CUDA: can I use placement new for device pointer on host directly?

In my program, I want to use two steps to new an object:

• use cudaMalloc to get a raw memory
• use placement new to construct the object at the raw pointer.

So I wonder whether this is feasible because the pointer returned by cudaMalloc is on the host and the placement new is also running on the host. If this is not OK, are there any other best practices to do this?

The purpose of this separation is that I want to use an Allocator class and its derived classes to management the raw memory for different devices such as CPU and GPU. And then construct objects by using these raw memory returned by Allocator.