Switch to KernelAbstractions.jl

[maleadt]

Continuation of #525. Separate PR because I reverted the launch_configuration removal that's on @leios' branch.

[maleadt]

One of the consequences here is that the backend function disappears, and is now part of KernelAbstractions.jl, so cc'ing people who were involved in #420 and JuliaGPU/Adapt.jl#50: @N5N3 @piever.

NeuralAttentionlib.jl also relies on GPUArraysCore.backend, so cc @chengchingwen.

I also see that Bennu.jl uses GPUArrays.launch_configuration directly, so cc @lcw.

[maleadt]

@vchuravy Does KA.jl already support recursing into wrapped arrays for get_backend queries?

[vchuravy]

Uhm I don't think so, but we can add that.

[maleadt]

I'm not entirely sure we want to; it would pull the whole Union mess of Adapt's WrappedArray into KA.jl. On the other hand, there's not much of an alternative right now if we want the ability to launch kernels on wrapped arrays...

[lcw]

Thanks for the ping. I have archived Bennu.jl as I have switched to developing Raven.jl. In Raven.jl I do use GPUArraysCore.backend here but I also already use KernelAbstractions.jl. So, I think it will be a simple transition.

[vchuravy]

This is clever xD

[maleadt]

Can you explain? I didn't get what this was for, and it seems unused?

[leios]

Unless I did something wrong, it's used for kernel configuration:

function (obj::Kernel{JLBackend})(args...; ndrange=nothing, workgroupsize=nothing)
    device_args = jlconvert.(args)
    new_obj = convert_to_cpu(obj)
    new_obj(device_args...; ndrange, workgroupsize)
end

And will essentially transform anything that has a JLBackend into a CPU(static) (default false) for KA execution so we can use the GPU kernels on CPU Arrays. It's needed because JLBackend is within KernelAbstractions.GPU, but needs to actually call CPU kernels:

struct JLBackend <: KernelAbstractions.GPU
    static::Bool
    JLBackend(;static::Bool=false) = new(static)
end

[maleadt]

Let's do this!

[leios]

Woo! Great work!