type_of(move) == NORMAL && !pos.see_ge(reverse_move(move))) - r -= 2; + r -= 2 + ss->ttPv - (type_of(movedPiece) == PAWN); ss->statScore = thisThread->mainHistory[us][from_to(move)] + (*contHist[0])[movedPiece][to_sq(move)] + (*contHist[1])[movedPiece][to_sq(move)] + (*contHist[3])[movedPiece][to_sq(move)] - - 4729; - - // Reset statScore to zero if negative and most stats shows >= 0 - if ( ss->statScore < 0 - && (*contHist[0])[movedPiece][to_sq(move)] >= 0 - && (*contHist[1])[movedPiece][to_sq(move)] >= 0 - && thisThread->mainHistory[us][from_to(move)] >= 0) - ss->statScore = 0; + - 4741; // Decrease/increase reduction by comparing opponent's stat score (~10 Elo) - if (ss->statScore >= -99 && (ss-1)->statScore < -116) + if (ss->statScore >= -89 && (ss-1)->statScore < -116) r--; - else if ((ss-1)->statScore >= -117 && ss->statScore < -144) + else if ((ss-1)->statScore >= -112 && ss->statScore < -100) r++; // Decrease/increase reduction for moves with a good/bad history (~30 Elo) - r -= ss->statScore / 16384; + // If we are not in check use statScore, but if we are in check we use + // the sum of main history and first continuation history with an offset. + if (ss->inCheck) + r -= (thisThread->mainHistory[us][from_to(move)] + + (*contHist[0])[movedPiece][to_sq(move)] - 3833) / 16384; + else + r -= ss->statScore / 14790; } - Depth d = clamp(newDepth - r, 1, newDepth); + // In general we want to cap the LMR depth search at newDepth. But for nodes + // close to the principal variation the cap is at (newDepth + 1), which will + // allow these nodes to be searched deeper than the pv (up to 4 plies deeper). + Depth d = std::clamp(newDepth - r, 1, newDepth + ((ss+1)->distanceFromPv <= 4)); value = -search(pos, ss+1, -(alpha+1), -alpha, d, true); - doFullDepthSearch = (value > alpha && d != newDepth), doLMR = true; + // If the son is reduced and fails high it will be re-searched at full depth + doFullDepthSearch = value > alpha && d < newDepth; + didLMR = true; } else - doFullDepthSearch = !PvNode || moveCount > 1, doLMR = false; + { + doFullDepthSearch = !PvNode || moveCount > 1; + didLMR = false; + } // Step 17. Full depth search when LMR is skipped or fails high if (doFullDepthSearch) { value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth, !cutNode); - if (doLMR && !captureOrPromotion) + // If the move passed LMR update its stats + if (didLMR && !captureOrPromotion) { int bonus = value > alpha ? stat_bonus(newDepth) : -stat_bonus(newDepth); - if (move == ss->killers[0]) - bonus += bonus / 4; - update_continuation_histories(ss, movedPiece, to_sq(move), bonus); } } @@ -1224,7 +1354,8 @@ moves_loop: // When in check, search starts from here (ss+1)->pv = pv; (ss+1)->pv[0] = MOVE_NONE; - value = -search(pos, ss+1, -beta, -alpha, newDepth, false); + value = -search(pos, ss+1, -beta, -alpha, + std::min(maxNextDepth, newDepth), false); } // Step 18. Undo move @@ -1257,8 +1388,7 @@ moves_loop: // When in check, search starts from here rm.pv.push_back(*m); // We record how often the best move has been changed in each - // iteration. This information is used for time management: When - // the best move changes frequently, we allocate some more time. + // iteration. This information is used for time management and LMR if (moveCount > 1) ++thisThread->bestMoveChanges; } @@ -1291,6 +1421,7 @@ moves_loop: // When in check, search starts from here } } + // If the move is worse than some previously searched move, remember it to update its stats later if (move != bestMove) { if (captureOrPromotion && captureCount < 32) @@ -1314,26 +1445,17 @@ moves_loop: // When in check, search starts from here // must be a mate or a stalemate. If we are in a singular extension search then // return a fail low score. - assert(moveCount || !inCheck || excludedMove || !MoveList(pos).size()); + assert(moveCount || !ss->inCheck || excludedMove || !MoveList(pos).size()); if (!moveCount) bestValue = excludedMove ? alpha - : inCheck ? mated_in(ss->ply) : VALUE_DRAW; + : ss->inCheck ? mated_in(ss->ply) : VALUE_DRAW; + + // If there is a move which produces search value greater than alpha we update stats of searched moves else if (bestMove) - { - // Quiet best move: update move sorting heuristics - if (!pos.capture_or_promotion(bestMove)) - update_quiet_stats(pos, ss, bestMove, quietsSearched, quietCount, - stat_bonus(depth + (bestValue > beta + PawnValueMg))); + update_all_stats(pos, ss, bestMove, bestValue, beta, prevSq, + quietsSearched, quietCount, capturesSearched, captureCount, depth); - update_capture_stats(pos, bestMove, capturesSearched, captureCount, stat_bonus(depth + 1)); - - // Extra penalty for a quiet TT or main killer move in previous ply when it gets refuted - if ( ((ss-1)->moveCount == 1 || ((ss-1)->currentMove == (ss-1)->killers[0])) - && !priorCapture) - update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -stat_bonus(depth + 1)); - - } // Bonus for prior countermove that caused the fail low else if ( (depth >= 3 || PvNode) && !priorCapture) @@ -1342,9 +1464,19 @@ moves_loop: // When in check, search starts from here if (PvNode) bestValue = std::min(bestValue, maxValue); - if (!excludedMove) + // If no good move is found and the previous position was ttPv, then the previous + // opponent move is probably good and the new position is added to the search tree. + if (bestValue <= alpha) + ss->ttPv = ss->ttPv || ((ss-1)->ttPv && depth > 3); + // Otherwise, a counter move has been found and if the position is the last leaf + // in the search tree, remove the position from the search tree. + else if (depth > 3) + ss->ttPv = ss->ttPv && (ss+1)->ttPv; + + // Write gathered information in transposition table + if (!excludedMove && !(rootNode && thisThread->pvIdx)) Cluster::save(thisThread, tte, - posKey, value_to_tt(bestValue, ss->ply), ttPv, + posKey, value_to_tt(bestValue, ss->ply), ss->ttPv, bestValue >= beta ? BOUND_LOWER : PvNode && bestMove ? BOUND_EXACT : BOUND_UPPER, depth, bestMove, ss->staticEval); @@ -1368,12 +1500,14 @@ moves_loop: // When in check, search starts from here Move pv[MAX_PLY+1]; StateInfo st; + ASSERT_ALIGNED(&st, Eval::NNUE::kCacheLineSize); + TTEntry* tte; Key posKey; Move ttMove, move, bestMove; Depth ttDepth; Value bestValue, value, ttValue, futilityValue, futilityBase, oldAlpha; - bool ttHit, pvHit, inCheck, givesCheck, captureOrPromotion, evasionPrunable; + bool pvHit, givesCheck, captureOrPromotion; int moveCount; if (PvNode) @@ -1386,30 +1520,30 @@ moves_loop: // When in check, search starts from here Thread* thisThread = pos.this_thread(); (ss+1)->ply = ss->ply + 1; bestMove = MOVE_NONE; - inCheck = pos.checkers(); + ss->inCheck = pos.checkers(); moveCount = 0; // Check for an immediate draw or maximum ply reached if ( pos.is_draw(ss->ply) || ss->ply >= MAX_PLY) - return (ss->ply >= MAX_PLY && !inCheck) ? evaluate(pos) : VALUE_DRAW; + return (ss->ply >= MAX_PLY && !ss->inCheck) ? evaluate(pos) : VALUE_DRAW; assert(0 <= ss->ply && ss->ply < MAX_PLY); // Decide whether or not to include checks: this fixes also the type of // TT entry depth that we are going to use. Note that in qsearch we use // only two types of depth in TT: DEPTH_QS_CHECKS or DEPTH_QS_NO_CHECKS. - ttDepth = inCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS + ttDepth = ss->inCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS : DEPTH_QS_NO_CHECKS; // Transposition table lookup posKey = pos.key(); - tte = TT.probe(posKey, ttHit); - ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE; - ttMove = ttHit ? tte->move() : MOVE_NONE; - pvHit = ttHit && tte->is_pv(); + tte = TT.probe(posKey, ss->ttHit); + ttValue = ss->ttHit ? value_from_tt(tte->value(), ss->ply, pos.rule50_count()) : VALUE_NONE; + ttMove = ss->ttHit ? tte->move() : MOVE_NONE; + pvHit = ss->ttHit && tte->is_pv(); if ( !PvNode - && ttHit + && ss->ttHit && tte->depth() >= ttDepth && ttValue != VALUE_NONE // Only in case of TT access race && (ttValue >= beta ? (tte->bound() & BOUND_LOWER) @@ -1417,14 +1551,14 @@ moves_loop: // When in check, search starts from here return ttValue; // Evaluate the position statically - if (inCheck) + if (ss->inCheck) { ss->staticEval = VALUE_NONE; bestValue = futilityBase = -VALUE_INFINITE; } else { - if (ttHit) + if (ss->ttHit) { // Never assume anything about values stored in TT if ((ss->staticEval = bestValue = tte->eval()) == VALUE_NONE) @@ -1436,16 +1570,19 @@ moves_loop: // When in check, search starts from here bestValue = ttValue; } else + // In case of null move search use previous static eval with a different sign + // and addition of two tempos ss->staticEval = bestValue = (ss-1)->currentMove != MOVE_NULL ? evaluate(pos) - : -(ss-1)->staticEval + 2 * Eval::Tempo; + : -(ss-1)->staticEval + 2 * Tempo; // Stand pat. Return immediately if static value is at least beta if (bestValue >= beta) { - if (!ttHit) + // Save gathered info in transposition table + if (!ss->ttHit) Cluster::save(thisThread, tte, - posKey, value_to_tt(bestValue, ss->ply), pvHit, BOUND_LOWER, + posKey, value_to_tt(bestValue, ss->ply), false, BOUND_LOWER, DEPTH_NONE, MOVE_NONE, ss->staticEval); return bestValue; @@ -1454,7 +1591,7 @@ moves_loop: // When in check, search starts from here if (PvNode && bestValue > alpha) alpha = bestValue; - futilityBase = bestValue + 153; + futilityBase = bestValue + 155; } const PieceToHistory* contHist[] = { (ss-1)->continuationHistory, (ss-2)->continuationHistory, @@ -1463,8 +1600,8 @@ moves_loop: // When in check, search starts from here // Initialize a MovePicker object for the current position, and prepare // to search the moves. Because the depth is <= 0 here, only captures, - // queen promotions and checks (only if depth >= DEPTH_QS_CHECKS) will - // be generated. + // queen and checking knight promotions, and other checks(only if depth >= DEPTH_QS_CHECKS) + // will be generated. MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, &thisThread->captureHistory, contHist, @@ -1480,13 +1617,15 @@ moves_loop: // When in check, search starts from here moveCount++; - // Futility pruning - if ( !inCheck + // Futility pruning and moveCount pruning + if ( bestValue > VALUE_TB_LOSS_IN_MAX_PLY && !givesCheck && futilityBase > -VALUE_KNOWN_WIN && !pos.advanced_pawn_push(move)) { - assert(type_of(move) != ENPASSANT); // Due to !pos.advanced_pawn_push + + if (moveCount > 2) + continue; futilityValue = futilityBase + PieceValue[EG][pos.piece_on(to_sq(move))]; @@ -1503,15 +1642,8 @@ moves_loop: // When in check, search starts from here } } - // Detect non-capture evasions that are candidates to be pruned - evasionPrunable = inCheck - && (depth != 0 || moveCount > 2) - && bestValue > VALUE_MATED_IN_MAX_PLY - && !pos.capture(move); - - // Don't search moves with negative SEE values - if ( (!inCheck || evasionPrunable) - && (!givesCheck || !