Do not install/load CUDA packages if no GPU is detected.

[ali-ramadhan]

This would be nice because most people might first install Oceananigans on their laptop without a GPU or CUDA toolkit, and are greeted with error messages that e.g. CUDAdrv failed to build.

While harmless errors, they don't look nice and make the package seem unstable. This would also speed up Travis CI and Appveyor builds.

---

Relevant bits from Pkg.instantiate():

┌ Error: Error building `CUDAdrv`: 
│ Could not find CUDA driver library.
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
┌ Error: Error building `CUDAnative`: 
│ Dependent package CUDAdrv.jl has not been built successfully.
│ 
│ This is not a fatal error, but GPU functionality will be unavailable.
│ If you expected this to work, please open a thread on
│ https://discourse.julialang.org/c/domain/gpu
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
┌ Error: Error building `CuArrays`: 
│ Dependent package CUDAdrv.jl has not been built successfully
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
┌ Error: Error building `CUDAdrv`: 
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
┌ Error: Error building `CUDAdrv`: 
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
  Building SpecialFunctions → `~/.julia/packages/SpecialFunctions/fvheQ/deps/build.log`
  Building LLVM ────────────→ `~/.julia/packages/LLVM/tg8MX/deps/build.log`
  Building CUDAnative ──────→ `~/.julia/packages/CUDAnative/B210M/deps/build.log`
Dependent package CUDAdrv.jl has not been built successfully.
This is not a fatal error, but GPU functionality will be unavailable.
If you expected this to work, please open a thread on
https://discourse.julialang.org/c/domain/gpu
┌ Error: Error building `CUDAnative`: 
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
┌ Error: Error building `CuArrays`: 
└ @ Pkg.Operations /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.1/Pkg/src/Operations.jl:1075
ERROR: LoadError: LoadError: UndefVarError: CUBLAS not defined
Stacktrace:
 [1] top-level scope at none:0 (repeats 2 times)
 [2] include at ./boot.jl:326 [inlined]
 [3] include_relative(::Module, ::String) at ./loading.jl:1038
 [4] include at ./sysimg.jl:29 [inlined]
 [5] include(::String) at /home/travis/.julia/packages/CuArrays/qZCAt/src/CuArrays.jl:3
 [6] top-level scope at none:0
 [7] include at ./boot.jl:326 [inlined]
 [8] include_relative(::Module, ::String) at ./loading.jl:1038
 [9] include(::Module, ::String) at ./sysimg.jl:29
 [10] top-level scope at none:2
 [11] eval at ./boot.jl:328 [inlined]
 [12] eval(::Expr) at ./client.jl:404
 [13] top-level scope at ./none:3

The cuBLAS error might be related to CuArrays: JuliaGPU/CuArrays.jl#255

[ali-ramadhan]

This might be an upcoming update to CUDAnative! JuliaGPU/CUDAnative.jl#392

[maleadt]

Although that PR makes CUDAnative behave a little better, it does not change the warnings. Ideally, Pkg would have conditional dependencies and a way for the build process to inform that the package is or is not available, but lacking that the warnings seem like the only way. Maybe you have an other idea?

Maybe you should comment at JuliaLang/Pkg.jl#977

[ali-ramadhan]

Thanks for the link, I wasn't aware it there was an open PR with a proposal.

Unfortunately it doesn't look like this feature will become available for a while and people have already tried to install Oceananigans and just gave up because of warnings/errors, so I'm tempted to just hack in a solution.

I'm thinking of removing the dependency on all CUDA packages, and have them get added/installed at run time if the Model constructor is called with arch=GPU(). I think this should give us exactly what we want for now.

[ali-ramadhan]

With new versions of the CUDA packages in PR #388 there are no more warnings so this isn't an issue anymore.