~4x faster than the C version.
The shuffles in the 15pt dim1 are seriously expensive. Not happy with it,
but I'm contempt.

Can be easily converted to pure AVX by removing all vpermpd/vpermps
instructions.