Fix GpuAdvancedIncSubtensor1_dev20 with negative index

theano/sandbox/cuda/basic_ops.py

@@ -2974,7 +2974,7 @@ class GpuAdvancedIncSubtensor1_dev20(GpuAdvancedIncSubtensor1):
         return Apply(self, [x_, y_, ilist_], [x_.type()])
 
     def c_code_cache_version(self):
-        return (3,)
+        return (4,)
 
     def c_code(self, node, name, inputs, outputs, sub):
         active_device_no = theano.sandbox.cuda.active_device_number()
@@ -3030,6 +3030,8 @@ class GpuAdvancedIncSubtensor1_dev20(GpuAdvancedIncSubtensor1):
                   for(int j = (threadIdx.x); j < numColsX;j += blockDim.x)
                   {
                       int x_row = d_indices_arr[i];
+                      if(x_row < 0)
+                          x_row += numRowsX;
                       int y_row = i;
                       atomicAdd(&X[(x_row * stridesX0) + (j * stridesX1)], Y[(y_row * stridesY0) + (j * stridesY1)]);
                   }

theano/tensor/tests/test_subtensor.py

@@ -437,7 +437,7 @@ class T_subtensor(unittest.TestCase, utt.TestOptimizationMixin):
 
     def test_ok_list(self):
         for data, idx in [(rand(4), [1, 0]),
-                          (rand(4, 5), [2, 3]),
+                          (rand(4, 5), [2, 3, -1]),
                           (rand(4, 2, 3), [0, 3]),
                           (rand(4, 2, 3), [3, 3, 1, 1, 2, 2, 0, 0]),
                           (rand(4, 2, 3), [3, 3, 1, 1, 2, 2, 0, 0,
@@ -479,6 +479,16 @@ class T_subtensor(unittest.TestCase, utt.TestOptimizationMixin):
                 out2 = test_out[0][0]
                 assert out1 is out2
 
+            # test the grad
+            gn = theano.grad(t.sum(), n)
+            g = self.function([], gn, op=self.adv_incsub1)
+            theano.printing.debugprint(g)
+            utt.verify_grad(lambda m: m[[1, 3]],
+                            [numpy.random.rand(5, 5).astype(self.dtype)])
+            g_0 = g()
+            utt.verify_grad(lambda m: m[idx],
+                            [data])
+
     def test_err_invalid_list(self):
         n = self.shared(numpy.asarray(5, dtype=self.dtype))
         self.assertRaises(TypeError, n.__getitem__, [0, 0])