Merge remote-tracking branch 'remotes/origin/master' into trainer

src/evaluate.cpp

@@ -56,7 +56,7 @@ namespace Eval {
     return UseNNUEMode::False;
   }
 
-  void init_NNUE() {
+  void NNUE::init() {
 
     useNNUE = nnue_mode_from_option(Options["Use NNUE"]);
     if (useNNUE == UseNNUEMode::False)
@@ -81,8 +81,8 @@ namespace Eval {
         }
   }
 
-  /// verify_NNUE() verifies that the last net used was loaded successfully
-  void verify_NNUE() {
+  /// NNUE::verify() verifies that the last net used was loaded successfully
+  void NNUE::verify() {
 
     string eval_file = string(Options["EvalFile"]);
 
@@ -984,24 +984,32 @@ make_v:
 /// evaluation of the position from the point of view of the side to move.
 
 Value Eval::evaluate(const Position& pos) {
-  if (useNNUE == UseNNUEMode::Pure) {
-      return NNUE::evaluate(pos);
-  }
 
-  // Use classical eval if there is a large imbalance
-  // If there is a moderate imbalance, use classical eval with probability (1/8),
-  // as derived from the node counter.
-  bool useClassical = abs(eg_value(pos.psq_score())) * 16 > NNUEThreshold1 * (16 + pos.rule50_count());
-  bool classical = (useNNUE == UseNNUEMode::False)
-                ||  useClassical
-                || (abs(eg_value(pos.psq_score())) > PawnValueMg / 4 && !(pos.this_thread()->nodes & 0xB));
-  Value v = classical ? Evaluation<NO_TRACE>(pos).value()
-                      : NNUE::evaluate(pos);
+  Value v;
 
-  if (   useClassical
-      && useNNUE != UseNNUEMode::False
-      && abs(v) * 16 < NNUEThreshold2 * (16 + pos.rule50_count()))
+  if (Eval::useNNUE == UseNNUEMode::Pure) {
       v = NNUE::evaluate(pos);
+  }
+  else if (Eval::useNNUE == UseNNUEMode::False)
+      v = Evaluation<NO_TRACE>(pos).value();
+  else
+  {
+      // scale and shift NNUE for compatibility with search and classical evaluation
+      auto  adjusted_NNUE = [&](){ return NNUE::evaluate(pos) * 5 / 4 + 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));
+
+      v = classical ? Evaluation<NO_TRACE>(pos).value() : adjusted_NNUE();
+
+      // if the classical eval is small and imbalance large, use NNUE nevertheless.
+      if (   largePsq
+          && abs(v) * 16 < NNUEThreshold2 * r50)
+          v = adjusted_NNUE();
+  }
 
   // Damp down the evaluation linearly when shuffling
   v = v * (100 - pos.rule50_count()) / 100;

src/evaluate.h

@@ -38,8 +38,6 @@ namespace Eval {
 
   extern UseNNUEMode useNNUE;
   extern std::string eval_file_loaded;
-  void init_NNUE();
-  void verify_NNUE();
 
   // 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
@@ -49,7 +47,9 @@ namespace Eval {
   namespace NNUE {
 
     Value evaluate(const Position& pos);
-    bool  load_eval(std::string streamName, std::istream& stream);
+    bool load_eval(std::string name, std::istream& stream);
+    void init();
+    void verify();
 
   } // namespace NNUE
 

src/learn/gensfen.cpp

@@ -1168,7 +1168,7 @@ namespace Learner
             << "  detect_draw_by_insufficient_mating_material = " << detect_draw_by_insufficient_mating_material << endl;
 
         // Show if the training data generator uses NNUE.
-        Eval::verify_NNUE();
+        Eval::NNUE::verify();
 
         Threads.main()->ponder = false;
 

src/learn/learn.cpp

@@ -1841,7 +1841,7 @@ namespace Learner
 
         if (use_convert_plain)
         {
-            Eval::init_NNUE();
+            Eval::NNUE::init();
             cout << "convert_plain.." << endl;
             convert_plain(filenames, output_file_name);
             return;
@@ -1849,7 +1849,7 @@ namespace Learner
 
         if (use_convert_bin)
         {
-            Eval::init_NNUE();
+            Eval::NNUE::init();
             cout << "convert_bin.." << endl;
             convert_bin(
                 filenames,
@@ -1870,7 +1870,7 @@ namespace Learner
 
         if (use_convert_bin_from_pgn_extract)
         {
-            Eval::init_NNUE();
+            Eval::NNUE::init();
             cout << "convert_bin_from_pgn-extract.." << endl;
             convert_bin_from_pgn_extract(
                 filenames,
@@ -1938,7 +1938,7 @@ namespace Learner
         cout << "init.." << endl;
 