(pos.blockers_for_king(~pos.side_to_move()) & from_sq(move))) + // Do not search moves with negative SEE values + if ( bestValue > VALUE_TB_LOSS_IN_MAX_PLY && !pos.see_ge(move)) continue; @@ -1526,11 +1658,18 @@ moves_loop: // When in check, search starts from here } ss->currentMove = move; - ss->continuationHistory = &thisThread->continuationHistory[inCheck] + ss->continuationHistory = &thisThread->continuationHistory[ss->inCheck] [captureOrPromotion] [pos.moved_piece(move)] [to_sq(move)]; + // CounterMove based pruning + if ( !captureOrPromotion + && bestValue > VALUE_TB_LOSS_IN_MAX_PLY + && (*contHist[0])[pos.moved_piece(move)][to_sq(move)] < CounterMovePruneThreshold + && (*contHist[1])[pos.moved_piece(move)][to_sq(move)] < CounterMovePruneThreshold) + continue; + // Make and search the move pos.do_move(move, st, givesCheck); value = -qsearch(pos, ss+1, -beta, -alpha, depth - 1); @@ -1558,11 +1697,16 @@ moves_loop: // When in check, search starts from here } } - // All legal moves have been searched. A special case: If we're in check + // All legal moves have been searched. A special case: if we're in check // and no legal moves were found, it is checkmate. - if (inCheck && bestValue == -VALUE_INFINITE) - return mated_in(ss->ply); // Plies to mate from the root + if (ss->inCheck && bestValue == -VALUE_INFINITE) + { + assert(!MoveList(pos).size()); + return mated_in(ss->ply); // Plies to mate from the root + } + + // Save gathered info in transposition table Cluster::save(thisThread, tte, posKey, value_to_tt(bestValue, ss->ply), pvHit, bestValue >= beta ? BOUND_LOWER : @@ -1575,28 +1719,47 @@ moves_loop: // When in check, search starts from here } - // value_to_tt() adjusts a mate score from "plies to mate from the root" to - // "plies to mate from the current position". Non-mate scores are unchanged. + // value_to_tt() adjusts a mate or TB score from "plies to mate from the root" to + // "plies to mate from the current position". Standard scores are unchanged. // The function is called before storing a value in the transposition table. Value value_to_tt(Value v, int ply) { assert(v != VALUE_NONE); - return v >= VALUE_MATE_IN_MAX_PLY ? v + ply - : v <= VALUE_MATED_IN_MAX_PLY ? v - ply : v; + return v >= VALUE_TB_WIN_IN_MAX_PLY ? v + ply + : v <= VALUE_TB_LOSS_IN_MAX_PLY ? v - ply : v; } - // value_from_tt() is the inverse of value_to_tt(): It adjusts a mate score - // from the transposition table (which refers to the plies to mate/be mated - // from current position) to "plies to mate/be mated from the root". + // value_from_tt() is the inverse of value_to_tt(): it adjusts a mate or TB score + // from the transposition table (which refers to the plies to mate/be mated from + // current position) to "plies to mate/be mated (TB win/loss) from the root". However, + // for mate scores, to avoid potentially false mate scores related to the 50 moves rule + // and the graph history interaction, we return an optimal TB score instead. - Value value_from_tt(Value v, int ply) { + Value value_from_tt(Value v, int ply, int r50c) { - return v == VALUE_NONE ? VALUE_NONE - : v >= VALUE_MATE_IN_MAX_PLY ? v - ply - : v <= VALUE_MATED_IN_MAX_PLY ? v + ply : v; + if (v == VALUE_NONE) + return VALUE_NONE; + + if (v >= VALUE_TB_WIN_IN_MAX_PLY) // TB win or better + { + if (v >= VALUE_MATE_IN_MAX_PLY && VALUE_MATE - v > 99 - r50c) + return VALUE_MATE_IN_MAX_PLY - 1; // do not return a potentially false mate score + + return v - ply; + } + + if (v <= VALUE_TB_LOSS_IN_MAX_PLY) // TB loss or worse + { + if (v <= VALUE_MATED_IN_MAX_PLY && VALUE_MATE + v > 99 - r50c) + return VALUE_MATED_IN_MAX_PLY + 1; // do not return a potentially false mate score + + return v + ply; + } + + return v; } @@ -1610,44 +1773,75 @@ moves_loop: // When in check, search starts from here } + // update_all_stats() updates stats at the end of search() when a bestMove is found + + void update_all_stats(const Position& pos, Stack* ss, Move bestMove, Value bestValue, Value beta, Square prevSq, + Move* quietsSearched, int quietCount, Move* capturesSearched, int captureCount, Depth depth) { + + int bonus1, bonus2; + Color us = pos.side_to_move(); + Thread* thisThread = pos.this_thread(); + CapturePieceToHistory& captureHistory = thisThread->captureHistory; + Piece moved_piece = pos.moved_piece(bestMove); + PieceType captured = type_of(pos.piece_on(to_sq(bestMove))); + + bonus1 = stat_bonus(depth + 1); + bonus2 = bestValue > beta + PawnValueMg ? bonus1 // larger bonus + : std::min(bonus1, stat_bonus(depth)); // smaller bonus + + if (!pos.capture_or_promotion(bestMove)) + { + // Increase stats for the best move in case it was a quiet move + update_quiet_stats(pos, ss, bestMove, bonus2, depth); + + // Decrease stats for all non-best quiet moves + for (int i = 0; i < quietCount; ++i) + { + thisThread->mainHistory[us][from_to(quietsSearched[i])] << -bonus2; + update_continuation_histories(ss, pos.moved_piece(quietsSearched[i]), to_sq(quietsSearched[i]), -bonus2); + } + } + else + // Increase stats for the best move in case it was a capture move + captureHistory[moved_piece][to_sq(bestMove)][captured] << bonus1; + + // Extra penalty for a quiet early move that was not a TT move or + // main killer move in previous ply when it gets refuted. + if ( ((ss-1)->moveCount == 1 + (ss-1)->ttHit || ((ss-1)->currentMove == (ss-1)->killers[0])) + && !pos.captured_piece()) + update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -bonus1); + + // Decrease stats for all non-best capture moves + for (int i = 0; i < captureCount; ++i) + { + moved_piece = pos.moved_piece(capturesSearched[i]); + captured = type_of(pos.piece_on(to_sq(capturesSearched[i]))); + captureHistory[moved_piece][to_sq(capturesSearched[i])][captured] << -bonus1; + } + } + + // update_continuation_histories() updates histories of the move pairs formed - // by moves at ply -1, -2, and -4 with current move. + // by moves at ply -1, -2, -4, and -6 with current move. void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus) { for (int i : {1, 2, 4, 6}) + { + // Only update first 2 continuation histories if we are in check + if (ss->inCheck && i > 2) + break; if (is_ok((ss-i)->currentMove)) (*(ss-i)->continuationHistory)[pc][to] << bonus; + } } - // update_capture_stats() updates move sorting heuristics when a new capture best move is found + // update_quiet_stats() updates move sorting heuristics - void update_capture_stats(const Position& pos, Move move, - Move* captures, int captureCount, int bonus) { - - CapturePieceToHistory& captureHistory = pos.this_thread()->captureHistory; - Piece moved_piece = pos.moved_piece(move); - PieceType captured = type_of(pos.piece_on(to_sq(move))); - - if (pos.capture_or_promotion(move)) - captureHistory[moved_piece][to_sq(move)][captured] << bonus; - - // Decrease all the other played capture moves - for (int i = 0; i < captureCount; ++i) - { - moved_piece = pos.moved_piece(captures[i]); - captured = type_of(pos.piece_on(to_sq(captures[i]))); - captureHistory[moved_piece][to_sq(captures[i])][captured] << -bonus; - } - } - - - // update_quiet_stats() updates move sorting heuristics when a new quiet best move is found - - void update_quiet_stats(const Position& pos, Stack* ss, Move move, - Move* quiets, int quietCount, int bonus) { + void update_quiet_stats(const Position& pos, Stack* ss, Move move, int bonus, int depth) { + // Update killers if (ss->killers[0] != move) { ss->killers[1] = ss->killers[0]; @@ -1659,21 +1853,20 @@ moves_loop: // When in check, search starts from here thisThread->mainHistory[us][from_to(move)] << bonus; update_continuation_histories(ss, pos.moved_piece(move), to_sq(move), bonus); + // Penalty for reversed move in case of moved piece not being a pawn if (type_of(pos.moved_piece(move)) != PAWN) thisThread->mainHistory[us][from_to(reverse_move(move))] << -bonus; + // Update countermove history if (is_ok((ss-1)->currentMove)) { Square prevSq = to_sq((ss-1)->currentMove); thisThread->counterMoves[pos.piece_on(prevSq)][prevSq] = move; } - // Decrease all the other played quiet moves - for (int i = 0; i < quietCount; ++i) - { - thisThread->mainHistory[us][from_to(quiets[i])] << -bonus; - update_continuation_histories(ss, pos.moved_piece(quiets[i]), to_sq(quiets[i]), -bonus); - } + // Update low ply history + if (depth > 11 && ss->ply < MAX_LPH) + thisThread->lowPlyHistory[ss->ply][from_to(move)] << stat_bonus(depth - 7); } // When playing with strength handicap, choose best move among a set of RootMoves @@ -1711,6 +1904,7 @@ moves_loop: // When in check, search starts from here } // namespace + /// MainThread::check_time() is used to print debug info and, more importantly, /// to detect when we are out of available time and thus stop the search. @@ -1762,15 +1956,18 @@ string UCI::pv(const Position& pos, Depth depth, Value alpha, Value beta) { for (size_t i = 0; i < multiPV; ++i) { - bool updated = (i <= pvIdx && rootMoves[i].score != -VALUE_INFINITE); + bool updated = rootMoves[i].score != -VALUE_INFINITE; - if (depth == 1 && !updated) + if (depth == 1 && !updated && i > 0) continue; - Depth d = updated ? depth : depth - 1; + Depth d = updated ? depth : std::max(1, depth - 1); Value v = updated ? rootMoves[i].score : rootMoves[i].previousScore; - bool tb = TB::RootInTB && abs(v) < VALUE_MATE - MAX_PLY; + if (v == -VALUE_INFINITE) + v = VALUE_ZERO; + + bool tb = TB::RootInTB && abs(v) < VALUE_MATE_IN_MAX_PLY; v = tb ? rootMoves[i].tbScore : v; if (ss.rdbuf()->in_avail()) // Not at first line @@ -1782,6 +1979,9 @@ string UCI::pv(const Position& pos, Depth depth, Value alpha, Value beta) { << " multipv " << i + 1 << " score " << UCI::value(v); + if (Options["UCI_ShowWDL"]) + ss << UCI::wdl(v, pos.game_ply()); + if (!tb && i == pvIdx) ss << (v >= beta ? " lowerbound" : v <= alpha ? " upperbound" : ""); @@ -1811,6 +2011,8 @@ string UCI::pv(const Position& pos, Depth depth, Value alpha, Value beta) { bool RootMove::extract_ponder_from_tt(Position& pos) { StateInfo st; + ASSERT_ALIGNED(&st, Eval::NNUE::kCacheLineSize); + bool ttHit; assert(pv.size() == 1); @@ -1864,7 +2066,7 @@ void Tablebases::rank_root_moves(Position& pos, Search::RootMoves& rootMoves) { if (RootInTB) { // Sort moves according to TB rank - std::sort(rootMoves.begin(), rootMoves.end(), + std::stable_sort(rootMoves.begin(), rootMoves.end(), [](const RootMove &a, const RootMove &b) { return a.tbRank > b.tbRank; } ); // Probe during search only if DTZ is not available and we are winning @@ -1878,3 +2080,5 @@ void Tablebases::rank_root_moves(Position& pos, Search::RootMoves& rootMoves) { m.tbRank = 0; } } + +} // namespace Stockfish diff --git a/src/search.h b/src/search.h index 45ec908a..a309070e 100644 --- a/src/search.h +++ b/src/search.