What is the difference between Kernel Input and Output?

[cadop]

In the runtime reference https://nvidia.github.io/warp/modules/runtime.html#kernels it only mentions

inputs (Sequence) – The input parameters to the kernel (optional)

outputs (Sequence) – The output parameters (optional)

In kernel basics https://nvidia.github.io/warp/basics.html#kernels there is no discussion about Outputs. Instead it says

Because Warp kernels are compiled .... Instead they must read and write data through their input parameters such as arrays.

What is the difference?

If I am trying to update an existing array, such as my_array[tid]+=1, is this an "output", since the array is modified? But it is also an "input" since I am passing an array of data. Is this just a useful name convention that passes arguments in order? The kernel signature doesn't seem to differentiate between them, just order.

[shi-eric]

The only place that I'm aware of where using outputs makes sense is when you're writing kernels that might participate in automatic differentiation. Using both the inputs and outputs keywords can be helpful here to avoid overwrite issues that would result in incorrect gradients. Some of the gradient debugging tools also require the use of the outputs keyword. If you aren't using automatic differentiation, then I would just stick with inputs.