         // Read evaluation function parameters
-        Eval::init_NNUE();
+        Eval::NNUE::init();
 
         Threads.main()->ponder = false;
 

src/learn/multi_think.cpp

@@ -12,7 +12,7 @@ void MultiThink::go_think()
 	// Read evaluation function, etc.
 	// In the case of the learn command, the value of the evaluation function may be corrected after reading the evaluation function, so
 	// Skip memory corruption check.
-	Eval::init_NNUE();
+	Eval::NNUE::init();
 
 	// Call the derived class's init().
 	init();

src/main.cpp

@@ -45,7 +45,7 @@ int main(int argc, char* argv[]) {
   Endgames::init();
   Threads.set(size_t(Options["Threads"]));
   Search::clear(); // After threads are up
-  Eval::init_NNUE();
+  Eval::NNUE::init();
 
   UCI::loop(argc, argv);
 

src/misc.cpp

@@ -357,27 +357,11 @@ void std_aligned_free(void* ptr) {
 #endif
 }
 
-/// aligned_ttmem_alloc() will return suitably aligned memory, if possible using large pages.
-/// The returned pointer is the aligned one, while the mem argument is the one that needs
-/// to be passed to free. With c++17 some of this functionality could be simplified.
+/// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
 
-#if defined(__linux__) && !defined(__ANDROID__)
+#if defined(_WIN32)
 
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
-
-  constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page sizes
-  size_t size = ((allocSize + alignment - 1) / alignment) * alignment; // multiple of alignment
-  if (posix_memalign(&mem, alignment, size))
-     mem = nullptr;
-#if defined(MADV_HUGEPAGE)
-  madvise(mem, allocSize, MADV_HUGEPAGE);
-#endif
-  return mem;
-}
-
-#elif defined(_WIN64)
-
-static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
+static void* aligned_large_pages_alloc_win(size_t allocSize) {
 
   HANDLE hProcessToken { };
   LUID luid { };
@@ -422,12 +406,13 @@ static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
   return mem;
 }
 
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
+void* aligned_large_pages_alloc(size_t allocSize) {
 
   static bool firstCall = true;
+  void* mem;
 
   // Try to allocate large pages
-  mem = aligned_ttmem_alloc_large_pages(allocSize);
+  mem = aligned_large_pages_alloc_win(allocSize);
 
   // Suppress info strings on the first call. The first call occurs before 'uci'
   // is received and in that case this output confuses some GUIs.
@@ -449,23 +434,31 @@ void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
 
 #else
 
-void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
+void* aligned_large_pages_alloc(size_t allocSize) {
 
-  constexpr size_t alignment = 64; // assumed cache line size
-  size_t size = allocSize + alignment - 1; // allocate some extra space
-  mem = malloc(size);
-  void* ret = reinterpret_cast<void*>((uintptr_t(mem) + alignment - 1) & ~uintptr_t(alignment - 1));
-  return ret;
+#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_ttmem_free() will free the previously allocated ttmem
+/// aligned_large_pages_free() will free the previously allocated ttmem
 
-#if defined(_WIN64)
+#if defined(_WIN32)
 
-void aligned_ttmem_free(void* mem) {
+void aligned_large_pages_free(void* mem) {
 
   if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
   {
@@ -478,8 +471,8 @@ void aligned_ttmem_free(void* mem) {
 
 #else
 
-void aligned_ttmem_free(void *mem) {
-  free(mem);
+void aligned_large_pages_free(void *mem) {
+  std_aligned_free(mem);
 }
 
 #endif

src/misc.h

@@ -39,8 +39,8 @@ void prefetch(void* addr);
 void start_logger(const std::string& fname);
 void* std_aligned_alloc(size_t alignment, size_t size);
 void std_aligned_free(void* ptr);
-void* aligned_ttmem_alloc(size_t size, void*& mem);
-void aligned_ttmem_free(void* mem); // nop if mem == nullptr
+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);

src/nnue/evaluate_nnue.cpp

@@ -30,7 +30,7 @@
 
 namespace Eval::NNUE {
 
-  uint32_t kpp_board_index[PIECE_NB][COLOR_NB] = {
+  const uint32_t kpp_board_index[PIECE_NB][COLOR_NB] = {
    // convention: W - us, B - them
    // viewed from other side, W and B are reversed
       { PS_NONE,     PS_NONE     },
@@ -52,7 +52,7 @@ namespace Eval::NNUE {
   };
 
