Problem using `@atomic` in a kernel

[emmabeniston]

Hello!

I'm trying to run some code on Apple M2 and am encountering errors which we think are due to @atomic. I'm not experienced enough to produce a MWE but the issue seems to be due to the @atomic lines in the following kernel:

@kernel inbounds = true function MassCGKernel!(M,MMass,@Const(JJ),@Const(w),@Const(Glob))
  I,J,Iz,IF = @index(Global, NTuple)

  N = @uniform @groupsize()[1]
  Nz = @uniform @ndrange()[3]
  NF = @uniform @ndrange()[4]

  if Iz <= Nz && IF <= NF
    ID = I + (J - 1) * N  
    ind = Glob[ID,IF]  
    @atomic :monotonic M[Iz,ind,1] += JJ[ID,1,Iz,IF]
    @atomic :monotonic M[Iz,ind,2] += JJ[ID,2,Iz,IF]
    @atomic :monotonic MMass[Iz,ind] += eltype(M)(0.5) * (JJ[ID,1,Iz,IF] + JJ[ID,2,Iz,IF]) * w[I] * w[J]
  end  
end

I get the following error message (not including the Stacktrace):

ERROR: LoadError: Compilation to native code failed; see below for details.
If you think this is a bug, please file an issue and attach /var/folders/hb/kf4qgnbd3mjb5gbb7nm8klmw0000gn/T/jl_gjS3EJcBoZ.metallib
caused by: NSError: Failed to materializeAll. (AGXMetalG14G, code 3)

Are there any suggestions for how to get this to run? Are there known to be any problems with using @atomic with Metal?

Code is courtesy of @OsKnoth who should be able to provide more details if needed. Related to this issue on oneAPI: JuliaGPU/oneAPI.jl#508

Thanks!