Construct pylibcudf columns from objects supporting __cuda_array_interface__
#15615
Conversation
brandon-b-miller
added
feature request
| data = gpumemoryview(obj) | ||
| iface = data.__cuda_array_interface__() | ||
| data_type = _data_type_from_iface(iface) | ||
| size = iface['shape'][0] |
There was a problem hiding this comment.
issue: Someone needs to check that the object being described is one-dimensional and contiguous. So we should probably:
| size = iface['shape'][0] | |
| (size,) = iface['shape'] |
And also assert that strides are appropriate (either None or (1,) I think?)
There was a problem hiding this comment.
We implemented something similar for buffer here, but it's in the cuDF python namespace. Would it be overkill to move the impl to a cython utility we can consume from both places?
There was a problem hiding this comment.
No, because we need to check here too, otherwise we're potentially producing a nonsense result.
Co-authored-by: Lawrence Mitchell <wence@gmx.li>
|
Some failing tests locally seem to be because our datetime column supports |
Not sure if it's intended but appears like an oversight. The interface for numeric and datetime columns are exactly the same, so I suppose timedelta should have support since it's a supported type code |
…15622) Column types that support CAI already have custom `NotImplementedError`s, and since the implementation is the same for datetime and numeric columns, moving their implementation to `ColumnBase` Should help address timedelta support in #15615 cc @brandon-b-miller Authors: - Matthew Roeschke (https://github.com/mroeschke) Approvers: - GALI PREM SAGAR (https://github.com/galipremsagar) URL: #15622
|
Timedelta columns should work now with #15622 in |
Thanks @mroeschke tests pass locally for me now. |
|
/merge |
This PR allows zero copy construction of
pylibcudfcolumns from device arrays via thegpumemoryviewclass. cc @mroeschke