What is rocFFT?#
The rocFFT library implements the discrete Fast Fourier Transform (FFT) in HIP for GPU devices. It provides a fast and accurate platform for calculating discrete FFTs. The source code can be found at ROCm/rocFFT.
rocFFT supports the following features:
Half (
FP16
), single, and double precision floating point formats1D, 2D, and 3D transforms
Computation of transforms in batches
Real and complex FFTs
Arbitrary lengths, with optimizations for combinations of powers of 2, 3, 5, 7, 11, 13, and 17
rocFFT also provides experimental support for:
Distributing transforms across multiple GPU devices in a single process
Distributing transforms across multiple MPI (Message Passing Interface) processes
For information about how rocFFT computes FFTs, see FFT computation.