New NNUE architecture and net

Introduces a new NNUE network architecture and associated network parameters,
as obtained by a new pytorch trainer.

The network is already very strong at short TC, without regression at longer TC,
and has potential for further improvements.

https://tests.stockfishchess.org/tests/view/60a159c65085663412d0921d
TC: 10s+0.1s, 1 thread
ELO: 21.74 +-3.4 (95%) LOS: 100.0%
Total: 10000 W: 1559 L: 934 D: 7507
Ptnml(0-2): 38, 701, 2972, 1176, 113

https://tests.stockfishchess.org/tests/view/60a187005085663412d0925b
TC: 60s+0.6s, 1 thread
ELO: 5.85 +-1.7 (95%) LOS: 100.0%
Total: 20000 W: 1381 L: 1044 D: 17575
Ptnml(0-2): 27, 885, 7864, 1172, 52

https://tests.stockfishchess.org/tests/view/60a2beede229097940a03806
TC: 20s+0.2s, 8 threads
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 34272 W: 1610 L: 1452 D: 31210
Ptnml(0-2): 30, 1285, 14350, 1439, 32

https://tests.stockfishchess.org/tests/view/60a2d687e229097940a03c72
TC: 60s+0.6s, 8 threads
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 45544 W: 1262 L: 1214 D: 43068
Ptnml(0-2): 12, 1129, 20442, 1177, 12

The network has been trained (by vondele) using the https://github.com/glinscott/nnue-pytorch/ trainer (started by glinscott),
specifically the branch https://github.com/Sopel97/nnue-pytorch/tree/experiment_56.
The data used are in 64 billion positions (193GB total) generated and scored with the current master net
d8: https://drive.google.com/file/d/1hOOYSDKgOOp38ZmD0N4DV82TOLHzjUiF/view?usp=sharing
d9: https://drive.google.com/file/d/1VlhnHL8f-20AXhGkILujnNXHwy9T-MQw/view?usp=sharing
d10: https://drive.google.com/file/d/1ZC5upzBYMmMj1gMYCkt6rCxQG0GnO3Kk/view?usp=sharing
fishtest_d9: https://drive.google.com/file/d/1GQHt0oNgKaHazwJFTRbXhlCN3FbUedFq/view?usp=sharing

This network also contains a few architectural changes with respect to the current master:

    Size changed from 256x2-32-32-1 to 512x2-16-32-1
        ~15-20% slower
        ~2x larger
        adds a special path for 16 valued ClippedReLU
        fixes affine transform code for 16 inputs/outputs, buy using InputDimensions instead of PaddedInputDimensions
            this is safe now because the inputs are processed in groups of 4 in the current affine transform code
    The feature set changed from HalfKP to HalfKAv2
        Includes information about the kings like HalfKA
        Packs king features better, resulting in 8% size reduction compared to HalfKA
    The board is flipped for the black's perspective, instead of rotated like in the current master
    PSQT values for each feature
        the feature transformer now outputs a part that is fowarded directly to the output and allows learning piece values more directly than the previous network architecture. The effect is visible for high imbalance positions, where the current master network outputs evaluations skewed towards zero.
        8 PSQT values per feature, chosen based on (popcount(pos.pieces()) - 1) / 4
        initialized to classical material values on the start of the training
    8 subnetworks (512x2->16->32->1), chosen based on (popcount(pos.pieces()) - 1) / 4
        only one subnetwork is evaluated for any position, no or marginal speed loss

A diagram of the network is available: https://user-images.githubusercontent.com/8037982/118656988-553a1700-b7eb-11eb-82ef-56a11cbebbf2.png
A more complete description: https://github.com/glinscott/nnue-pytorch/blob/master/docs/nnue.md

closes https://github.com/official-stockfish/Stockfish/pull/3474

Bench: 3806488

src/Makefile

@@ -41,7 +41,7 @@ endif
 SRCS = benchmark.cpp bitbase.cpp bitboard.cpp endgame.cpp evaluate.cpp main.cpp \
 	material.cpp misc.cpp movegen.cpp movepick.cpp pawns.cpp position.cpp psqt.cpp \
 	search.cpp thread.cpp timeman.cpp tt.cpp uci.cpp ucioption.cpp tune.cpp syzygy/tbprobe.cpp \
-	nnue/evaluate_nnue.cpp nnue/features/half_kp.cpp
+	nnue/evaluate_nnue.cpp nnue/features/half_ka_v2.cpp
 
 OBJS = $(notdir $(SRCS:.cpp=.o))
 

src/evaluate.cpp

@@ -1116,10 +1116,8 @@ Value Eval::evaluate(const Position& pos) {
       // Scale and shift NNUE for compatibility with search and classical evaluation
       auto  adjusted_NNUE = [&]()
       {
-         int material = pos.non_pawn_material() + 4 * PawnValueMg * pos.count<PAWN>();
+
-         int scale =  580
+         int scale = 903 + 28 * pos.count<PAWN>() + 28 * pos.non_pawn_material() / 1024;
-                    + material / 32
-                    - 4 * pos.rule50_count();
 