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -28,6 +26,8 @@ #include "types.h" #include "cluster.h" +namespace Stockfish { + class Position; namespace Search { @@ -50,6 +50,10 @@ struct Stack { Value staticEval; int statScore; int moveCount; + int distanceFromPv; + bool inCheck; + bool ttPv; + bool ttHit; }; @@ -71,7 +75,6 @@ struct RootMove { Value previousScore = -VALUE_INFINITE; int selDepth = 0; int tbRank = 0; - int bestMoveCount = 0; Value tbScore; std::vector pv; }; @@ -91,7 +94,7 @@ struct LimitsType { } bool use_time_management() const { - return Cluster::is_root() && !(mate | movetime | depth | nodes | perft | infinite); + return Cluster::is_root() && (time[WHITE] || time[BLACK]); } std::vector searchmoves; @@ -107,4 +110,6 @@ void clear(); } // namespace Search +} // namespace Stockfish + #endif // #ifndef SEARCH_H_INCLUDED diff --git a/src/syzygy/tbprobe.cpp b/src/syzygy/tbprobe.cpp index b9315795..5d8f4106 100644 --- a/src/syzygy/tbprobe.cpp +++ b/src/syzygy/tbprobe.cpp @@ -1,7 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (c) 2013 Ronald de Man - Copyright (C) 2016-2019 Marco Costalba, Lucas Braesch + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -52,22 +51,23 @@ #include #endif -using namespace Tablebases; +using namespace Stockfish::Tablebases; -int Tablebases::MaxCardinality; +int Stockfish::Tablebases::MaxCardinality; + +namespace Stockfish { namespace { constexpr int TBPIECES = 7; // Max number of supported pieces enum { BigEndian, LittleEndian }; -enum TBType { KEY, WDL, DTZ }; // Used as template parameter +enum TBType { WDL, DTZ }; // Used as template parameter // Each table has a set of flags: all of them refer to DTZ tables, the last one to WDL tables enum TBFlag { STM = 1, Mapped = 2, WinPlies = 4, LossPlies = 8, Wide = 16, SingleValue = 128 }; inline WDLScore operator-(WDLScore d) { return WDLScore(-int(d)); } -inline Square operator^=(Square& s, int i) { return s = Square(int(s) ^ i); } inline Square operator^(Square s, int i) { return Square(int(s) ^ i); } const std::string PieceToChar = " PNBRQK pnbrqk"; @@ -226,7 +226,9 @@ public: *mapping = statbuf.st_size; *baseAddress = mmap(nullptr, statbuf.st_size, PROT_READ, MAP_SHARED, fd, 0); +#if defined(MADV_RANDOM) madvise(*baseAddress, statbuf.st_size, MADV_RANDOM); +#endif ::close(fd); if (*baseAddress == MAP_FAILED) @@ -405,7 +407,17 @@ TBTable::TBTable(const TBTable& wdl) : TBTable() { // at init time, accessed at probe time. class TBTables { - typedef std::tuple*, TBTable*> Entry; + struct Entry + { + Key key; + TBTable* wdl; + TBTable* dtz; + + template + TBTable* get() const { + return (TBTable*)(Type == WDL ? (void*)wdl : (void*)dtz); + } + }; static constexpr int Size = 1 << 12; // 4K table, indexed by key's 12 lsb static constexpr int Overflow = 1; // Number of elements allowed to map to the last bucket @@ -417,12 +429,12 @@ class TBTables { void insert(Key key, TBTable* wdl, TBTable* dtz) { uint32_t homeBucket = (uint32_t)key & (Size - 1); - Entry entry = std::make_tuple(key, wdl, dtz); + Entry entry{ key, wdl, dtz }; // Ensure last element is empty to avoid overflow when looking up for (uint32_t bucket = homeBucket; bucket < Size + Overflow - 1; ++bucket) { - Key otherKey = std::get(hashTable[bucket]); - if (otherKey == key || !std::get(hashTable[bucket])) { + Key otherKey = hashTable[bucket].key; + if (otherKey == key || !hashTable[bucket].get()) { hashTable[bucket] = entry; return; } @@ -431,7 +443,7 @@ class TBTables { // insert here and search for a new spot for the other element instead. uint32_t otherHomeBucket = (uint32_t)otherKey & (Size - 1); if (otherHomeBucket > homeBucket) { - swap(entry, hashTable[bucket]); + std::swap(entry, hashTable[bucket]); key = otherKey; homeBucket = otherHomeBucket; } @@ -444,8 +456,8 @@ public: template TBTable* get(Key key) { for (const Entry* entry = &hashTable[(uint32_t)key & (Size - 1)]; ; ++entry) { - if (std::get(*entry) == key || !std::get(*entry)) - return std::get(*entry); + if (entry->key == key || !entry->get()) + return entry->get(); } } @@ -522,7 +534,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // I(k) = k * d->span + d->span / 2 (1) // First step is to get the 'k' of the I(k) nearest to our idx, using definition (1) - uint32_t k = idx / d->span; + uint32_t k = uint32_t(idx / d->span); // Then we read the corresponding SparseIndex[] entry uint32_t block = number(&d->sparseIndex[k].block); @@ -568,7 +580,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // All the symbols of a given length are consecutive integers (numerical // sequence property), so we can compute the offset of our symbol of // length len, stored at the beginning of buf64. - sym = (buf64 - d->base64[len]) >> (64 - len - d->minSymLen); + sym = Sym((buf64 - d->base64[len]) >> (64 - len - d->minSymLen)); // Now add the value of the lowest symbol of length len to get our symbol sym += number(&d->lowestSym[len]); @@ -684,7 +696,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu bool blackStronger = (pos.material_key() != entry->key); int flipColor = (symmetricBlackToMove || blackStronger) * 8; - int flipSquares = (symmetricBlackToMove || blackStronger) * 070; + int flipSquares = (symmetricBlackToMove || blackStronger) * 56; int stm = (symmetricBlackToMove || blackStronger) ^ pos.side_to_move(); // For pawns, TB files store 4 separate tables according if leading pawn is on @@ -707,9 +719,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu std::swap(squares[0], *std::max_element(squares, squares + leadPawnsCnt, pawns_comp)); - tbFile = file_of(squares[0]); - if (tbFile > FILE_D) - tbFile = file_of(squares[0] ^ 7); // Horizontal flip: SQ_H1 -> SQ_A1 + tbFile = File(edge_distance(file_of(squares[0]))); } // DTZ tables are one-sided, i.e. they store positions only for white to @@ -733,8 +743,8 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // Then we reorder the pieces to have the same sequence as the one stored // in pieces[i]: the sequence that ensures the best compression. - for (int i = leadPawnsCnt; i < size; ++i) - for (int j = i; j < size; ++j) + for (int i = leadPawnsCnt; i < size - 1; ++i) + for (int j = i + 1; j < size; ++j) if (d->pieces[i] == pieces[j]) { std::swap(pieces[i], pieces[j]); @@ -746,14 +756,14 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // the triangle A1-D1-D4. if (file_of(squares[0]) > FILE_D) for (int i = 0; i < size; ++i) - squares[i] ^= 7; // Horizontal flip: SQ_H1 -> SQ_A1 + squares[i] = flip_file(squares[i]); // Encode leading pawns starting with the one with minimum MapPawns[] and // proceeding in ascending order. if (entry->hasPawns) { idx = LeadPawnIdx[leadPawnsCnt][squares[0]]; - std::sort(squares + 1, squares + leadPawnsCnt, pawns_comp); + std::stable_sort(squares + 1, squares + leadPawnsCnt, pawns_comp); for (int i = 1; i < leadPawnsCnt; ++i) idx += Binomial[i][MapPawns[squares[i]]]; @@ -765,7 +775,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // piece is below RANK_5. if (rank_of(squares[0]) > RANK_4) for (int i = 0; i < size; ++i) - squares[i] ^= 070; // Vertical flip: SQ_A8 -> SQ_A1 + squares[i] = flip_rank(squares[i]); // Look for the first piece of the leading group not on the A1-D4 diagonal // and ensure it is mapped below the diagonal. @@ -773,7 +783,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu if (!off_A1H8(squares[i])) continue; - if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C3 + if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C1 for (int j = i; j < size; ++j) squares[j] = Square(((squares[j] >> 3) | (squares[j] << 3)) & 63); break; @@ -854,7 +864,7 @@ encode_remaining: while (d->groupLen[++next]) { - std::sort(groupSq, groupSq + d->groupLen[next]); + std::stable_sort(groupSq, groupSq + d->groupLen[next]); uint64_t n = 0; // Map down a square if "comes later" than a square in the previous @@ -978,7 +988,7 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { d->sizeofBlock = 1ULL << *data++; d->span = 1ULL << *data++; - d->sparseIndexSize = (tbSize + d->span - 1) / d->span; // Round up + d->sparseIndexSize = size_t((tbSize + d->span - 1) / d->span); // Round up auto padding = number(data++); d->blocksNum = number(data); data += sizeof(uint32_t); d->blockLengthSize = d->blocksNum + padding; // Padded to ensure SparseIndex[] @@ -993,7 +1003,7 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { // so that d->lowestSym[i] >= d->lowestSym[i+1] (when read as LittleEndian). // Starting from this we compute a base64[] table indexed by symbol length // and containing 64 bit values so that d->base64[i] >= d->base64[i+1]. - // See http://www.eecs.harvard.edu/~michaelm/E210/huffman.pdf + // See https://en.wikipedia.org/wiki/Huffman_coding for (int i = d->base64.size() - 2; i >= 0; --i) { d->base64[i] = (d->base64[i + 1] + number(&d->lowestSym[i]) - number(&d->lowestSym[i + 1])) / 2; @@ -1063,8 +1073,8 @@ void set(T& e, uint8_t* data) { enum { Split = 1, HasPawns = 2 }; - assert(e.hasPawns == !!(*data & HasPawns)); - assert((e.key != e.key2) == !!