Closed wangpinzhi closed 3 weeks ago
I think it looks okay. You can pass None
to an adj_input
/ adj_output
when it hasn't been allocated / should not participate in backpropagation. You can also pass None
if the gradient exists and is already associated with its corresponding Warp array. In this case you allocate a gradient array for grid_v_in
and that gets mapped to the corresponding Warp array here:
ctx.grid_v_in = wp.from_torch(grid_v_in, dtype=wp.vec3)
Thank you for your wonderful work. I have a question about the warp. launch function. When there are none warp. array variables int the inputs/outputs, what should be the corresponding _adjinputs and _adjoutputs? I have an example, can you tell me if the code is correct?