Updated comments following code review.

theano/sandbox/cuda/opt.py

@@ -142,7 +142,8 @@ def local_gpu_elemwise_0(node):
     if isinstance(node.op, tensor.Elemwise) and dtype_in_elemwise_supported(node.op):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu) for i in node.inputs]):
             if numpy.all([o.type.dtype == 'float32' for o in node.outputs]):
-                #don't set any inplace pattern. gpu_inplace_optimizer will do it later
+                # Don't set any inplace pattern.
+                # gpu_inplace_elemwise_optimizer will do it later
                 new_op = GpuElemwise(node.op.scalar_op)
 
                 #   first establish that float32 can store all inputs
@@ -185,7 +186,8 @@ def local_gpu_elemwise_1(node):
             dtype_in_elemwise_supported(node.op)):
 
             elemwise_node = host_i.owner
-            #don't set any inplace pattern. gpu_inplace_optimizer will do it later
+            # Don't set any inplace pattern.
+            # gpu_inplace_elemwise_optimizer will do it later
             new_op = GpuElemwise(elemwise_node.op.scalar_op)
             if all([i.dtype=='float32' for i in elemwise_node.inputs]):
                 gpu_elemwise = new_op(*[gpu_from_host(i) for i in elemwise_node.inputs])