Recent Intel's compiler are able to automatically vectorize math functions such as cos, exp, etc. through calls to SVML. Apparently, SVML versions are faster than ours, and therefore disabling Eigen's vectorization might lead to faster code...
One possible strategy would be to perform ourself explicit calls to SVML in our pexp, p* functions when the user compiles with -DEIGEN_USE_SVML.
Just a note here:
Visual Studio 2019 ships with SVML intrinsics (svmlintrin.h included by immintrin.h).
Would be nice if Eigen would call those functions using visual studio.
*** Bug 1681 has been marked as a duplicate of this bug. ***