New NNUE architecture and net

[stockfish] / src / nnue / nnue_common.h

diff --git a/src/nnue/nnue_common.h b/src/nnue/nnue_common.h
index a9d8e5af98d1f0756f517a969d6d17a8add32fb5..dc70006120dca4bef8502d7e6428c63425f106ce 100644 (file)
--- a/src/nnue/nnue_common.h
+++ b/src/nnue/nnue_common.h
@@ -1,6 +1,6 @@
 /*
   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file)
 
   Stockfish is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
@@ -43,56 +43,33 @@
 #include <arm_neon.h>
 #endif
 
-// HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Otherwise a binary
-//       compiled with older g++ crashes because the output memory is not aligned
-//       even though alignas is specified.
-#if defined(USE_AVX2)
-#if defined(__GNUC__ ) && (__GNUC__ < 9) && defined(_WIN32) && !defined(__clang__)
-#define _mm256_loadA_si256  _mm256_loadu_si256
-#define _mm256_storeA_si256 _mm256_storeu_si256
-#else
-#define _mm256_loadA_si256  _mm256_load_si256
-#define _mm256_storeA_si256 _mm256_store_si256
-#endif
-#endif
-
-#if defined(USE_AVX512)
-#if defined(__GNUC__ ) && (__GNUC__ < 9) && defined(_WIN32) && !defined(__clang__)
-#define _mm512_loadA_si512   _mm512_loadu_si512
-#define _mm512_storeA_si512  _mm512_storeu_si512
-#else
-#define _mm512_loadA_si512   _mm512_load_si512
-#define _mm512_storeA_si512  _mm512_store_si512
-#endif
-#endif
-
-namespace Eval::NNUE {
+namespace Stockfish::Eval::NNUE {
 
   // Version of the evaluation file
-  constexpr std::uint32_t kVersion = 0x7AF32F16u;
+  constexpr std::uint32_t Version = 0x7AF32F20u;
 
   // Constant used in evaluation value calculation
-  constexpr int FV_SCALE = 16;
-  constexpr int kWeightScaleBits = 6;
+  constexpr int OutputScale = 16;
+  constexpr int WeightScaleBits = 6;
 
   // Size of cache line (in bytes)
-  constexpr std::size_t kCacheLineSize = 64;
+  constexpr std::size_t CacheLineSize = 64;
 
   // SIMD width (in bytes)
   #if defined(USE_AVX2)
-  constexpr std::size_t kSimdWidth = 32;
+  constexpr std::size_t SimdWidth = 32;
 
   #elif defined(USE_SSE2)
-  constexpr std::size_t kSimdWidth = 16;
+  constexpr std::size_t SimdWidth = 16;
 
   #elif defined(USE_MMX)
-  constexpr std::size_t kSimdWidth = 8;
+  constexpr std::size_t SimdWidth = 8;
 
   #elif defined(USE_NEON)
-  constexpr std::size_t kSimdWidth = 16;
+  constexpr std::size_t SimdWidth = 16;
   #endif
 
-  constexpr std::size_t kMaxSimdWidth = 32;
+  constexpr std::size_t MaxSimdWidth = 32;
 
   // Type of input feature after conversion
   using TransformedFeatureType = std::uint8_t;
@@ -100,7 +77,7 @@ namespace Eval::NNUE {
 
   // Round n up to be a multiple of base
   template <typename IntType>
-  constexpr IntType CeilToMultiple(IntType n, IntType base) {
+  constexpr IntType ceil_to_multiple(IntType n, IntType base) {
       return (n + base - 1) / base * base;
   }
 
@@ -122,6 +99,24 @@ namespace Eval::NNUE {
       return result;
   }
 
-}  // namespace Eval::NNUE
+  template <typename IntType>
+  inline void write_little_endian(std::ostream& stream, IntType value) {
+
+      std::uint8_t u[sizeof(IntType)];
+      typename std::make_unsigned<IntType>::type v = value;
+
+      std::size_t i = 0;
+      // if constexpr to silence the warning about shift by 8
+      if constexpr (sizeof(IntType) > 1) {
+        for (; i + 1 < sizeof(IntType); ++i) {
+            u[i] = v;
+            v >>= 8;
+        }
+      }
+      u[i] = v;
+
+      stream.write(reinterpret_cast<char*>(u), sizeof(IntType));
+  }
+}  // namespace Stockfish::Eval::NNUE
 
 #endif // #ifndef NNUE_COMMON_H_INCLUDED