Use numpy.nbytes rather than memoryview.nbytes

[mrocklin]

This avoids a python 2 issue

[pitrou]

Looks fine to me.

[mrocklin]

Running into other issues here that you might be able to help with actually @pitrou. I like memoryviews, but I'm having trouble creating a very simple memory view with a linear shape and itemsize=1.

I can get to itemsize=1 with my_numpy_array.view('u1').data, but I'm not sure how to flatten out a memoryview's shape in a way that doesn't involve a copy on the numpy side.

[mrocklin]

I want this:

x = np.ones((10, 10, 10, 10, 10, 10, 10))
x.flatten().view('u1').data

But flatten involves a copy. I have yet to find find API either in numpy or the Python memoryview that lets me do this without a copy.

[mrocklin]

Ah, ravel

[pitrou]

The cast() method should work (but perhaps not on Python 2):

>>> x = np.ones((10, 10, 10, 10, 10, 10, 10))
>>> m = memoryview(x)
>>> m.shape, m.nbytes, m.itemsize
((10, 10, 10, 10, 10, 10, 10), 80000000, 8)
>>> m = m.cast('B')
>>> m.shape, m.nbytes, m.itemsize
((80000000,), 80000000, 1)

[pitrou]

array.ravel() will make a copy if necessary (if the array isn't contiguous), while memoryview.cast() will refuse to work on non-contiguous arrays.

[mrocklin]

I think we want ravel then. We're willing to copy. #833