Leo Fang
[We don't guarantee ordering/strides](https://docs.cupy.dev/en/stable/user_guide/difference.html#contiguity-strides) even when an API has a NumPy counterpart, not to mention this is semi-private API. Furthermore, CuPy should take care of array access regardless of the...
I think devil's in the details. In CuPy, array operations like transpose, swap_axes, etc are just simple metadata manipulations (recalculating the strides without actually copying data around), same as in...
They look like pure Python code (NumPy/CuPy-based) to me. I think this discussion would likely lead to a premature optimization. Why don't you proceed to complete a drop-in replacement, before...
Numba's CUDA target is capable of handling F-order arrays too. Strides are always passed to the device as part of the internal kernel launch arguments.
It has never been implemented since the debut, so it is not a bug: https://github.com/cupy/cupy/blob/f05ce6a7b64ad35253a95d6d02e7481a4f3e700d/cupyx/scipy/sparse/_base.py#L43-L46
I've been hoping to refactor this page. I think CAI and DLPack should be introduced as two separate sections, followed by examples of how they can be used for "CuPy...
Sorry for delay, will try to finish my review in a day or two!
cc: @grlee77 for vis
Rendered docs: https://cupy--9514.org.readthedocs.build/en/9514/user_guide/performance.html#profiling-kernels-with-nsight-compute
It is quite shocking that the Copilot CLI can't even search GitHub's own issues and PRs. I would suggest to prioritize accessing GitHub's own resources, like what the web chat...