Provide vectorized NNUE code for SSE2 and MMX targets

[stockfish] / src / nnue / nnue_common.h

diff --git a/src/nnue/nnue_common.h b/src/nnue/nnue_common.h
index 972ef3e50c66291cb50039741e5a24f0c91ba6a1..cb1251c58aa37923a23e87fb4e6dbaa668d24b8d 100644 (file)
--- a/src/nnue/nnue_common.h
+++ b/src/nnue/nnue_common.h
@@ -33,10 +33,36 @@
 #elif defined(USE_SSE2)
 #include <emmintrin.h>
 
+#elif defined(USE_MMX)
+#include <mmintrin.h>
+
 #elif defined(USE_NEON)
 #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)
+#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)
+#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 {
 
   // Version of the evaluation file
@@ -56,6 +82,9 @@ namespace Eval::NNUE {
   #elif defined(USE_SSE2)
   constexpr std::size_t kSimdWidth = 16;
 
+  #elif defined(USE_MMX)
+  constexpr std::size_t kSimdWidth = 8;
+
   #elif defined(USE_NEON)
   constexpr std::size_t kSimdWidth = 16;
   #endif