Introduce a function to compute NNUE accumulator

This patch introduces `hint_common_parent_position()` to signal that potentially several child nodes will require an NNUE eval. By populating explicitly the accumulator, these subsequent evaluations can be performed more efficiently.

This was based on the observation that calculating the evaluation in an excluded move position yielded a significant Elo gain, even though the evaluation itself was already available (work by pb00067).

Sopel wrote the code to perform just the accumulator update. This PR is based on cleaned up code that

passed STC:
https://tests.stockfishchess.org/tests/view/63f62f9be74a12625bcd4aa0
 LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 110368 W: 29607 L: 29167 D: 51594
Ptnml(0-2): 41, 10551, 33572, 10967, 53

and in an the earlier (equivalent) version

passed STC:
https://tests.stockfishchess.org/tests/view/63f3c3fee74a12625bcce2a6
LLR: 2.95 (-2.94,2.94) <0.00,2.00>
Total: 47552 W: 12786 L: 12467 D: 22299
Ptnml(0-2): 120, 5107, 12997, 5438, 114

passed LTC:
https://tests.stockfishchess.org/tests/view/63f45cc2e74a12625bccfa63
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 110368 W: 29607 L: 29167 D: 51594
Ptnml(0-2): 41, 10551, 33572, 10967, 53

closes https://github.com/official-stockfish/Stockfish/pull/4402

Bench: 3726250

src/evaluate.h

@@ -45,6 +45,7 @@ namespace Eval {
 
     std::string trace(Position& pos);
     Value evaluate(const Position& pos, bool adjusted = false, int* complexity = nullptr);
+    void hint_common_parent_position(const Position& pos);
 
     void init();
     void verify();

src/nnue/evaluate_nnue.cpp

@@ -136,6 +136,11 @@ namespace Stockfish::Eval::NNUE {
     return (bool)stream;
   }
 
+  void hint_common_parent_position(const Position& pos) {
+    if (Eval::useNNUE)
+        featureTransformer->hint_common_access(pos);
+  }
+
   // Evaluation function. Perform differential calculation.
   Value evaluate(const Position& pos, bool adjusted, int* complexity) {
 

src/nnue/evaluate_nnue.h

@@ -31,6 +31,7 @@ namespace Stockfish::Eval::NNUE {
   constexpr std::uint32_t HashValue =
       FeatureTransformer::get_hash_value() ^ Network::get_hash_value();
 
+
   // Deleter for automating release of memory area
   template <typename T>
   struct AlignedDeleter {

src/nnue/nnue_feature_transformer.h

@@ -25,6 +25,7 @@
 #include "nnue_architecture.h"
 
 #include <cstring> // std::memset()
+#include <utility> // std::pair
 
 namespace Stockfish::Eval::NNUE {
 
@@ -332,27 +333,16 @@ namespace Stockfish::Eval::NNUE {
 #endif
 
       return psqt;
-
     } // end of function transform()
 
-
+    void hint_common_access(const Position& pos) const {
+      hint_common_access_for_perspective<WHITE>(pos);
+      hint_common_access_for_perspective<BLACK>(pos);
+    }
 
    private:
     template<Color Perspective>
-    void update_accumulator(const Position& pos) const {
-
-      // The size must be enough to contain the largest possible update.
-      // That might depend on the feature set and generally relies on the
-      // feature set's update cost calculation to be correct and never
-      // allow updates with more added/removed features than MaxActiveDimensions.
-
-  #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[NumRegs];
-      psqt_vec_t psqt[NumPsqtRegs];
-  #endif
-
+    [[nodiscard]] std::pair<StateInfo*, StateInfo*> try_find_computed_accumulator(const Position& pos) const {
       // 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;
@@ -367,31 +357,61 @@ namespace Stockfish::Eval::NNUE {
         next = st;
         st = st->previous;
       }
+      return { st, next };
+    }
 
-      if (st->accumulator.computed[Perspective])
-      {
-        if (next == nullptr)
+    // NOTE: The parameter states_to_update is an array of position states, ending with nullptr.
+    //       All states must be sequential, that is states_to_update[i] must either be reachable
+    //       by repeatedly applying ->previous from states_to_update[i+1] or states_to_update[i] == nullptr.
+    //       computed_st must be reachable by repeatadly applying ->previous on states_to_update[0], if not nullptr.
+    template<Color Perspective, size_t N>
+    void update_accumulator_incremetal(const Position& pos, StateInfo* computed_st, StateInfo* states_to_update[N]) const {
+      static_assert(N > 0);
+      assert(states_to_update[N-1] == nullptr);
+
+  #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[NumRegs];
+      psqt_vec_t psqt[NumPsqtRegs];
+  #endif
+
+      if (states_to_update[0] == nullptr)
         return;
 
-        // Update incrementally in two steps. First, we update the "next"
-        // accumulator. Then, we update the current accumulator (pos.state()).
+      // Update incrementally going back through states_to_update.
 
