RFC: Support Adapt.AbstractGPUDevice

[oschulz]

See JuliaGPU/Adapt.jl#52

Not functional like this yet, I don't know AMDGPU well enough and don't have a device to test, currently. See JuliaGPU/CUDA.jl#1520 for the CUDA.jl counterpart.

[jpsamaroo]

Thanks for putting this together! CI is down while I'm out of the country, so it'll have to wait another week before I can get CI back up.

[oschulz]

Thanks for putting this together! CI is down while I'm out of the country

Thanks @jpsamaroo! When you're back, I'll need some hints from you on the AMDGPU side of this (I don't really know what the equivalent of CuDevice is, and so on). :-)

[jpsamaroo]

Can you instead change

AMDGPU.jl/src/agent.jl

Line 3 in 99967b7

mutable struct HSAAgent

to mutable struct HSAAgent <: AbstractGPUDevice? We use HSAAgent for most operations involving devices.

[oschulz]

Thanks, done! How do I do

Adapt.adapt_storage(device::HSAAgent, x) = adapt_storage(ROCArray, x)

correctly (i.e. send the array to the right device)?

[jpsamaroo]

@oschulz can you rebase this on AMDGPU master? HSAAgent is now named ROCDevice.

[oschulz]

Will do, I need to rebase the other related PRs as well anyway.

[jpsamaroo]

@oschulz bump

[oschulz]

@oschulz bump

Sorry for the delay, it's not forgotten, but holidays got in the way. :-) Also looks like the whole thing will now include ArrayInterfaceCore as well (recent discussion on JuliaGPU/Adapt.jl#52). Gimme a bit of time to adapt to that, I'll modify and rebase the whole PR-complex in the process.

[oschulz]

Closed, but not forgotten, to be reincarnated in the future (see JuliaGPU/Adapt.jl#64).