Why Can I Not Replicate the NumPy v1.13.dev0 Manual Examples

Question

As an example I am reading through the following:

https://docs.scipy.org/doc/numpy-dev/neps/new-iterator-ufunc.html

So I run a bit of code on my computer using iPython shown for example in the manual as:

   def iter_add_itview(x, y, out=None):
        it = np.nditer([x,y,out], [],
                    [['readonly'],['readonly'],['writeonly','allocate']])

        (a, b, c) = it.itviews
        np.add(a, b, c)

        return it.operands[2]

Which per the example results in tests cases as follows:

In [10]: a = np.arange(1000000,dtype='f4').reshape(100,100,100).T
In [12]: b = np.arange(10000,dtype='f4').reshape(100,100,1).T
In [11]: c = np.arange(10000,dtype='f4').reshape(1,100,100).T

In [4]: timeit np.add(np.add(np.add(a,b), c), a)
1 loops, best of 3: 99.5 ms per loop

In [9]: timeit iter_add_itview(iter_add_itview(iter_add_itview(a,b), c), a)
10 loops, best of 3: 29.3 ms per loop

SO naturally I want to try this excitement for myself on using NumPy1.12.1 in Python 2.7 on Linux with an Intel chipset, the only problem is I consistently get null results for the exact same experimental setups as in the example above:

In [12]: timeit np.add(np.add(np.add(a,b), c), a)
100 loops, best of 3: 10.7 ms per loop

In [13]: timeit iter_add_itview(iter_add_itview(iter_add_itvie
    ...: w(a,b), c), a)
100 loops, best of 3: 10.7 ms per loop

In this case I was supposed to be seeing improvements from buffer cache optimization.

Why is it that I am not able to replicate a section of the dev manual for a updated NumPy version using a legacy version of the code?

Those improvements have probably been folded into the standard ufunc implementation already. — user2357112
– user2357112, Commented Apr 28, 2017 at 20:31
They don't take NEPs out of the docs just because they've been implemented. They just put them under the "Implemented NEPs" section, like this one is. — user2357112
– user2357112, Commented Apr 28, 2017 at 20:36
Go to the NEP overview page, and you'll see this one in the Implemented section. — user2357112
– user2357112, Commented Apr 28, 2017 at 20:38
Note that the document you're reading is from 2010. It's over 6 years old. — user2357112
– user2357112, Commented Apr 28, 2017 at 20:39

hpaulj · Accepted Answer · 2017-04-29 00:07:31Z

2

Just to be clear the In[4] and [9] times are copied from that 6 yr old document. The [12][13] times are from your own tests (and basically the same as what I get). I haven't studied that old dev document, but have run the np.nditer examples on

https://docs.scipy.org/doc/numpy/reference/arrays.nditer.html

and know that nditer does not normally speed code. This iterator is meant to be used in C level code, and its exposure at the Python level is a convenience. It lets us test ideas in Python code before moving them to C or Cython. Note the Cython example at the end. Even so, I found I could get better speed in a simple multiplication case by using Cython memoryviews.

np.ndindex is one of the few numpy functions that uses np.nditer in Python code. I've occasionally suggested a similar pattern to produce depth limited iteration.

So don't worry too much about mastering np.nditer.

edited Apr 29, 2017 at 0:07

answered Apr 28, 2017 at 22:21

hpaulj

233k14 gold badges260 silver badges392 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Why Can I Not Replicate the NumPy v1.13.dev0 Manual Examples

1 Answer 1

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related