   // Input feature converter
-  AlignedPtr<FeatureTransformer> feature_transformer;
+  LargePagePtr<FeatureTransformer> feature_transformer;
 
   // Evaluation function
   AlignedPtr<Network> network;
@@ -79,14 +79,22 @@ namespace Eval::NNUE {
     std::memset(pointer.get(), 0, sizeof(T));
   }
 
+  template <typename T>
+  void Initialize(LargePagePtr<T>& pointer) {
+
+    static_assert(alignof(T) <= 4096, "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
+    pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
+    std::memset(pointer.get(), 0, sizeof(T));
+  }
+
   // Read evaluation function parameters
   template <typename T>
-  bool ReadParameters(std::istream& stream, const AlignedPtr<T>& pointer) {
+  bool ReadParameters(std::istream& stream, T& reference) {
 
     std::uint32_t header;
     header = read_little_endian<std::uint32_t>(stream);
     if (!stream || header != T::GetHashValue()) return false;
-    return pointer->ReadParameters(stream);
+    return reference.ReadParameters(stream);
   }
 
   // write evaluation function parameters
@@ -97,6 +105,13 @@ namespace Eval::NNUE {
     return pointer->WriteParameters(stream);
   }
 
+  template <typename T>
+  bool WriteParameters(std::ostream& stream, const LargePagePtr<T>& pointer) {
+    constexpr std::uint32_t header = T::GetHashValue();
+    stream.write(reinterpret_cast<const char*>(&header), sizeof(header));
+    return pointer->WriteParameters(stream);
+  }
+
   }  // namespace Detail
 
   // Initialize the evaluation function parameters
@@ -138,8 +153,8 @@ namespace Eval::NNUE {
     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;
+    if (!Detail::ReadParameters(stream, *feature_transformer)) return false;
+    if (!Detail::ReadParameters(stream, *network)) return false;
     return stream && stream.peek() == std::ios::traits_type::eof();
   }
   // write evaluation function parameters
@@ -162,7 +177,7 @@ namespace Eval::NNUE {
   }
 
   // Load eval, from a file stream or a memory stream
-  bool load_eval(std::string streamName, std::istream& stream) {
+  bool load_eval(std::string name, std::istream& stream) {
 
     Initialize();
 
@@ -171,7 +186,7 @@ namespace Eval::NNUE {
       std::cout << "info string SkipLoadingEval set to true, Net not loaded!" << std::endl;
       return true;
     }
-    fileName = streamName;
+    fileName = name;
     return ReadParameters(stream);
   }
 

src/nnue/evaluate_nnue.h

@@ -40,11 +40,22 @@ namespace Eval::NNUE {
     }
   };
 
+  template <typename T>
+  struct LargePageDeleter {
+    void operator()(T* ptr) const {
+      ptr->~T();
+      aligned_large_pages_free(ptr);
+    }
+  };
+
   template <typename T>
   using AlignedPtr = std::unique_ptr<T, AlignedDeleter<T>>;
 
+  template <typename T>
+  using LargePagePtr = std::unique_ptr<T, LargePageDeleter<T>>;
+
   // Input feature converter
-  extern AlignedPtr<FeatureTransformer> feature_transformer;
+  extern LargePagePtr<FeatureTransformer> feature_transformer;
 
   // Evaluation function
   extern AlignedPtr<Network> network;

src/nnue/nnue_common.h

@@ -113,7 +113,7 @@ namespace Eval::NNUE {
     PS_END2     = 12 * SQUARE_NB + 1
   };
 
-  extern uint32_t kpp_board_index[PIECE_NB][COLOR_NB];
+  extern const uint32_t kpp_board_index[PIECE_NB][COLOR_NB];
 
   // Type of input feature after conversion
   using TransformedFeatureType = std::uint8_t;

src/search.cpp

@@ -217,7 +217,7 @@ void MainThread::search() {
   Time.init(Limits, us, rootPos.game_ply());
   TT.new_search();
 
-  Eval::verify_NNUE();
+  Eval::NNUE::verify();
 
   if (rootMoves.empty())
   {
@@ -454,10 +454,7 @@ void Thread::search() {
                   ++failedHighCnt;
               }
               else
-              {
-                  ++rootMoves[pvIdx].bestMoveCount;
                   break;
-              }
 
               delta += delta / 4 + 5;
 
