MemoryPointer: ensure CUdeviceptr used with NVIDIA binding (#328)

* `MmemoryPointer` ensure `CUdeviceptr` used with NVIDIA binding

The docs state that `pointer` should be a `c_void_p`, but the rest of
the driver binding code expects the pointer to be represented as a
`CUdeviceptr` when the NVIDIA bindings are in use.

If we're given a `c_void_p` by the user when using the NVIDIA bindings,
we can work around this by converting it to a `CUdeviceptr`.

This was first noticed when testing PyArrow with Numba-CUDA. See
apache/arrow#47128

* Correct device pointer value usage in EMM test

Author: gmarkall

numba_cuda/numba/cuda/cudadrv/driver.py

@@ -2062,6 +2062,9 @@ class MemoryPointer(object):
     __cuda_memory__ = True
 
     def __init__(self, context, pointer, size, owner=None, finalizer=None):
+        if USE_NV_BINDING and isinstance(pointer, ctypes.c_void_p):
+            pointer = binding.CUdeviceptr(pointer.value)
+
         self.context = context
         self.device_pointer = pointer
         self.size = size

numba_cuda/numba/cuda/tests/cudadrv/test_emm_plugins.py

@@ -84,7 +84,7 @@ def get_ipc_handle(self, memory):
             # the tests don't try to do too much with it (e.g. open / close
             # it).
             self.get_ipc_handle_called = True
-            return "Dummy IPC handle for alloc %s" % memory.device_pointer.value
+            return "Dummy IPC handle for alloc %s" % memory.device_pointer_value
 
         @property
         def interface_version(self):