AOCL-FFTW is AMD optimized version of FFTW implementation targeted for AMD EPYC CPUs. It is developed on top of FFTW (version fftw-3.3.10). All known features and functionalities of FFTW are retained and supported as it is with this AMD optimized FFTW library.
AOCL-FFTW achieves high performance as a result of its various optimizations involving improved SIMD Kernel functions, improved copy functions (cpy2d and cpy2d_pair used in rank-0 transform and buffering plan), improved 256-bit kernels selection by Planner and an optional in-place transpose for large problem sizes. AOCL-FFTW improves the performance of in-place MPI FFTs by employing a faster in-place MPI transpose function. AOCL-FFTW provides a new fast planner mode as an extension to the original planner that improves planning time of various planning modes in general and PATIENT mode in particular. Another new planning mode called Top N planner is also available that minimizes single-threaded run-to-run variations. AOCL-FFTW has a feature called AMD's application optimization layer that speeds up HPC and scientific applications. AOCL-FFTW implements the dynamic dispatcher feature that can build a single portable optimized library for execution on a wide range of x86 CPU architectures.
FFTW is a free collection of fast C routines for computing the Discrete Fourier Transform and various special cases thereof in one or more dimensions. It includes complex, real, symmetric, and parallel transforms, and can handle arbitrary array sizes efficiently.
The doc/ directory contains the manual in texinfo, PDF, info, and HTML formats. Frequently asked questions and answers can be found in the doc/FAQ/ directory in ASCII and HTML.
For a quick introduction to calling FFTW, see the "Tutorial" section of the manual.
INSTALLATION FROM AOCL-FFTW GIT REPOSITORY:
After downloading the latest stable release from the git repository, https://github.com/amd/amd-fftw, follow the below steps to configure and build it for AMD EPYC processor based on Naples, Rome, Milan and future generation architectures.
./configure --enable-sse2 --enable-avx --enable-avx2 --enable-avx512
--enable-mpi --enable-openmp --enable-shared
--enable-amd-opt --enable-amd-mpifft
--enable-dynamic-dispatcher
--prefix=<your-install-dir>
make
make install
The configure option "--enable-amd-opt" enables all the improvements and optimizations targeted for AMD EPYC CPUs. For enabling various optional configure options provided for AMD EPYC CPUs, the master optimization switch "--enable-amd-opt" must be kept enabled.
When enabling configure option "--enable-amd-opt", do not use the configure option "--enable-generic-simd128" or "--enable-generic-simd256".
The optional configure option "--enable-amd-mpifft" enables the MPI FFT related optimizations.
An optional configure option "--enable-amd-mpi-vader-limit" is supported that controls enabling of AMD's new MPI transpose algorithms. When using this configure option, the user needs to set --mca btl_vader_eager_limit appropriately (current preference is 65536) in the MPIRUN command.
The new fast planner can be enabled using optional configure option "--enable-amd-fast-planner". It is supported in single and double precisions.
Top N planner mode can be enabled using optional configure option "--enable-amd-top-n-planner" to minimize run-to-run variations in performance. It is supported in single-threaded execution in single and double precisions.
An optional configure option "AMD_ARCH" is supported that can be set to CPU architecture values like "auto" or "znver1" or "znver2" or "znver3" or "znver4" for AMD EPYC processors.
The optional configure option "--enable-amd-app-opt" turns on AMD's application optimization layer to benefit performance of HPC and scientific applications. Currently it is developed for complex and real (r2c and c2r) DFT problem types in double and single precisions. It is not supported for MPI FFTs, r2r real DFT problem types, Quad or Long double precisions, and split array format.
Dynamic dispatcher achieves Function Multi-versioning by using compiler's attributes. Use "--enable-dynamic-dispatcher" configure option to enable this feature. It is supported for Linux based systems for now. The set of x86 CPUs on which the single portable library can work depends upon the highest level of CPU SIMD instruction set with which it is configured.
An optional configure option "--enable-amd-trans" is provided that may benefit the performance of transpose operations in case of very large FFT problem sizes. This is by default not enabled and provided as an experimental optional switch.
By default, configure script enables double-precision mode. User should pass appropriate configure options to enable the single-precision or quad-precision or long-double mode.
AOCL-FFTW is developed and maintained by AMD. For support of these libraries and the other tools of AMD Zen Software Studio, see https://www.amd.com/en/developer/aocc/compiler-technical-support.html
FFTW was developed by Matteo Frigo and Steven G. Johnson. We thank Matteo Frigo for his support provided to us.