The purpose is to let compiler optimize/vectorize operations It's expected to obtain a perf comparable to that in SIMD mode