Merge branch 'master' into release/0.26

.circleci/Dockerfile

@@ -2,7 +2,7 @@ FROM floopcz/tensorflow_cc:ubuntu-shared-cuda
 ARG DEBIAN_FRONTEND=noninteractive
 
 RUN wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB && apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB && sh -c 'echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list' && apt-get update && apt-get install -y intel-mkl-64bit-2018.2-046
-RUN apt-get install -y clang ninja-build python3-pip nvidia-opencl-dev libopenblas-dev libboost-dev libgtest-dev git ssh tar gzip ca-certificates sudo
+RUN apt-get install -y clang ninja-build python3-pip nvidia-opencl-dev libopenblas-dev libboost-dev libgtest-dev git ssh tar gzip ca-certificates libeigen3-dev sudo
 RUN apt-get install -y g++-8
 RUN apt-get install -y cuda
 RUN pip3 install meson

.circleci/config.yml

@@ -2,7 +2,7 @@ version: 2
 jobs:
   build:
     docker:
-      - image: danieluranga/leela_chess_zero-lc0_ubuntu_builder:0.0.7
+      - image: danieluranga/leela_chess_zero-lc0_ubuntu_builder:0.0.8
     steps:
       - checkout
       - run:

appveyor.yml

@@ -26,8 +26,8 @@ install:
 - cmd: set BLAS=false
 - cmd: set GTEST=false
 - cmd: set ANDROID=false
-- cmd: set NET=703350
-- cmd: set NET_HASH=feb6dcb84697b55c3573f503dc9fca2783478ef5f3021669ad854129aaab0683
+- cmd: set NET=703810
+- cmd: set NET_HASH=b30e742bcfd905815e0e7dbd4e1bafb41ade748f85d006b8e28758f1a3107ae3
 - cmd: IF %NAME%==android set ANDROID=true
 - cmd: IF %NAME%==gpu-nvidia-cuda set CUDA=true
 - cmd: IF %NAME%==gpu-dx12 set DX=true
@@ -53,7 +53,7 @@ install:
 - cmd: IF %CUDA%==true IF NOT EXIST C:\cache\cuda appveyor DownloadFile http://developer.download.nvidia.com/compute/redist/cudnn/v7.4.2/cudnn-10.0-windows10-x64-v7.4.2.24.zip
 - cmd: IF %CUDA%==true IF NOT EXIST C:\cache\cuda 7z x cudnn-10.0-windows10-x64-v7.4.2.24.zip -oC:\cache
 - cmd: set PATH=C:\Python36;C:\Python36\scripts;%PATH%
-- cmd: pip3 install --upgrade meson==0.54.3
+- cmd: pip3 install --upgrade meson
 - cmd: call "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsall.bat" amd64
 - cmd: IF %NAME%==android IF NOT EXIST C:\ndk\android-ndk-r19c\toolchains\llvm\prebuilt\windows-x86_64 appveyor DownloadFile https://dl.google.com/android/repository/android-ndk-r19c-windows-x86_64.zip
 - cmd: IF %NAME%==android IF NOT EXIST C:\ndk\android-ndk-r19c\toolchains\llvm\prebuilt\windows-x86_64 7z x android-ndk-r19c-windows-x86_64.zip -oC:\ndk
@@ -88,7 +88,9 @@ before_build:
 - cmd: IF %DX%==true SET BUILD_BLAS=true
 - cmd: SET EMBED=false
 - cmd: IF %APPVEYOR_REPO_TAG%==true IF %ANDROID%==true SET EMBED=true
-- cmd: IF %ANDROID%==false meson build --backend vs2017 --buildtype release -Dgtest=%GTEST% -Dopencl=%OPENCL% -Dblas=%BUILD_BLAS% -Ddnnl=true -Ddx=%DX% -Dcudnn=%CUDA% -Dispc_native_only=false -Dpopcnt=false -Dcudnn_include="%CUDA_PATH%\include","%PKG_FOLDER%\cuda\include" -Dcudnn_libdirs="%CUDA_PATH%\lib\x64","%PKG_FOLDER%\cuda\lib\x64" -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\dist64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\dist64\lib" -Ddnnl_dir="%PKG_FOLDER%\dnnl_win_1.5.0_cpu_vcomp" -Dopencl_include="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\include" -Dopencl_libdirs="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\lib\x64" -Ddefault_library=static
+- cmd: SET EXTRA=
+- cmd: IF %NAME%==cpu-dnnl SET EXTRA=-Db_vscrt=md
+- cmd: IF %ANDROID%==false meson build --backend vs2017 --buildtype release -Dgtest=%GTEST% -Dopencl=%OPENCL% -Dblas=%BUILD_BLAS% -Ddnnl=true -Ddx=%DX% -Dcudnn=%CUDA% -Dispc_native_only=false -Dpopcnt=false -Dcudnn_include="%CUDA_PATH%\include","%PKG_FOLDER%\cuda\include" -Dcudnn_libdirs="%CUDA_PATH%\lib\x64","%PKG_FOLDER%\cuda\lib\x64" -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\dist64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\dist64\lib" -Ddnnl_dir="%PKG_FOLDER%\dnnl_win_1.5.0_cpu_vcomp" -Dopencl_include="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\include" -Dopencl_libdirs="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\lib\x64" -Ddefault_library=static %EXTRA%
 - cmd: IF %ANDROID%==true meson arm64-v8a --buildtype release -Dgtest=false -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\android-aarch64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\android-aarch64\lib" -Dembed=%EMBED% -Ddefault_library=static --cross-file crossfile-aarch64
 - cmd: IF %ANDROID%==true meson armeabi-v7a --buildtype release -Dgtest=false -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\android-armv7a\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\android-armv7a\lib" -Dembed=%EMBED% -Ddefault_library=static --cross-file crossfile-armv7a -Dispc=false
 build_script:

