Ending of Phase-1..

Hey! This is the third blog-post in my series of GSoC experience. Phase-1 has come to its end. I’ll be sharing my experiences of third and fourth week of GSoC.

Third week

In third week of my coding, I did many minor fixes and improvements. It includes _RecursiveAttr.__repr__, better error handling and other changes suggested by my mentor. Thanks to Imanishi-san, he provided ideas to make existing code cleaner. So, there was some re-implementation of _Recursiveattr.__getattr__ and data transfer methods. fallback_mode now supports direct retrieval of NumPy scalars. I also implemented some tests for fallback_mode.

Fourth week

In this week, I again did some minor fixes, moved fallback_mode to cupyx/fallback_mode and implemented a test decorator for testing fallback methods. PR #2229 got merged and fallback_mode is added for release in cupy v7.0.0b1. I am super happy, as this was my first true opensource contribution.

Now, I am currently working on next goal, that is to implement fallback_mode.ndarray. It’s needed in order to apply fallback on ndarrays. Supporting methods such as ndarray.min(), ndarray.reshape() is easy. But, supporting magic methods such as __add__, __neg__ won’t work, as they bypasses __getattr__ in new-style classes. So, for the last couple of days, I was researching about all this and found out, that this could be overcame by proxying magic methods of cupy.ndarray to fallback_mode.ndarray. So, for now, working on it.

Summary

fallback_mode now has __repr__ and basic tests are done. It now supports methods of following types:

• result = numpy.func(*args, **kwargs)
• result = numpy.module.func(*args, **kwargs)
  where, execution steps are performed on args, kwargs and result is returned.

Following is how the fallback_mode currently works.

>>> from cupyx.fallback_mode import numpy
>>> numpy.ones
<numpy = <function ones at 0x000002AC4E0AC598>, cupy = <function ones at 0x000002AC4EB977B8>>
>>> numpy.nanargmin
<numpy = <function nanargmin at 0x000002AC4E1FA7B8>, cupy = None>
>>> numpy.pi
3.141592653589793
>>> numpy.nan
nan
>>> a = numpy.array([[1,numpy.nan,2],[3,4,numpy.nan]])
>>> type(a)
<class cupy.core.core.ndarray>
>>> res = numpy.nanargmin(a, axis=0)
>>> res
array([0, 1, 0], dtype=int64)
>>> type(res)
<class cupy.core.core.ndarray>
>>> numpy.linalg
<numpy = module numpy.linalg, cupy = module cupy.linalg>
>>> numpy.matrixlib
<numpy = module numpy.matrixlib, cupy = module None>

Future work

Future work will include:

• Support fallback for fallback_mode.ndarray.
• Notification mechanism, similar to that of numpy.seterr.
• Supporting methods such as ndarray.resize() where nothing is returned, but references are modified.