(*data & Split)); + assert(e.hasPawns == bool(*data & HasPawns)); + assert((e.key != e.key2) == bool(*data & Split)); data++; // First byte stores flags @@ -1134,7 +1144,7 @@ void* mapped(TBTable& e, const Position& pos) { if (e.ready.load(std::memory_order_acquire)) return e.baseAddress; // Could be nullptr if file does not exist - std::unique_lock lk(mutex); + std::scoped_lock lk(mutex); if (e.ready.load(std::memory_order_relaxed)) // Recheck under lock return e.baseAddress; @@ -1194,7 +1204,7 @@ WDLScore search(Position& pos, ProbeState* result) { auto moveList = MoveList(pos); size_t totalCount = moveList.size(), moveCount = 0; - for (const Move& move : moveList) + for (const Move move : moveList) { if ( !pos.capture(move) && (!CheckZeroingMoves || type_of(pos.moved_piece(move)) != PAWN)) @@ -1347,7 +1357,7 @@ void Tablebases::init(const std::string& paths) { if (leadPawnsCnt == 1) { MapPawns[sq] = availableSquares--; - MapPawns[sq ^ 7] = availableSquares--; // Horizontal flip + MapPawns[flip_file(sq)] = availableSquares--; } LeadPawnIdx[leadPawnsCnt][sq] = idx; idx += Binomial[leadPawnsCnt - 1][MapPawns[sq]]; @@ -1356,7 +1366,7 @@ void Tablebases::init(const std::string& paths) { LeadPawnsSize[leadPawnsCnt][f] = idx; } - // Add entries in TB tables if the corresponding ".rtbw" file exsists + // Add entries in TB tables if the corresponding ".rtbw" file exists for (PieceType p1 = PAWN; p1 < KING; ++p1) { TBTables.add({KING, p1, KING}); @@ -1434,7 +1444,7 @@ WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) { // If n = 100 immediately after a capture or pawn move, then the position // is also certainly a win, and during the whole phase until the next // capture or pawn move, the inequality to be preserved is -// dtz + 50-movecounter <= 100. +// dtz + 50-move-counter <= 100. // // In short, if a move is available resulting in dtz + 50-move-counter <= 99, // then do not accept moves leading to dtz + 50-move-counter == 100. @@ -1464,7 +1474,7 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { StateInfo st; int minDTZ = 0xFFFF; - for (const Move& move : MoveList(pos)) + for (const Move move : MoveList(pos)) { bool zeroing = pos.capture(move) || type_of(pos.moved_piece(move)) == PAWN; @@ -1604,3 +1614,5 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) { return true; } + +} // namespace Stockfish diff --git a/src/syzygy/tbprobe.h b/src/syzygy/tbprobe.h index 264f6e84..56734af9 100644 --- a/src/syzygy/tbprobe.h +++ b/src/syzygy/tbprobe.h @@ -1,7 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (c) 2013 Ronald de Man - Copyright (C) 2016-2019 Marco Costalba, Lucas Braesch + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -24,7 +23,7 @@ #include "../search.h" -namespace Tablebases { +namespace Stockfish::Tablebases { enum WDLScore { WDLLoss = -2, // Loss @@ -74,6 +73,6 @@ inline std::ostream& operator<<(std::ostream& os, const ProbeState v) { return os; } -} +} // namespace Stockfish::Tablebases #endif diff --git a/src/thread.cpp b/src/thread.cpp index 86d2009a..1316453f 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -28,6 +26,8 @@ #include "syzygy/tbprobe.h" #include "tt.h" +namespace Stockfish { + ThreadPool Threads; // Global object @@ -52,15 +52,6 @@ Thread::~Thread() { stdThread.join(); } -/// Thread::bestMoveCount(Move move) return best move counter for the given root move - -int Thread::best_move_count(Move move) { - - auto rm = std::find(rootMoves.begin() + pvIdx, - rootMoves.begin() + pvLast, move); - - return rm != rootMoves.begin() + pvLast ? rm->bestMoveCount : 0; -} /// Thread::clear() reset histories, usually before a new game @@ -68,19 +59,20 @@ void Thread::clear() { counterMoves.fill(MOVE_NONE); mainHistory.fill(0); + lowPlyHistory.fill(0); captureHistory.fill(0); for (bool inCheck : { false, true }) - for (StatsType c : { NoCaptures, Captures }) - for (auto& to : continuationHistory[inCheck][c]) - for (auto& h : to) - h->fill(0); - - for (bool inCheck : { false, true }) - for (StatsType c : { NoCaptures, Captures }) - continuationHistory[inCheck][c][NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1); + for (StatsType c : { NoCaptures, Captures }) + { + for (auto& to : continuationHistory[inCheck][c]) + for (auto& h : to) + h->fill(0); + continuationHistory[inCheck][c][NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1); + } } + /// Thread::start_searching() wakes up the thread that will start the search void Thread::start_searching() { @@ -151,7 +143,7 @@ void ThreadPool::set(size_t requested) { clear(); // Reallocate the hash with the new threadpool size - TT.resize(Options["Hash"]); + TT.resize(size_t(Options["Hash"])); // Adjust cluster buffers Cluster::ttSendRecvBuff_resize(requested); @@ -161,7 +153,8 @@ void ThreadPool::set(size_t requested) { } } -/// ThreadPool::clear() sets threadPool data to initial values. + +/// ThreadPool::clear() sets threadPool data to initial values void ThreadPool::clear() { @@ -169,10 +162,11 @@ void ThreadPool::clear() { th->clear(); main()->callsCnt = 0; - main()->previousScore = VALUE_INFINITE; + main()->bestPreviousScore = VALUE_INFINITE; main()->previousTimeReduction = 1.0; } + /// 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. @@ -182,6 +176,7 @@ void ThreadPool::start_thinking(Position& pos, StateListPtr& states, main()->wait_for_search_finished(); main()->stopOnPonderhit = stop = false; + increaseDepth = true; main()->ponder = ponderMode; Search::Limits = limits; Search::RootMoves rootMoves; @@ -203,22 +198,72 @@ void ThreadPool::start_thinking(Position& pos, StateListPtr& states, // We use Position::set() to set root position across threads. But there are // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot - // be deduced from a fen string, so set() clears them and to not lose the info - // we need to backup and later restore setupStates->back(). Note that setupStates - // is shared by threads but is accessed in read-only mode. - StateInfo tmp = setupStates->back(); - + // be deduced from a fen string, so set() clears them and they are set from + // setupStates->back() later. The rootState is per thread, earlier states are shared + // since they are read-only. for (Thread* th : *this) { - th->shuffleExts = th->nodes = th->tbHits = th->TTsaves = th->nmpMinPly = 0; + th->nodes = th->tbHits = th->TTsaves = th->nmpMinPly = th->bestMoveChanges = 0; th->rootDepth = th->completedDepth = 0; th->rootMoves = rootMoves; - th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th); + th->rootPos.set(pos.fen(), pos.is_chess960(), &th->rootState, th); + th->rootState = setupStates->back(); } - setupStates->back() = tmp; - Cluster::signals_init(); main()->start_searching(); } + +Thread* ThreadPool::get_best_thread() const { + + Thread* bestThread = front(); + std::map votes; + Value minScore = VALUE_NONE; + + // Find minimum score of all threads + for (Thread* th: *this) + minScore = std::min(minScore, th->rootMoves[0].score); + + // Vote according to score and depth, and select the best thread + for (Thread* th : *this) + { + votes[th->rootMoves[0].pv[0]] += + (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth); + + if (abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY) + { + // Make sure we pick the shortest mate / TB conversion or stave off mate the longest + if (th->rootMoves[0].score > bestThread->rootMoves[0].score) + bestThread = th; + } + else if ( th->rootMoves[0].score >= VALUE_TB_WIN_IN_MAX_PLY + || ( th->rootMoves[0].score > VALUE_TB_LOSS_IN_MAX_PLY + && votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]])) + bestThread = th; + } + + return bestThread; +} + + +/// Start non-main threads + +void ThreadPool::start_searching() { + + for (Thread* th : *this) + if (th != front()) + th->start_searching(); +} + + +/// Wait for non-main threads + +void ThreadPool::wait_for_search_finished() const { + + for (Thread* th : *this) + if (th != front()) + th->wait_for_search_finished(); +} + +} // namespace Stockfish diff --git a/src/thread.h b/src/thread.h index efb0f8de..9ae74448 100644 --- a/src/thread.h +++ b/src/thread.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -35,6 +33,7 @@ #include "search.h" #include "thread_win32_osx.h" +namespace Stockfish { /// Thread class keeps together all the thread-related stuff. We use /// per-thread pawn and material hash tables so that once we get a @@ -57,24 +56,26 @@ public: void idle_loop(); void start_searching(); void wait_for_search_finished(); - int best_move_count(Move move); Pawns::Table pawnsTable; Material::Table materialTable; - size_t pvIdx, pvLast, shuffleExts; + size_t pvIdx, pvLast; + uint64_t ttHitAverage; int selDepth, nmpMinPly; Color nmpColor; std::atomic nodes, tbHits, TTsaves, bestMoveChanges; Position rootPos; + StateInfo rootState; Search::RootMoves rootMoves; Depth rootDepth, completedDepth; CounterMoveHistory counterMoves; ButterflyHistory mainHistory; + LowPlyHistory lowPlyHistory; CapturePieceToHistory captureHistory; ContinuationHistory continuationHistory[2][2]; Score contempt; - + int failedHighCnt; #ifdef USE_MPI struct { std::mutex mutex; @@ -94,7 +95,8 @@ struct MainThread : public Thread { void check_time(); double previousTimeReduction; - Value previousScore; + Value bestPreviousScore; + Value iterValue[4]; int callsCnt; bool stopOnPonderhit; std::atomic_bool ponder; @@ -115,8 +117,11 @@ struct ThreadPool : public std::vector { uint64_t nodes_searched() const { return accumulate(&Thread::nodes); } uint64_t tb_hits() const { return accumulate(&Thread::tbHits); } uint64_t TT_saves() const { return accumulate(&Thread::TTsaves); } + Thread* get_best_thread() const; + void start_searching(); + void wait_for_search_finished() const; - std::atomic_bool stop; + std::atomic_bool stop, increaseDepth; private: StateListPtr setupStates; @@ -132,4 +137,6 @@ private: extern ThreadPool Threads; +} // namespace Stockfish + #endif // #ifndef THREAD_H_INCLUDED diff --git a/src/thread_win32_osx.h b/src/thread_win32_osx.h index f8cb466b..a21674cc 100644 --- a/src/thread_win32_osx.h +++ b/src/thread_win32_osx.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -29,10 +27,12 @@ /// The implementation calls pthread_create() with the stack size parameter /// equal to the linux 8MB default, on platforms that support it. -#if defined(__APPLE__) || defined(__MINGW32__) || defined(__MINGW64__) +#if defined(__APPLE__) || defined(__MINGW32__) || defined(__MINGW64__) || defined(USE_PTHREADS) #include +namespace Stockfish { + static const size_t TH_STACK_SIZE = 8 * 1024 * 1024; template > @@ -59,10 +59,16 @@ public: void join() { pthread_join(thread, NULL); } }; +} // namespace Stockfish + #else // Default case: use STL classes +namespace Stockfish { + typedef std::thread NativeThread; +} // namespace Stockfish + #endif #endif // #ifndef THREAD_WIN32_OSX_H_INCLUDED diff --git a/src/timeman.cpp b/src/timeman.cpp index 484aaa65..f742d1e4 100644 --- a/src/timeman.cpp +++ b/src/timeman.