[BUG] The libcudf device atomic operations cause memcheck errors on types smaller than 4 bytes

[davidwendt]

The device_atomics implementation in libcudf are failing memchecks when used on types smaller than 4 bytes.

cudf/cpp/include/cudf/detail/utilities/device_atomics.cuh

Lines 65 to 69 in c929a84

	auto* address_uint32 = reinterpret_cast<T_int*>(addr - (reinterpret_cast<size_t>(addr) & 3));
	T_int shift = ((reinterpret_cast<size_t>(addr) & 3) * 8);
	T_int old = *address_uint32;
	T_int assumed;

This was found while investigating some memcheck errors that occurred on cudf::reduce where the aggregation return type is bool (1 byte) and an atomicOr was used to produce the output. Reference: #13574

The solution is to use the libcudaxx cuda::atomic and cuda::atomic_ref in place of these implementations and potentially remove the device_atomics.cuh altogether.

[PointKernel]

Depends on NVIDIA/cccl#1024

[PointKernel]

Update: atomic_ref is limited to hardware-supported types, so it doesn't work properly with types like booleans, timestamps, and durations. As a result, we can't fully eliminate device_atomics.cuh. This issue should be addressed alternatively.

[davidwendt]

Closed by #13583 and NVIDIA/cccl#1024