[FIXED] np.argsort() implementation is not found

Issue

I would like to see how numpy.argsort() works.

1. In the documentation, the source for numpy.argsort() is numpy.core.fromnumeric.py. This is understandable. https://numpy.org/doc/stable/reference/generated/numpy.argsort.html

2. core.fromnumeric.argsort() is a bit more complicated.
   Ignoring decorators, if fromnumeric.argsort(arr) returns _wrapfunc(arr, "argsort"), which returns arr.argsort(). This is not a problem.
   Assuming arr is numpy.ndarray, it might be in array_api.__init__.py. https://github.com/numpy/numpy/blob/v1.21.0/numpy/core/fromnumeric.py

3. array_api.argsort() is from array_api._sorting_functions.argsort(). OK.https://github.com/numpy/numpy/blob/main/numpy/array_api/__init__.py

4. _sorting_functions.argsort() calls numpy.argsort(). That is what I was looking for at first. It is circular. https://github.com/numpy/numpy/blob/main/numpy/array_api/_sorting_functions.py

Extra

5. In numpy.__init__.pyi, numpy.argsort() is from core.fromnumeric https://github.com/numpy/numpy/blob/main/numpy/__init__.pyi

   1. and 5. are the same thing.

Are these circular references? Of course I know these work. Is it might be in array_api.__init__.py. in 2. wrong? So where is the actual location of its implementation?

---

Background on this issue

I noticed that np.unique is slow when return_index=True. I wanted to run np.unique on the sorted array, but found that np.unique calls np.argsort. So I tried to find out the difference between np.argsort and np.sort and needed to know more about np.argsort.

Solution

Why do you want to see the source? To implement it in your own c code project? I don't think it will help you use it more effectively in python. In an Ipython session I use ??

In [22]: np.argsort??
...
return _wrapfunc(a, 'argsort', axis=axis, kind=kind, order=order)

OK, that's the typical case of a function passing the buck to the method. The function version will convert the input to array if necessary, and then call the array's method. Typically the function version has a more complete documentation, but the functionality is basically the same.

In [21]: arr.argsort??
Type:      builtin_function_or_method

Usually that's the end of the story.

The other route is to click the [source] link on the documentation. Here that leads to the same thing.

Notice:

@array_function_dispatch(_argsort_dispatcher)

recent versions have added this dispatch layer; check the release notes for more details. In my experience that just makes searching for code harder.

The other step is to go to github and do a search. Sometimes that turns up some useful bit, but often it's a wild-goose-chase.

As a user I don't need to know the "how" details. It's easy enough to read the docs, and then do some experiments if I still have questions. Digging into the c code will not help be use it better.

As for your added question:

All ndarray objects are "multiarray", with anything from 0 to 32 dimensions.

github

On numpy github I searched for argsort, and chose the most promising file, numpy/core/src/multiarray/methods.c

This has function

array_argsort(PyArrayObject *self,
        PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)

Skipping over code that appears to handle the input arguments, it looks the work is done in the

res = PyArray_ArgSort(self, axis, sortkind);

That appears to be defined in numpy/core/src/multiarray/item_selection.c

 PyArray_ArgSort(PyArrayObject *op, int axis, NPY_SORTKIND which)
 ...
 if (argsort == NULL) {
    if (PyArray_DESCR(op)->f->compare) {
        switch (which) {
            default:
            case NPY_QUICKSORT:
                argsort = npy_aquicksort;
                break;
            case NPY_HEAPSORT:
                argsort = npy_aheapsort;
                break;
            case NPY_STABLESORT:
                argsort = npy_atimsort;
                break;
   ...
   ret = _new_argsortlike(op2, axis, argsort, NULL, NULL, 0);

and so on ....

None of that helps me use it any better.

Answered By - hpaulj