Hi all,
I’ll be presenting some slides at this public dev meeting discussing my Numpy subclassing PR. The purpose of this PR is to allow Numba to operate on ndarray and potentially multiple other subclasses of ndarray within the same jitted function while preserving their individual types. The PR implements the Numpy array_ufunc API in Numba so that you can control which array type is generated for different ufuncs for different combinations of array types. Also, this PR changes the memory subsystem and allows you to register a custom allocator for a given Numpy subclass. New arrays of that type would use the custom allocator and the object carries information about the allocator used so that at the time of destruction we are able to call the associated custom deallocator.
I welcome anyone who is working on related issues to come to the meeting and contribute to the discussion so that these additional Numba capabilities can generalize to help everyone and not just our particular use-case.