Add large page support for NNUE weights and simplify TT mem management

Use TT memory functions to allocate memory for the NNUE weights. This
should provide a small speed-up on systems where large pages are not
automatically used, including Windows and some Linux distributions.

Further, since we now have a wrapper for std::aligned_alloc(), we can
simplify the TT memory management a bit:

- We no longer need to store separate pointers to the hash table and
  its underlying memory allocation.
- We also get to merge the Linux-specific and default implementations
  of aligned_ttmem_alloc().

Finally, we'll enable the VirtualAlloc code path with large page
support also for Win32.

STC: https://tests.stockfishchess.org/tests/view/5f66595823a84a47b9036fba
LLR: 2.94 (-2.94,2.94) {-0.25,1.25}
Total: 14896 W: 1854 L: 1686 D: 11356
Ptnml(0-2): 65, 1224, 4742, 1312, 105

closes https://github.com/official-stockfish/Stockfish/pull/3081

No functional change.

README.md

@@ -152,7 +152,7 @@ to find the best move. The classical evaluation computes this value as a functio
 of various chess concepts, handcrafted by experts, tested and tuned using fishtest.
 The NNUE evaluation computes this value with a neural network based on basic
 inputs (e.g. piece positions only). The network is optimized and trained
-on the evalutions of millions of positions at moderate search depth.
+on the evaluations of millions of positions at moderate search depth.
 
 The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward.
 It can be evaluated efficiently on CPUs, and exploits the fact that only parts

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

@@ -33,8 +33,8 @@ 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

@@ -52,7 +52,7 @@ namespace Eval::NNUE {
   };
 
   // Input feature converter
-  AlignedPtr<FeatureTransformer> feature_transformer;
+  LargePagePtr<FeatureTransformer> feature_transformer;
 
   // Evaluation function
   AlignedPtr<Network> network;
@@ -70,14 +70,22 @@ 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
@@ -110,8 +118,8 @@ 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();
   }
 

src/nnue/evaluate_nnue.h

@@ -40,9 +40,20 @@ namespace Eval::NNUE {
     }
   };
 
+  template <typename T>
+  struct TtmemDeleter {
+    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, TtmemDeleter<T>>;
+
 }  // namespace Eval::NNUE
 
 #endif // #ifndef NNUE_EVALUATE_NNUE_H_INCLUDED

src/tt.cpp

@@ -62,11 +62,12 @@ 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;
@@ -89,7 +89,6 @@ private:
 
   size_t clusterCount;
   Cluster* table;
-  void* mem;
   uint8_t generation8; // Size must be not bigger than TTEntry::genBound8
 };