Revert "Avoid special casing for MinGW"

[stockfish] / src / nnue / layers / affine_transform.h

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index ecc3008a6cc8a728927011ba4b6154a232e76994..b585bc87819d23c808ce66a472c4ffba59e47072 100644 (file)
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -104,8 +104,13 @@ namespace Eval::NNUE::Layers {
         __m512i sum = _mm512_setzero_si512();
         const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
         for (IndexType j = 0; j < kNumChunks; ++j) {
-            __m512i product = _mm512_maddubs_epi16(
-              _mm512_load_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            __m512i product = _mm512_maddubs_epi16(_mm512_loadu_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #else
+            __m512i product = _mm512_maddubs_epi16(_mm512_load_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #endif
+
             product = _mm512_madd_epi16(product, kOnes);
             sum = _mm512_add_epi32(sum, product);
         }
@@ -120,8 +125,12 @@ namespace Eval::NNUE::Layers {
             const auto row_256 = reinterpret_cast<const __m256i*>(&weights_[offset]);
             int j = kNumChunks * 2;
 
-            __m256i sum256 = _mm256_maddubs_epi16(
-              _mm256_load_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+  #if defined(__MINGW32__) || defined(__MINGW64__)  // See HACK comment below in AVX2.
+            __m256i sum256 = _mm256_maddubs_epi16(_mm256_loadu_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+  #else
+            __m256i sum256 = _mm256_maddubs_epi16(_mm256_load_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
+  #endif
+
             sum256 = _mm256_madd_epi16(sum256, _mm256_set1_epi16(1));
             sum256 = _mm256_hadd_epi32(sum256, sum256);
             sum256 = _mm256_hadd_epi32(sum256, sum256);
@@ -135,7 +144,17 @@ namespace Eval::NNUE::Layers {
         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
         for (IndexType j = 0; j < kNumChunks; ++j) {
           __m256i product = _mm256_maddubs_epi16(
-            _mm256_load_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
+
+  #if defined(__MINGW32__) || defined(__MINGW64__)
+            // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
+            //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
+            //       even though alignas is specified.
+            _mm256_loadu_si256
+  #else
+            _mm256_load_si256
+  #endif
+
+            (&input_vector[j]), _mm256_load_si256(&row[j]));
           product = _mm256_madd_epi16(product, kOnes);
           sum = _mm256_add_epi32(sum, product);
         }