UHFFT: A High Performance DFT Framework

We present a framework for algorithm selection, code generation and optimization for the Fast Fourier Transform. The framework consists of two main parts – the DFT Code Generator (FftGen) and the Execution Scheduler (FftPlanner). The library generator outputs a highly optimized set of small to modest size DFT codelets written in ‘C’. The FftPlanner searches an exponential space of possible algorithms, factors and schedules to select the best computation plan at run time. UHFFT is a high performance and portable framework that adapts to underlying architecture. We present two sets of results on two different architectures; Itanium2 and Opteron. We also compare our performance against other vendor and non-vendor FFT computation libraries.