scatter!() on CuArrays now triggers scalar operations

[dfdx]

MWE:

using NNlib   # v0.7.17
using CUDA

CUDA.allowscalar(false)

scatter!(+, cu(rand(5)), cu([2.0]), cu([1]))
# ==> ERROR: scalar getindex is disallowed

Note that this works in ScatterNNlib (destination slightly modified to conform to older API):

ScatterNNlib.scatter_add!(cu(rand(5, 1)), cu([2.0]), cu([1]))
# ==> ok

From the recent PRs I see that only the generic (i.e. CPU) version of scatter!() is implemented. Is the CUDA version implemented somewhere, planned or is explicitly unsupported?

cc: @yuehhua

[ToucheSir]

#296?

[yuehhua]

As you see, we are currently migrating functionality of ScatterNNlib.jl to NNlib.jl. scatter!() for generic use is merged, while for cuda is WIP. scatter!(), scatter(), gather!() and gather() will be supported for CPU and CUDA, as well as their AD in the future.

[dfdx]

Perfect, thank you!