The same exercise can be done using the streaming interface. Here the compiler has no clue about the size of the input and output frame (the inner loop can be an infinite loop), and you have to specify the repetition count for each kernel.