Implement more array creation & serialization methods in CuPy

Currently, the following NumPy-compatible APIs are missing. It would be great to have them in CuPy.

These APIs are relatively simple, so it would be a good starting point for those thinking of contributing to CuPy.

APIs

Array creation functions

• numpy.frombuffer
• numpy.fromfunction
• ~numpy.fromregex~ (not for CuPy, as CuPy does not support structured array yet)
• numpy.fromstring
• numpy.genfromtxt
• numpy.loadtxt

Array serialization functions

• numpy.array2string
• numpy.savetxt

Steps to contribute

• Fork the CuPy repository.
• Implement a function. You can implement these APIs in a similar way as cupy.fromfile: cupy/_creation/from_data.py
• Add the function to cupy/__init__.py.
• Write a test code. (example: cupy.fromfile test: tests/cupy_tests/creation_tests/test_from_data.py)
• Build CuPy locally, then run the test code to confirm that the function runs fine. (in short: pip install -e . && pytest tests/cupy_tests/creation_tests/test_from_data.py. See the Contribution Guide for details.)
• Add the function to the API reference. (example: cupy.fromfile is at docs/source/reference/creation.rst)
• Submit a pull-request.

[SaharCarmel]

I’m having trouble to decide where to place all the info in the .rst files. Since fromstring and from regex can be regarded as a creation function from existing data. What do you think?

Ah, sorry I overlooked that they are listed twice in the NumPy API reference… Could you put fromregex, fromstring and loadtxt in cupy/_creation/from_data.py in that case, and list it twice in both rst files?

Sure.

[kmaehashi]

About the Array serialization functions. Where those functions should be written at?

You can tell where to implement a function from a filename and a section name of the NumPy API Reference. For example, numpy.array2string is documented at https://numpy.org/doc/stable/reference/routines.io.html (io) and is under String formatting section, so cupy/_io/formatting.py is the right place to put it.