meson.build

@@ -15,7 +15,7 @@
 # along with Leela Chess.  If not, see <http://www.gnu.org/licenses/>.
 
 project('lc0', 'cpp',
-        default_options : ['cpp_std=c++17', 'b_ndebug=if-release', 'warning_level=3', 'b_lto=true'],
+        default_options : ['cpp_std=c++17', 'b_ndebug=if-release', 'warning_level=3', 'b_lto=true', 'b_vscrt=mt'],
         meson_version: '>=0.52')
 
 cc = meson.get_compiler('cpp')
@@ -418,11 +418,14 @@ if get_option('build_backends')
     cuda_arguments = ['@EXTRA_ARGS@', '-c', '@INPUT@', '-o', '@OUTPUT@',
                       '-I', meson.current_source_dir() + '/src']
     if host_machine.system() == 'windows'
-      # meson bug fixed in v0.51
-      if meson.version().version_compare('<0.51') or get_option('default_library') != 'static'
-        cuda_arguments += ['-Xcompiler', '-MD']
-      else
+      if get_option('b_vscrt') == 'mt'
         cuda_arguments += ['-Xcompiler', '-MT']
+      elif get_option('b_vscrt') == 'mtd'
+        cuda_arguments += ['-Xcompiler', '-MTd']
+      elif get_option('b_vscrt') == 'mdd' or (get_option('b_vscrt') == 'from_buildtype' and get_option('buildtype') == 'debug')
+        cuda_arguments += ['-Xcompiler', '-MDd']
+      elif get_option('b_vscrt') != 'none'
+        cuda_arguments += ['-Xcompiler', '-MD']
       endif
     else
       cuda_arguments += ['--std=c++14', '-Xcompiler', '-fPIC']
@@ -591,4 +594,4 @@ if get_option('python_bindings')
     [py_files + files],
     include_directories: [includes],
     dependencies: [cpython] + deps)
-endif
+endif

src/chess/pgn.h

@@ -74,6 +74,7 @@ class PgnReader {
     bool in_comment = false;
     bool started = false;
     while (GzGetLine(file, line)) {
+      if (!line.empty() && line.back() == '\r') line.pop_back();
       // TODO: support line breaks in tags to ensure they are properly ignored.
       if (line.empty() || line[0] == '[') {
         if (started) {

src/chess/position.cc

@@ -91,17 +91,21 @@ void PositionHistory::Append(Move m) {
   //                has a bug in implementation of emplace_back, when
   //                reallocation happens. (it also reallocates Last())
   positions_.push_back(Position(Last(), m));
-  positions_.back().SetRepetitions(ComputeLastMoveRepetitions());
+  int cycle_length;
+  int repetitions = ComputeLastMoveRepetitions(&cycle_length);
+  positions_.back().SetRepetitions(repetitions, cycle_length);
 }
 
