Zygote+Views+CuArrays=Bad News Bears (silently gives incorrect results)

[clintonTE]

Found this in the context of trying to adapt a Flux model for gpu usage.

using Revise, Zygote, BenchmarkTools, CuArrays, CUDAnative, Flux
function badnewsbears(N)

  A = reshape(collect(1:N^2), (N,N))
  Acu = A |> gpu

  x = collect(1:N)
  xcu = x|>gpu

  f(A,x)=sum(view(A, :, (i->(i-1)%N+1).(1:N^2)) .* x)

  @info "cpu value: $(f(A,x)); gpu value: $(f(Acu,xcu))"

  gs = gradient((x,A)->f(A,x), x, A)
  @info "check cpu gradient: $(gs)"

  gscu = gradient((xcu,Acu)->f(Acu, xcu), xcu, Acu)
  @info "check gpu gradient: $(gscu)"
end

badnewsbears(3)

Output:

[ Info: cpu value: 288; gpu value: 288
[ Info: check cpu gradient: ([36, 45, 54], [3 3 3; 6 6 6; 9 9 9])
[ Info: check gpu gradient: ([36, 45, 54], [1 1 1; 2 2 2; 3 3 3])