       // Gather all features to be updated.
       const Square ksq = pos.square<KING>(Perspective);
-        FeatureSet::IndexList removed[2], added[2];
-        FeatureSet::append_changed_indices<Perspective>(
-          ksq, next->dirtyPiece, removed[0], added[0]);
-        for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
-          FeatureSet::append_changed_indices<Perspective>(
-            ksq, st2->dirtyPiece, removed[1], added[1]);
 
-        // Mark the accumulators as computed.
-        next->accumulator.computed[Perspective] = true;
-        pos.state()->accumulator.computed[Perspective] = true;
+      // The size must be enough to contain the largest possible update.
+      // That might depend on the feature set and generally relies on the
+      // feature set's update cost calculation to be correct and never
+      // allow updates with more added/removed features than MaxActiveDimensions.
+      FeatureSet::IndexList removed[N-1], added[N-1];
+
+      {
+        int i = N-2; // last potential state to update. Skip last element because it must be nullptr.
+        while (states_to_update[i] == nullptr)
+          --i;
+
+        StateInfo *st2 = states_to_update[i];
+
+        for (; i >= 0; --i)
+        {
+          states_to_update[i]->accumulator.computed[Perspective] = true;
+
+          StateInfo* end_state = i == 0 ? computed_st : states_to_update[i - 1];
+
+          for (; st2 != end_state; st2 = st2->previous)
+            FeatureSet::append_changed_indices<Perspective>(
+              ksq, st2->dirtyPiece, removed[i], added[i]);
+        }
+      }
+
+      StateInfo* st = computed_st;
 
       // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
-        StateInfo *states_to_update[3] =
-          { next, next == pos.state() ? nullptr : pos.state(), nullptr };
 #ifdef VECTOR
       for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
       {
@@ -502,10 +522,24 @@ namespace Stockfish::Eval::NNUE {
         }
       }
 #endif
+
+  #if defined(USE_MMX)
+      _mm_empty();
+  #endif
     }
-      else
-      {
+
+    template<Color Perspective>
+    void update_accumulator_refresh(const Position& pos) 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[NumRegs];
+      psqt_vec_t psqt[NumPsqtRegs];
+  #endif
+
       // Refresh the accumulator
+      // Could be extracted to a separate function because it's done in 2 places,
+      // but it's unclear if compilers would correctly handle register allocation.
       auto& accumulator = pos.state()->accumulator;
       accumulator.computed[Perspective] = true;
       FeatureSet::IndexList active;
@@ -572,13 +606,64 @@ namespace Stockfish::Eval::NNUE {
           accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
       }
 #endif
-      }
 
   #if defined(USE_MMX)
       _mm_empty();
   #endif
     }
 
+    template<Color Perspective>
+    void hint_common_access_for_perspective(const Position& pos) const {
+
+      // Works like update_accumulator, but performs less work.
+      // Updates ONLY the accumulator for pos.
+
+      // Look for a usable accumulator of an earlier position. We keep track
+      // of the estimated gain in terms of features to be added/subtracted.
+      // Fast early exit.
+      if (pos.state()->accumulator.computed[Perspective])
+        return;
+
+      auto [oldest_st, _] = try_find_computed_accumulator<Perspective>(pos);
+
+      if (oldest_st->accumulator.computed[Perspective])
+      {
+        // Only update current position accumulator to minimize work.
+        StateInfo* states_to_update[2] = { pos.state(), nullptr };
+        update_accumulator_incremetal<Perspective, 2>(pos, oldest_st, states_to_update);
+      }
+      else
+      {
+        update_accumulator_refresh<Perspective>(pos);
+      }
+    }
+
+    template<Color Perspective>
+    void update_accumulator(const Position& pos) const {
+
+      auto [oldest_st, next] = try_find_computed_accumulator<Perspective>(pos);
+
+      if (oldest_st->accumulator.computed[Perspective])
+      {
+        if (next == nullptr)
+          return;
+
+        // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
+        // Currently we update 2 accumulators.
+        //     1. for the current position
+        //     2. the next accumulator after the computed one
+        // The heuristic may change in the future.
+        StateInfo *states_to_update[3] =
+          { next, next == pos.state() ? nullptr : pos.state(), nullptr };
+
+        update_accumulator_incremetal<Perspective, 3>(pos, oldest_st, states_to_update);
+      }
+      else
+      {
+        update_accumulator_refresh<Perspective>(pos);
+      }
+    }
+
     alignas(CacheLineSize) BiasType biases[HalfDimensions];
     alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
     alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];

src/search.cpp

@@ -730,10 +730,10 @@ namespace {
         goto moves_loop;
     }
     else if (excludedMove) {
-        // excludeMove implies that we had a ttHit on the containing non-excluded search with ss->staticEval filled from TT
-        // However static evals from the TT aren't good enough (-13 elo), presumably due to changing optimism context
-        // Recalculate value with current optimism (without updating thread avgComplexity)
-        ss->staticEval = eval = evaluate(pos, &complexity);
+        // Providing the hint that this node's accumulator will be used often brings significant Elo gain (13 elo)
+        Eval::NNUE::hint_common_parent_position(pos);
+        eval = ss->staticEval;
+        complexity = abs(ss->staticEval - pos.psq_eg_stm());
     }
     else if (ss->ttHit)
     {