2

I have a 1D numpy array containing complex numbers (eg. numpy.complex64). I need to create views on this array, but I don't know how to create a view given a list or range of indices to include in the view.

>>> myArray = np.ndarray(shape=(1000,), dtype=np.complex64)

I know how to create a view on consecutive items, for example the first 100 elements:

>>> myView = myArray[:100]

As I experimented, it's not possible to create a view on a singular item, like myArray[2] because if I modified that value, it wouldn't change the underlying array. This is OK, but I hope there is a way to construct a view from multiple arbitrary indices, so a function like this would be great:

>>> myView = createView(myArray, indices=(0, 1, 6, 7, 13))

which would return a view that points to those indices given in a list (or any iterable form), and if I change myView, it would change myArray like it should.

Is this possible, or is there a sane workaround? Thanks

edit:

To explain why I need this: I want to copy the array to OpenCL device memory (and back). I need a different order of the elements on the device than in the original array, and these shorter arrays would be treated as vectors, and matrix multiplication will take place in the OpenCL kernel. If the elements are not copied to the device memory in the order that is required, memory coalescing wouldn't be possible, and significant performance decrease would occur, not to mention the additional logic that would be needed on OpenCL side.

In my case, it is possible to do this with evenly spaced views, but I wonder if there's a more general way. A small CPU-side performance decrease is OK if a more general implementation can be done.

Improve this question
2
  • It's not really possible to make an unevenly-spaced view. But it is possible to use a list of arbitrary indices in various ways. What do you need it for? Commented Nov 27, 2011 at 12:01
  • Actually I use up all the indices of the original array once, and want to create multiple arrays that have shorter length, and reference to the original array. (It can be done with evenly spaced views in my case, but I wanted to know if it's possible in a more general way.) I want to copy the data to an OpenCL device memory, and use PyOpenCL InOut for the filtered arrays, to change the correct indices in the original array. Commented Nov 27, 2011 at 12:37

3 Answers 3

Reset to default
2

As far as I know, you can't have a view on arbitrary indices; only on consecutive indices or regularly spaced indices. This is because the elements in the view's underlying memory store have to be separated by a constant number of bytes, otherwise all the fast NumPy routines don't work.

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

1 Comment

I want to copy "the views" to an OpenCL device, and do computations on the data, so in my case the CPU side of the code need not to be perfectly optimized. As Alexandru suggested, I'll look into filters or masks to achieve this.
1

The following demonstrates the OP's request:

>>> a = numpy.array(xrange(103,114))
>>> a
array([103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113])
>>> view = numpy.array([2,6,7,9])
>>> a[view]
array([105, 109, 110, 112])
>>> a[view] += 9999
>>> a
array([  103,   104, 10104,   106,   107,   108, 10108, 10109,   111,
       10111,   113])
Improve this answer

Comments

0

As far as I know views on ndarray are generated like this: myArray.view(dtype). Here is the documentation: http://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.view.html

It does not answer to your non-consecutive items problem, but I think it's a good start.

I think you should create another filtered array with your indexes. Because I view doesn't really have to do with filtering, but rather with different representation of the same data.

Improve this answer

Comments

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.