I was running some tests with dlange with the first argument 'F' (Frobenius norm) for 1000x1000 matrices and I noticed that it is twice as slower as using the same matrix (as a vector) with dnrm2.
Is it expected? I use MKL 11.3.2.
no, this is not expected behavior. is that hotspot for your application?
Zbigniew, the problem is fixed in 2017 u2. this update is released a few days ago. Could you check and keep us inform how this fix works on your side.
For more complete information about compiler optimizations, see our Optimization Notice.