The following figure shows a diagram of how the “vector x matrix” multiplication form of the IDFT is vectorized and mapped to the AI Engine array of 4 x 4 = 16 tiles. The figure shows two consecutive IDFT transforms, one above the other. Recall each full transform is performed over two cycles. The operation of the design is outlined as follows:
The design consists of a four x four array of tiles. Each tile performs two [1x2] x [2x4] operations over two cycles. Each row of tiles passes its computed outputs to the tile below in the same column using the cascade stream.
Four samples are input on each of two input streams for each tile. The same data is broadcast to each tile in the row. For example, the orange input samples are broadcast to all tiles in the orange row, whereas the purple input samples are broadcast to all tiles in the purple row.
Notice how the four input samples on a given stream span particular consecutive samples of a pair of transform inputs. For example, the four orange inputs on stream “ss0” contain the first two samples in the top (current) and bottom (next) input vector. Similarly, the four left-most purple samples on (unlabelled) stream “ss4” contain the 9th and 10th samples in the top and bottom input vectors.
The array combines outputs top-to-bottom (in the diagram) using the cascade streams. The four tiles in the bottom row produce the outputs, writing four samples every four cycles on both streams in each tile. Note in the physical array, the cascade streams run horizontally left to right — the physical layout is rotated 90 degrees from the diagram in the following figure.
Each full compute takes two cycles, with throughput sustained at that rate with 100% efficient compute in each AI Engine tile.
