Trying to make it work on CPU

[tkf]

No description provided.

[vchuravy]

Suggested change

	igl = @uniform @index(Group, Linear)
	igl = @index(Group, Linear)

[tkf]

I thought I might need to put it in @uniform since I'm using it after @synchronize?

FoldsKernelAbstractions.jl/src/transduce.jl

Lines 173 to 175 in 18c7b8f

	if t == 1
	@inbounds dest[igl] = shared[1]
	end

I can re-fetch the index, there, though

[vchuravy]

@index expression are always valid (and automatically inserted into the loop body).

[tkf]

Without @uniform, I'm now getting UndefVarError: igl not defined: https://github.com/JuliaFolds/FoldsKernelAbstractions.jl/pull/6/checks?check_run_id=1958277481

Actually, I just remember putting it in @uniform since it was the dependency of other @uniform variables:

FoldsKernelAbstractions.jl/src/transduce.jl

Lines 151 to 152 in 064aba9

	offsetb = (igl - 1) * groupsize()[1]
	bound = max(0, nbasecases - offsetb)

[vchuravy]

I suspect this form of @private is rather slow until KernelCompiler.jl lands. Using the older form is likely faster

[piever]

Not sure if this is helpful, but I've also played with block-level reduction with KernelAbstractions. I could get it to work on GPU (code similar to the one here), but not on CPU.

I tried replacing s and c with @private Int (1,) as well, but somehow using c[1] != 0 as a condition to break the while loop errors (KernelAbstractions still seems to think that s[1] is a vector rather than a scalar for some reason). Would love to know if a solution can be found here!

[tkf]

Yeah, I wonder if we need to tweak something in KernelAbstractions.jl or maybe just wait for KernelCompiler.jl

cc @vchuravy