Also check against CPU implementation.

theano/gpuarray/tests/test_dnn.py

@@ -1621,6 +1621,7 @@ def test_dnn_reduction_axis_size_one():
             sum_squares = (x**2).sum(axis=axis)
             sum_abs = abs(x).sum(axis=axis)
             absmax = abs(x).max(axis=axis)
+            cpu_f = theano.function([x], [sum, sum_squares, sum_abs, absmax], mode=mode_without_gpu)
             f1 = theano.function([x], sum, mode=mode_with_gpu)
             f2 = theano.function([x], sum_squares, mode=mode_with_gpu)
             f3 = theano.function([x], sum_abs, mode=mode_with_gpu)
@@ -1631,6 +1632,11 @@ def test_dnn_reduction_axis_size_one():
             xval_reshaped = xval.reshape(shape[:axis] + shape[(axis + 1):])
             test_val = abs(xval_reshaped)
             val_sum, val_sum_squares, val_sum_abs, val_absmax = f1(xval), f2(xval), f3(xval), f4(xval)
+            cpu_val_sum, cpu_val_sum_squares, cpu_val_sum_abs, cpu_val_absmax = cpu_f(xval)
+            utt.assert_allclose(cpu_val_sum, val_sum)
+            utt.assert_allclose(cpu_val_sum_squares, val_sum_squares)
+            utt.assert_allclose(cpu_val_sum_abs, val_sum_abs)
+            utt.assert_allclose(cpu_val_absmax, val_absmax)
             utt.assert_allclose(xval_reshaped, val_sum)
             utt.assert_allclose(test_val**2, val_sum_squares)
             utt.assert_allclose(test_val, val_sum_abs)