Merge pull request #5684 from lamblin/fix_5670

theano/gpuarray/subtensor.py

@@ -300,8 +300,9 @@ class GpuIncSubtensor(IncSubtensor):
         size_t dims[%(view_ndim)s];
         for(int i=0; i<%(view_ndim)s; i++)
             dims[i] = xview_dims[i];
+
         zview = pygpu_fromgpudata(%(x)s->ga.data,
-                                  xview_offset,
+                                  %(x)s->ga.offset + xview_offset,
                                   %(x)s->ga.typecode,
                                   %(view_ndim)s,
                                   dims,
@@ -396,7 +397,7 @@ int sub_setarray(GpuArray *dst, GpuArray *src) {
         parent_version = super(GpuIncSubtensor, self).c_code_cache_version()
         if not parent_version:
             return
-        return parent_version + (8,)
+        return parent_version + (9,)
 
 
 class GpuAdvancedSubtensor1(HideC, tensor.AdvancedSubtensor1):

theano/gpuarray/tests/test_subtensor.py

@@ -6,6 +6,7 @@ import theano
 from theano import tensor
 from theano.compile import DeepCopyOp
 from theano.tensor.tests import test_subtensor
+from theano.tests import unittest_tools as utt
 
 from ..basic_ops import HostFromGpu, GpuFromHost, GpuContiguous
 from ..elemwise import GpuDimShuffle
@@ -136,6 +137,22 @@ def test_incsub_f16():
     assert np.allclose(rval, rep)
 
 
+def test_incsub_offset():
+    # Test for https://github.com/Theano/Theano/issues/5670
+
+    # Build a GPU variable which value will have an offset (x1)
+    x = gpuarray_shared_constructor(np.zeros(5, dtype=theano.config.floatX))
+    x1 = x[1:]
+    # Use inc_subtensor on it
+    y = tensor.vector()
+    z = tensor.inc_subtensor(x1[2:], y)
+    # Use updates so that inc_subtensor can happen inplace
+    f = theano.function([y], z, updates={x: z}, mode=mode_with_gpu)
+    utt.assert_allclose(
+        f([1, 2]),
+        np.array([0, 0, 1, 2], dtype=theano.config.floatX))
+
+
 class G_advancedsubtensor(test_subtensor.TestAdvancedSubtensor):
     def shortDescription(self):
         return None