Issue 23632: memoryview doesn't allow tuple-indexing

It is a bit of pity. A least non-sliced indexing should be able to work:

>>> import numpy as np
>>> a = np.arange(10)
>>> memoryview(a)[0]
0
>>> a = np.arange(10).reshape((2,5))
>>> a[0,1]
1
>>> memoryview(a)[0,1]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: memoryview: invalid slice key

Here is a patch.

Aren't there wider implications of python starting to support tuple indexing?  If we make this work, aren't people going to expect [[1,2], [3,4]][0,1] to work?

> If we make this work, aren't people going to expect [[1,2], [3,4]][0,1] to work?

Or even [[1,2], [3,4]][*(0,1)] :-)

But seriously, I don't know. memorview is a pretty specialized object, its semantics have more to do with Numpy (which has been a major inspiration and use case for the buffer protocol's design) than standard Python containers. I wouldn't expect a list of lists to support tuple indexing.

Multi-dimensional slicing is explicitly mentioned in PEP-3118, so
I guess the intention was to cover this use case as well.

To be sure, the PEP contains some odd proposals like "Unpacking a long-double will return a decimal object", but the one in this issue seems reasonable.

May be add support of multi-dimensional sub-views?

>>> m = memoryview(bytearray(range(12)))
>>> m2 = m.cast('B', (3,4))
>>> m2[0]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NotImplementedError: multi-dimensional sub-views are not implemented

Then we could use m2[0][1].

Sub-views are not as easily implemented. Besides, they create an intermediate object which is immediately thrown away. And tuple indexing is a very common idiom in the Numpy universe.

sub-views and multi-dimensional slicing are not that bad: They're already implemented in _testbuffer (indeed with intermediate objects though).

Well, we could certainly do both :) I don't think we should weigh this issue with separate features, though.

Yes, to be clear I'm +1 on this specific feature -- and separate issues. :)

It turns out that msg237933 was too simplistic: There are some
intricacies of numpy array indexing that *are* very complex:

  http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html


In particular, x[::2, 1, ::2] is not necessarily equal to
x[::2][1][::2].

Do you want to review the patch?

I left few nitpicks on Rietveld, but in general the patch LGTM.

Oh, needed a documentation.

Updated patch addressing Serhiy's comments.

It would be good to add examples with multidimensional memoryviews. For example:

>>> a = array.array('l', [-11111111, 22222222, -33333333, 44444444])
>>> m = memoryview(a).cast('B').cast('l', [2, 2])
>>> m[0, 0]
-11111111
>>> m[1, 0]
-33333333
>>> m[-1, -1]
44444444

New changeset 7d4eb5902f82 by Antoine Pitrou in branch 'default':
Issue #23632: Memoryviews now allow tuple indexing (including for multi-dimensional memoryviews).
https://hg.python.org/cpython/rev/7d4eb5902f82

I didn't add the 2d example as the double cast is quite ugly.  Probably we should allow casting when the format stays the same.

stage: patch review -> resolved

messages: + msg238479