The performance of the inverse 1D-FFT routines varies drastically depending on the number of points to be transformed. The following graphs summarizes this seemingly erratic behavior.
The following graph is obtained by assuming that an inverse complex to complex 1D-FFT on n data points takes 3nlog2(n) operations.
As expected, the best performance is obtained when the number of points is a power of two. Since the performance of the inverse 1D-FFT is not a simple function of the number of points to be transformed and is machine dependent, we do not attempt to mathematically model the run time of the inverse 1D-FFT but rely on empirical measurements. For our implementation, we actually enlarge the cube to the next reasonably good size for the 1D-FFTs.