The DOA subgraph estimates the MUSIC Spectrum $\hat{\textbf{P}}_m$ defined earlier at \(256\) equally spaced bins. In order to achieve the target throughput of 1 $\mu s\(, this workload is partitioned across a number of tiles where each tile computes the spectrum for \)L$ consecutive points. This is shown in the diagram below. A value of \(L=4\) is required to meet the throughput; this is equivalent to \(64\) AI Engine tiles.
The following diagram shows the AI Engine physical array view for the DOA subgraph. The data flow graph is linear with \(64\) total tiles to support evaluation of the MUSIC spectrum over \(256\) equally spaced points. The incidence signals used for the MUSIC workloads are pre-computed at compile time based on the array steering vector and stored in lookup tables. Data flow proceeds from tile to tile, each one passing the noise subspace basis vectors and number of sources to the next tile in the graph. Each tile computes \(L=4\) spectrum bins and passes them down the pipeline. Bins are passed with cfloat
data type with the imaginary part set to zero.