Description

S3L_fft performs a simple FFT on the complex parallel array a. The same FFT operation is performed along all axes of the array.

Both power-of-two and arbitrary radix FFTs are supported. The 1D parallel FFT can be used for sizes that are a multiple of the square of the number of processes. The 2D and 3D FFTs can be used for arbitrary sizes and distributions.

The S3L_fft routine computes a multidimensional transform by performing a one-dimensional transform along each axis in turn.

The sign of the twiddle factor exponents determines the direction of an FFT. Twiddle factors with a negative exponent imply a forward transform, and twiddle factors with positive exponents are used for an inverse transform.

For the 2D FFT, a more efficient transpose algorithm will be used if the blocksizes along each dimension are equal to the extents divided by the number of processes, resulting in significant performance improvements.

S3L_fft (and S3L_ifft) can only be used for complex and double complex data types. To compute a real-data forward FFT, use S3L_rc_fft. This performs a forward FFT on the real data, yielding packed representation of the complex results. To compute the corresponding inverse FFT, use S3L_cr_fft, which will perform an inverse FFT on the complex data, overwriting the original real array with real-valued results of the inverse FFT.

The floating-point precision of the result always matches that of the input.

S3L_fft and S3L_ifft do not perform any scaling. Consequently, when a forward FFT is followed by an inverse FFT, the original data will be scaled by the product of the extents of the array.