Restrict cuDNN reduction to contiguous inputs.

theano/gpuarray/c_code/dnn_redux.c

@@ -61,6 +61,11 @@ int APPLY_SPECIFIC(dnn_redux)(PyGpuArrayObject *input,
   static float fbeta = 0.0f;
   static double dbeta = 0.0;
 
+  if (!GpuArray_IS_C_CONTIGUOUS(&input->ga)) {
+    PyErr_SetString(PyExc_ValueError, "Only contiguous inputs are supported.");
+    return 1;
+  }
+
   if (c_set_tensorNd(input, APPLY_SPECIFIC(input)) != 0)
     return 1;
 

theano/gpuarray/dnn.py

@@ -1682,6 +1682,7 @@ class GpuDnnReduction(DnnBase):
     def make_node(self, inp):
         ctx_name = infer_context_name(inp)
         inp = as_gpuarray_variable(inp, ctx_name)
+        inp = gpu_contiguous(inp)
         if inp.ndim > 8:
             raise ValueError("cuDNN reduction doesn't support nd > 8")
         assert inp.dtype in ['float16', 'float32', 'float64']