+
+# 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