Preferences.jl-based method to define where CUDA is installed

[giordano]

When I use CUDA.jl on clusters it happens occasionally that I forget to load the needed module and CUDA.jl yells at me because it can't find libcuda & friends. It'd be nice if I could stick all the information about where CUDA is installed in a LocalPreferences.toml file and don't need to remember to load the module beforehand.

Maybe one day I'll have the time to tackle this issue, in the meantime I'm opening the ticket in case someone else gets nerdsniped inspired and wants to do the same.

[maleadt]

Sounds like a good idea. The easy fix would be to add a preference for the CUDA_HOME etc env var, at

CUDA_Runtime_Discovery.jl/src/CUDA_Runtime_Discovery.jl

Lines 315 to 318 in afb1cc0

	# look for environment variables to override discovery
	envvars = ["CUDA_PATH", "CUDA_HOME", "CUDA_ROOT", "NVHPC_ROOT"]
	filter!(var -> haskey(ENV, var) && ispath(ENV[var]), envvars)
	@debug "Looking for CUDA toolkit via environment variables $(join(envvars, ", "))"

A more flexible solution would be to (also) support _path preferences like other JLLs do, but with the note that the package here can also be used generically, e.g., libcudnn isn't listed here and CUDNN.jl directly calls find_cuda_library.