Move data between CPU and GPU

A field, field vector, or space lives on one device. ClimaCore.to_device copies it to another device and ClimaCore.to_cpu to the CPU; both return a copy, so the original is unchanged and out === x does not hold.

Prerequisites

CUDA.jl loaded (ClimaComms.@import_required_backends with CLIMACOMMS_DEVICE=CUDA) for anything involving CUDADevice().

Steps

1. Build or receive an object on one device:

   using ClimaComms
   ClimaComms.@import_required_backends
   import ClimaCore
   using ClimaCore.CommonSpaces
   cpu_space = ColumnSpace(; z_elem = 10, z_min = 0.0, z_max = 10.0, staggering = CellCenter())

2. Move it to the GPU. Every array inside the space's grid becomes a CuArray, and fields created on the moved space live on the GPU:

   cuda_space = ClimaCore.to_device(ClimaComms.CUDADevice(), cpu_space)

3. Move a result back for host-side work (a plotting library, a test comparison):

   field_on_cpu = ClimaCore.to_cpu(field)

Notes

• A field moved on its own carries a moved copy of its space; a field and a space moved separately are on different space objects, and broadcasting between them raises a mismatched-spaces error. Move the state once and derive everything from it.
• Remapping.interpolate returns a host array, so plotting from a GPU run needs no explicit transfer.
• to_device targets CPUSingleThreaded() and CUDADevice().