Sharing Run-Time Parameters Across Multiple Kernels - 2021.2 English

Versal ACAP AI Engine Programming Environment User Guide (UG1076)

Document ID
Release Date
2021.2 English

The run-time parameter to switch channels in the previous graph is shared by the bypass encapsulator and the mixer kernel. Both entities need to see the same switched value at the same data boundary. When the nodes sharing the run-time parameter are mapped to the same AI Engine, switching of the parameter value is synchronized because each node, mapped to the same processor, processes the current set of data before any node processes the next set of data. However, when the sharing kernels are mapped to different processors, they can execute in a pipelined fashion on different sets of data, as shown in the following figure. Then, the run-time parameter should be pipelined along with the data.

Figure 1. Pipelined Run-Time Parameters

In the current release, you need to pipeline the control parameter through the kernel by making it an inout parameter on the producing kernel connected to an input parameter on the consuming kernel. Pipelining across processors can be intermixed with a one-to-many broadcast connection within a single processor to create arbitrary control topologies.