CUDA Dispatcher: use default finalizer

- For device functions, this works exactly like on the CPU target because device functions are CompileResult objects that are inserted into the target context (and therefore need removing by the finalizer). - For kernels, we give them a dummy entry point because they were never inserted and don't need removing (similar to with object mode functions).
sklam · Feb 18, 2022 · Feb 4, 2022 · Feb 4, 2022 · Feb 4, 2022 · Feb 4, 2022
commit 5d6384c58e531509fca7a190fc17357397ec2558
@@ -50,6 +50,14 @@ def __init__(self, py_func, argtypes, link=None, debug=False,
         # target with slight modifications.
         self.objectmode = False
 
+        # The finalizer constructed by _DispatcherBase._make_finalizer also
+        # expects overloads to be a CompileResult. It uses the entry_point to
+        # remove a CompileResult from a target context. However, since we never
+        # insert kernels into a target context (there is no need because they
+        # cannot be called by other functions, only through the dispatcher) it
+        # suffices to pretend we have an entry point of None.
+        self.entry_point = None
+
         self.py_func = py_func
         self.argtypes = argtypes
         self.debug = debug
@@ -485,13 +493,6 @@ def __init__(self, py_func, targetoptions, pipeline_class=CUDACompiler):
         # argument types
         self.specializations = {}
 
-    def _make_finalizer(self):
-        # Dummy finalizer whilst _DispatcherBase assumes the existence of a
-        # finalizer
-        def finalizer():
-            pass
-        return finalizer
-
     @property
     def _numba_type_(self):
         return cuda_types.CUDADispatcher(self)