Added NotImplemented error to Array.concatenate

[panchoop]

As suggested in this issue, a warning was added both in the code and in the documentation.

[inducer]

This isn't correct. This should instead be based on contiguity. Column-major arrays for example allow concatenation along the last axis.

[panchoop]

I noticed that the code failed the tests because I missed a syntax error. Sorry for the noise.

About contiguity, I did some tests, and the function does not work in any axis for colum-major arrays. In line 2571 of the Array.concatenate function, the output array result is always a row-major array, leading to the same NotImlementedError due to miss-matching strides.

Simple example:

import pyopencl as cl
platform = cl.get_platforms()[0]
device = platform.get_devices()[0]
context = cl.create_some_context([device])
queue = cl.CommandQueue(context)

a = np.random.rand(2,2)
b = np.random.rand(2,2)

a_cl = clarray.to_device(queue, np.require(a, np.float64, 'F'))
b_cl = clarray.to_device(queue, np.require(b, np.float64, 'F'))

c_cl = clarray.concatenate( (a_cl, b_cl), axis=1)

It shouldn't be hard to allow column-major arrays, it is just a matter of initializing the result array with order="F" in these cases. The issue I see is that the Array class has no member that indicates the ordering. One could deduce the order by looking at the strides (if Array.strides[0] > Array.strides[-1]: row-major, else: column-major) but maybe it is too hacky. what do you think?

Another approach is the one taken in the Array.stack function, in which when the axis != 0, it is assumed that the input is column mayor. But it would also lead to unclear error messages.

[inducer]

Thanks for pointing out the issue with concatenate and column-major arrays, we should look into fixing that. I agree with your assessment about inspection of the strides, it's not quite elegant.

FWIW, CI is still not passing on this branch.

[panchoop]

Oh yeah! I forgot to push. It should work now.