          Value nnue = NNUE::evaluate(pos) * scale / 1024 + Time.tempoNNUE;
 
@@ -1134,7 +1132,7 @@ Value Eval::evaluate(const Position& pos) {
       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));
+      bool  classical = largePsq;
 
       // Use classical evaluation for really low piece endgames.
       // One critical case is the draw for bishop + A/H file pawn vs naked king.
@@ -1151,8 +1149,7 @@ Value Eval::evaluate(const Position& pos) {
           && !lowPieceEndgame
           && (   abs(v) * 16 < NNUEThreshold2 * r50
               || (   pos.opposite_bishops()
-                  && abs(v) * 16 < (NNUEThreshold1 + pos.non_pawn_material() / 64) * r50
+                  && abs(v) * 16 < (NNUEThreshold1 + pos.non_pawn_material() / 64) * r50)))
-                  && !(pos.this_thread()->nodes & 0xB))))
           v = adjusted_NNUE();
   }
 

src/evaluate.h

@@ -39,7 +39,7 @@ namespace Eval {
   // 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"
+  #define EvalFileDefaultName   "nn-8a08400ed089.nnue"
 
   namespace NNUE {
 

src/nnue/evaluate_nnue.cpp

@@ -35,7 +35,7 @@ namespace Stockfish::Eval::NNUE {
   LargePagePtr<FeatureTransformer> featureTransformer;
 
   // Evaluation function
-  AlignedPtr<Network> network;
+  AlignedPtr<Network> network[LayerStacks];
 
   // Evaluation function file name
   std::string fileName;
@@ -83,7 +83,8 @@ namespace Stockfish::Eval::NNUE {
   void initialize() {
 
     Detail::initialize(featureTransformer);
-    Detail::initialize(network);
+    for (std::size_t i = 0; i < LayerStacks; ++i)
+      Detail::initialize(network[i]);
   }
 
   // Read network header
@@ -92,7 +93,7 @@ namespace Stockfish::Eval::NNUE {
     std::uint32_t version, size;
 
     version     = read_little_endian<std::uint32_t>(stream);
-    *hashValue = read_little_endian<std::uint32_t>(stream);
+    *hashValue  = read_little_endian<std::uint32_t>(stream);
     size        = read_little_endian<std::uint32_t>(stream);
     if (!stream || version != Version) return false;
     desc->resize(size);
@@ -117,7 +118,8 @@ namespace Stockfish::Eval::NNUE {
     if (!read_header(stream, &hashValue, &netDescription)) return false;
     if (hashValue != HashValue) return false;
     if (!Detail::read_parameters(stream, *featureTransformer)) return false;
-    if (!Detail::read_parameters(stream, *network)) return false;
+    for (std::size_t i = 0; i < LayerStacks; ++i)
+      if (!Detail::read_parameters(stream, *(network[i]))) return false;
     return stream && stream.peek() == std::ios::traits_type::eof();
   }
 
@@ -126,7 +128,8 @@ namespace Stockfish::Eval::NNUE {
 
     if (!write_header(stream, HashValue, netDescription)) return false;
     if (!Detail::write_parameters(stream, *featureTransformer)) return false;
-    if (!Detail::write_parameters(stream, *network)) return false;
+    for (std::size_t i = 0; i < LayerStacks; ++i)
+      if (!Detail::write_parameters(stream, *(network[i]))) return false;
     return (bool)stream;
   }
 
@@ -154,10 +157,15 @@ namespace Stockfish::Eval::NNUE {
     ASSERT_ALIGNED(transformedFeatures, alignment);
     ASSERT_ALIGNED(buffer, alignment);
 
-    featureTransformer->transform(pos, transformedFeatures);
+    const std::size_t bucket = (pos.count<ALL_PIECES>() - 1) / 4;
-    const auto output = network->propagate(transformedFeatures, buffer);
 
-    return static_cast<Value>(output[0] / OutputScale);
+    const auto [psqt, lazy] = featureTransformer->transform(pos, transformedFeatures, bucket);
+    if (lazy) {
+      return static_cast<Value>(psqt / OutputScale);
+    } else {
+      const auto output = network[bucket]->propagate(transformedFeatures, buffer);
+      return static_cast<Value>((output[0] + psqt) / OutputScale);
+    }
   }
 
   // Load eval, from a file stream or a memory stream

src/nnue/features/half_ka_v2.cpp

@@ -16,32 +16,32 @@
   along with this program.  If not, see <http://www.gnu.org/licenses/>.
 */
 
-//Definition of input features HalfKP of NNUE evaluation function
+//Definition of input features HalfKAv2 of NNUE evaluation function
 
-#include "half_kp.h"
+#include "half_ka_v2.h"
 
