GEMM: Substantially improved pure CPU-based implementation.

About a factor of 20 faster than previous implementation.
I estimate that more microtuning can get another factor of 2.
Higher performance gains will most likely require intrinsics.