-int PositionHistory::ComputeLastMoveRepetitions() const {
+int PositionHistory::ComputeLastMoveRepetitions(int* cycle_length) const {
+  *cycle_length = 0;
   const auto& last = positions_.back();
   // TODO(crem) implement hash/cache based solution.
   if (last.GetRule50Ply() < 4) return 0;
 
   for (int idx = positions_.size() - 3; idx >= 0; idx -= 2) {
     const auto& pos = positions_[idx];
     if (pos.GetBoard() == last.GetBoard()) {
+      *cycle_length = positions_.size() - 1 - idx;
       return 1 + pos.GetRepetitions();
     }
     if (pos.GetRule50Ply() < 2) return 0;

src/chess/position.h

@@ -49,9 +49,15 @@ class Position {
   // How many time the same position appeared in the game before.
   int GetRepetitions() const { return repetitions_; }
 
+  // How many half-moves since the same position appeared in the game before.
+  int GetPliesSincePrevRepetition() const { return cycle_length_; }
+
   // Someone outside that class knows better about repetitions, so they can
   // set it.
-  void SetRepetitions(int repetitions) { repetitions_ = repetitions; }
+  void SetRepetitions(int repetitions, int cycle_length) {
+    repetitions_ = repetitions;
+    cycle_length_ = cycle_length;
+  }
 
   // Number of ply with no captures and pawn moves.
   int GetRule50Ply() const { return rule50_ply_; }
@@ -73,6 +79,8 @@ class Position {
   int rule50_ply_ = 0;
   // How many repetitions this position had before. For new positions it's 0.
   int repetitions_;
+  // How many half-moves since the position was repeated or 0.
+  int cycle_length_;
   // number of half-moves since beginning of the game.
   int ply_count_ = 0;
 };
@@ -125,7 +133,7 @@ class PositionHistory {
   bool DidRepeatSinceLastZeroingMove() const;
 
  private:
-  int ComputeLastMoveRepetitions() const;
+  int ComputeLastMoveRepetitions(int* cycle_length) const;
 
   std::vector<Position> positions_;
 };

src/mcts/node.cc

@@ -293,7 +293,7 @@ void Node::SortEdges() {
 }
 
 void Node::MakeTerminal(GameResult result, float plies_left, Terminal type) {
-  SetBounds(result, result);
+  if (type != Terminal::TwoFold) SetBounds(result, result);
   terminal_type_ = type;
   m_ = plies_left;
   if (result == GameResult::DRAW) {
@@ -375,11 +375,37 @@ void Node::AdjustForTerminal(float v, float d, float m, int multivisit) {
   d_ += multivisit * d / n_;
   m_ += multivisit * m / n_;
   // Best child is potentially no longer valid. This shouldn't be needed since
-  // AjdustForTerminal is always called immediately after FinalizeScoreUpdate,
+  // AdjustForTerminal is always called immediately after FinalizeScoreUpdate,
   // but for safety in case that changes.
   best_child_cached_ = nullptr;
 }
 
+void Node::RevertTerminalVisits(float v, float d, float m, int multivisit) {
+  // Compute new n_ first, as reducing a node to 0 visits is a special case.
+  const int n_new = n_ - multivisit;
+  if (n_new <= 0) {
+    if (parent_ != nullptr) {
+      // To keep consistency with FinalizeScoreUpdate() expanding a node again,
+      // we need to reduce the parent's visited policy.
+      parent_->visited_policy_ -= parent_->edges_[index_].GetP();
+    }
+    // If n_new == 0, reset all relevant values to 0.
+    wl_ = 0.0;
+    d_ = 1.0;
+    m_ = 0.0;
+    n_ = 0;
+  } else {
+    // Recompute Q and M.
+    wl_ -= multivisit * (v - wl_) / n_new;
+    d_ -= multivisit * (d - d_) / n_new;
+    m_ -= multivisit * (m - m_) / n_new;
+    // Decrement N.
+    n_ -= multivisit;
+  }
+  // Best child is potentially no longer valid.
+  best_child_cached_ = nullptr;
+}
+
 void Node::UpdateBestChild(const Iterator& best_edge, int visits_allowed) {
   best_child_cached_ = best_edge.node();
   // An edge can point to an unexpanded node with n==0. These nodes don't

src/mcts/node.h

@@ -119,7 +119,7 @@ class Node {
   using Iterator = Edge_Iterator<false>;
   using ConstIterator = Edge_Iterator<true>;
 
-  enum class Terminal : uint8_t { NonTerminal, EndOfGame, Tablebase };
+  enum class Terminal : uint8_t { NonTerminal, EndOfGame, Tablebase, TwoFold };
 
