3.2.2. Floating-Point Variable Streaming FFTs
The FFT has ceiling(log 4 (N)) stages. If transform length is an integral power of four, a radix-4 FFT implements all of the log 4 (N) stages. If transform length is not an integral power of four, the FFT implements ceiling(log 4 (N)) 1 of the stages in a radix-4, and implements the remaining stage using a radix-2.
Each stage contains a single butterfly unit and a feedback delay unit. The feedback delay unit delays the incoming data by a specified number of cycles; in each stage the number of cycles of delay is one quarter of the number of cycles of delay in the previous stage. The delays align the butterfly input samples correctly for the butterfly calculations. The output of the pipeline is in index-reversed order.