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,68 +24,25 @@ #include "timeman.h" #include "uci.h" +namespace Stockfish { + TimeManagement Time; // Our global time management object -namespace { - enum TimeType { OptimumTime, MaxTime }; - - constexpr int MoveHorizon = 50; // Plan time management at most this many moves ahead - constexpr double MaxRatio = 7.3; // When in trouble, we can step over reserved time with this ratio - constexpr double StealRatio = 0.34; // However we must not steal time from remaining moves over this ratio - - - // move_importance() is a skew-logistic function based on naive statistical - // analysis of "how many games are still undecided after n half-moves". Game - // is considered "undecided" as long as neither side has >275cp advantage. - // Data was extracted from the CCRL game database with some simple filtering criteria. - - double move_importance(int ply) { - - constexpr double XScale = 6.85; - constexpr double XShift = 64.5; - constexpr double Skew = 0.171; - - return pow((1 + exp((ply - XShift) / XScale)), -Skew) + DBL_MIN; // Ensure non-zero - } - - template - TimePoint remaining(TimePoint myTime, int movesToGo, int ply, TimePoint slowMover) { - - constexpr double TMaxRatio = (T == OptimumTime ? 1.0 : MaxRatio); - constexpr double TStealRatio = (T == OptimumTime ? 0.0 : StealRatio); - - double moveImportance = (move_importance(ply) * slowMover) / 100.0; - double otherMovesImportance = 0.0; - - for (int i = 1; i < movesToGo; ++i) - otherMovesImportance += move_importance(ply + 2 * i); - - double ratio1 = (TMaxRatio * moveImportance) / (TMaxRatio * moveImportance + otherMovesImportance); - double ratio2 = (moveImportance + TStealRatio * otherMovesImportance) / (moveImportance + otherMovesImportance); - - return TimePoint(myTime * std::min(ratio1, ratio2)); // Intel C++ asks for an explicit cast - } - -} // namespace - - -/// init() is called at the beginning of the search and calculates the allowed -/// thinking time out of the time control and current game ply. We support four -/// different kinds of time controls, passed in 'limits': -/// -/// inc == 0 && movestogo == 0 means: x basetime [sudden death!] -/// inc == 0 && movestogo != 0 means: x moves in y minutes -/// inc > 0 && movestogo == 0 means: x basetime + z increment -/// inc > 0 && movestogo != 0 means: x moves in y minutes + z increment +/// TimeManagement::init() is called at the beginning of the search and calculates +/// the bounds of time allowed for the current game ply. We currently support: +// 1) x basetime (+ z increment) +// 2) x moves in y seconds (+ z increment) void TimeManagement::init(Search::LimitsType& limits, Color us, int ply) { - TimePoint minThinkingTime = Options["Minimum Thinking Time"]; - TimePoint moveOverhead = Options["Move Overhead"]; - TimePoint slowMover = Options["Slow Mover"]; - TimePoint npmsec = Options["nodestime"]; - TimePoint hypMyTime; + TimePoint moveOverhead = TimePoint(Options["Move Overhead"]); + TimePoint slowMover = TimePoint(Options["Slow Mover"]); + TimePoint npmsec = TimePoint(Options["nodestime"]); + + // optScale is a percentage of available time to use for the current move. + // maxScale is a multiplier applied to optimumTime. + double optScale, maxScale; // If we have to play in 'nodes as time' mode, then convert from time // to nodes, and use resulting values in time management formulas. @@ -105,29 +60,42 @@ void TimeManagement::init(Search::LimitsType& limits, Color us, int ply) { } startTime = limits.startTime; - optimumTime = maximumTime = std::max(limits.time[us], minThinkingTime); - const int maxMTG = limits.movestogo ? std::min(limits.movestogo, MoveHorizon) : MoveHorizon; + // Maximum move horizon of 50 moves + int mtg = limits.movestogo ? std::min(limits.movestogo, 50) : 50; - // We calculate optimum time usage for different hypothetical "moves to go" values - // and choose the minimum of calculated search time values. Usually the greatest - // hypMTG gives the minimum values. - for (int hypMTG = 1; hypMTG <= maxMTG; ++hypMTG) + // Make sure timeLeft is > 0 since we may use it as a divisor + TimePoint timeLeft = std::max(TimePoint(1), + limits.time[us] + limits.inc[us] * (mtg - 1) - moveOverhead * (2 + mtg)); + + // A user may scale time usage by setting UCI option "Slow Mover" + // Default is 100 and changing this value will probably lose elo. + timeLeft = slowMover * timeLeft / 100; + + // x basetime (+ z increment) + // If there is a healthy increment, timeLeft can exceed actual available + // game time for the current move, so also cap to 20% of available game time. + if (limits.movestogo == 0) { - // Calculate thinking time for hypothetical "moves to go"-value - hypMyTime = limits.time[us] - + limits.inc[us] * (hypMTG - 1) - - moveOverhead * (2 + std::min(hypMTG, 40)); - - hypMyTime = std::max(hypMyTime, TimePoint(0)); - - TimePoint t1 = minThinkingTime + remaining(hypMyTime, hypMTG, ply, slowMover); - TimePoint t2 = minThinkingTime + remaining(hypMyTime, hypMTG, ply, slowMover); - - optimumTime = std::min(t1, optimumTime); - maximumTime = std::min(t2, maximumTime); + optScale = std::min(0.0084 + std::pow(ply + 3.0, 0.5) * 0.0042, + 0.2 * limits.time[us] / double(timeLeft)); + maxScale = std::min(7.0, 4.0 + ply / 12.0); } + // x moves in y seconds (+ z increment) + else + { + optScale = std::min((0.8 + ply / 128.0) / mtg, + 0.8 * limits.time[us] / double(timeLeft)); + maxScale = std::min(6.3, 1.5 + 0.11 * mtg); + } + + // Never use more than 80% of the available time for this move + optimumTime = TimePoint(optScale * timeLeft); + maximumTime = TimePoint(std::min(0.8 * limits.time[us] - moveOverhead, maxScale * optimumTime)); + if (Options["Ponder"]) optimumTime += optimumTime / 4; } + +} // namespace Stockfish diff --git a/src/timeman.h b/src/timeman.h index 7060f065..c3cb37d0 100644 --- a/src/timeman.h +++ b/src/timeman.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,6 +23,8 @@ #include "search.h" #include "cluster.h" +namespace Stockfish { + /// The TimeManagement class computes the optimal time to think depending on /// the maximum available time, the game move number and other parameters. @@ -46,4 +46,6 @@ private: extern TimeManagement Time; +} // namespace Stockfish + #endif // #ifndef TIMEMAN_H_INCLUDED diff --git a/src/tt.cpp b/src/tt.cpp index d3cd094e..1f495ca9 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -28,29 +26,32 @@ #include "tt.h" #include "uci.h" +namespace Stockfish { + TranspositionTable TT; // Our global transposition table -/// TTEntry::save populates the TTEntry with a new node's data, possibly +/// 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) { // Preserve any existing move for the same position - if (m || (k >> 48) != key16) + if (m || (uint16_t)k != key16) move16 = (uint16_t)m; - // Overwrite less valuable entries - if ( (k >> 48) != key16 - || d - DEPTH_OFFSET > depth8 - 4 - || b == BOUND_EXACT) + // Overwrite less valuable entries (cheapest checks first) + if (b == BOUND_EXACT + || (uint16_t)k != key16 + || d - DEPTH_OFFSET > depth8 - 4) { - assert(d >= DEPTH_OFFSET); + assert(d > DEPTH_OFFSET); + assert(d < 256 + DEPTH_OFFSET); - key16 = (uint16_t)(k >> 48); + key16 = (uint16_t)k; + depth8 = (uint8_t)(d - DEPTH_OFFSET); + genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b); value16 = (int16_t)v; eval16 = (int16_t)ev; - genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b); - depth8 = (uint8_t)(d - DEPTH_OFFSET); } } @@ -63,19 +64,18 @@ void TranspositionTable::resize(size_t mbSize) { Threads.main()->wait_for_search_finished(); + aligned_large_pages_free(table); + clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster); - free(mem); - mem = malloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1); - - if (!mem) + table = static_cast(aligned_large_pages_alloc(clusterCount * sizeof(Cluster))); + if (!table) { std::cerr << "Failed to allocate " << mbSize << "MB for transposition table." << std::endl; exit(EXIT_FAILURE); } - table = (Cluster*)((uintptr_t(mem) + CacheLineSize - 1) & ~(CacheLineSize - 1)); clear(); } @@ -96,8 +96,8 @@ void TranspositionTable::clear() { WinProcGroup::bindThisThread(idx); // Each thread will zero its part of the hash table - const size_t stride = clusterCount / Options["Threads"], - start = stride * idx, + const size_t stride = size_t(clusterCount / Options["Threads"]), + start = size_t(stride * idx), len = idx != Options["Threads"] - 1 ? stride : clusterCount - start; @@ -105,10 +105,11 @@ void TranspositionTable::clear() { }); } - for (std::thread& th: threads) + for (std::thread& th : threads) th.join(); } + /// TranspositionTable::probe() looks up the current position in the transposition /// table. It returns true and a pointer to the TTEntry if the position is found. /// Otherwise, it returns false and a pointer to an empty or least valuable TTEntry @@ -119,25 +120,26 @@ void TranspositionTable::clear() { TTEntry* TranspositionTable::probe(const Key key, bool& found) const { TTEntry* const tte = first_entry(key); - const uint16_t key16 = key >> 48; // Use the high 16 bits as key inside the cluster + const uint16_t key16 = (uint16_t)key; // Use the low 16 bits as key inside the cluster for (int i = 0; i < ClusterSize; ++i) - if (!tte[i].key16 || tte[i].key16 == key16) + if (tte[i].key16 == key16 || !tte[i].depth8) { - tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & 0x7)); // Refresh + tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & (GENERATION_DELTA - 1))); // Refresh - return found = (bool)tte[i].key16, &tte[i]; + return found = (bool)tte[i].depth8, &tte[i]; } // Find an entry to be replaced according to the replacement strategy TTEntry* replace = tte; for (int i = 1; i < ClusterSize; ++i) // Due to our packed storage format for generation and its cyclic - // nature we add 263 (256 is the modulus plus 7 to keep the unrelated - // lowest three bits from affecting the result) to calculate the entry - // age correctly even after generation8 overflows into the next cycle. - if ( replace->depth8 - ((263 + generation8 - replace->genBound8) & 0xF8) - > tte[i].depth8 - ((263 + generation8 - tte[i].genBound8) & 0xF8)) + // nature we add GENERATION_CYCLE (256 is the modulus, plus what + // is needed to keep the unrelated lowest n bits from affecting + // the result) to calculate the entry age correctly even after + // generation8 overflows into the next cycle. + if ( replace->depth8 - ((GENERATION_CYCLE + generation8 - replace->genBound8) & GENERATION_MASK) + > tte[i].depth8 - ((GENERATION_CYCLE + generation8 - tte[i].genBound8) & GENERATION_MASK)) replace = &tte[i]; return found = false, replace; @@ -150,9 +152,11 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) const { int TranspositionTable::hashfull() const { int cnt = 0; - for (int i = 0; i < 1000 / ClusterSize; ++i) + for (int i = 0; i < 1000; ++i) for (int j = 0; j < ClusterSize; ++j) - cnt += (table[i].entry[j].genBound8 & 0xF8) == generation8; + cnt += table[i].entry[j].depth8 && (table[i].entry[j].genBound8 & GENERATION_MASK) == generation8; - return cnt * 1000 / (ClusterSize * (1000 / ClusterSize)); + return cnt / ClusterSize; } + +} // namespace Stockfish diff --git a/src/tt.h b/src/tt.h index 2ba10647..9db52b62 100644 --- a/src/tt.h +++ b/src/tt.