To illustrate throughput and latency considerations when working with FIR filters, the following example shows a channel filter followed by an interpolating half-band filter.
- Implementation 1: Channel and Interpolation filter kernels placed
into their own individual AI Engine tiles.
The ping-pong buffers are just one deep to manage the data across the simple example.Figure 2. Graph View from Vitis Analyzer
- Implementation 2: Channel filter is split across four AI Engines to improve the throughput.
Notice how the data on the input needs additional buffering, including four ping-pong pairs, (green box) to service these four AI Engine kernels. However, the overall throughput increases as a result of this partitioning, and there is a trade-off to be made with respect to AI Engine resource utilization.Figure 3. Graph View Showing Channel Filter Split Across Four AI Engine Tiles