Channel in global memory

You may be able to create your own FIFO in global memory with a global array and two scalars indices--front and back--defined at program scope. Then you could use the atomic integer functions to manipulate the front and back indices as needed.

If you need an extremely large FIFO, i.e., writing to it much more often than reading from it, then you may be better off having the second kernel wait for the first kernel to complete and use global memory buffers in the traditional way. Maybe that doesn't cover your use case, so I'd be interested if you could share more details of the problem you're trying to solve?