Well, as you can see in the

Well, as you can see in the attachment, there is no AVX to SSE and SSE to AVX conversions in the __svml_ functions, so I'm still wondering what the reason of __svml_ slowness is, as the non-vec functions are fast.