 #include "../../position.h"
 
 namespace Stockfish::Eval::NNUE::Features {
 
   // Orient a square according to perspective (rotates by 180 for black)
-  inline Square HalfKP::orient(Color perspective, Square s) {
+  inline Square HalfKAv2::orient(Color perspective, Square s) {
-    return Square(int(s) ^ (bool(perspective) * 63));
+    return Square(int(s) ^ (bool(perspective) * 56));
   }
 
   // Index of a feature for a given king position and another piece on some square
-  inline IndexType HalfKP::make_index(Color perspective, Square s, Piece pc, Square ksq) {
+  inline IndexType HalfKAv2::make_index(Color perspective, Square s, Piece pc, Square ksq) {
     return IndexType(orient(perspective, s) + PieceSquareIndex[perspective][pc] + PS_NB * ksq);
   }
 
   // Get a list of indices for active features
-  void HalfKP::append_active_indices(
+  void HalfKAv2::append_active_indices(
     const Position& pos,
     Color perspective,
     ValueListInserter<IndexType> active
   ) {
     Square ksq = orient(perspective, pos.square<KING>(perspective));
-    Bitboard bb = pos.pieces() & ~pos.pieces(KING);
+    Bitboard bb = pos.pieces();
     while (bb)
     {
       Square s = pop_lsb(bb);
@@ -52,7 +52,7 @@ namespace Stockfish::Eval::NNUE::Features {
 
   // append_changed_indices() : get a list of indices for recently changed features
 
-  void HalfKP::append_changed_indices(
+  void HalfKAv2::append_changed_indices(
     Square ksq,
     StateInfo* st,
     Color perspective,
@@ -63,7 +63,6 @@ namespace Stockfish::Eval::NNUE::Features {
     Square oriented_ksq = orient(perspective, ksq);
     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, oriented_ksq));
       if (dp.to[i] != SQ_NONE)
@@ -71,15 +70,15 @@ namespace Stockfish::Eval::NNUE::Features {
     }
   }
 
-  int HalfKP::update_cost(StateInfo* st) {
+  int HalfKAv2::update_cost(StateInfo* st) {
     return st->dirtyPiece.dirty_num;
   }
 
-  int HalfKP::refresh_cost(const Position& pos) {
+  int HalfKAv2::refresh_cost(const Position& pos) {
-    return pos.count<ALL_PIECES>() - 2;
+    return pos.count<ALL_PIECES>();
   }
 
-  bool HalfKP::requires_refresh(StateInfo* st, Color perspective) {
+  bool HalfKAv2::requires_refresh(StateInfo* st, Color perspective) {
     return st->dirtyPiece.piece[0] == make_piece(perspective, KING);
   }
 

src/nnue/features/half_ka_v2.h

@@ -18,8 +18,8 @@
 
 //Definition of input features HalfKP of NNUE evaluation function
 
-#ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED
+#ifndef NNUE_FEATURES_HALF_KA_V2_H_INCLUDED
-#define NNUE_FEATURES_HALF_KP_H_INCLUDED
+#define NNUE_FEATURES_HALF_KA_V2_H_INCLUDED
 
 #include "../nnue_common.h"
 
@@ -32,33 +32,34 @@ namespace Stockfish {
 
 namespace Stockfish::Eval::NNUE::Features {
 
-  // Feature HalfKP: Combination of the position of own king
+  // Feature HalfKAv2: Combination of the position of own king
-  // and the position of pieces other than kings
+  // and the position of pieces
-  class HalfKP {
+  class HalfKAv2 {
 
     // unique number for each piece type on each square
     enum {
       PS_NONE     =  0,
-      PS_W_PAWN   =  1,
+      PS_W_PAWN   =  0,
-      PS_B_PAWN   =  1 * SQUARE_NB + 1,
+      PS_B_PAWN   =  1 * SQUARE_NB,
-      PS_W_KNIGHT =  2 * SQUARE_NB + 1,
+      PS_W_KNIGHT =  2 * SQUARE_NB,
-      PS_B_KNIGHT =  3 * SQUARE_NB + 1,
+      PS_B_KNIGHT =  3 * SQUARE_NB,
-      PS_W_BISHOP =  4 * SQUARE_NB + 1,
+      PS_W_BISHOP =  4 * SQUARE_NB,
-      PS_B_BISHOP =  5 * SQUARE_NB + 1,
+      PS_B_BISHOP =  5 * SQUARE_NB,
-      PS_W_ROOK   =  6 * SQUARE_NB + 1,
+      PS_W_ROOK   =  6 * SQUARE_NB,
-      PS_B_ROOK   =  7 * SQUARE_NB + 1,
+      PS_B_ROOK   =  7 * SQUARE_NB,
-      PS_W_QUEEN  =  8 * SQUARE_NB + 1,
+      PS_W_QUEEN  =  8 * SQUARE_NB,
-      PS_B_QUEEN  =  9 * SQUARE_NB + 1,
+      PS_B_QUEEN  =  9 * SQUARE_NB,
-      PS_NB = 10 * SQUARE_NB + 1
+      PS_KING     =  10 * SQUARE_NB,
+      PS_NB = 11 * SQUARE_NB
     };
 
