Un-intended behaviour? Should Flux be able to reduce StaticArrays?

[AhmedSalih3d]

Hello

I am trying:

using CUDA
S = rand(SVector{3,Float32},5)
DST = zeros(SVector{3,Float32},5)
I = [3,1,2,5,4]

Works great on CPU

@CUDA.time NNlib.scatter!(+, DST,S,I)
0.000004 seconds
5-element Vector{SVector{3, Float32}}:
[1.3667885, 1.081403, 1.1134366]
[1.6543723, 0.50424564, 1.1448298]
[0.8695842, 1.8623418, 1.398939]
[1.0094697, 0.0052466393, 0.09184897]
[1.9389011, 0.3291297, 1.2308507]

Bugs out on GPU

@CUDA.time NNlib.scatter!(+, CuArray(DST),CuArray(S),CuArray(I))
ERROR: InvalidIRError: compiling kernel #scatter_kernel!(typeof(+), CuDeviceVector{SVector{3, Float32}, 1}, CuDeviceVector{SVector{3, Float32}, 1}, CuDeviceVector{Int64, 1}) resulted in invalid LLVM IR
Reason: unsupported dynamic function invocation (call to atomic_cas!)

Anyone knows why?

This is an example, I need it to work on GPU for a more complex case

Kind regards

[mcabbott]

This is FluxML/NNlib.jl#507 right?

You could try pinging those who wrote the scatter code, e.g. going by https://github.com/FluxML/NNlibCUDA.jl/pulls?q=is%3Apr+scatter

[AhmedSalih3d]

This is FluxML/NNlib.jl#507 right?

You could try pinging those who wrote the scatter code, e.g. going by https://github.com/FluxML/NNlibCUDA.jl/pulls?q=is%3Apr+scatter

I am not sure, I just went "using Flux" and then it exposes NNlib to me. I suppose the GPU version might be using NNlibCUDA in the background, but I have not tested that

[AhmedSalih3d]

To be honest I had forgotten I had made the first issue.

I made an updated comment there, thank you

EDIT: I think it is fine to close this issue, since it is not Flux I think? But the underlying library not supporting it?

[ToucheSir]

Let's not duplicate this discussion between 2 GitHub issues and Discourse.