Fix crashes probably due to nvcc not handling unsigned int as we expect

theano/sandbox/cuda/basic_ops.py

@@ -1831,7 +1831,7 @@ class GpuCAReduce(GpuOp):
         """ % locals(), file=sio)
 
     def c_code_cache_version_apply(self, node):
-        version = [14]  # the version corresponding to the c code in this Op
+        version = [15]  # the version corresponding to the c code in this Op
 
         # now we insert versions for the ops on which we depend...
         scalar_node = Apply(self.scalar_op,
@@ -2163,7 +2163,7 @@ class GpuCAReduce(GpuOp):
             #      memory (a segment of a column).
             reducebuf = self._k_reduce_buf('Z[blockIdx.x * sZ0]', node, nodename, sub={})
             reduce_fct = self._assign_reduce(node, nodename, "myresult",
-                                             "A[i0 * sA0 + i1 * sA1 + blockIdx.x * sA2]",
+                                             "A[i0 * sA0 + i1 * sA1 + ((int)blockIdx.x) * sA2]",
                                              {}, True)
             reduce_init = self._assign_init("A[blockIdx.x * sA2]")
             print("""