Automated Placement for Parallelized FPGA FFTs

We present an automated placement algorithm for a high-bandwidth, streaming fast Fourier transform (FFT) generated using an open-source core generator. For radio spectral line analysis, we generate a 2048-point FFT for a standard commercial FPGA high-bandwidth spectrometer, which can process 16 digital samples per cycle simultaneously (streaming). Our placement algorithm enables this system to operate at 375 MHz and process a 6 GHz sampling rate, nearly 85% of the theoretical maximum FPGA clock rate. This algorithm is a special case of the junction tree algorithm from machine learning, and allows our design to support a 25% faster clock than the standard Xilinx tool flow implementation while using 50% less CPU time to compile. This gain is equivalent to gains induced by floor planning with no hand placement.