     static constexpr IndexType PieceSquareIndex[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_NONE, PS_NONE,
+      { PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_KING, PS_NONE,
-        PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_NONE, PS_NONE },
+        PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_KING, PS_NONE },
-      { PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_NONE, PS_NONE,
+      { PS_NONE, PS_B_PAWN, PS_B_KNIGHT, PS_B_BISHOP, PS_B_ROOK, PS_B_QUEEN, PS_KING, PS_NONE,
-        PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_NONE, PS_NONE }
+        PS_NONE, PS_W_PAWN, PS_W_KNIGHT, PS_W_BISHOP, PS_W_ROOK, PS_W_QUEEN, PS_KING, PS_NONE }
     };
 
     // Orient a square according to perspective (rotates by 180 for black)
@@ -69,17 +70,17 @@ namespace Stockfish::Eval::NNUE::Features {
 
    public:
     // Feature name
-    static constexpr const char* Name = "HalfKP(Friend)";
+    static constexpr const char* Name = "HalfKAv2(Friend)";
 
     // Hash value embedded in the evaluation file
-    static constexpr std::uint32_t HashValue = 0x5D69D5B8u;
+    static constexpr std::uint32_t HashValue = 0x5f234cb8u;
 
     // Number of feature dimensions
     static constexpr IndexType Dimensions =
         static_cast<IndexType>(SQUARE_NB) * static_cast<IndexType>(PS_NB);
 
-    // Maximum number of simultaneously active features. 30 because kins are not included.
+    // Maximum number of simultaneously active features.
-    static constexpr IndexType MaxActiveDimensions = 30;
+    static constexpr IndexType MaxActiveDimensions = 32;
 
     // Get a list of indices for active features
     static void append_active_indices(
@@ -107,4 +108,4 @@ namespace Stockfish::Eval::NNUE::Features {
 
 }  // namespace Stockfish::Eval::NNUE::Features
 
-#endif // #ifndef NNUE_FEATURES_HALF_KP_H_INCLUDED
+#endif // #ifndef NNUE_FEATURES_HALF_KA_V2_H_INCLUDED

src/nnue/layers/affine_transform.h

@@ -69,62 +69,15 @@ namespace Stockfish::Eval::NNUE::Layers {
       if (!previousLayer.read_parameters(stream)) return false;
       for (std::size_t i = 0; i < OutputDimensions; ++i)
         biases[i] = read_little_endian<BiasType>(stream);
-#if !defined (USE_SSSE3)
       for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
+#if !defined (USE_SSSE3)
         weights[i] = read_little_endian<WeightType>(stream);
 #else
-      std::unique_ptr<uint32_t[]> indexMap = std::make_unique<uint32_t[]>(OutputDimensions * PaddedInputDimensions);
+        weights[
-      for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) {
-        const uint32_t scrambledIdx =
           (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
           i / PaddedInputDimensions * 4 +
-          i % 4;
+          i % 4
-        weights[scrambledIdx] = read_little_endian<WeightType>(stream);
+        ] = read_little_endian<WeightType>(stream);
-        indexMap[scrambledIdx] = i;
-      }
-
-      // 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 (OutputDimensions > 1 && !stream.fail())
-      {
-          canSaturate16.count = 0;
-#if !defined(USE_VNNI)
-          for (IndexType i = 0; i < PaddedInputDimensions; i += 16)
-              for (IndexType j = 0; j < OutputDimensions; ++j)
-                  for (int x = 0; x < 2; ++x)
-                  {
-                      WeightType* w = &weights[i * OutputDimensions + j * 4 + x * 2];
-                      int sum[2] = {0, 0};
-                      for (int k = 0; k < 8; ++k)
-                      {
-                          IndexType idx = k / 2 * OutputDimensions * 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 * OutputDimensions * 4 + k % 2;
-                                  if (maxW < sign * w[idx])
-                                      maxK = k, maxW = sign * w[idx];
-                              }
-
-                              IndexType idx = maxK / 2 * OutputDimensions * 4 + maxK % 2;
-                              sum[sign == -1] -= w[idx];
-                              const uint32_t scrambledIdx = idx + i * OutputDimensions + j * 4 + x * 2;
-                              canSaturate16.add(j, i + maxK / 2 * 4 + maxK % 2 + x * 2, w[idx], indexMap[scrambledIdx]);
-                              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();
@@ -148,8 +101,6 @@ namespace Stockfish::Eval::NNUE::Layers {
                 i % 4
               ];
       }
-      for (int i = 0; i < canSaturate16.count; ++i)
-          unscrambledWeights[canSaturate16.ids[i].wIdx] = canSaturate16.ids[i].w;
 
       for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
           write_little_endian<WeightType>(stream, unscrambledWeights[i]);
@@ -194,11 +145,11 @@ namespace Stockfish::Eval::NNUE::Layers {
         __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);
+        product0 = _mm512_adds_epi16(product0, product1);
-        product2 = _mm512_add_epi16(product2, product3);
-        product0 = _mm512_add_epi16(product0, product2);
         product0 = _mm512_madd_epi16(product0, Ones512);
-        acc = _mm512_add_epi32(acc, product0);
+        product2 = _mm512_adds_epi16(product2, product3);
+        product2 = _mm512_madd_epi16(product2, Ones512);
+        acc = _mm512_add_epi32(acc, _mm512_add_epi32(product0, product2));
 #endif
       };
 