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -24,21 +22,23 @@ #include "misc.h" #include "types.h" +namespace Stockfish { + +//void Cluster::init(); namespace Cluster { void init(); } -//void Cluster::init(); /// TTEntry struct is the 10 bytes transposition table entry, defined as below: /// /// key 16 bit -/// move 16 bit -/// value 16 bit -/// eval value 16 bit +/// depth 8 bit /// generation 5 bit /// pv node 1 bit /// bound type 2 bit -/// depth 8 bit +/// move 16 bit +/// value 16 bit +/// eval value 16 bit struct TTEntry { @@ -46,7 +46,7 @@ struct TTEntry { Value value() const { return (Value)value16; } Value eval() const { return (Value)eval16; } Depth depth() const { return (Depth)depth8 + DEPTH_OFFSET; } - bool is_pv() const { return (bool)(genBound8 & 0x4); } + bool is_pv() const { return (bool)(genBound8 & 0x4); } Bound bound() const { return (Bound)(genBound8 & 0x3); } void save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev); @@ -55,46 +55,49 @@ private: friend void Cluster::init(); uint16_t key16; + uint8_t depth8; + uint8_t genBound8; uint16_t move16; int16_t value16; int16_t eval16; - uint8_t genBound8; - uint8_t depth8; }; -/// A TranspositionTable consists of a power of 2 number of clusters and each -/// cluster consists of ClusterSize number of TTEntry. Each non-empty entry -/// contains information of exactly one position. The size of a cluster should -/// divide the size of a cache line size, to ensure that clusters never cross -/// cache lines. This ensures best cache performance, as the cacheline is -/// prefetched, as soon as possible. +/// A TranspositionTable is an array of Cluster, of size clusterCount. Each +/// cluster consists of ClusterSize number of TTEntry. Each non-empty TTEntry +/// contains information on exactly one position. The size of a Cluster should +/// divide the size of a cache line for best performance, as the cacheline is +/// prefetched when possible. class TranspositionTable { friend void Cluster::init(); - static constexpr int CacheLineSize = 64; static constexpr int ClusterSize = 3; struct Cluster { TTEntry entry[ClusterSize]; - char padding[2]; // Align to a divisor of the cache line size + char padding[2]; // Pad to 32 bytes }; - static_assert(CacheLineSize % sizeof(Cluster) == 0, "Cluster size incorrect"); + static_assert(sizeof(Cluster) == 32, "Unexpected Cluster size"); + + // Constants used to refresh the hash table periodically + static constexpr unsigned GENERATION_BITS = 3; // nb of bits reserved for other things + static constexpr int GENERATION_DELTA = (1 << GENERATION_BITS); // increment for generation field + static constexpr int GENERATION_CYCLE = 255 + (1 << GENERATION_BITS); // cycle length + static constexpr int GENERATION_MASK = (0xFF << GENERATION_BITS) & 0xFF; // mask to pull out generation number public: - ~TranspositionTable() { free(mem); } - void new_search() { generation8 += 8; } // Lower 3 bits are used by PV flag and Bound + ~TranspositionTable() { aligned_large_pages_free(table); } + void new_search() { generation8 += GENERATION_DELTA; } // Lower bits are used for other things TTEntry* probe(const Key key, bool& found) const; int hashfull() const; void resize(size_t mbSize); void clear(); - // The 32 lowest order bits of the key are used to get the index of the cluster TTEntry* first_entry(const Key key) const { - return &table[(uint32_t(key) * uint64_t(clusterCount)) >> 32].entry[0]; + return &table[mul_hi64(key, clusterCount)].entry[0]; } private: @@ -102,10 +105,11 @@ private: size_t clusterCount; Cluster* table; - void* mem; uint8_t generation8; // Size must be not bigger than TTEntry::genBound8 }; extern TranspositionTable TT; +} // namespace Stockfish + #endif // #ifndef TT_H_INCLUDED diff --git a/src/tune.cpp b/src/tune.cpp new file mode 100644 index 00000000..d9618efc --- /dev/null +++ b/src/tune.cpp @@ -0,0 +1,152 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +#include +#include +#include + +#include "types.h" +#include "misc.h" +#include "uci.h" + +using std::string; + +namespace Stockfish { + +bool Tune::update_on_last; +const UCI::Option* LastOption = nullptr; +BoolConditions Conditions; +static std::map TuneResults; + +string Tune::next(string& names, bool pop) { + + string name; + + do { + string token = names.substr(0, names.find(',')); + + if (pop) + names.erase(0, token.size() + 1); + + std::stringstream ws(token); + name += (ws >> token, token); // Remove trailing whitespace + + } while ( std::count(name.begin(), name.end(), '(') + - std::count(name.begin(), name.end(), ')')); + + return name; +} + +static void on_tune(const UCI::Option& o) { + + if (!Tune::update_on_last || LastOption == &o) + Tune::read_options(); +} + +static void make_option(const string& n, int v, const SetRange& r) { + + // Do not generate option when there is nothing to tune (ie. min = max) + if (r(v).first == r(v).second) + return; + + if (TuneResults.count(n)) + v = TuneResults[n]; + + Options[n] << UCI::Option(v, r(v).first, r(v).second, on_tune); + LastOption = &Options[n]; + + // Print formatted parameters, ready to be copy-pasted in Fishtest + std::cout << n << "," + << v << "," + << r(v).first << "," << r(v).second << "," + << (r(v).second - r(v).first) / 20.0 << "," + << "0.0020" + << std::endl; +} + +template<> void Tune::Entry::init_option() { make_option(name, value, range); } + +template<> void Tune::Entry::read_option() { + if (Options.count(name)) + value = int(Options[name]); +} + +template<> void Tune::Entry::init_option() { make_option(name, value, range); } + +template<> void Tune::Entry::read_option() { + if (Options.count(name)) + value = Value(int(Options[name])); +} + +template<> void Tune::Entry::init_option() { + make_option("m" + name, mg_value(value), range); + make_option("e" + name, eg_value(value), range); +} + +template<> void Tune::Entry::read_option() { + if (Options.count("m" + name)) + value = make_score(int(Options["m" + name]), eg_value(value)); + + if (Options.count("e" + name)) + value = make_score(mg_value(value), int(Options["e" + name])); +} + +// Instead of a variable here we have a PostUpdate function: just call it +template<> void Tune::Entry::init_option() {} +template<> void Tune::Entry::read_option() { value(); } + + +// Set binary conditions according to a probability that depends +// on the corresponding parameter value. + +void BoolConditions::set() { + + static PRNG rng(now()); + static bool startup = true; // To workaround fishtest bench + + for (size_t i = 0; i < binary.size(); i++) + binary[i] = !startup && (values[i] + int(rng.rand() % variance) > threshold); + + startup = false; + + for (size_t i = 0; i < binary.size(); i++) + sync_cout << binary[i] << sync_endl; +} + +} // namespace Stockfish + + +// Init options with tuning session results instead of default values. Useful to +// get correct bench signature after a tuning session or to test tuned values. +// Just copy fishtest tuning results in a result.txt file and extract the +// values with: +// +// cat results.txt | sed 's/^param: \([^,]*\), best: \([^,]*\).*/ TuneResults["\1"] = int(round(\2));/' +// +// Then paste the output below, as the function body + +#include + +namespace Stockfish { + +void Tune::read_results() { + + /* ...insert your values here... */ +} + +} // namespace Stockfish diff --git a/src/tune.h b/src/tune.h new file mode 100644 index 00000000..c904c09d --- /dev/null +++ b/src/tune.h @@ -0,0 +1,197 @@ +/* + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) + + Stockfish is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + Stockfish is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +*/ + +#ifndef TUNE_H_INCLUDED +#define TUNE_H_INCLUDED + +#include +#include +#include +#include + +namespace Stockfish { + +typedef std::pair Range; // Option's min-max values +typedef Range (RangeFun) (int); + +// Default Range function, to calculate Option's min-max values +inline Range default_range(int v) { + return v > 0 ? Range(0, 2 * v) : Range(2 * v, 0); +} + +struct SetRange { + explicit SetRange(RangeFun f) : fun(f) {} + SetRange(int min, int max) : fun(nullptr), range(min, max) {} + Range operator()(int v) const { return fun ? fun(v) : range; } + + RangeFun* fun; + Range range; +}; + +#define SetDefaultRange SetRange(default_range) + + +/// BoolConditions struct is used to tune boolean conditions in the +/// code by toggling them on/off according to a probability that +/// depends on the value of a tuned integer parameter: for high +/// values of the parameter condition is always disabled, for low +/// values is always enabled, otherwise it is enabled with a given +/// probability that depnends on the parameter under tuning. + +struct BoolConditions { + void init(size_t size) { values.resize(size, defaultValue), binary.resize(size, 0); } + void set(); + + std::vector binary, values; + int defaultValue = 465, variance = 40, threshold = 500; + SetRange range = SetRange(0, 1000); +}; + +extern BoolConditions Conditions; + +inline void set_conditions() { Conditions.set(); } + + +/// Tune class implements the 'magic' code that makes the setup of a fishtest +/// tuning session as easy as it can be. Mainly you have just to remove const +/// qualifiers from the variables you want to tune and flag them for tuning, so +/// if you have: +/// +/// const Score myScore = S(10, 15); +/// const Value myValue[][2] = { { V(100), V(20) }, { V(7), V(78) } }; +/// +/// If you have a my_post_update() function to run after values have been updated, +/// and a my_range() function to set custom Option's min-max values, then you just +/// remove the 'const' qualifiers and write somewhere below in the file: +/// +/// TUNE(SetRange(my_range), myScore, myValue, my_post_update); +/// +/// You can also set the range directly, and restore the default at the end +/// +/// TUNE(SetRange(-100, 100), myScore, SetDefaultRange); +/// +/// In case update function is slow and you have many parameters, you can add: +/// +/// UPDATE_ON_LAST(); +/// +/// And the values update, including post update function call, will be done only +/// once, after the engine receives the last UCI option, that is the one defined +/// and created as the last one, so the GUI should