cuda-python-bugfix: allow passage of cuda stream objects or pointers #419
Conversation
Pointers are dangerous because they aren't reference counted and so can be deleted before cufinufft cleans up (flatironinstitute#417). This patch allows the user to pass either a stream object OR a pointer in the `gpu_stream` argument of Plan
|
Hi @janden - any chance you can check over this PR sometime? Thanks, Alex |
There was a problem hiding this comment.
Sorry for the delay. This looks great!
If I understand correctly, we still have an issue if a user passes a raw pointer (since we can't make sure it's properly reference-counted)? Then yes, it would make sense to remove this possibility (can't imagine that anyone would use it, especially given that we require the input array to be from one of the frameworks). However, since it's already part of the 2.2 API, maybe we deprecate for now and remove in 2.4?
I've added a deprecation warning when it finds a raw pointer. I'll merge tomorrow barring any objections |
|
👍 |
|
closes #417 |
Pointers are dangerous because they aren't reference counted and so can be deleted before cufinufft cleans up (#417). This patch allows the user to pass either a stream object OR a pointer in the
gpu_streamargument of PlanMight warrant discussion of if we should just scrap the raw pointer altogether