@@ -236,11 +187,11 @@ namespace Stockfish::Eval::NNUE::Layers {
         __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);
+        product0 = _mm256_adds_epi16(product0, product1);
-        product2 = _mm256_add_epi16(product2, product3);
-        product0 = _mm256_add_epi16(product0, product2);
         product0 = _mm256_madd_epi16(product0, Ones256);
-        acc = _mm256_add_epi32(acc, product0);
+        product2 = _mm256_adds_epi16(product2, product3);
+        product2 = _mm256_madd_epi16(product2, Ones256);
+        acc = _mm256_add_epi32(acc, _mm256_add_epi32(product0, product2));
 #endif
       };
 
@@ -267,11 +218,11 @@ namespace Stockfish::Eval::NNUE::Layers {
         __m128i product1 = _mm_maddubs_epi16(a1, b1);
         __m128i product2 = _mm_maddubs_epi16(a2, b2);
         __m128i product3 = _mm_maddubs_epi16(a3, b3);
-        product0 = _mm_add_epi16(product0, product1);
+        product0 = _mm_adds_epi16(product0, product1);
-        product2 = _mm_add_epi16(product2, product3);
-        product0 = _mm_add_epi16(product0, product2);
         product0 = _mm_madd_epi16(product0, Ones128);
-        acc = _mm_add_epi32(acc, product0);
+        product2 = _mm_adds_epi16(product2, product3);
+        product2 = _mm_madd_epi16(product2, Ones128);
+        acc = _mm_add_epi32(acc, _mm_add_epi32(product0, product2));
       };
 
 #endif
@@ -300,6 +251,8 @@ namespace Stockfish::Eval::NNUE::Layers {
 #endif
 
 #if defined (USE_SSSE3)
+      // Different layout, we process 4 inputs at a time, always.
+      static_assert(InputDimensions % 4 == 0);
 
       const auto output = reinterpret_cast<OutputType*>(buffer);
       const auto inputVector = reinterpret_cast<const vec_t*>(input);
@@ -310,7 +263,7 @@ namespace Stockfish::Eval::NNUE::Layers {
       // because then it is also an input dimension.
       if constexpr (OutputDimensions % OutputSimdWidth == 0)
       {
-          constexpr IndexType NumChunks = PaddedInputDimensions / 4;
+          constexpr IndexType NumChunks = InputDimensions / 4;
 
           const auto input32 = reinterpret_cast<const std::int32_t*>(input);
           vec_t* outptr = reinterpret_cast<vec_t*>(output);
@@ -329,8 +282,6 @@ namespace Stockfish::Eval::NNUE::Layers {
               for (int j = 0; j * OutputSimdWidth < OutputDimensions; ++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 (OutputDimensions == 1)
       {
@@ -377,17 +328,21 @@ namespace Stockfish::Eval::NNUE::Layers {
       auto output = reinterpret_cast<OutputType*>(buffer);
 
 #if defined(USE_SSE2)
-      constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
+      // At least a multiple of 16, with SSE2.
+      static_assert(InputDimensions % SimdWidth == 0);
+      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
       const __m128i Zeros = _mm_setzero_si128();
       const auto inputVector = reinterpret_cast<const __m128i*>(input);
 
 #elif defined(USE_MMX)
-      constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
+      static_assert(InputDimensions % SimdWidth == 0);
+      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
       const __m64 Zeros = _mm_setzero_si64();
       const auto inputVector = reinterpret_cast<const __m64*>(input);
 
 #elif defined(USE_NEON)
-      constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
+      static_assert(InputDimensions % SimdWidth == 0);
+      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
       const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
 #endif
 
@@ -473,25 +428,6 @@ namespace Stockfish::Eval::NNUE::Layers {
 
     alignas(CacheLineSize) BiasType biases[OutputDimensions];
     alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
-#if defined (USE_SSSE3)
-    struct CanSaturate {
-        int count;
-        struct Entry {
-            uint32_t wIdx;
-            uint16_t out;
-            uint16_t in;
-            int8_t w;
-        } ids[PaddedInputDimensions * OutputDimensions * 3 / 4];
-
-        void add(int i, int j, int8_t w, uint32_t wIdx) {
-            ids[count].wIdx = wIdx;
-            ids[count].out = i;
-            ids[count].in = j;
-            ids[count].w = w;
-            ++count;
-        }
-    } canSaturate16;
-#endif
   };
 
