With the input data rate being 2.5 Gsps on each AI Engine and the filter with 32 taps, the data stream and coefficient can be split into eight phases as each AI Engine is capable of 4-tap filter processing. This leads to 2.5 Gsps x 8 Phases = 20 Gsps input sample rate. You will now design the maximum performance for the filter.
The same recommendations as in the previous section will apply:
Each AI Engine in a column should receive the same data.
One row in every two has a cascade stream in the other direction, leading to a differentiated set of stream for even and odd rows.
The kernels above the diagonal (lower left to upper right) should discard one element in the stream.