Repository navigation
Add cuda.core.utils.get_aligned_dtype #734
Description
Many CUDA Python projects require the ability to declare a custom alignment for a give dtype. While NumPy builtin/structure dtypes work in most cases, this is one of the few issues that it cannot fulfill today. @seberg has figured out a prototype to address this issue (numpy/numpy#28318 (comment)), and the plan is for cuda.core to provide this infrastructure, let other teams start using it, gather feedbacks, and eventually upstream this back to NumPy.
Metadata
Metadata
Assignees
Type
Projects
- Status
Activity
leofang commentedon Jun 27, 2025
cc @shwina @tpn @ZzEeKkAa @isVoid for vis
leofang commentedon Aug 20, 2025
@ZzEeKkAa does Sebastian's prototype work for you out of box, or does it require additional tweaks?
ZzEeKkAa commentedon Sep 9, 2025
Sebastian's prototype works out of the box as an intermediate solution for
nvmath. Howevercomplex*types as well as half precision packed complex type are required3 remaining items
seberg commentedon May 29, 2026
We now have
cupy.make_aligned_dtypeto help with structured dtypes.(This may not solve all issues around this e.g. for the numba side of things, but should help.)