 }  // namespace Stockfish::Eval::NNUE::Layers

src/nnue/layers/clipped_relu.h

@@ -72,22 +72,42 @@ namespace Stockfish::Eval::NNUE::Layers {
       const auto output = reinterpret_cast<OutputType*>(buffer);
 
   #if defined(USE_AVX2)
-      constexpr IndexType NumChunks = InputDimensions / SimdWidth;
+      if constexpr (InputDimensions % SimdWidth == 0) {
-      const __m256i Zero = _mm256_setzero_si256();
+        constexpr IndexType NumChunks = InputDimensions / SimdWidth;
-      const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+        const __m256i Zero = _mm256_setzero_si256();
-      const auto in = reinterpret_cast<const __m256i*>(input);
+        const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
-      const auto out = reinterpret_cast<__m256i*>(output);
+        const auto in = reinterpret_cast<const __m256i*>(input);
-      for (IndexType i = 0; i < NumChunks; ++i) {
+        const auto out = reinterpret_cast<__m256i*>(output);
-        const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
+        for (IndexType i = 0; i < NumChunks; ++i) {
-            _mm256_load_si256(&in[i * 4 + 0]),
+          const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
-            _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
+              _mm256_load_si256(&in[i * 4 + 0]),
-        const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
+              _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
-            _mm256_load_si256(&in[i * 4 + 2]),
+          const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
-            _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
+              _mm256_load_si256(&in[i * 4 + 2]),
-        _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
+              _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
-            _mm256_packs_epi16(words0, words1), Zero), Offsets));
+          _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
+              _mm256_packs_epi16(words0, words1), Zero), Offsets));
+        }
+      } else {
+        constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
+        const __m128i Zero = _mm_setzero_si128();
+        const auto in = reinterpret_cast<const __m128i*>(input);
+        const auto out = reinterpret_cast<__m128i*>(output);
+        for (IndexType i = 0; i < NumChunks; ++i) {
+          const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
+              _mm_load_si128(&in[i * 4 + 0]),
+              _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
+          const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
+              _mm_load_si128(&in[i * 4 + 2]),
+              _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
+          const __m128i packedbytes = _mm_packs_epi16(words0, words1);
+          _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
+        }
       }
-      constexpr IndexType Start = NumChunks * SimdWidth;
+      constexpr IndexType Start =
+        InputDimensions % SimdWidth == 0
+        ? InputDimensions / SimdWidth * SimdWidth
+        : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
 
   #elif defined(USE_SSE2)
       constexpr IndexType NumChunks = InputDimensions / SimdWidth;

src/nnue/nnue_accumulator.h

@@ -30,8 +30,8 @@ namespace Stockfish::Eval::NNUE {
 
   // Class that holds the result of affine transformation of input features
   struct alignas(CacheLineSize) Accumulator {
-    std::int16_t
+    std::int16_t accumulation[2][TransformedFeatureDimensions];
-        accumulation[2][TransformedFeatureDimensions];
+    std::int32_t psqtAccumulation[2][PSQTBuckets];
     AccumulatorState state[2];
   };
 

src/nnue/nnue_architecture.h

@@ -23,7 +23,7 @@
 
 #include "nnue_common.h"
 
-#include "features/half_kp.h"
+#include "features/half_ka_v2.h"
 
 #include "layers/input_slice.h"
 #include "layers/affine_transform.h"
@@ -32,16 +32,18 @@
 namespace Stockfish::Eval::NNUE {
 
   // Input features used in evaluation function
-  using FeatureSet = Features::HalfKP;
+  using FeatureSet = Features::HalfKAv2;
 
   // Number of input feature dimensions after conversion
-  constexpr IndexType TransformedFeatureDimensions = 256;
+  constexpr IndexType TransformedFeatureDimensions = 512;
+  constexpr IndexType PSQTBuckets = 8;
+  constexpr IndexType LayerStacks = 8;
 
   namespace Layers {
 
     // Define network structure
     using InputLayer = InputSlice<TransformedFeatureDimensions * 2>;
-    using HiddenLayer1 = ClippedReLU<AffineTransform<InputLayer, 32>>;
+    using HiddenLayer1 = ClippedReLU<AffineTransform<InputLayer, 16>>;
     using HiddenLayer2 = ClippedReLU<AffineTransform<HiddenLayer1, 32>>;
     using OutputLayer = AffineTransform<HiddenLayer2, 1>;
 

src/nnue/nnue_common.h

@@ -46,7 +46,7 @@
 namespace Stockfish::Eval::NNUE {
 
   // Version of the evaluation file
-  constexpr std::uint32_t Version = 0x7AF32F16u;
+  constexpr std::uint32_t Version = 0x7AF32F20u;
 
