Last month of GSoC..

05 Aug 2019

Reading time ~2 minutes

Hey! This is the sixth blog-post in the series of GSoC blog-posts. 2^nd evaluations are done and this is going to be the last month of GSoC. I think, this month will be the best of all. I am going to go deeper into CuPy for creation of view of ndarray (ndarray sharing same memory space with another ndarray) and deal with core ndarray implementation, cuda.MemoryPointer, __cuda_array_interface__(CuPy), __array_interface__(NumPy) etc. for dealing with memory location and stuff. It’s going to be amazing!

Ninth week

I was on leave for about 3-4 days due to my college re-opening. Still, I was able to do some work. I implemented cupy.nanvar and cupy.nanstd using ReductionKernel. It was something new so, I enjoyed it.
Then, I implemented fallback.vectorize (a wrapper around numpy.vectorize). I faced some difficulty for the elimination of recursion error, but it was resolved without taking too much time.

Tenth week

This week I implemented support for in-place operations. It was decided to update all args and kwargs as an initial solution. I created some tests to cover special cases too such as numpy.putmask, ndarray.byteswap etc. and it’s working as expected.
Then, I went ahead for adding support for view. I was able to create a view of cupy.ndarray with desired shape[0]. Manipulating zero-axis was quite straight forward, but it will be way much harder to support all axes. Following is the prototype:

import cupy as cp
import numpy as np
​
a = np.ones((3, 4))
b = a[1:]
​
x = cp.ones((3, 4))
# we need to create view wrt x, as b is a view wrt a
​
def getptr(a):
    if isinstance(a, np.ndarray):
        return a.__array_interface__['data'][0]
    if isinstance(a, cp.ndarray):
        return a.data.ptr
    raise TypeError
​
if b.base is a:
    offset = getptr(b) - getptr(a)
    memptr = x.data + offset
    y = cp.ndarray(b.shape, x.dtype, memptr)
​
y[0][0] = 0
print(x)
'''
[[1. 1. 1. 1.]
 [0. 1. 1. 1.]
 [1. 1. 1. 1.]]
'''

Summary

Implemented support for in-place operations.
PR #2350 for fallback.vectorize got merged.
PR #2344 for cupy.nanvar and cupy.nanstd got merged.

Future work

Future work will consist of following:

Supporting view of ndarray.
Supporting of numpy.ndarray variants such as numpy.ma.MaskedArray, numpy.chararray etc.