Top-level statement has no effect on execution of user code?

[vchuravy]

I just noticed these two lines:

WaterLily.jl/ext/WaterLilyCUDAExt.jl

Line 19 in c4d3500

CUDA.allowscalar(false) # disallow scalar operations on GPU

WaterLily.jl/ext/WaterLilyAMDGPUExt.jl

Line 19 in c4d3500

AMDGPU.allowscalar(false) # disallow scalar operations on GPU

Since precompilation executes top-level statements in isolated processes they have no effect?

Similarly the __init__ codes would break some use-cases, (like creating a docker container with a precompilation cache and loading it on a GPU enabled machine later)

[b-fg]

Ah, maybe we did not notice because these are the default options anyways? I have seen the scalar-indexing error correctly popping up when looping through a GPU array outside of a kernel. What would be better solution if these cannot be included in __init__? Since false is already the default flag, we could just delete them I guess?