I use _mm256_fmadd_ps to perform a * b and accumulate it to the result c, like c=a*b+c. It is found that under certain circumstances, fmadd operations will cause precision loss compared with those that mul first and then add, especially when c already has a non-zero value.