GPU Support in Field Calculations

[lmh91]

GPU Support for the SOR Field Calculation

In this PR I added GPU Support for the field calculations via KernelAbstractions.jl and it will close #60.

So far I just implemented it for cylindrical coordinates and depletion handling is not yet enabled for the GPU backend.

using CUDAKernels, SolidStateDetectors # CUDAKernels has to be loaded before SolidStateDetectors
using CUDA # -> CuArray
using Plots

T = Float32
sim = Simulation{T}(SSD_examples[:Coax])
calculate_electric_potential!( 
    sim, 
    device_array_type = CuArray, # CuArray for NVIDIA GPU / Array for CPU / ROCArray for AMD GPU
    convergence_limit = 0, # For this we have to probably add an addition kernel function 
    max_n_iterations = 10_000, # For now just perform always `max_n_iterations` iterations
    refinement_limits = [0.2, 0.1, 0.05, 0.02, 0.01]
)     

plot(sim.electric_potential)

ToDo for this PR:

• Support cylindrical coordinates
• Support Cartesian coordinates
• Support Depletion handling
• Add GPU documentation (Doc of this PR)
• Add tests

Not necessarily in this PR

• Test it on AMD GPUs
• How to choose a specific GPU device (-> using multiple GPUs in parallel)
• Add convergence check for GPU backend

Note:

• There was also one change regarding the CPU implementation of the SOR:
  After each iteration all grid points values at r = 0 were averaged (since they are in principle the same points).
  Now, this is only done once at the very end.
  I don't think doing it after each iteration is not necessary (tests still pass) (they anyhow converge against the same value within numerical precision) and I now removed it since otherwise I would need to add an extra kernel function (GPU) for it.

• It includes also a small bugfix in the depletion handling on the CPU backend.