I am studying different SIMD extensions on convolution network using MKL. I was able to set different SIMD extension using MKL_CBWR. I am wondering is it possible to direct the MKL to use no-vectorization code. I want to quantify the effect of using SIMD extensions against no vectorization.
For that purpose (if you want single thread comparison), you might build public source with no vectorization set for your compiler. You will need to decide about unroll flags. For example, non vector unroll by 4 might be relevant.