X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fnnue%2Fevaluate_nnue.cpp;h=b0ed7d2f5a4f78e0092603edf4756402c69de2ae;hb=3f6451eff7c62e8d4a33c5b11f055a81b3da8387;hp=d6ac9894cbbd4203303478c554a0cc5c1ce1ba78;hpb=406979ea12ee7828e079871b0f9f3dc8f127a741;p=stockfish

diff --git a/src/nnue/evaluate_nnue.cpp b/src/nnue/evaluate_nnue.cpp
index d6ac9894..b0ed7d2f 100644
--- a/src/nnue/evaluate_nnue.cpp
+++ b/src/nnue/evaluate_nnue.cpp
@@ -25,12 +25,13 @@
 #include "../position.h"
 #include "../misc.h"
 #include "../uci.h"
+#include "../types.h"
 
 #include "evaluate_nnue.h"
 
 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 +53,7 @@ namespace Eval::NNUE {
   };
 
   // Input feature converter
-  AlignedPtr<FeatureTransformer> feature_transformer;
+  LargePagePtr<FeatureTransformer> feature_transformer;
 
   // Evaluation function
   AlignedPtr<Network> network;
@@ -70,14 +71,22 @@ 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);
   }
 
   }  // namespace Detail
@@ -110,59 +119,47 @@ 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();
   }
 
-  // Proceed with the difference calculation if possible
-  static void UpdateAccumulatorIfPossible(const Position& pos) {
+  // Evaluation function. Perform differential calculation.
+  Value evaluate(const Position& pos) {
 
-    feature_transformer->UpdateAccumulatorIfPossible(pos);
-  }
+    // We manually align the arrays on the stack because with gcc < 9.3
+    // overaligning stack variables with alignas() doesn't work correctly.
 
-  // Calculate the evaluation value
-  static Value ComputeScore(const Position& pos, bool refresh) {
+    constexpr uint64_t alignment = kCacheLineSize;
 
-    auto& accumulator = pos.state()->accumulator;
-    if (!refresh && accumulator.computed_score) {
-      return accumulator.score;
-    }
+#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
+    TransformedFeatureType transformed_features_unaligned[
+      FeatureTransformer::kBufferSize + alignment / sizeof(TransformedFeatureType)];
+    char buffer_unaligned[Network::kBufferSize + alignment];
 
-    alignas(kCacheLineSize) TransformedFeatureType
-        transformed_features[FeatureTransformer::kBufferSize];
-    feature_transformer->Transform(pos, transformed_features, refresh);
-    alignas(kCacheLineSize) char buffer[Network::kBufferSize];
-    const auto output = network->Propagate(transformed_features, buffer);
+    auto* transformed_features = align_ptr_up<alignment>(&transformed_features_unaligned[0]);
+    auto* buffer = align_ptr_up<alignment>(&buffer_unaligned[0]);
+#else
+    alignas(alignment)
+      TransformedFeatureType transformed_features[FeatureTransformer::kBufferSize];
+    alignas(alignment) char buffer[Network::kBufferSize];
+#endif
 
-    auto score = static_cast<Value>(output[0] / FV_SCALE);
+    ASSERT_ALIGNED(transformed_features, alignment);
+    ASSERT_ALIGNED(buffer, alignment);
 
-    accumulator.score = score;
-    accumulator.computed_score = true;
-    return accumulator.score;
+    feature_transformer->Transform(pos, transformed_features);
+    const auto output = network->Propagate(transformed_features, buffer);
+
+    return static_cast<Value>(output[0] / FV_SCALE);
   }
 
   // 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();
-    fileName = streamName;
+    fileName = name;
     return ReadParameters(stream);
   }
 
-  // Evaluation function. Perform differential calculation.
-  Value evaluate(const Position& pos) {
-    return ComputeScore(pos, false);
-  }
-
-  // Evaluation function. Perform full calculation.
-  Value compute_eval(const Position& pos) {
-    return ComputeScore(pos, true);
-  }
-
-  // Proceed with the difference calculation if possible
-  void update_eval(const Position& pos) {
-    UpdateAccumulatorIfPossible(pos);
-  }
-
 } // namespace Eval::NNUE