no method matching GPU or "KernelAbstractions" not found

[XinghaoJ]

When I use Julia/1.10.9 run my code, there are some errors showing up.

using Random
using Printf
using Oceananigans
using Oceananigans.Units: minute, minutes, hour, hours, days,day
....
grid = RectilinearGrid(GPU(), size = (Nx, Ny, Nz),
x = (0, Lx),
y = (0, Ly),
z = z_faces)
....
u_bsc = FieldBoundaryConditions(top=FluxBoundaryCondition(τx));
v_bsc = FieldBoundaryConditions(top=FluxBoundaryCondition(τy));
b_bsc = FieldBoundaryConditions(top=FluxBoundaryCondition(B_surf),bottom = GradientBuondaryCondition(dbdz_interior));
boundary_conditions=(u=u_bsc, v=v_bsc, b=b_bsc);

Run above code and I got
ERROR: LoadError: MethodError: no method matching GPU().
So I add "using CUDA", but then I got
ERROR: LoadError: UndefVarError: GradientBuondaryCondition not defined
But when I changed Julia/1.10.9 to Julia/1.10.2 with same enviroment(v1.10), and run this code, I got
ERROR: LoadError: KeyError: key "KernelAbstractions" not found

[simone-silvestri]

It seems that you have a typo in GradientBuondaryCondition, it should be GradientBoundaryCondition.
Because of a bug in the extensions in julia 1.10.2, Oceananigans runs on versions strictly larger than 1.10.2 (we should probably output a warning)

[XinghaoJ]

Thank you! I also noticed that in the previous version 1.10.2, I didn't need to add "using CUDA" to run my simulations on GPUs, but now in 1.10.9 or later, I have to; otherwise, I get a "no method matching GPU" error. This update from 1.10.2 to 1.10.9 seems to complicate my programming work.

[simone-silvestri]

We have moved all CUDA-related methods to a CUDA extension so that using a GPU on a CUDA device requires using CUDA in every script. It was necessary to have clean internals while supporting GPU computation on machines from other vendors (AMD, Metal, and Intel), which would have cluttered the source code otherwise.

For this reason, users need to add a using CUDA before arch = GPU() in every script. This is definitely a complication from the user side, but the cost is minimal, and, as a positive, users can also do

using AMDGPU
arch = GPU()

And the script will run on AMD GPUs.

[XinghaoJ]

Cool!