Nested conversion of Float64 to Float32

[cortner]

When moving an Array{Float64, N} to the device, the floating point type is automatically converted,

using Metal
mtl( rand(1000) )  
# MtlVector{Float32, ...}

But for an array of static arrays, or dual numbers, etc, this conversion fails:

using StaticArrays 
mtl( rand( SVector{2, Float64}, 1000 ) )
# ERROR: Metal does not support Float64 values, try using Float32 instead

Conversion of arrays of static arrays does work if the number type is Float32,

mtl( rand( SVector{2, Float32}, 1000 ) )
# MtlVector{SVector{2, Float32}, ...}

I don't particularly mind whether I have to manually ensure that I use the correct number types, or whether Metal auto-converts for me. I see advantages either way. But the current behaviour is inconsistent.

I may be able to help with a PR that tackles StaticArrays, ForwardDiff (dual numbers) and HyperDuals but would probably need some pointers on how to approach this correctly without wasting everybody's time.

[maleadt]

Approaching this generically would involve doing what CUDA.jl does with Adapt: https://github.com/JuliaGPU/CUDA.jl/blob/205c238e5d37b74d32e24977f65aed74d7eb4e7a/src/array.jl#L744-L819

[papo1011]

Hi @maleadt , I've opened a PR #748 to fix this inconsistency