Fixed meshgrid to return arrays with same dtype as arguments.

[idfah]

Fixes issue with meshgrid discussed in issue #5297.

A few things to note:

• meshgrid takes multiple arrays as arguments and returns multiple arrays. With this proposed change the dtype of each returned array will have the same dtype as its corresponding argument. An alternative would have been to use the common dtype returned by np.return_type but this didn't seem to make as much sense to me.
• I added a test case to verify that the dtypes of the returned arrays match the arguments.
• I added a test case to test_no_input because this test case would have failed before my patch if copy=False had been set.

[charris]

Might want subok=True here to match the asanyarray above. Otherwise LGTM.

[charris]

Should probably add a check that the input arrays are all 1-D.

[charris]

NVM, looks like that inputs are all flattened, at least if they aren't matrices. That should be in the documentation somewhere.

[charris]

LGTM in general, just need to decide on the subok option for the broadcast values. I'm thinking of arrays with units that some folks are implementing.

[seberg]

Makes sense, but I wonder if we should put it into the release notes. If you used to have an integer arange input and now use meshgrid, it seems like you used to get floats. That could create annoying casting issues.

[wjakob]

I just ran into this and was about to file a bug. What's keeping this from being merged?

[idfah]

I added subok=True in the call to broadcast_arrays as per charris' comments. Can someone please look at merging this? Please let me know if there is anything else I can do.

[rgommers]

Makes sense, but I wonder if we should put it into the release notes. If you used to have an integer arange input and now use meshgrid, it seems like you used to get floats. That could create annoying casting issues.

That doesn't seem to be the case, with numpy 1.11.1:

>>> x = np.arange(2)
>>> y = np.arange(2.)
>>> np.meshgrid(x, y)[0].dtype
dtype('int32')
>>> np.meshgrid(x, y)[1].dtype
dtype('float64')

Changing output dtypes of meshgrid seems like an unhealthy idea.

[rgommers]

Okay, int was treated fine but gh-5297 discussed an issue with float32. Copying here because most of gh-5297 is about something else:

>>> ax, ay = np.meshgrid(a,a)
>>> ax.dtype
dtype('float64')

>>> ax, ay = np.meshgrid(a,a, copy=False)
>>> ax.dtype
dtype('float32')

[rgommers]

makes sense to change that.

[rgommers]

Merged, thanks @idfah