The most commonly used FFT algorithms by far are the power-of-two-length FFT algorithms. The Prime Factor Algorithm (PFA) and Winograd Fourier Transform Algorithm (WFTA) require somewhat fewer multiplies, but the overall difference usually isn't sufficient to warrant the extra difficulty. This is particularly true now that most processors have single-cycle pipelined hardware multipliers, so the total operation count is more relevant. As can be seen from the following table, for similar lengths the split-radix algorithm is comparable in total operations to the Prime Factor Algorithm, and is considerably better than the WFTA, although the PFA and WTFA require fewer multiplications and more additions. Many processors now support single cycle multiply-accumulate (MAC) operations; in the power-of-two algorithms all multiplies can be combined with adds in MACs, so the number of additions is the most relevant indicator of computational cost. The Winograd Fourier Transform Algorithm is particularly difficult to program and is rarely used in practice. For applications in which the transform length is somewhat arbitrary (such as fast convolution or general spectrum analysis), the length is usually chosen to be a power of two. When a particular length is required (for example, in the USA each carrier has exactly 416 frequency channels in each band in the AMPS cellular telephone standard), a Prime Factor Algorithm for all the relatively prime terms is preferred, with a Common Factor Algorithm for other non-prime lengths. Winograd's short-length modules should be used for the prime-length factors that are not powers of two. The chirp z-transform offers a universal way to compute any length DFT (for example, Matlab reportedly uses this method for lengths other than a power of two), at a few times higher cost than that of a CFA or PFA optimized for that specific length. The chirp z-transform, along with Rader's conversion, assure us that algorithms of ONlogN O N N complexity exist for any DFT length N N.

The choice of a power-of-two algorithm may not just depend on computational complexity. The latest extensions of the split-radix algorithm offer the lowest known power-of-two FFT operation counts, but the 10%-30% difference may not make up for other factors such as regularity of structure or data flow, FFT programming tricks, or special hardware features. For example, the decimation-in-time radix-2 FFT is the fastest FFT on Texas Instruments' TMS320C54x DSP microprocessors, because this processor family has special assembly-language instructions that accelerate this particular algorithm. On other hardware, radix-4 algorithms may be more efficient. Some devices, such as AMI Semiconductor's Toccata ultra-low-power DSP microprocessor family, have on-chip FFT accelerators; it is always faster and more power-efficient to use these accelerators and whatever radix they prefer. For fast convolution, the decimation-in-frequency algorithms may be preferred because the bit-reversing can be bypassed; however, most DSP microprocessors provide zero-overhead bit-reversed indexing hardware and prefer decimation-in-time algorithms, so this may not be true for such machines. Good, compiler- or hardware-friendly programming always matters more than modest differences in raw operation counts, so manufacturers' or good third-party FFT libraries are often the best choice. The module FFT programming tricks references some good, free FFT software (including the FFTW package) that is carefully coded to be compiler-friendly; such codes are likely to be considerably faster than codes written by the casual programmer.

Multi-dimensional FFTs pose additional possibilities and problems. The orthogonality and separability of multi-dimensional DFTs allows them to be efficiently computed by a series of one-dimensional FFTs along each dimension. (For example, a two-dimensional DFT can quickly be computed by performing FFTs of each row of the data matrix followed by FFTs of all columns, or vice-versa.) Vector-radix FFTs have been developed with higher efficiency per sample than row-column algorithms. Multi-dimensional datasets, however, are often large and frequently exceed the cache size of the processor, and excessive cache misses may increase the computational time greatly, thus overwhelming any minor complexity reduction from a vector-radix algorithm. Either vector-radix FFTs must be carefully programmed to match the cache limitations of a specific processor, or a row-column approach should be used with matrix transposition in between to ensure data locality for high cache utilization throughout the computation.

FFT algorithms gain their efficiency through intermediate computations that can be reused to compute many DFT frequency samples at once. Some applications require only a handful of frequency samples to be computed; when that number is of order less than OlogN O N , direct computation of those values via Goertzel's algorithm is faster. This has the additional advantage that any frequency, not just the equally-spaced DFT frequency samples, can be selected. Sorensen and Burrus developed algorithms for when most input samples are zero or only a block of DFT frequencies are needed, but the computational cost is of the same order.

Some applications, such as time-frequency analysis via the short-time Fourier transform or spectrogram, require DFTs of overlapped blocks of discrete-time samples. When the step-size between blocks is less than OlogN O N , the running FFT will be most efficient. (Note that any window must be applied via frequency-domain convolution, which is quite efficient for sinusoidal windows such as the Hamming window.) For step-sizes of OlogN O N or greater, computation of the DFT of each successive block via an FFT is faster.