Seems reminiscent of assignment 2's different types of threadpools

While stream looks much more efficient compared to pipeline, it has the restriction that communication between controllers is done through either streaming interfaces or queues.

@jessiexu yes stream can be more efficient than pipeline if the program is amenable to this kind of scheduling. Stream operates with each unit being connected to queues so that they can start operating as soon as their data arrives.

Could someone explain the fork-join diagram?

@pmp I think it is similar to the multithreading concept of forking and joining. At the beginning of each iteration all sram loads are dispatched concurrently. However, load 0 cannot proceed to the next iteration even if it returns earlier than load 1. It has to wait until all loads have returned.

@pmp Maybe you can also think it as putting a "barricade" at the end of each group of loading.