I am writing a Python module in Rust using PyO3. In one of these Rust functions, I am calling a python function, my_py_func, which returns a numpy.ndarray, with dtype=int64 entries, and shape=(k,2) for some dynamically determined k.
My rust code looks like this.
let my_py_func_rs = PyModule::import(py, "__main__").unwrap().getattr("my_py_func").unwrap().into();
// input argument to the python function
// my_input_array is a ndarray::Array1<u8>
let args = PyTuple::new(py, &[my_input_array.into_pyarray(py)]);
let output_array: Vec<Vec<usize>> = my_py_func_rs.call1(py, args).unwrap().extract(py).unwrap();
This works! Compiles and runs as expected.
But I have a feeling that it's not correct or optimal to use vector of vectors as the data structure. The subsequent operation I want to do on output_array is iterating over the rows and doing something based on each row. I want to extract every bit of performance, because the final code will run for many days/weeks.
Will it be faster to use Array2? I can't test this because I don't know how to convert the python function output to Array2. I am also using the numpy crate elsewhere in my code, if that is relevant.
What is the fastest datastructure for output_array?