Versal Adaptive SoC AIE-ML Architecture Manual (AM020)

Document ID
Release Date
1.3 English

In the context of task-completion tokens, a task refers to a unit of work that is pushed into the input task queue of an actor (such as an AIE-ML tile, AIE-ML memory, and AIE-ML array interface ) through the memory mapped interface. The AIE-ML tile, AIE-ML memory, and AIE-ML array interfaces are referred to as actors, each equipped with an input task queue. A task can be placed in the queue through the AXI4 interface, managed by a controller. The actor processes the task until it is completed before proceeding to the next one. Once all tasks in the queue are finished, the actor enters an idle state, awaiting the insertion of a new task.

Upon the completion of a task, the actor triggers the tile task completion unit to send a task completion token to the stream switch through the control-packet response port. In AIE-ML, these task completion tokens are exposed at the AXIS PLIO interface located at the perimeter of the AI Engine array and are intended to be used through PL.

Figure 1. Task-Complete-Tokens

The figure shows the controller implemented in soft logic in the PL, however, a valid design methodology is to have the front end of the application run on the PS and only a minimum back-end IP on the PL to service the token streams. The task-complete-token consists of a single stream word. This token is routed back to the issuing controller through the standard stream-switch network. There is an eight-bit controller ID register that determines the ownership of each actor. The lower five bits of the controller ID register are used as the stream packet ID, and the upper three bits are available for further routing in the PL.