send the options in the same +/// order in which have been defined. + +class Tune { + + typedef void (PostUpdate) (); // Post-update function + + Tune() { read_results(); } + Tune(const Tune&) = delete; + void operator=(const Tune&) = delete; + void read_results(); + + static Tune& instance() { static Tune t; return t; } // Singleton + + // Use polymorphism to accomodate Entry of different types in the same vector + struct EntryBase { + virtual ~EntryBase() = default; + virtual void init_option() = 0; + virtual void read_option() = 0; + }; + + template + struct Entry : public EntryBase { + + static_assert(!std::is_const::value, "Parameter cannot be const!"); + + static_assert( std::is_same::value + || std::is_same::value + || std::is_same::value + || std::is_same::value, "Parameter type not supported!"); + + Entry(const std::string& n, T& v, const SetRange& r) : name(n), value(v), range(r) {} + void operator=(const Entry&) = delete; // Because 'value' is a reference + void init_option() override; + void read_option() override; + + std::string name; + T& value; + SetRange range; + }; + + // Our facility to fill the container, each Entry corresponds to a parameter + // to tune. We use variadic templates to deal with an unspecified number of + // entries, each one of a possible different type. + static std::string next(std::string& names, bool pop = true); + + int add(const SetRange&, std::string&&) { return 0; } + + template + int add(const SetRange& range, std::string&& names, T& value, Args&&... args) { + list.push_back(std::unique_ptr(new Entry(next(names), value, range))); + return add(range, std::move(names), args...); + } + + // Template specialization for arrays: recursively handle multi-dimensional arrays + template + int add(const SetRange& range, std::string&& names, T (&value)[N], Args&&... args) { + for (size_t i = 0; i < N; i++) + add(range, next(names, i == N - 1) + "[" + std::to_string(i) + "]", value[i]); + return add(range, std::move(names), args...); + } + + // Template specialization for SetRange + template + int add(const SetRange&, std::string&& names, SetRange& value, Args&&... args) { + return add(value, (next(names), std::move(names)), args...); + } + + // Template specialization for BoolConditions + template + int add(const SetRange& range, std::string&& names, BoolConditions& cond, Args&&... args) { + for (size_t size = cond.values.size(), i = 0; i < size; i++) + add(cond.range, next(names, i == size - 1) + "_" + std::to_string(i), cond.values[i]); + return add(range, std::move(names), args...); + } + + std::vector> list; + +public: + template + static int add(const std::string& names, Args&&... args) { + return instance().add(SetDefaultRange, names.substr(1, names.size() - 2), args...); // Remove trailing parenthesis + } + static void init() { for (auto& e : instance().list) e->init_option(); read_options(); } // Deferred, due to UCI::Options access + static void read_options() { for (auto& e : instance().list) e->read_option(); } + static bool update_on_last; +}; + +// Some macro magic :-) we define a dummy int variable that compiler initializes calling Tune::add() +#define STRINGIFY(x) #x +#define UNIQUE2(x, y) x ## y +#define UNIQUE(x, y) UNIQUE2(x, y) // Two indirection levels to expand __LINE__ +#define TUNE(...) int UNIQUE(p, __LINE__) = Tune::add(STRINGIFY((__VA_ARGS__)), __VA_ARGS__) + +#define UPDATE_ON_LAST() bool UNIQUE(p, __LINE__) = Tune::update_on_last = true + +// Some macro to tune toggling of boolean conditions +#define CONDITION(x) (Conditions.binary[__COUNTER__] || (x)) +#define TUNE_CONDITIONS() int UNIQUE(c, __LINE__) = (Conditions.init(__COUNTER__), 0); \ + TUNE(Conditions, set_conditions) + +} // namespace Stockfish + +#endif // #ifndef TUNE_H_INCLUDED diff --git a/src/types.h b/src/types.h index cc4008b3..efebce1a 100644 --- a/src/types.h +++ b/src/types.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -40,7 +38,6 @@ #include #include -#include #include #include #include @@ -60,6 +57,12 @@ /// _WIN32 Building on Windows (any) /// _WIN64 Building on Windows 64 bit +#if defined(__GNUC__ ) && (__GNUC__ < 9 || (__GNUC__ == 9 && __GNUC_MINOR__ <= 2)) && defined(_WIN32) && !defined(__clang__) +#define ALIGNAS_ON_STACK_VARIABLES_BROKEN +#endif + +#define ASSERT_ALIGNED(ptr, alignment) assert(reinterpret_cast(ptr) % alignment == 0) + #if defined(_WIN64) && defined(_MSC_VER) // No Makefile used # include // Microsoft header for _BitScanForward64() # define IS_64BIT @@ -80,6 +83,8 @@ # define pext(b, m) 0 #endif +namespace Stockfish { + #ifdef USE_POPCNT constexpr bool HasPopCnt = true; #else @@ -110,7 +115,7 @@ constexpr int MAX_PLY = 246; /// bit 6-11: origin square (from 0 to 63) /// bit 12-13: promotion piece type - 2 (from KNIGHT-2 to QUEEN-2) /// bit 14-15: special move flag: promotion (1), en passant (2), castling (3) -/// NOTE: EN-PASSANT bit is set only when a pawn can be captured +/// NOTE: en passant bit is set only when a pawn can be captured /// /// Special cases are MOVE_NONE and MOVE_NULL. We can sneak these in because in /// any normal move destination square is always different from origin square @@ -124,7 +129,7 @@ enum Move : int { enum MoveType { NORMAL, PROMOTION = 1 << 14, - ENPASSANT = 2 << 14, + EN_PASSANT = 2 << 14, CASTLING = 3 << 14 }; @@ -176,14 +181,17 @@ enum Value : int { VALUE_INFINITE = 32001, VALUE_NONE = 32002, - VALUE_MATE_IN_MAX_PLY = VALUE_MATE - 2 * MAX_PLY, - VALUE_MATED_IN_MAX_PLY = -VALUE_MATE + 2 * MAX_PLY, + VALUE_TB_WIN_IN_MAX_PLY = VALUE_MATE - 2 * MAX_PLY, + VALUE_TB_LOSS_IN_MAX_PLY = -VALUE_TB_WIN_IN_MAX_PLY, + VALUE_MATE_IN_MAX_PLY = VALUE_MATE - MAX_PLY, + VALUE_MATED_IN_MAX_PLY = -VALUE_MATE_IN_MAX_PLY, - PawnValueMg = 128, PawnValueEg = 213, - KnightValueMg = 782, KnightValueEg = 865, - BishopValueMg = 830, BishopValueEg = 918, - RookValueMg = 1289, RookValueEg = 1378, - QueenValueMg = 2529, QueenValueEg = 2687, + PawnValueMg = 126, PawnValueEg = 208, + KnightValueMg = 781, KnightValueEg = 854, + BishopValueMg = 825, BishopValueEg = 915, + RookValueMg = 1276, RookValueEg = 1380, + QueenValueMg = 2538, QueenValueEg = 2682, + Tempo = 28, MidgameLimit = 15258, EndgameLimit = 3915 }; @@ -196,23 +204,28 @@ enum PieceType { enum Piece { NO_PIECE, - W_PAWN = 1, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING, - B_PAWN = 9, B_KNIGHT, B_BISHOP, B_ROOK, B_QUEEN, B_KING, + W_PAWN = PAWN, W_KNIGHT, W_BISHOP, W_ROOK, W_QUEEN, W_KING, + B_PAWN = PAWN + 8, B_KNIGHT, B_BISHOP, B_ROOK, B_QUEEN, B_KING, PIECE_NB = 16 }; -extern Value PieceValue[PHASE_NB][PIECE_NB]; +constexpr Value PieceValue[PHASE_NB][PIECE_NB] = { + { VALUE_ZERO, PawnValueMg, KnightValueMg, BishopValueMg, RookValueMg, QueenValueMg, VALUE_ZERO, VALUE_ZERO, + VALUE_ZERO, PawnValueMg, KnightValueMg, BishopValueMg, RookValueMg, QueenValueMg, VALUE_ZERO, VALUE_ZERO }, + { VALUE_ZERO, PawnValueEg, KnightValueEg, BishopValueEg, RookValueEg, QueenValueEg, VALUE_ZERO, VALUE_ZERO, + VALUE_ZERO, PawnValueEg, KnightValueEg, BishopValueEg, RookValueEg, QueenValueEg, VALUE_ZERO, VALUE_ZERO } +}; typedef int Depth; enum : int { - DEPTH_QS_CHECKS = 0, DEPTH_QS_NO_CHECKS = -1, DEPTH_QS_RECAPTURES = -5, DEPTH_NONE = -6, - DEPTH_OFFSET = DEPTH_NONE, + + DEPTH_OFFSET = -7 // value used only for TT entry occupancy check }; enum Square : int { @@ -226,7 +239,8 @@ enum Square : int { SQ_A8, SQ_B8, SQ_C8, SQ_D8, SQ_E8, SQ_F8, SQ_G8, SQ_H8, SQ_NONE, - SQUARE_NB = 64 + SQUARE_ZERO = 0, + SQUARE_NB = 64 }; enum Direction : int { @@ -249,6 +263,21 @@ enum Rank : int { RANK_1, RANK_2, RANK_3, RANK_4, RANK_5, RANK_6, RANK_7, RANK_8, RANK_NB }; +// Keep track of what a move changes on the board (used by NNUE) +struct DirtyPiece { + + // Number of changed pieces + int dirty_num; + + // Max 3 pieces can change in one move. A promotion with capture moves + // both the pawn and the captured piece to SQ_NONE and the piece promoted + // to from SQ_NONE to the capture square. + Piece piece[3]; + + // From and to squares, which may be SQ_NONE + Square from[3]; + Square to[3]; +}; /// Score enum stores a middlegame and an endgame value in a single integer (enum). /// The least significant 16 bits are used to store the middlegame value and the @@ -274,11 +303,11 @@ inline Value mg_value(Score s) { } #define ENABLE_BASE_OPERATORS_ON(T) \ -constexpr T operator+(T d1, T d2) { return T(int(d1) + int(d2)); } \ -constexpr T operator-(T d1, T d2) { return T(int(d1) - int(d2)); } \ +constexpr T operator+(T d1, int d2) { return T(int(d1) + d2); } \ +constexpr T operator-(T d1, int d2) { return T(int(d1) - d2); } \ constexpr T operator-(T d) { return T(-int(d)); } \ -inline T& operator+=(T& d1, T d2) { return d1 = d1 + d2; } \ -inline T& operator-=(T& d1, T d2) { return d1 = d1 - d2; } +inline T& operator+=(T& d1, int d2) { return d1 = d1 + d2; } \ +inline T& operator-=(T& d1, int d2) { return d1 = d1 - d2; } #define ENABLE_INCR_OPERATORS_ON(T) \ inline T& operator++(T& d) { return d = T(int(d) + 1); } \ @@ -296,8 +325,8 @@ inline T& operator/=(T& d, int i) { return d = T(int(d) / i); } ENABLE_FULL_OPERATORS_ON(Value) ENABLE_FULL_OPERATORS_ON(Direction) -ENABLE_INCR_OPERATORS_ON(PieceType) ENABLE_INCR_OPERATORS_ON(Piece) +ENABLE_INCR_OPERATORS_ON(PieceType) ENABLE_INCR_OPERATORS_ON(Square) ENABLE_INCR_OPERATORS_ON(File) ENABLE_INCR_OPERATORS_ON(Rank) @@ -308,12 +337,6 @@ ENABLE_BASE_OPERATORS_ON(Score) #undef ENABLE_INCR_OPERATORS_ON #undef ENABLE_BASE_OPERATORS_ON -/// Additional operators to add integers to a Value -constexpr Value operator+(Value v, int i) { return Value(int(v) + i); } -constexpr Value operator-(Value v, int i) { return Value(int(v) - i); } -inline Value& operator+=(Value& v, int i) { return v = v + i; } -inline Value& operator-=(Value& v, int i) { return v = v - i; } - /// Additional operators to add a Direction to a Square constexpr Square operator+(Square s, Direction d) { return Square(int(s) + int(d)); } constexpr Square operator-(Square s, Direction d) { return Square(int(s) - int(d)); } @@ -341,25 +364,25 @@ inline Score operator*(Score s, int i) { return result; } -/// Multiplication of a Score by an boolean +/// Multiplication of a Score by a boolean inline Score operator*(Score s, bool b) { - return Score(int(s) * int(b)); + return b ? s : SCORE_ZERO; } constexpr Color operator~(Color c) { return Color(c ^ BLACK); // Toggle color } -constexpr Square operator~(Square s) { - return Square(s ^ SQ_A8); // Vertical flip SQ_A1 -> SQ_A8 +constexpr Square flip_rank(Square s) { // Swap A1 <-> A8 + return Square(s ^ SQ_A8); +} + +constexpr Square flip_file(Square s) { // Swap A1 <-> H1 + return Square(s ^ SQ_H1); } constexpr Piece operator~(Piece pc) { - return Piece(pc ^ 8); // Swap color of piece B_KNIGHT -> W_KNIGHT -} - -inline File map_to_queenside(File f) { - return std::min(f, File(FILE_H - f)); // Map files ABCDEFGH to files ABCDDCBA + return Piece(pc ^ 8); // Swap color of piece B_KNIGHT <-> W_KNIGHT } constexpr CastlingRights operator&(Color c, CastlingRights cr) { @@ -456,4 +479,13 @@ constexpr bool is_ok(Move m) { return from_sq(m) != to_sq(m); // Catch MOVE_NULL and MOVE_NONE } +/// Based on a congruential pseudo random number generator +constexpr Key make_key(uint64_t seed) { + return seed * 6364136223846793005ULL + 1442695040888963407ULL; +} + +} // namespace Stockfish + #endif // #ifndef TYPES_H_INCLUDED + +#include "tune.h" // Global visibility to tuning setup diff --git a/src/uci.cpp b/src/uci.cpp index 00e8ff54..20f6f1ff 100644 --- a/src/uci.cpp +++ b/src/uci.