SIMD code

[msricher]

The code can be sped up by using SIMD instructions. This is something we should explicitly implement.

[PaulWAyers]

Do you think it will be a big gain over what is achieved with an optimizing compliler @msricher ?

[msricher]

I actualy checked this last night, dumped the assembly and looked at it. With make BUILD_NATIVE=1, it makes good use of SIMD already and does go quite a bit faster. I'm sure we could do even better if we did it by hand... but I'm not so sure it's necessary anymore, this feels good enough.