   // Takes pointer to a parent node and own index in a parent.
   Node(Node* parent, uint16_t index)
@@ -165,6 +165,7 @@ class Node {
   // Returns whether the node is known to be draw/lose/win.
   bool IsTerminal() const { return terminal_type_ != Terminal::NonTerminal; }
   bool IsTbTerminal() const { return terminal_type_ == Terminal::Tablebase; }
+  bool IsTwoFoldTerminal() const { return terminal_type_ == Terminal::TwoFold; }
   typedef std::pair<GameResult, GameResult> Bounds;
   Bounds GetBounds() const { return {lower_bound_, upper_bound_}; }
   uint8_t GetNumEdges() const { return num_edges_; }
@@ -191,6 +192,8 @@ class Node {
   void FinalizeScoreUpdate(float v, float d, float m, int multivisit);
   // Like FinalizeScoreUpdate, but it updates n existing visits by delta amount.
   void AdjustForTerminal(float v, float d, float m, int multivisit);
+  // Revert visits to a node which ended in a now reverted terminal.
+  void RevertTerminalVisits(float v, float d, float m, int multivisit);
   // When search decides to treat one visit as several (in case of collisions
   // or visiting terminal nodes several times), it amplifies the visit by
   // incrementing n_in_flight.

src/mcts/params.cc

@@ -91,6 +91,11 @@ const OptionId SearchParams::kRootHasOwnCpuctParamsId{
     "If enabled, cpuct parameters for root node are taken from *AtRoot "
     "parameters. Otherwise, they are the same as for the rest of nodes. "
     "Temporary flag for transition to a new version."};
+const OptionId SearchParams::kTwoFoldDrawsId{
+    "two-fold-draws", "TwoFoldDraws",
+    "Evaluates twofold repetitions in the search tree as draws. Visits to "
+    "these positions are reverted when the first occurrence is played "
+    "and not in the search tree anymore."};
 const OptionId SearchParams::kTemperatureId{
     "temperature", "Temperature",
     "Tau value from softmax formula for the first move. If equal to 0, the "
@@ -283,6 +288,7 @@ void SearchParams::Populate(OptionsParser* options) {
   options->Add<FloatOption>(kCpuctFactorId, 0.0f, 1000.0f) = 2.815f;
   options->Add<FloatOption>(kCpuctFactorAtRootId, 0.0f, 1000.0f) = 2.815f;
   options->Add<BoolOption>(kRootHasOwnCpuctParamsId) = true;
+  options->Add<BoolOption>(kTwoFoldDrawsId) = true;
   options->Add<FloatOption>(kTemperatureId, 0.0f, 100.0f) = 0.0f;
   options->Add<IntOption>(kTempDecayMovesId, 0, 100) = 0;
   options->Add<IntOption>(kTempDecayDelayMovesId, 0, 100) = 0;

src/mcts/params.h

@@ -53,6 +53,7 @@ class SearchParams {
   float GetCpuctFactor(bool at_root) const {
     return at_root ? kCpuctFactorAtRoot : kCpuctFactor;
   }
+  bool GetTwoFoldDraws() const { return options_.Get<bool>(kTwoFoldDrawsId); }
   float GetTemperature() const { return options_.Get<float>(kTemperatureId); }
   float GetTemperatureVisitOffset() const {
     return options_.Get<float>(kTemperatureVisitOffsetId);
@@ -121,6 +122,7 @@ class SearchParams {
   static const OptionId kCpuctFactorId;
   static const OptionId kCpuctFactorAtRootId;
   static const OptionId kRootHasOwnCpuctParamsId;
+  static const OptionId kTwoFoldDrawsId;
   static const OptionId kTemperatureId;
   static const OptionId kTempDecayMovesId;
   static const OptionId kTempDecayDelayMovesId;

src/mcts/search.cc

@@ -1044,7 +1044,7 @@ void SearchWorker::GatherMinibatch() {
     // of the game), it means that we already visited this node before.
     if (picked_node.IsExtendable()) {
       // Node was never visited, extend it.
-      ExtendNode(node);
+      ExtendNode(node, picked_node.depth);
 
