On Sun, 2017-05-28 at 14:53 -0600, Charles R Harris wrote:
Hi All, This post is to open a discussion of the future of ufuncs. There are two contradictory ideas that have floated about regarding ufuncs evolution. One is to generalize ufuncs to operate on buffers, essentially separating them from their current entanglement with ndarrays. The other is to accept that they are fundamentally part of the ndarray universe and move them into the multiarray module, thus avoiding the odd overloading of functions in the multiarray module. The first has been a long time proposal that I once thought sounded good, but I've come to prefer the second. That change of mind was driven by the resulting code simplification and the removal of a dependence on a Python feature, buffers, that we cannot easily modify to adapt to changing needs and new dtypes. Because I'd like to move the ufuncs, if we decide to move them, sometime after NumPy 1.14 is released, now seems a good time to decide the issue. Thoughts?
I did not think about it much. But I agree that the dtypes are probably the biggest issue, also I am not sure anymore if there is much of a gain on having ufuncs work on buffers in any case? The dtype thing goes a bit back to ideas like the datashape things and trying to make the dtypes somewhat separate from numpy? Though I doubt I would want to make that an explicit goal. I wonder how much of the C-loops and type resolving we could/should expose? What I mean is that ufuncs are: * type resolving (somewhat ufunc specific) * outer loops (normal, reduce, etc.) using nditer (buffering) * inner 1d loops It is a bit more complicating, but just wondering if it might make sense to try and expose the individual ufunc things (type resolving and 1d loop) but not all the outer loop nditer setup which is ndarray specific in any case (honestly, I am not sure it is entirely possible it is already exposed). - Sebastian
Chuck _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion