Cleanup and optimize SSE/AVX code

AVX512 +4% faster
AVX2 +1% faster
SSSE3 +5% faster

passed non-regression STC:
STC https://tests.stockfishchess.org/tests/view/5f31249f90816720665374f6
LLR: 2.96 (-2.94,2.94) {-1.50,0.50}
Total: 17576 W: 2344 L: 2245 D: 12987
Ptnml(0-2): 127, 1570, 5292, 1675, 124

closes https://github.com/official-stockfish/Stockfish/pull/2962

No functional change

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 20ec2f1234befe5a67f6ab2d3c70c42fabac08c4..89cfaad7dbd8ff847bbccf8bd451c08927860292 100644 (file)
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -108,24 +108,19 @@ namespace Eval::NNUE::Layers {
             product = _mm512_madd_epi16(product, kOnes);
             sum = _mm512_add_epi32(sum, product);
         }
-        output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
 
         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
         // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit)
         // and we have to do one more 256bit chunk.
         if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2)
         {
-            const auto iv_256  = reinterpret_cast<const __m256i*>(input);
-            const auto row_256 = reinterpret_cast<const __m256i*>(&weights_[offset]);
-            int j = kNumChunks * 2;
-            __m256i sum256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv_256[j]), _mm256_load_si256(&row_256[j]));
-            sum256 = _mm256_madd_epi16(sum256, _mm256_set1_epi16(1));
-            sum256 = _mm256_hadd_epi32(sum256, sum256);
-            sum256 = _mm256_hadd_epi32(sum256, sum256);
-            const __m128i lo = _mm256_extracti128_si256(sum256, 0);
-            const __m128i hi = _mm256_extracti128_si256(sum256, 1);
-            output[i] += _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi);
+            const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
+            const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
+            __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
+            product256 = _mm256_madd_epi16(product256, _mm256_set1_epi16(1));
+            sum = _mm512_add_epi32(sum, _mm512_zextsi256_si512(product256));
         }
+        output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
 
   #elif defined(USE_AVX2)
         __m256i sum = _mm256_setzero_si256();
@@ -135,23 +130,30 @@ namespace Eval::NNUE::Layers {
           product = _mm256_madd_epi16(product, kOnes);
           sum = _mm256_add_epi32(sum, product);
         }
-        sum = _mm256_hadd_epi32(sum, sum);
-        sum = _mm256_hadd_epi32(sum, sum);
-        const __m128i lo = _mm256_extracti128_si256(sum, 0);
-        const __m128i hi = _mm256_extracti128_si256(sum, 1);
-        output[i] = _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi) + biases_[i];
+        __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
+        sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
+        sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
+        output[i] = _mm_cvtsi128_si32(sum128) + biases_[i];
 
   #elif defined(USE_SSSE3)
-        __m128i sum = _mm_cvtsi32_si128(biases_[i]);
+        __m128i sum = _mm_setzero_si128();
         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
-        for (IndexType j = 0; j < kNumChunks; ++j) {
-          __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
+        for (int j = 0; j < (int)kNumChunks - 1; j += 2) {
+          __m128i product0 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
+          product0 = _mm_madd_epi16(product0, kOnes);
+          sum = _mm_add_epi32(sum, product0);
+          __m128i product1 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j+1]), _mm_load_si128(&row[j+1]));
+          product1 = _mm_madd_epi16(product1, kOnes);
+          sum = _mm_add_epi32(sum, product1);
+        }
+        if (kNumChunks & 0x1) {
+          __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[kNumChunks-1]), _mm_load_si128(&row[kNumChunks-1]));
           product = _mm_madd_epi16(product, kOnes);
           sum = _mm_add_epi32(sum, product);
         }
-        sum = _mm_hadd_epi32(sum, sum);
-        sum = _mm_hadd_epi32(sum, sum);
-        output[i] = _mm_cvtsi128_si32(sum);
+        sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
+        sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
+        output[i] = _mm_cvtsi128_si32(sum) + biases_[i];
 
   #elif defined(USE_NEON)
         int32x4_t sum = {biases_[i]};

diff --git a/src/nnue/nnue_accumulator.h b/src/nnue/nnue_accumulator.h
index 2a354a3c2844a29a7bcec9d9a5ce2a12822eb203..69dfaad214795deb4155890922b4925c99714646 100644 (file)
--- a/src/nnue/nnue_accumulator.h
+++ b/src/nnue/nnue_accumulator.h
@@ -26,7 +26,7 @@
 namespace Eval::NNUE {
 
   // Class that holds the result of affine transformation of input features
-  struct alignas(32) Accumulator {
+  struct alignas(kCacheLineSize) Accumulator {
     std::int16_t
         accumulation[2][kRefreshTriggers.size()][kTransformedFeatureDimensions];
     Value score;

diff --git a/src/nnue/nnue_common.h b/src/nnue/nnue_common.h
index e7ce84f7b9f420eaaa3dbadb4302eeb242315c4f..ff33cc7974b12d616d6f6f49c3dcbcaa5d252e5e 100644 (file)
--- a/src/nnue/nnue_common.h
+++ b/src/nnue/nnue_common.h
@@ -52,9 +52,11 @@
 
 #if defined(USE_AVX512)
 #if defined(__GNUC__ ) && (__GNUC__ < 9)
-#define _mm512_loadA_si512  _mm512_loadu_si512
+#define _mm512_loadA_si512   _mm512_loadu_si512
+#define _mm512_storeA_si512  _mm512_storeu_si512
 #else
-#define _mm512_loadA_si512  _mm512_load_si512
+#define _mm512_loadA_si512   _mm512_load_si512
+#define _mm512_storeA_si512  _mm512_store_si512
 #endif
 #endif
 

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
 
         }