   // Constant used in evaluation value calculation
   constexpr int OutputScale = 16;

src/nnue/nnue_feature_transformer.h

@@ -35,45 +35,82 @@ namespace Stockfish::Eval::NNUE {
   // vector registers.
   #define VECTOR
 
+  static_assert(PSQTBuckets == 8, "Assumed by the current choice of constants.");
+
   #ifdef USE_AVX512
   typedef __m512i vec_t;
+  typedef __m256i psqt_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)
+  #define vec_load_psqt(a) _mm256_load_si256(a)
+  #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
+  #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
+  #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
+  #define vec_zero_psqt() _mm256_setzero_si256()
   static constexpr IndexType NumRegs = 8; // only 8 are needed
+  static constexpr IndexType NumPsqtRegs = 1;
 
   #elif USE_AVX2
   typedef __m256i vec_t;
+  typedef __m256i psqt_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)
+  #define vec_load_psqt(a) _mm256_load_si256(a)
+  #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
+  #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
+  #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
+  #define vec_zero_psqt() _mm256_setzero_si256()
   static constexpr IndexType NumRegs = 16;
+  static constexpr IndexType NumPsqtRegs = 1;
 
   #elif USE_SSE2
   typedef __m128i vec_t;
+  typedef __m128i psqt_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)
+  #define vec_load_psqt(a) (*(a))
+  #define vec_store_psqt(a,b) *(a)=(b)
+  #define vec_add_psqt_32(a,b) _mm_add_epi32(a,b)
+  #define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b)
+  #define vec_zero_psqt() _mm_setzero_si128()
   static constexpr IndexType NumRegs = Is64Bit ? 16 : 8;
+  static constexpr IndexType NumPsqtRegs = 2;
 
   #elif USE_MMX
   typedef __m64 vec_t;
+  typedef std::int32_t psqt_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)
+  #define vec_load_psqt(a) (*(a))
+  #define vec_store_psqt(a,b) *(a)=(b)
+  #define vec_add_psqt_32(a,b) a+b
+  #define vec_sub_psqt_32(a,b) a-b
+  #define vec_zero_psqt() 0
   static constexpr IndexType NumRegs = 8;
+  static constexpr IndexType NumPsqtRegs = 8;
 
   #elif USE_NEON
   typedef int16x8_t vec_t;
+  typedef int32x4_t psqt_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)
+  #define vec_load_psqt(a) (*(a))
+  #define vec_store_psqt(a,b) *(a)=(b)
+  #define vec_add_psqt_32(a,b) vaddq_s32(a,b)
+  #define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
+  #define vec_zero_psqt() psqt_vec_t{0}
   static constexpr IndexType NumRegs = 16;
+  static constexpr IndexType NumPsqtRegs = 2;
 
   #else
   #undef VECTOR
@@ -87,9 +124,13 @@ namespace Stockfish::Eval::NNUE {
     // Number of output dimensions for one side
     static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
 
+    static constexpr int LazyThreshold = 1400;
+
     #ifdef VECTOR
     static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
+    static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4;
     static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
+    static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets");
     #endif
 
    public:
@@ -115,6 +156,8 @@ namespace Stockfish::Eval::NNUE {
         biases[i] = read_little_endian<BiasType>(stream);
       for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i)
         weights[i] = read_little_endian<WeightType>(stream);
+      for (std::size_t i = 0; i < PSQTBuckets * InputDimensions; ++i)
+        psqtWeights[i] = read_little_endian<PSQTWeightType>(stream);
       return !stream.fail();
     }
 
@@ -128,11 +171,21 @@ namespace Stockfish::Eval::NNUE {
     }
 
     // Convert input features
-    void transform(const Position& pos, OutputType* output) const {
+    std::pair<std::int32_t, bool> transform(const Position& pos, OutputType* output, int bucket) const {
       update_accumulator(pos, WHITE);
       update_accumulator(pos, BLACK);
 
+      const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
       const auto& accumulation = pos.state()->accumulator.accumulation;
+      const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
+
+      const auto psqt = (
+            psqtAccumulation[static_cast<int>(perspectives[0])][bucket]
+          - psqtAccumulation[static_cast<int>(perspectives[1])][bucket]
+        ) / 2;
+
+      if (abs(psqt) > LazyThreshold * OutputScale)
+        return { psqt, true };
 
   #if defined(USE_AVX512)
       constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2);
@@ -163,7 +216,6 @@ namespace Stockfish::Eval::NNUE {
       const int8x8_t Zero = {0};
   #endif
 
-      const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
       for (IndexType p = 0; p < 2; ++p) {
         const IndexType offset = HalfDimensions * p;
 
@@ -240,6 +292,8 @@ namespace Stockfish::Eval::NNUE {
   #if defined(USE_MMX)
       _mm_empty();
   #endif
+
+      return { psqt, false };
     }
 
    private:
@@ -255,6 +309,7 @@ namespace Stockfish::Eval::NNUE {
       // 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
 