       // Only send non-terminal nodes to a neural network.
       if (!node->IsTerminal()) {
@@ -1145,14 +1145,54 @@ SearchWorker::NodeToProcess SearchWorker::PickNodeToExtend(
       }
       return NodeToProcess::Collision(node, depth, collision_limit);
     }
-    // Either terminal or unexamined leaf node -- the end of this playout.
-    if (node->IsTerminal() || !node->HasChildren()) {
+    // If terminal, we either found a twofold draw to be reverted, or
+    // reached the end of this playout.
+    if (node->IsTerminal()) {
+      // Probably best place to check for two-fold draws consistently.
+      // Depth starts with 1 at root, so real depth is depth - 1.
+      // Check whether first repetition was before root. If yes, remove
+      // terminal status of node and revert all visits in the tree.
+      // Length of repetition was stored in m_. This code will only do
+      // something when tree is reused and twofold visits need to be reverted.
+      if (node->IsTwoFoldTerminal() && depth - 1 < node->GetM()) {
+        int depth_counter = 0;
+        // Cache node's values as we reset them in the process. We could
+        // manually set wl and d, but if we want to reuse this for reverting
+        // other terminal nodes this is the way to go.
+        const auto wl = node->GetWL();
+        const auto d = node->GetD();
+        const auto m = node->GetM();
+        const auto terminal_visits = node->GetN();
+        for (Node* node_to_revert = node; node_to_revert != nullptr;
+             node_to_revert = node_to_revert->GetParent()) {
+          // Revert all visits on twofold draw when making it non terminal.
+          node_to_revert->RevertTerminalVisits(wl, d, m + (float)depth_counter,
+                                               terminal_visits);
+          depth_counter++;
+          // Even if original tree still exists, we don't want to revert more
+          // than until new root.
+          if (depth_counter > depth - 1) break;
+          // If wl != 0, we would have to switch signs at each depth.
+        }
+        // Mark the prior twofold draw as non terminal to extend it again.
+        node->MakeNotTerminal();
+        // When reverting the visits, we also need to revert the initial
+        // visits, as we reused fewer nodes than anticipated.
+        search_->initial_visits_ -= terminal_visits;
+        // Max depth doesn't change when reverting the visits, and cum_depth_
+        // only counts the average depth of new nodes, not reused ones.
+      } else {
+        return NodeToProcess::Visit(node, depth);
+      }
+    }
+    // If unexamined leaf node -- the end of this playout.
+    if (!node->HasChildren()) {
       return NodeToProcess::Visit(node, depth);
     }
     Node* possible_shortcut_child = node->GetCachedBestChild();
     if (possible_shortcut_child) {
-      // Add two here to reverse the conservatism that goes into calculating the
-      // remaining cache visits.
+      // Add two here to reverse the conservatism that goes into calculating
+      // the remaining cache visits.
       collision_limit =
           std::min(collision_limit, node->GetRemainingCacheVisits() + 2);
       is_root_node = false;
@@ -1239,7 +1279,7 @@ SearchWorker::NodeToProcess SearchWorker::PickNodeToExtend(
   }
 }
 
-void SearchWorker::ExtendNode(Node* node) {
+void SearchWorker::ExtendNode(Node* node, int depth) {
   // Initialize position sequence with pre-move position.
   history_.Trim(search_->played_history_.GetLength());
   std::vector<Move> to_add;
@@ -1291,9 +1331,20 @@ void SearchWorker::ExtendNode(Node* node) {
       return;
     }
 
-    if (history_.Last().GetRepetitions() >= 2) {
+    const auto repetitions = history_.Last().GetRepetitions();
+    // Mark two-fold repetitions as draws according to settings.
+    // Depth starts with 1 at root, so number of plies in PV is depth - 1.
+    if (repetitions >= 2) {
       node->MakeTerminal(GameResult::DRAW);
       return;
+    } else if (repetitions == 1 && depth - 1 >= 4 &&
+               params_.GetTwoFoldDraws() &&
+               depth - 1 >= history_.Last().GetPliesSincePrevRepetition()) {
+      const auto cycle_length = history_.Last().GetPliesSincePrevRepetition();
+      // use plies since first repetition as moves left; exact if forced draw.
+      node->MakeTerminal(GameResult::DRAW, (float)cycle_length,
+                         Node::Terminal::TwoFold);
+      return;
     }
 
     // Neither by-position or by-rule termination, but maybe it's a TB position.

src/mcts/search.h

@@ -172,7 +172,7 @@ class Search {
   const SearchParams params_;
   const MoveList searchmoves_;
   const std::chrono::steady_clock::time_point start_time_;
-  const int64_t initial_visits_;
+  int64_t initial_visits_;
   // tb_hits_ must be initialized before root_move_filter_.
   std::atomic<int> tb_hits_{0};
   const MoveList root_move_filter_;
@@ -303,7 +303,7 @@ class SearchWorker {
   };
 
   NodeToProcess PickNodeToExtend(int collision_limit);
-  void ExtendNode(Node* node);
+  void ExtendNode(Node* node, int depth);
   bool AddNodeToComputation(Node* node, bool add_if_cached, int* transform_out);
   int PrefetchIntoCache(Node* node, int budget, bool is_odd_depth);
   void FetchSingleNodeResult(NodeToProcess* node_to_process,