Fix problems.

theano/gpuarray/dnn.py

@@ -1620,7 +1620,7 @@ class GpuDnnReduction(DnnBase):
 
         bcast = []
         for i in range(inp.ndim):
-            if not (self.axis & (1 << i)):
+            if not (self.c_axis & (1 << i)):
                 bcast.append(inp.broadcastable[i])
         outs = [inp.type.clone(dtype=self.dtype, broadcastable=bcast)()]
         if self.arg:

theano/gpuarray/tests/test_opt.py

@@ -17,6 +17,7 @@ from ..basic_ops import (
 from ..blas import GpuGemm
 from ..elemwise import (
     GpuCAReduceCuda, GpuCAReduceCPY, GpuElemwise, Elemwise, max_inputs_to_GpuElemwise)
+from ..dnn import GpuDnnReduction
 from ..subtensor import GpuSubtensor
 from ..linalg import GpuCusolverSolve, cusolver_available, GpuCholesky