Architecture and SIMD agnostic about exapf.jl

[frapac]

I investigate more this issue. I think at some point we should be able to find a better abstraction model in ExaPF.jl. Thinking more about it, I am wondering if we could add a dependency to KernelAbstractions.jl. The idea would be to define the code we are deporting to GPU in a @kernel macro, and then dispatch to CPU/GPU with a dedicated statement, e.g.

 if isa(a, Array)
     kernel! = powerflow_kernel!(CPU(), 4)
 else
     kernel! = powerflow_kernel!(CUDADevice(), 256)
 end
 kernel!(pf_inputs, ndrange=a_size_for_pf) 

(this snippet is adapted from this example).

However, I am wondering whether KernelAbstractions is mature enough. Maybe we could use the JuliaCon as an opportunity to discuss with the developers?

edit: I checked, and the talk about KernelAbstractions is just after yours, Michel :-)

[frapac]

Solved by #15