rocFFT is a software library for computing Fast Fourier Transforms (FFT) written in HIP. It is part of AMD's software ecosystem based on ROCm. In addition to AMD GPU devices, the library can also be compiled with the CUDA compiler using HIP tools for running on Nvidia GPU devices.
Download pre-built packages either from ROCm's package servers or by clicking the github releases tab and downloading the source, which could be more recent than the pre-build packages. Release notes are available for each release on the releases tab.
sudo apt update && sudo apt install rocfft
rocFFT is compiled with hcc and uses cmake. To compile the library one calls, for example, the following commands:
mkdir build && cd build
cmake -DCMAKE_CXX_COMPILER=hcc ..
A static library can be compiled by using the option -DBUILD_SHARED_LIBS=off
The hip-clang compiler (https://github.com/ROCm-Developer-Tools/HIP/blob/master/INSTALL.md#hip-clang), one must specify
-DUSE_HIP_CLANG=ON -DHIP_COMPILER=clang to cmake.
One can use nvcc as a backend compiler by passing the option -DUSE_CUDA=yes
and setting HIP_PLATFORM=nvcc in your environment.
rocfft-rider is a client which will run general transforms and is
useful for performance analysis. Compilation is enabled via the
-DBUILD_CLIENTS_RIDER=on cmake option. rocfft-rider uses boost
program options.
rocfft-test runs functionality tests and uses FFTW, Google test, and
boost program options. Compilation is enabled by calling cmake with the -DBUILD_CLIENTS_TESTS=on option.
To install the clients depencencies on Ubuntu, run
sudo apt install libgtest-dev libfftw3-dev libboost-program-options-dev
The file install.sh is a bash script that is a wrapper for the cmake
script, which also install dependencies on certain Linux
distributions. The preferred method for compiling rocFFT is to call
cmake directly.
Please refer to the Library documentation for current documentation.
Examples may be found in the clients/samples subdirectory.