Support VNNI on 256bit vectors

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

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 89cfaad7dbd8ff847bbccf8bd451c08927860292..94d0b5a9494644e574cd111104943d18667c9196 100644 (file)
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -62,11 +62,10 @@ namespace Eval::NNUE::Layers {
    // Read network parameters
     bool ReadParameters(std::istream& stream) {
       if (!previous_layer_.ReadParameters(stream)) return false;
-      stream.read(reinterpret_cast<char*>(biases_),
-                  kOutputDimensions * sizeof(BiasType));
-      stream.read(reinterpret_cast<char*>(weights_),
-                  kOutputDimensions * kPaddedInputDimensions *
-                  sizeof(WeightType));
+      for (std::size_t i = 0; i < kOutputDimensions; ++i)
+        biases_[i] = read_little_endian<BiasType>(stream);
+      for (std::size_t i = 0; i < kOutputDimensions * kPaddedInputDimensions; ++i)
+        weights_[i] = read_little_endian<WeightType>(stream);
       return !stream.fail();
     }
 
@@ -79,19 +78,32 @@ namespace Eval::NNUE::Layers {
 
   #if defined(USE_AVX512)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
-      const __m512i kOnes = _mm512_set1_epi16(1);
       const auto input_vector = reinterpret_cast<const __m512i*>(input);
+  #if !defined(USE_VNNI)
+      const __m512i kOnes = _mm512_set1_epi16(1);
+  #endif
 
   #elif defined(USE_AVX2)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
-      const __m256i kOnes = _mm256_set1_epi16(1);
       const auto input_vector = reinterpret_cast<const __m256i*>(input);
+  #if !defined(USE_VNNI)
+      const __m256i kOnes = _mm256_set1_epi16(1);
+  #endif
 
-  #elif defined(USE_SSSE3)
+  #elif defined(USE_SSE2)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+  #ifndef USE_SSSE3
+      const __m128i kZeros = _mm_setzero_si128();
+  #else
       const __m128i kOnes = _mm_set1_epi16(1);
+  #endif
       const auto input_vector = reinterpret_cast<const __m128i*>(input);
 
+  #elif defined(USE_MMX)
+      constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
+      const __m64 kZeros = _mm_setzero_si64();
+      const auto input_vector = reinterpret_cast<const __m64*>(input);
+
   #elif defined(USE_NEON)
       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
       const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
@@ -104,9 +116,13 @@ 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) {
+  #if defined(USE_VNNI)
+            sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
+  #else
             __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
             product = _mm512_madd_epi16(product, kOnes);
             sum = _mm512_add_epi32(sum, product);
+  #endif
         }
 
         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
@@ -116,9 +132,14 @@ namespace Eval::NNUE::Layers {
         {
             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
+  #if defined(USE_VNNI)
+            __m256i product256 = _mm256_dpbusd_epi32(
+                _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
+            sum = _mm512_inserti32x8(sum, product256, 0);
+  #else
             __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));
+            sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256));
+  #endif
         }
         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
 
@@ -126,9 +147,13 @@ namespace Eval::NNUE::Layers {
         __m256i sum = _mm256_setzero_si256();
         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
         for (IndexType j = 0; j < kNumChunks; ++j) {
+  #if defined(USE_VNNI)
+          sum = _mm256_dpbusd_epi32(sum, _mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
+  #else
           __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
           product = _mm256_madd_epi16(product, kOnes);
           sum = _mm256_add_epi32(sum, product);
+  #endif
         }
         __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));
@@ -155,6 +180,51 @@ namespace Eval::NNUE::Layers {
         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
         output[i] = _mm_cvtsi128_si32(sum) + biases_[i];
 
+  #elif defined(USE_SSE2)
+        __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]);
+        __m128i sum_hi = kZeros;
+        const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m128i row_j = _mm_load_si128(&row[j]);
+          __m128i input_j = _mm_load_si128(&input_vector[j]);
+          __m128i row_signs = _mm_cmpgt_epi8(kZeros, row_j);
+          __m128i extended_row_lo = _mm_unpacklo_epi8(row_j, row_signs);
+          __m128i extended_row_hi = _mm_unpackhi_epi8(row_j, row_signs);
+          __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros);
+          __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros);
+          __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo);
+          __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi);
+          sum_lo = _mm_add_epi32(sum_lo, product_lo);
+          sum_hi = _mm_add_epi32(sum_hi, product_hi);
+        }
+        __m128i sum = _mm_add_epi32(sum_lo, sum_hi);
+        __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
+        sum = _mm_add_epi32(sum, sum_high_64);
+        __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
+        sum = _mm_add_epi32(sum, sum_second_32);
+        output[i] = _mm_cvtsi128_si32(sum);
+
+  #elif defined(USE_MMX)
+        __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]);
+        __m64 sum_hi = kZeros;
+        const auto row = reinterpret_cast<const __m64*>(&weights_[offset]);
+        for (IndexType j = 0; j < kNumChunks; ++j) {
+          __m64 row_j = row[j];
+          __m64 input_j = input_vector[j];
+          __m64 row_signs = _mm_cmpgt_pi8(kZeros, row_j);
+          __m64 extended_row_lo = _mm_unpacklo_pi8(row_j, row_signs);
+          __m64 extended_row_hi = _mm_unpackhi_pi8(row_j, row_signs);
+          __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros);
+          __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros);
+          __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo);
+          __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi);
+          sum_lo = _mm_add_pi32(sum_lo, product_lo);
+          sum_hi = _mm_add_pi32(sum_hi, product_hi);
+        }
+        __m64 sum = _mm_add_pi32(sum_lo, sum_hi);
+        sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
+        output[i] = _mm_cvtsi64_si32(sum);
+
   #elif defined(USE_NEON)
         int32x4_t sum = {biases_[i]};
         const auto row = reinterpret_cast<const int8x8_t*>(&weights_[offset]);
@@ -174,6 +244,9 @@ namespace Eval::NNUE::Layers {
   #endif
 
       }
+  #if defined(USE_MMX)
+      _mm_empty();
+  #endif
       return output;
     }