If possible, I think we should avoid updating input arrays in place in the pyccelized kernels. This would make the cupy/numpy interoperability more streamlined since we just need to convert to returned values back to cupy arrays, and we don't have to rely on possibly changed arrays in the arguments (which need to be converted from cupy to numpy before they are passed to a pyccelized kernel).
If possible, I think we should avoid updating input arrays in place in the pyccelized kernels. This would make the cupy/numpy interoperability more streamlined since we just need to convert to returned values back to cupy arrays, and we don't have to rely on possibly changed arrays in the arguments (which need to be converted from cupy to numpy before they are passed to a pyccelized kernel).