Jago Strong-Wright

Yeah I think so

I get this with the default free surface model (which I think can be solved fairly easily): ```julia model = HydrostaticFreeSurfaceModel(; grid) ERROR: Metal does not support Float64 values, try...

Found where the promotion is occurring in the former which solves that error, and get us back to the FFT transform planning error that the nonhydostatic model gives

And if I modify the default free surface model to `free_surface = ImplicitFreeSurface(gravitational_acceleration=eltype(grid)(g_Earth))` I also get the same error (the docstring is also wrong for what the default free surface...

Ah yeah that also gives the fft error

Ooh so with: ```julia using Oceananigans using Oceananigans.BuoyancyModels: g_Earth grid = RectilinearGrid(Oceananigans.Architectures.MetalBackend(), Float32; size = (3, 3, 50), extent=(10, 10, 500), halo = (3, 3, 3)) advection = Centered(Float32) free_surface...

With a much bigger grid (512, 512, 512): Metal: ``` @btime time_step!(model, 1) 629.676 ms (8811 allocations: 813.31 KiB) ``` (This is big enough to pin the Metal processor) CPU...

Ran the baroclinical adjustement example: ``` using Oceananigans using Oceananigans.Units Lx = 1000kilometers # east-west extent [m] Ly = 1000kilometers # north-south extent [m] Lz = 1kilometers # depth [m]...

MetalBacked is what KA calls it but MetalGPU sounds better

I'm not sure why the TimeInterval thing is being passed into any kernels?