VectorAPI: SubAll intrinsics for byte, short, float and double

[Paul Sandoz]

> On May 1, 2018, at 4:46 PM, Lupusoru, Razvan A <razvan.a.lupusoru at intel.com> wrote:
> 
> Just to chime in - I completely agree that for FP we could choose doing parallel instead of sequential operations. In fact, I imagine that in the coming future we will definitely have a data parallel implementation - so it would be good to think of how API would allow that to happen (either by default or explicitly).
> 
> There are a couple of reasons we have not done the data parallel approach yet:
> - We are reusing reduction nodes that vectorizer uses and those are required to be sequential. We would have to add additional metadata to reduction nodes to keep track of how reduction is intended to be done
> - Since our priority right now is to have some vector implementation for each method in API, it is easier as first step to do the sequential order for FP (since there are existing examples in code that we can use as example)
> 

Ok, that makes sense. It’s not entirely clear to me yet whether the "strictness" should be a property of the operation or something global via a command line switch, i suspect it may have to be per operation.

Paul.

> Thanks again for your thorough review!
> 
> —Razvan
>