Tell the limitation of the pooling on the GPU.

theano/tensor/signal/downsample.py

@@ -80,13 +80,16 @@ def max_pool_2d(input, ds, ignore_border=None, st=None, padding=(0, 0),
     if input.ndim < 2:
         raise NotImplementedError('max_pool_2d requires a dimension >= 2')
     if ignore_border is None:
-        warnings.warn("max_pool_2d() will have the parameter ignore_border"
-                      " default value changed to True (currently"
-                      " False). To have consistent behavior with all Theano"
-                      " version, explicitly add the parameter"
-                      "  ignore_border=True. (this is also faster than"
-                      " ignore_border=False)",
-                      stacklevel=2)
+        warnings.warn(
+            "max_pool_2d() will have the parameter ignore_border"
+            " default value changed to True (currently"
+            " False). To have consistent behavior with all Theano"
+            " version, explicitly add the parameter ignore_border=True."
+            " On the GPU, using ignore_border=False allow to use cudnn."
+            " Without ignore_border=True and cudnn, the only GPU combination"
+            " supported is when the `ds == st and padding == (0, 0) and"
+            " mode == 'max'`",
+            stacklevel=2)
         ignore_border = False
     if input.ndim == 4:
         op = DownsampleFactorMax(ds, ignore_border, st=st, padding=padding,