[Tracker] Implement all `numpy.*` APIs in CuPy

Implement GPU version of numpy.* functions in cupy.* namespace.

This is a tracker issue that lists the remaining numpy.* APIs (see also: comparison table). I’ve categorized them based on difficulty so that new contributors can pick the right task. Your contribution is highly welcomed and appreciated!

List of APIs

Very Easy

• numpy.asfarray (#6085)
• numpy.byte_bounds (hint: ndarray.data.ptr) (#7015)
• numpy.format_float_positional (#6308)
• numpy.format_float_scientific (#6474)
• numpy.ndarray.searchsorted (#7059)

Easy

• numpy.array_equiv (#6254)
• numpy.asarray_chkfinite (#6275)
• numpy.fabs (#6282)
• numpy.float_power (#6371)
• numpy.heaviside (#6798)
• numpy.isneginf (#6089)
• numpy.isposinf (#6089)
• numpy.mask_indices (#6156)
• numpy.ndarray.itemset
• numpy.real_if_close (#6475)
• numpy.setdiff1d (#6433)
• numpy.setxor1d (#6582)
• numpy.tril_indices & numpy.tril_indices_from (#6305)
• numpy.triu_indices & numpy.triu_indices_from (#6316)
• numpy.union1d (#6357)
• numpy.lib.format.*
• numpy.lib.stride_tricks.sliding_window_view (#6956)
• numpy.emath.* (numpy.lib.scimath.*)

Low priority (iterator functions):

• numpy.ndenumerate
• numpy.nditer
• numpy.nested_iters

Low priority (help functions):

• numpy.info
• numpy.lookfor
• numpy.source

Low priority (internal functions):

• numpy.deprecate
• numpy.deprecate_with_doc

Medium

• numpy.apply_over_axes
• numpy.block
• numpy.delete
• numpy.geomspace
• numpy.insert (#6597)
• numpy.put_along_axis
• numpy.row_stack (#6312)
• numpy.spacing
• numpy.vander (#6279)

Low priority (dtype APIs - need to filter types unsupported by CuPy):

• numpy.maximum_sctype
• numpy.cast (undocumented API) (hint: see numpy/core/numerictypes.py)
• numpy.typecodes (undocumented API)
• numpy.sctypeDict (undocumented API)
• numpy.sctypes (undocumented API)
• numpy.nbytes (undocumented API)

Low priority (rarely used APIs):

• numpy.einsum_path (#6723)
• numpy.frompyfunc (maybe just call GUFunc?)

Medium to Hard

• numpy.histogram_bin_edges
• numpy.ediff1d (#6280)
• numpy.intersect1d (#6402, #6407)
• numpy.nanpercentile
• numpy.nanquantile
• numpy.ndarray.ctypes (note: needs design discussion)
• numpy.ndarray.getfield
• numpy.ndarray.resize
• numpy.ndarray.setfield
• numpy.polynomial.*
• numpy.poly (#3547, #6697)
• numpy.polyder (#6469)
• numpy.polydiv (#3780)
• numpy.polyint
• numpy.trapz (#6107)

Linear Algebra APIs:

• numpy.linalg.cond
• numpy.linalg.eig
• numpy.linalg.eigvals
• numpy.linalg.multi_dot (#6358)

Random Sampling APIs:

• numpy.random.Generator.* (see the dedicated tracker issue for details: https://github.com/cupy/cupy/issues/4557)

Steps to Contribute

1. Fork and star ⭐ the CuPy repository 😉

2. Pick a function you want to work on. You can find the function in the NumPy API Reference to understand what should be implemented.

3. Implement a function in your branch. If you need help, join Gitter or just ask for help in this issue.

4. Don’t forget to write a test code!

5. Build CuPy and run tests to confirm that the function runs fine: pip install -e . && pytest tests/cupy_tests/PATH_TO_YOUR_TEST See the Contribution Guide for details.

6. Submit a pull-request to the master branch.

Note that you will need a GPU environment to develop CuPy.

See also:

• #6324
• #6692

[kmaehashi]

@rgommers Thank you for all the information. I’ve removed deprecated ones from the list:

• numpy.loads
• numpy.ndfromtxt
• numpy.alen
• numpy.asscalar
• numpy.ndarray.tostring
• numpy.set_numeric_ops

Also marked dtype-related functions as low priority as there’s an ongoing discussion in NumPy.

[rgommers]

loads and ndfromtxt were already removed in https://github.com/numpy/numpy/pull/19615, so can be removed from the list.

The medium & medium to hard seem ok though (with the exception of the typing stuff)

With “typing stuff” do you mean the dtype-related functions (cast, typecodes, etc.)? If so, then I agree - that’s questionable. Also xref https://github.com/numpy/numpy/issues/17325 for that.

The iterating functions (particularly on GPUs) I’m not sure about, but maybe Leo has thoughts there.

Anything else stick out to you, Ralf? Does the above sound roughly correct to you or are there other things that stick out?

That sounds about right. The random and linalg stuff is important, as are some of the medium-hard functions like put_along_axis, nanquantile, nanpercentile, etc.

All of the undocumented ones I’d like to get rid of in NumPy.