Problem with a simple add vecs for some arrays sizes on CUDA

Created by: jkozdon

So I am not sure what I am doing wrong:

using KernelAbstractions
using CUDA
using Printf
using Random

@kernel function ka_knl_add!(C, A, B)
    I = @index(Global)
    @inbounds C[I] = A[I] + B[I]
end

function main(device)
  @show device
  ka_add! = ka_knl_add!(device, device isa CPU ? 4 : 1024)
  for k = 2:2:28
    N = 2^k

    A = rand(N)
    B = rand(N)
    C = similar(A)
    if device isa CUDADevice
      A = CuArray(A)
      B = CuArray(B)
      C = CuArray(C)
    end

    wait(ka_add!(C, A, B, ndrange = N))
    if C ≠ A + B
      println()
      println("2^$k = $N")
      @show C == A + B
      @show C ≈ A + B
      @show extrema(Array(C - A - B))
    end
  end
end

Random.seed!(77)
main(CPU())
main(CUDADevice())

But for some arrays sizes on a V100 this produces:

device = CPU()

device = CUDADevice()

2^16 = 65536
C == A + B = false
C ≈ A + B = false
extrema(Array((C - A) - B)) = (-5.447940142435828, 0.13998980843263142)

2^18 = 262144
C == A + B = false
C ≈ A + B = false
extrema(Array((C - A) - B)) = (-1.9034060006220799, 1.8951141473846453)