Fix local_gpua_advanced_incsubtensor to return variable on the GPU. op_lifer…

theano/gpuarray/opt.py

@@ -1056,7 +1056,8 @@ def local_gpua_advanced_incsubtensor(op, context_name, inputs, outputs):
         x = x.dimshuffle(0, 'x')
         y = y.dimshuffle('x', 'x')
         ret = GpuAdvancedIncSubtensor1_dev20(
-            set_instead_of_inc=set_instead_of_inc)(x, y, ilist).dimshuffle(0)
+            set_instead_of_inc=set_instead_of_inc)(x, y, ilist)
+        ret = GpuDimShuffle(ret.type.broadcastable, [0])(ret)
         return ret
     elif compute_capability < 2 or x.ndim != 2 or y.ndim != 2:
         return GpuAdvancedIncSubtensor1(

theano/gpuarray/tests/test_opt.py

@@ -560,3 +560,13 @@ def test_local_lift_solve():
     A_val = np.random.uniform(-0.4, 0.4, (5, 5)).astype("float32")
     b_val = np.random.uniform(-0.4, 0.4, (5, 3)).astype("float32")
     utt.assert_allclose(f_cpu(A_val, b_val), f_gpu(A_val, b_val))
+
+
+def test_local_gpua_advanced_incsubtensor():
+    # test a corner case reported at gh-5589
+    target = tensor.ftensor4()
+    y = target.dimshuffle(1, 0, 2, 3).flatten(ndim=1)
+    w = tensor.ones_like(y)
+    w = tensor.set_subtensor(w[tensor.eq(y, 1.0).nonzero()], 100)
+    w = tensor.set_subtensor(w[tensor.eq(y, -1.0).nonzero()], 0)
+    theano.function([target], w)