add implicit float conversion to math functions

[CarloLucibello]

add implicit float conversion to math functions

[maleadt]

which is also what Base does

Wait, what? No, Base doesn't do any eager conversions. Having cu([1]) == [1f0] seems like a recipe for disaster, why do you want that conversion back? The fact that you were silently dispatching to floating-point operations when asking for (unsupported) integer versions is an argument for the current behavior, to me.

[CarloLucibello]

sorry I meant Base converts when doing mathematical operations such as log.([1, 1]), so I would expect the same for log.(cu([1, 1]))

[CarloLucibello]

I see now that on CuArrays 1.7 cu performed the casting, I agree that should be avoided and current behavior for cu is preferable. On the other hand, I'd like all mathematical operations to work on integer cuarrays through conversion

[maleadt]

I'd like all mathematical operations to work on integer cuarrays through conversion

Yeah, understandably. We're in the annoying situation now where we've been hoping to reuse Base's conversions through Cassette, only redefining the bottommost calls to incompatible CPU library intrinsics with NVIDIA ones, but progress on Cassette inferability has been... slow. In the meantime, we don't have a good policy or place to add higher-level functions like log accepting integers. I guess we should just add them to CUDAnative for now, e.g. like JuliaGPU/CUDAnative.jl#445. If you have the need, and the time to create a PR, feel free to add some implementations there!