Cleanup and optimize SSE/AVX code

[stockfish] / src / nnue / nnue_feature_transformer.h

diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h
index cbcc26f3efae9f592eead48230d153c93ddd1301..3818e444b6af9710110dff8eba49b4148d55b53b 100644 (file)
--- a/src/nnue/nnue_feature_transformer.h
+++ b/src/nnue/nnue_feature_transformer.h
@@ -169,38 +169,41 @@ namespace Eval::NNUE {
                    kHalfDimensions * sizeof(BiasType));
         for (const auto index : active_indices[perspective]) {
           const IndexType offset = kHalfDimensions * index;
+  #if defined(USE_AVX512)
+          auto accumulation = reinterpret_cast<__m512i*>(
+              &accumulator.accumulation[perspective][i][0]);
+          auto column = reinterpret_cast<const __m512i*>(&weights_[offset]);
+          constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
+          for (IndexType j = 0; j < kNumChunks; ++j)
+            _mm512_storeA_si512(&accumulation[j], _mm512_add_epi16(_mm512_loadA_si512(&accumulation[j]), column[j]));
 
-  #if defined(USE_AVX2)
+  #elif defined(USE_AVX2)
           auto accumulation = reinterpret_cast<__m256i*>(
               &accumulator.accumulation[perspective][i][0]);
           auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
-          for (IndexType j = 0; j < kNumChunks; ++j) {
+          for (IndexType j = 0; j < kNumChunks; ++j)
             _mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j]));
-          }
 
   #elif defined(USE_SSE2)
           auto accumulation = reinterpret_cast<__m128i*>(
               &accumulator.accumulation[perspective][i][0]);
           auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
-          for (IndexType j = 0; j < kNumChunks; ++j) {
+          for (IndexType j = 0; j < kNumChunks; ++j)
             accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
-          }
 
   #elif defined(USE_NEON)
           auto accumulation = reinterpret_cast<int16x8_t*>(
               &accumulator.accumulation[perspective][i][0]);
           auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
-          for (IndexType j = 0; j < kNumChunks; ++j) {
+          for (IndexType j = 0; j < kNumChunks; ++j)
             accumulation[j] = vaddq_s16(accumulation[j], column[j]);
-          }
 
   #else
-          for (IndexType j = 0; j < kHalfDimensions; ++j) {
+          for (IndexType j = 0; j < kHalfDimensions; ++j)
             accumulator.accumulation[perspective][i][j] += weights_[offset + j];
-          }
   #endif
 
         }