It seems that MKL_DIRECT_CALL does not apply to zgemm3 and cgemm3, only zgemm and cgemm
It seems slightly annoying that this is not the case as zgemm3 is significantly more numerically efficient for large matrices and is therefore a replacement for zgemm and it would have been helpful for MKL_DIRECT_CALL to switch calls of zgemm3 as well as zgemm for small matrices