@@ -1146,7 +1143,7 @@ moves_loop: // When in check, search starts from here
       // Step 16. Reduced depth search (LMR, ~200 Elo). If the move fails high it will be
       // re-searched at full depth.
       if (    depth >= 3
-          &&  moveCount > 1 + 2 * rootNode + 2 * (PvNode && abs(bestValue) < 2)
+          &&  moveCount > 1 + 2 * rootNode
           && (  !captureOrPromotion
               || moveCountPruning
               || ss->staticEval + PieceValue[EG][pos.captured_piece()] <= alpha
@@ -1213,14 +1210,14 @@ moves_loop: // When in check, search starts from here
           }
           else
           {
-            // Increase reduction for captures/promotions if late move and at low depth
-            if (depth < 8 && moveCount > 2)
-                r++;
+              // Increase reduction for captures/promotions if late move and at low depth
+              if (depth < 8 && moveCount > 2)
+                  r++;
 
-            // Unless giving check, this capture is likely bad
-            if (   !givesCheck
-                && ss->staticEval + PieceValue[EG][pos.captured_piece()] + 213 * depth <= alpha)
-                r++;
+              // Unless giving check, this capture is likely bad
+              if (   !givesCheck
+                  && ss->staticEval + PieceValue[EG][pos.captured_piece()] + 213 * depth <= alpha)
+                  r++;
           }
 
           Depth d = std::clamp(newDepth - r, 1, newDepth);
@@ -1567,6 +1564,7 @@ moves_loop: // When in check, search starts from here
                                                                 [pos.moved_piece(move)]
                                                                 [to_sq(move)];
 
+      // CounterMove based pruning
       if (  !captureOrPromotion
           && Search::prune_at_shallow_depth
           && moveCount

src/search.h

@@ -73,7 +73,6 @@ struct RootMove {
   Value previousScore = -VALUE_INFINITE;
   int selDepth = 0;
   int tbRank = 0;
-  int bestMoveCount = 0;
   Value tbScore;
   std::vector<Move> pv;
 };

src/thread.cpp

@@ -236,16 +236,16 @@ Thread* ThreadPool::get_best_thread() const {
         votes[th->rootMoves[0].pv[0]] +=
             (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth);
 
-          if (abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY)
-          {
-              // 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;
+        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;

src/tt.cpp

@@ -67,11 +67,12 @@ void TranspositionTable::resize(size_t mbSize) {
 
   Threads.main()->wait_for_search_finished();
 
-  aligned_ttmem_free(mem);
+  aligned_large_pages_free(table);
 
   clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
-  table = static_cast<Cluster*>(aligned_ttmem_alloc(clusterCount * sizeof(Cluster), mem));
-  if (!mem)
+
+  table = static_cast<Cluster*>(aligned_large_pages_alloc(clusterCount * sizeof(Cluster)));
+  if (!table)
   {
       std::cerr << "Failed to allocate " << mbSize
                 << "MB for transposition table." << std::endl;

src/tt.h

@@ -73,7 +73,7 @@ class TranspositionTable {
   static_assert(sizeof(Cluster) == 32, "Unexpected Cluster size");
 
 public:
- ~TranspositionTable() { aligned_ttmem_free(mem); }
+ ~TranspositionTable() { aligned_large_pages_free(table); }
   void new_search() { generation8 += 8; } // Lower 3 bits are used by PV flag and Bound
   TTEntry* probe(const Key key, bool& found) const;
   int hashfull() const;
@@ -91,7 +91,6 @@ private:
 
   size_t clusterCount;
   Cluster* table;
-  void* mem;
   uint8_t generation8; // Size must be not bigger than TTEntry::genBound8
 };
 

src/uci.cpp

@@ -47,7 +47,7 @@ const char* StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
 void test_cmd(Position& pos, istringstream& is)
 {
     // Initialize as it may be searched.
-    Eval::init_NNUE();
+    Eval::NNUE::init();
 
     std::string param;
     is >> param;
@@ -100,7 +100,7 @@ namespace {
     Position p;
     p.set(pos.fen(), Options["UCI_Chess960"], &states->back(), Threads.main());
 
-    Eval::verify_NNUE();
+    Eval::NNUE::verify();
 
     sync_cout << "\n" << Eval::trace(p) << sync_endl;
   }

src/ucioption.cpp

@@ -41,8 +41,8 @@ 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(size_t(o)); }
 void on_tb_path(const Option& o) { Tablebases::init(o); }
-void on_use_NNUE(const Option& ) { Eval::init_NNUE(); }
-void on_eval_file(const Option& ) { Eval::init_NNUE(); }
+void on_use_NNUE(const Option& ) { Eval::NNUE::init(); }
+void on_eval_file(const Option& ) { Eval::NNUE::init(); }
 void on_prune_at_shallow_depth(const Option& o) {
     Search::prune_at_shallow_depth = o;
 }