Comment by cl3misch

Isn't the input of the layer fundamentally a part of the gradient computation? So even in this case (inspecting LLVM code) the computation still needs to look at the input.