Override perform() (otherwise DebugMode and py linker would use it)

theano/sandbox/gpuarray/elemwise.py

@@ -596,6 +596,8 @@ class GpuCAReduce(GpuOp):
         # been added to the python bindings yet
         z[0] = x.reduce_sum(self.reduce_mask)
     """
+    def perform(self, node, inp, out):
+        raise MethodNotDefined("")
 
     def supports_c_code(self, inputs):
         """ Returns True if the current op and reduce pattern

theano/sandbox/gpuarray/tests/test_elemwise.py

@@ -72,3 +72,9 @@ class test_GpuCAReduce(test_GpuCAReduceCPY):
     bin_dtypes = ["uint8", "int8"]
     op = GpuCAReduce
     reds = [scalar.add, scalar.mul]
+
+    def test_perform(self):
+        return
+
+    def test_perform_nan(self):
+        return