       // Look for a usable accumulator of an earlier position. We keep track
@@ -333,12 +388,52 @@ namespace Stockfish::Eval::NNUE {
           }
         }
 
+        for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
+        {
+          // Load accumulator
+          auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
+            &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+            psqt[k] = vec_load_psqt(&accTilePsqt[k]);
+
+          for (IndexType i = 0; states_to_update[i]; ++i)
+          {
+            // Difference calculation for the deactivated features
+            for (const auto index : removed[i])
+            {
+              const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
+              auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
+              for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+                psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
+            }
+
+            // Difference calculation for the activated features
+            for (const auto index : added[i])
+            {
+              const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
+              auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
+              for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+                psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
+            }
+
+            // Store accumulator
+            accTilePsqt = reinterpret_cast<psqt_vec_t*>(
+              &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+            for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+              vec_store_psqt(&accTilePsqt[k], psqt[k]);
+          }
+        }
+
   #else
         for (IndexType i = 0; states_to_update[i]; ++i)
         {
           std::memcpy(states_to_update[i]->accumulator.accumulation[perspective],
               st->accumulator.accumulation[perspective],
               HalfDimensions * sizeof(BiasType));
+
+          for (std::size_t k = 0; k < PSQTBuckets; ++k)
+            states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k];
+
           st = states_to_update[i];
 
           // Difference calculation for the deactivated features
@@ -348,6 +443,9 @@ namespace Stockfish::Eval::NNUE {
 
             for (IndexType j = 0; j < HalfDimensions; ++j)
               st->accumulator.accumulation[perspective][j] -= weights[offset + j];
+
+            for (std::size_t k = 0; k < PSQTBuckets; ++k)
+              st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k];
           }
 
           // Difference calculation for the activated features
@@ -357,6 +455,9 @@ namespace Stockfish::Eval::NNUE {
 
             for (IndexType j = 0; j < HalfDimensions; ++j)
               st->accumulator.accumulation[perspective][j] += weights[offset + j];
+
+            for (std::size_t k = 0; k < PSQTBuckets; ++k)
+              st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
           }
         }
   #endif
@@ -392,16 +493,42 @@ namespace Stockfish::Eval::NNUE {
             vec_store(&accTile[k], acc[k]);
         }
 
+        for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
+        {
+          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+            psqt[k] = vec_zero_psqt();
+
+          for (const auto index : active)
+          {
+            const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
+            auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
+
+            for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+              psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
+          }
+
+          auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
+            &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
+          for (std::size_t k = 0; k < NumPsqtRegs; ++k)
+            vec_store_psqt(&accTilePsqt[k], psqt[k]);
+        }
+
   #else
         std::memcpy(accumulator.accumulation[perspective], biases,
             HalfDimensions * sizeof(BiasType));
 
+        for (std::size_t k = 0; k < PSQTBuckets; ++k)
+          accumulator.psqtAccumulation[perspective][k] = 0;
+
         for (const auto index : active)
         {
           const IndexType offset = HalfDimensions * index;
 
           for (IndexType j = 0; j < HalfDimensions; ++j)
             accumulator.accumulation[perspective][j] += weights[offset + j];
+
+          for (std::size_t k = 0; k < PSQTBuckets; ++k)
+            accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
         }
   #endif
       }
@@ -413,9 +540,11 @@ namespace Stockfish::Eval::NNUE {
 
     using BiasType = std::int16_t;
     using WeightType = std::int16_t;
+    using PSQTWeightType = std::int32_t;
 
     alignas(CacheLineSize) BiasType biases[HalfDimensions];
     alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
+    alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
   };
 
 }  // namespace Stockfish::Eval::NNUE

src/search.cpp

@@ -66,7 +66,7 @@ namespace {
 
   // Futility margin
   Value futility_margin(Depth d, bool improving) {
-    return Value(234 * (d - improving));
+    return Value(231 * (d - improving));
   }
 
   // Reductions lookup table, initialized at startup
@@ -801,7 +801,7 @@ namespace {
         && (ss-1)->statScore < 24185
         &&  eval >= beta
         &&  eval >= ss->staticEval
-        &&  ss->staticEval >= beta - 24 * depth - 34 * improving + 162 * ss->ttPv + 159
+        &&  ss->staticEval >= beta - 22 * depth - 34 * improving + 162 * ss->ttPv + 159
         && !excludedMove
         &&  pos.non_pawn_material(us)
         && (ss->ply >= thisThread->nmpMinPly || us != thisThread->nmpColor))
@@ -1172,7 +1172,7 @@ moves_loop: // When in check, search starts from here
                              + (*contHist[0])[movedPiece][to_sq(move)]
                              + (*contHist[1])[movedPiece][to_sq(move)]
                              + (*contHist[3])[movedPiece][to_sq(move)]
-                             - 4741;
+                             - 4791;
 
               // Decrease/increase reduction for moves with a good/bad history (~30 Elo)
               if (!ss->inCheck)