Use Adapt.jl for backend array conversion

[michel2323]

@sshin23 , This is equivalent to #235, but uses Adapt.

[michel2323]

The MacOS test is failing because findall is not implemented in the backend. However, it worked before, so I wonder what was actually used for the tests on a Mac. One should either use Metal.jl for the GPU or just plain CPU without OpenCL. OpenCL is deprecated on MacOS.

[michel2323]

@sshin23 Previously, the macOS runner was set to macos-latest and x64, and it used OpenCL. This has two issues:

• macos-latest + x64 will run Julia in Rosetta emulation
• OpenCL is officially deprecated on macOS

I added a macOS runner set to macos-latest and aarch64. Julia will now run natively on an M1 chip. The backend is set to Metal.jl, which is supported through KA.

[michel2323]

Myea, Metal doesn't support Float64 😕.

[klamike]

I gave metal support a shot a while ago. I don't think it's complete but maybe it can be useful for you: main...klamike:ExaModels.jl:mk/metal

[michel2323]

I gave metal support a shot a while ago. I don't think it's complete but maybe it can be useful for you: main...klamike:ExaModels.jl:mk/metal

I've reverted now in this PR. Using Metal should work out of the box, but to get the tests working with Float32 seems hard.

[sshin23]

We have some NaN <: Float64 in our code that will cause some issues
https://github.com/search?q=repo%3Aexanauts%2FExaModels.jl%20NaN&type=code

[klamike]

Nice! Yeah looks like most of those changes were just casting to Float32 all over the place.

[sshin23]

I wonder what the benefit of adapt is compared to using allocate. Given that we need an extremely simple feature, I'm more inclined to just using allocate as in #235 to reduce one dependency

[michel2323]

I wonder what the benefit of adapt is compared to using allocate. Given that we need an extremely simple feature, I'm more inclined to just using allocate as in #235 to reduce one dependency

ExaModels.convert_array(v, backend) = KernelAbstractions.allocate(backend, eltype(v), length(v))

allocate allocates uninitialized memory, whereas Adapt is closer to convert. It performs the appropriate conversion for the backend. In addition, adapt is supposed to preserve properties from the original array, such as unified, device, host, or something similar. It also does things like a noop when the array is already on the right backend.

Adapt is a dependency of all GPU packages and KA. It's a very light package that basically defines an API.

[sshin23]

Great. Thank you, @michel2323!

[michel2323]

Maybe waiting for this JuliaGPU/OpenCL.jl#423 so we can remove OpenCL too.

[michel2323]

The OpenCL PR takes too long. Merging.