X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Flayers%2Faffine_transform.h;fp=src%2Fnnue%2Flayers%2Faffine_transform.h;h=af85c817c2aa0765e54aa14364c14ed7fb4a3e2d;hp=61cdb7818661286cde3c2be0c55f87035746fcad;hb=97f706ecc11459c8d0aa1901134d12fba00b4b15;hpb=0e32287af470dee230a30d9f513682c3ce798668

diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h
index 61cdb781..af85c817 100644
--- a/src/nnue/layers/affine_transform.h
+++ b/src/nnue/layers/affine_transform.h
@@ -45,6 +45,7 @@ namespace Stockfish::Eval::NNUE::Layers {
   template <IndexType InputDimensions, IndexType PaddedInputDimensions, IndexType OutputDimensions>
   static void affine_transform_non_ssse3(std::int32_t* output, const std::int8_t* weights, const std::int32_t* biases, const std::uint8_t* input)
   {
+# if defined(USE_SSE2) || defined(USE_MMX) || defined(USE_NEON_DOTPROD) || defined(USE_NEON)
 # if defined(USE_SSE2)
     // At least a multiple of 16, with SSE2.
     constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
@@ -129,18 +130,25 @@ namespace Stockfish::Eval::NNUE::Layers {
       }
       output[i] = sum[0] + sum[1] + sum[2] + sum[3];
 
-# else
-      std::int32_t sum = biases[i];
-      for (IndexType j = 0; j < InputDimensions; ++j) {
-        sum += weights[offset + j] * input[j];
-      }
-      output[i] = sum;
 # endif
     }
 
 # if defined(USE_MMX)
     _mm_empty();
 # endif
+
+# else
+  std::memcpy(output, biases, sizeof(std::int32_t) * OutputDimensions);
+
+  // Traverse weights in transpose order to take advantage of input sparsity
+  for (IndexType i = 0; i < InputDimensions; ++i)
+      if (input[i]) {
+          const std::int8_t* w = &weights[i];
+          const int in = input[i];
+          for (IndexType j = 0; j < OutputDimensions; ++j)
+              output[j] += w[j * PaddedInputDimensions] * in;
+      }
+# endif
   }
 #endif