Method: for each split, a view is created and then copied into an output.

To come up:
* Add C code for GPU.
* Check if `self.len_splits` can be defined as a param op.