fix way that gpu view is created

theano/sandbox/cuda/basic_ops.py

@@ -2448,7 +2448,7 @@ class GpuIncSubtensor(tensor.IncSubtensor, GpuOp):
         """
         return """(CudaNdarray*) CudaNdarray_Copy(%(x)s)""" % locals()
 
-    def make_view_buffer(self, x, view_ndim):
+    def make_view_array(self, x, view_ndim):
         """
             x: a string identifying an array to be viewed
             view_ndim: a string specifying the number of dimensions
@@ -2484,6 +2484,33 @@ class GpuIncSubtensor(tensor.IncSubtensor, GpuOp):
     def define_set_data(self):
         return _define_set_data
 
+    def link_view_array(self, x, fail):
+
+        return """
+        if (CudaNdarray_set_device_data(xview, CudaNdarray_DEV_DATA(%(x)s),
+                                       (PyObject*) NULL))
+        {
+            PyErr_Format(PyExc_RuntimeError,
+                         "GpuSubtensor is not able to set the"
+                         " devdata field of the view");
+            Py_XDECREF(xview);
+            %(fail)s;
+        cnda_mark_dev_structure_dirty(xview);
+        }""" % locals()
+
+    def set_view_base(self, x, fail):
+        return """
+        //Set the base only now
+
+        if(CudaNdarray_set_device_data(xview, CudaNdarray_DEV_DATA(xview),
+                                    %(x)s)){
+            PyErr_Format(PyExc_RuntimeError,
+                         "GpuSubtensor is not able to set"
+                         " the base of the view array");
+            Py_XDECREF(xview);
+            %(fail)s;
+        }""" % locals()
+
     def c_code_cache_version(self):
         # TODO: cooperate with parent class' C code
         return ()

theano/sandbox/cuda/cuda_ndarray.cuh

@@ -437,7 +437,7 @@ DllExport int CudaNdarray_dimshuffle(CudaNdarray * self, unsigned int len, const
 DllExport PyObject*
 CudaNdarray_TakeFrom(CudaNdarray * self, PyObject *args);
 
-static int fprint_CudaNdarray(FILE * fd, const CudaNdarray *self);
+int fprint_CudaNdarray(FILE * fd, const CudaNdarray *self);
 
 
 PyObject * CudaNdarray_View(const CudaNdarray * self);

theano/tensor/basic.py

@@ -4616,7 +4616,14 @@ class IncSubtensor(Op):
         }
         """ % locals()
 
-        alloc_view_of_z = self.make_view_buffer(z, view_ndim)
+        # IG: Note: this makes a variable called "xview"
+        # even though it is a view of z.
+        # I assume this is because IncSubtensor was written
+        # by copy-pasting Subtensor and in Subtensor you make
+        # a view of x.
+        alloc_view_of_z = self.make_view_array(z, view_ndim)
+        # On GPU, it takes two steps to make a view
+        link_view_of_z = self.link_view_array(z, fail);
 
         #Make a first view on the output, as we will write into it.
         build_view = """
@@ -4627,6 +4634,7 @@ class IncSubtensor(Op):
         {
             %(fail)s;
         }
+        %(link_view_of_z)s;
         """ % locals()
         # make xview actually a view of %(z)s
         get_xview = self.define_set_data() + \
@@ -4713,7 +4721,7 @@ class IncSubtensor(Op):
         return """(PyArrayObject*)PyArray_FromAny(py_%(x)s, NULL, 0, 0,
                 NPY_ARRAY_ENSURECOPY, NULL)""" % locals()
 
-    def make_view_buffer(x, view_ndim):
+    def make_view_array(x, view_ndim):
         """
             x: a string identifying an array to be viewed
             view_ndim: a string specifying the number of dimensions
@@ -4754,6 +4762,19 @@ class IncSubtensor(Op):
         set data argument to the helper C code. """
         return ""
 
+    def link_view_array(self, x, fail):
+        """ Returns code to complete making xview a view of x"""
+
+        # On CPU there is nothing to do, make_view_array already did this
+        return ""
+
+    def set_view_base(self, x, fail):
+        """ Returns code to make xview be a correct view of x,
+        after helper_c_code is done messing with x"""
+
+        # On CPU there is nothing to do
+        return ""
+
 
     def infer_shape(self, node, shapes):
         return [shapes[0]]