Separate LoopVectorization.vsum from VectorizationBase.vsum

[brenhinkeller]

Seems like it might be better to have LoopVectorization.vsum != VectorizationBase.vsum esp. if we might want to extend the former (but not the latter) in other packages. I think this isn't technically breaking and hopefully it doesn't break anything if I did it right, but will see how tests look

[chriselrod]

Thanks for addressing this/taking over from #515.
Note that LoopVectorization itself calls VectorizationBase.vsum.
Double check whether it namespacing them properly (VectorizationBase.vsum), or being lazy (LoopVectorization.vsum and relying on the using VectorizationBase: vsum that you removed here).

[brenhinkeller]

Oh oops I hadn’t noticed #515! Looks like this is still breaking something, but I don't think I understand why..

[chriselrod]

Oh oops I hadn’t noticed #515! Looks like this is still breaking something, but I don't think I understand why..

Some failures are because LV uses VectorizationBase.vsum, without saying VectorizationBase.vsum.

[brenhinkeller]

That would make sense but I can't find any other usages of vsum when I search the package...

[chriselrod]

That would make sense but I can't find any other usages of vsum when I search the package...

E.g.

LoopVectorization.jl/src/modeling/costs.jl

Lines 599 to 601 in d2f749d

	reduction_to_scalar(x::Float64) =
	if x == ADDITIVE_IN_REDUCTIONS
	:vsum

LoopVectorization.jl/src/modeling/operations.jl

Lines 564 to 565 in d2f749d

	reduction_to_scalar(op::Operation)::GlobalRef =
	lv(reduction_to_scalar(instruction(op)))

LV doesn't call the code, it generates expressions with calls.

[brenhinkeller]

Ooh, so is it as simple as replacing that with :(VectorizationBase.vsum) or such? Trying that locally now and looks a bit better but still some local test failures

[chriselrod]

Ooh, so is it as simple as replacing that with :(VectorizationBase.vsum) or such? Trying that locally now and looks a bit better but still some local test failures

Note that lv(x) (called in one of those code snippets) does GlobalRef(LoopVectorization, x). We need VectorizationBase instead.