rocm

I believe this is undocumented behaviour.

import numba as nb


@nb.njit
def f1():
    for i in nb.prange(1):
        print(type(i))  # >>> int64


@nb.njit(parallel=True)
def f2():
    for i in nb.prange(1):
        print(type(i))  # >>> uint64


f1()
f2()

This caused a nasty bug in my own code that was hard to debug as the problem did not exist without `parallel=Tr

Please go to https://github.com/deepmodeling/deepmd-kit/discussions to ask questions. Before asking questions, you can

• search the previous discussions
• check the document, especially training parameters

Please provide necessary information including the version of software and installation way,

Bug summary
There is evidence that sub_group::get_group_id() does not return the same value as threadIdx.x / warpSize (assuming 1D kernel), as expected on CUDA. We should check the implementation of this function. Our implementation of this function performs bit manipulation magic, presumably the optimization went to far...

To Reproduce
Compare sub_group{}.get_group_id() or `sub

Since arrays may not actually be modified by a given operation, or might only be partially modified (or the user has some other way to ensure correctness).