Valentin Churavy
Valentin Churavy
`KAROCDevice` reads absolutly horrid. You also need to update ROCKernels to 0.4 since this is a breaking change. Maybe `ROCmDevice`? Can we cheat and just re-use `ROCDevice` from `AMDGPU.jl`
I don't think we use that dispatch rule for much...
Can you remove the Manidest.toml you added?
Sure! On Mon, Aug 8, 2022, 10:35 Julian Samaroo ***@***.***> wrote: > @vchuravy the ROCDevice? > > — > Reply to this email directly, view it on GitHub > ,...
@lcw is that roughly what you were thinking? I am questioning whether CPU code should have ranges auto-injected or if I should leave that up to the user.
`@localmem` => `cuStaticSharedMem`. `@private` all names are bad :), but yes I hope to excise it eventually.
Also `@private` is no-op on the GPU. Well pretty much, you could use a `MArray` if you actually needed multidimensional scratch spaces
You would need to use multiple Julia tasks. One for doing the compute and one for doing the copy/communication. On Thu, Sep 2, 2021 at 3:45 PM Lucas C Wilcox...
Now I remember why we have https://github.com/JuliaGPU/KernelAbstractions.jl/blob/0337e2f1012b6014eac24f61720dcc04da68bf25/src/compiler.jl#L49-L62 Is this on 1.4 or 1.5? We probably need to overdub that definition to remove the string creation.
Ah, can you try 1.4 I haven't checked & fixed Cassette on 1.5