Accelerating 3D-FFT Using Hard Embedded Blocks in FPGAs

Abstract

Three dimensional Fast Fourier Transform (3D-FFT) is popularly used in many scientific applications in various domains like image processing, bioinformatics and molecular dynamics. Typically 3D-FFT computation takes significant part of the execution time of these applications. In order to speedup these applications, it becomes necessary to accelerate 3D-FFT computation. 3D-FFT can be accelerated using Field Programmable Gate Array (FPGA) based accelerators. But speedup always may not be possible as FPGAs run at slower clock frequency vis-a-vis processors and the resources available in an FPGA device might not be sufficient for the implementation of a sufficient number of copies of the processing elements to compensate for the loss of clock frequency. FPGAs with heterogeneous mix of coarse grained hard blocks along with programmable soft logic, can facilitate implementing a much larger number of processing elements and thus achieve much higher speedups. Modern FPGAs do consist of different heterogeneous hard embedded blocks (HEBs) like multipliers, DSP blocks and memory units. It is easy to predict that many more such hard blocks will be embedded into future FPGAs. The evaluation approach to identify and incorporate HEBs is complex as there are many parameters and constraints like area, granularity routing resources, etc. that need to be considered in an integrated manner to get an efficient implementation. In this paper we show acceleration of 3D-FFT using future fabrics incorporating HEBs. By using these fabrics we show speedups of upto 1900x for 2048 point FFT. We also present an evaluation methodology to design future FPGA fabrics incorporating accelerators as hard embedded blocks. This methodology will be useful for i selection of blocks to be embedded into the fabric and ii evaluating the performance gain that can be achieved by such an embedding.