Re-put test_dnn_reduction_sum_squares into test_dnn.

theano/gpuarray/tests/test_dnn.py

@@ -1569,6 +1569,20 @@ def test_dnn_reduction_opt():
         yield dnn_reduction, 2, idtype, adtype, odtype
 
 
+def test_dnn_reduction_sum_squares():
+    if not dnn.dnn_available(test_ctx_name) or dnn.version(raises=False) < 6000:
+        raise SkipTest(dnn.dnn_available.msg)
+
+    M = T.matrix()
+    for axis in (None, 0, 1):
+        out = (M**2).sum(axis=axis)
+        f = theano.function([M], out, mode=mode_with_gpu)
+        assert any(isinstance(node.op, dnn.GpuDnnReduction) and node.op.red_op == 'norm2'
+                   for node in f.maker.fgraph.apply_nodes)
+        M_val = np.random.random((4, 5)).astype(theano.config.floatX)
+        utt.assert_allclose((M_val**2).sum(axis=axis), f(M_val))
+
+
 def dnn_reduction_strides(shp, shuffle, slice):
     utt.fetch_seed()
     inp = GpuArrayType('float32', (False,) * len(shp),

theano/gpuarray/tests/test_opt.py

@@ -382,21 +382,6 @@ def test_local_gpu_elemwise_careduce():
     utt.assert_allclose(f(data), (data * data).sum(axis=1))
 
 
-def test_dnn_reduction_sum_squares():
-    if not dnn.dnn_available(test_ctx_name) or dnn.version(raises=False) < 6000:
-        raise SkipTest(dnn.dnn_available.msg)
-
-    M = tensor.matrix()
-    for axis in (None, 0, 1):
-        out = (M**2).sum(axis=axis)
-        f = theano.function([M], out, mode=mode_with_gpu)
-        assert any(isinstance(node.op, dnn.GpuDnnReduction) and node.op.red_op == 'norm2'
-                   for node in f.maker.fgraph.apply_nodes)
-        assert _check_stack_trace(f)
-        M_val = np.random.random((4, 5)).astype(theano.config.floatX)
-        utt.assert_allclose((M_val**2).sum(axis=axis), f(M_val))
-
-
 def test_local_lift_dot22scalar():
     x = tensor.matrix()
     y = tensor.matrix()