use SIMD.jl for explicit vectorization of partial operations

[KristofferC]

Alternative to #555

Co-authored-by: Yingbo Ma [email protected]

Using the benchmark in #555

I extracted the code from the benchmark in #555 (https://gist.github.com/KristofferC/a3915eec407491b7a4fc8a7d81c16fbe) and put it into https://uica.uops.info/. The result is (with some various tools) in cycles per iteration:

Explicit:

uiCA | 265.50
IACA 3.0 | 245.00
llvm-mca | 210.06

Master:

uiCA | 371.00
IACA 3.0 | 318.00
llvm-mca | 263.11

So it does seem that the strategy in #555 is quite good.

[KristofferC]

Since this seems to have the same performance improvements as #555 (and also some better compile-time behavior) as well as not having to duplicate all the llvmcall logic that exists in SIMD.jl I think it is better to go with this one.