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -19,6 +17,7 @@ */ #include +#include #include #include #include @@ -36,6 +35,8 @@ using namespace std; +namespace Stockfish { + extern vector setup_bench(const Position&, istream&); namespace { @@ -78,6 +79,20 @@ namespace { } } + // trace_eval() prints the evaluation for the current position, consistent with the UCI + // options set so far. + + void trace_eval(Position& pos) { + + StateListPtr states(new std::deque(1)); + Position p; + p.set(pos.fen(), Options["UCI_Chess960"], &states->back(), Threads.main()); + + Eval::NNUE::verify(); + + sync_cout << "\n" << Eval::trace(p) << sync_endl; + } + // setoption() is called when engine receives the "setoption" UCI command. The // function updates the UCI option ("name") to the given value ("value"). @@ -116,7 +131,7 @@ namespace { limits.startTime = now(); // As early as possible! while (is >> token) - if (token == "searchmoves") + if (token == "searchmoves") // Needs to be the last command on the line while (is >> token) limits.searchmoves.push_back(UCI::to_move(pos, token)); @@ -147,7 +162,7 @@ namespace { uint64_t num, nodes = 0, cnt = 1; vector list = setup_bench(pos, args); - num = count_if(list.begin(), list.end(), [](string s) { return s.find("go ") == 0; }); + num = count_if(list.begin(), list.end(), [](string s) { return s.find("go ") == 0 || s.find("eval") == 0; }); TimePoint elapsed = now(); @@ -156,13 +171,19 @@ namespace { istringstream is(cmd); is >> skipws >> token; - if (token == "go") + if (token == "go" || token == "eval") { if (Cluster::is_root()) - cerr << "\nPosition: " << cnt++ << '/' << num << endl; - go(pos, is, states); - Threads.main()->wait_for_search_finished(); - nodes += Cluster::nodes_searched(); + cerr << "\nPosition: " << cnt++ << '/' << num << " (" << pos.fen() << ")" << endl; + + if (token == "go") + { + go(pos, is, states); + Threads.main()->wait_for_search_finished(); + nodes += Threads.nodes_searched(); + } + else if (Cluster::is_root()) + trace_eval(pos); } else if (token == "setoption") setoption(is); else if (token == "position") position(pos, is, states); @@ -180,6 +201,28 @@ namespace { << "\nNodes/second : " << 1000 * nodes / elapsed << 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) { + + // The model captures only up to 240 plies, so limit input (and rescale) + double m = std::min(240, ply) / 64.0; + + // Coefficients of a 3rd order polynomial fit based on fishtest data + // for two parameters needed to transform eval to the argument of a + // logistic function. + double as[] = {-8.24404295, 64.23892342, -95.73056462, 153.86478679}; + double bs[] = {-3.37154371, 28.44489198, -56.67657741, 72.05858751}; + double a = (((as[0] * m + as[1]) * m + as[2]) * m) + as[3]; + double b = (((bs[0] * m + bs[1]) * m + bs[2]) * m) + bs[3]; + + // Transform eval to centipawns with limited range + double x = 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))); + } + } // namespace @@ -234,13 +277,15 @@ void UCI::loop(int argc, char* argv[]) { // Additional custom non-UCI commands, mainly for debugging. // Do not use these commands during a search! - else if (token == "flip") pos.flip(); - else if (token == "bench") bench(pos, is, states); + else if (token == "flip") pos.flip(); + else if (token == "bench") bench(pos, is, states); else if (token == "d" && Cluster::is_root()) sync_cout << pos << sync_endl; else if (token == "eval" && Cluster::is_root()) - sync_cout << Eval::trace(pos) << sync_endl; - else if (Cluster::is_root()) + trace_eval(pos); + else if (token == "compiler" && Cluster::is_root()) + sync_cout << compiler_info() << sync_endl; + else if (!token.empty() && token[0] != '#' && Cluster::is_root()) sync_cout << "Unknown command: " << cmd << sync_endl; } while (token != "quit" && argc == 1); // Command line args are one-shot @@ -260,7 +305,7 @@ string UCI::value(Value v) { stringstream ss; - if (abs(v) < VALUE_MATE - MAX_PLY) + if (abs(v) < VALUE_MATE_IN_MAX_PLY) ss << "cp " << v * 100 / PawnValueEg; else ss << "mate " << (v > 0 ? VALUE_MATE - v + 1 : -VALUE_MATE - v) / 2; @@ -269,6 +314,22 @@ string UCI::value(Value v) { } +/// UCI::wdl() report WDL statistics given an evaluation and a game ply, based on +/// data gathered for fishtest LTC games. + +string UCI::wdl(Value v, int ply) { + + stringstream ss; + + int wdl_w = win_rate_model( v, ply); + int wdl_l = win_rate_model(-v, ply); + int wdl_d = 1000 - wdl_w - wdl_l; + ss << " wdl " << wdl_w << " " << wdl_d << " " << wdl_l; + + return ss.str(); +} + + /// UCI::square() converts a Square to a string in algebraic notation (g1, a7, etc.) std::string UCI::square(Square s) { @@ -318,3 +379,5 @@ Move UCI::to_move(const Position& pos, string& str) { return MOVE_NONE; } + +} // namespace Stockfish diff --git a/src/uci.h b/src/uci.h index 31b63e2f..d3160109 100644 --- a/src/uci.h +++ b/src/uci.h @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -26,6 +24,8 @@ #include "types.h" +namespace Stockfish { + class Position; namespace UCI { @@ -73,10 +73,13 @@ std::string value(Value v); std::string square(Square s); 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); } // namespace UCI extern UCI::OptionsMap Options; +} // namespace Stockfish + #endif // #ifndef UCI_H_INCLUDED diff --git a/src/ucioption.cpp b/src/ucioption.cpp index 23c0c480..d59c0100 100644 --- a/src/ucioption.cpp +++ b/src/ucioption.cpp @@ -1,8 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad - Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad + Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -23,6 +21,7 @@ #include #include +#include "evaluate.h" #include "misc.h" #include "search.h" #include "thread.h" @@ -32,17 +31,20 @@ using std::string; +namespace Stockfish { + UCI::OptionsMap Options; // Global object namespace UCI { /// 'On change' actions, triggered by an option's value change void on_clear_hash(const Option&) { Search::clear(); } -void on_hash_size(const Option& o) { TT.resize(o); } +void on_hash_size(const Option& o) { TT.resize(size_t(o)); } void on_logger(const Option& o) { start_logger(o); } -void on_threads(const Option& o) { Threads.set(o); } +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(); } /// Our case insensitive less() function as required by UCI protocol bool CaseInsensitiveLess::operator() (const string& s1, const string& s2) const { @@ -52,12 +54,11 @@ bool CaseInsensitiveLess::operator() (const string& s1, const string& s2) const } -/// init() initializes the UCI options to their hard-coded default values +/// UCI::init() initializes the UCI options to their hard-coded default values void init(OptionsMap& o) { - // at most 2^32 clusters. - constexpr int MaxHashMB = Is64Bit ? 131072 : 2048; + constexpr int MaxHashMB = Is64Bit ? 33554432 : 2048; o["Debug Log File"] << Option("", on_logger); o["Contempt"] << Option(24, -100, 100); @@ -68,18 +69,20 @@ void init(OptionsMap& o) { o["Ponder"] << Option(false); o["MultiPV"] << Option(1, 1, 500); o["Skill Level"] << Option(20, 0, 20); - o["Move Overhead"] << Option(30, 0, 5000); - o["Minimum Thinking Time"] << Option(20, 0, 5000); - o["Slow Mover"] << Option(84, 10, 1000); + o["Move Overhead"] << Option(10, 0, 5000); + o["Slow Mover"] << Option(100, 10, 1000); o["nodestime"] << Option(0, 0, 10000); o["UCI_Chess960"] << Option(false); o["UCI_AnalyseMode"] << Option(false); o["UCI_LimitStrength"] << Option(false); o["UCI_Elo"] << Option(1350, 1350, 2850); + o["UCI_ShowWDL"] << Option(false); o["SyzygyPath"] << Option("", on_tb_path); 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["EvalFile"] << Option(EvalFileDefaultName, on_eval_file); } @@ -189,3 +192,5 @@ Option& Option::operator=(const string& v) { } } // namespace UCI + +} // namespace Stockfish diff --git a/tests/instrumented.sh b/tests/instrumented.sh index ae6d5c4b..bfb50e94 100755 --- a/tests/instrumented.sh +++ b/tests/instrumented.sh @@ -20,7 +20,7 @@ case $1 in --valgrind-thread) echo "valgrind-thread testing started" prefix='' - exeprefix='valgrind --error-exitcode=42' + exeprefix='valgrind --fair-sched=try --error-exitcode=42' postfix='1>/dev/null' threads="2" ;; @@ -39,16 +39,16 @@ case $1 in threads="2" cat << EOF > tsan.supp -race:TTEntry::move -race:TTEntry::depth -race:TTEntry::bound -race:TTEntry::save -race:TTEntry::value -race:TTEntry::eval -race:TTEntry::is_pv +race:Stockfish::TTEntry::move +race:Stockfish::TTEntry::depth +race:Stockfish::TTEntry::bound +race:Stockfish::TTEntry::save +race:Stockfish::TTEntry::value +race:Stockfish::TTEntry::eval +race:Stockfish::TTEntry::is_pv -race:TranspositionTable::probe -race:TranspositionTable::hashfull +race:Stockfish::TranspositionTable::probe +race:Stockfish::TranspositionTable::hashfull EOF @@ -70,7 +70,7 @@ for args in "eval" \ "go depth 10" \ "go movetime 1000" \ "go wtime 8000 btime 8000 winc 500 binc 500" \ - "bench 128 $threads 10 default depth" + "bench 128 $threads 8 default depth" do echo "$prefix $exeprefix ./stockfish $args $postfix" @@ -80,7 +80,7 @@ done # more general testing, following an uci protocol exchange cat << EOF > game.exp - set timeout 10 + set timeout 240 spawn $exeprefix ./stockfish send "uci\n" @@ -98,7 +98,7 @@ cat << EOF > game.exp expect "bestmove" send "position fen 5rk1/1K4p1/8/8/3B4/8/8/8 b - - 0 1\n" - send "go depth 30\n" + send "go depth 20\n" expect "bestmove" send "quit\n" @@ -121,7 +121,7 @@ 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 10 default depth\n" + send "bench 128 1 8 default depth\n" send "quit\n" expect eof