FFTFPGA is an OpenCL based library for Fast Fourier Transformations for FPGAs. This repository provides OpenCL host code in the form of FFTW like APIs, which can be used to offload existing FFT routines to FPGAs with minimal effort. It also provides OpenCL kernels that can be synthesized to bitstreams, which the APIs can utilize.
- 1D, 2D and 3D Transforms
- Input sizes of powers of 2
- Single Precision (32 bit floating point)
- C2C: Complex input to complex output
- Out-of-place transforms
- Batched 3D transforms
- OpenCL Shared Virtual Memory (SVM) extensions for data transfers
This library has been tested using the following FPGAs present in the Noctua cluster of the Paderborn Center for Parallel Computing (PC2) at Paderborn University:
- Bittware 520N card with Intel Stratix 10 GX 2800 FPGA
- Intel FPGA PAC D5005 card with Intel Stratix 10 SX 2800 FPGA
- CP2K: the quantum chemistry software package has an interface to offload 3d FFTs to Intel FPGAs that uses the OpenCL kernel designs of FFTFPGA.
Firstly, dependencies for building the system
- CMake >= 3.10
- C++ compiler with C++11 support (GCC 4.9.0+)
- Intel FPGA SDK for OpenCL
- FFTW3
Once you have this covered, execute the following:
mkdir build && cd build
cmake ..
makeYou have built the API i.e., the OpenCL host code that invokes different transformations correctly are packed into a static library. This must be linked to an application.
You have also compiled a sample application that helps invoke these APIs.
Strictly said, you have done the following:
fftfpgastatic library, linked such as-lfftfpgafftfpga/fftfpga.hheader filefft- a sample application which links and includes the above two.
Now onto synthesizing the OpenCL FFT kernels. These can be synthesized to run on software emulation or on hardware as bitstreams.
- Emulation
make <kernel_name>_emu
make fft3d_ddr_emulate- Hardware Bitstream
make <kernel_name>_syn
make fft3d_ddr_synPutting them all together, in order to execute the required FFT, set the path to the synthesized bitstream along with other correct configurations as command line parameters to the sample application generated.
./fft --num=64 --dim=3 --path=fft3d_ddr_128.aocxTip: for emulation, use the --emulate command line parameter.
| Kernel Name | Description | |
|---|---|---|
| 1D | fft1d | OpenCL design provided by Intel |
| 2D | fft2d_ddr | DDR memory is used for 2D Transpose |
| fft2d_bram | BRAM is used for 2D Transpose | |
| 3D | fft3d_ddr | DDR memory is used for 3D Transpose |
| fft3d_bram | BRAM is used for 3D Transpose |
These kernels can be synthesized by appending _emulate or _syn to its suffix such as fft1d_emulate.
Please checkout the User Guide for more information such as configuration options etc.
FFTFPGA has been cited in the following publications:
-
Evaluating the Design Space for Offloading 3D FFT Calculations to an FPGA for High-Performance Computing : https://doi.org/10.1007/978-3-030-79025-7_21
-
CP2K: An electronic structure and molecular dynamics software package - Quickstep: Efficient and accurate electronic structure calculations: https://doi.org/10.1063/5.0007045
-
Efficient Ab-Initio Molecular Dynamic Simulations by Offloading Fast Fourier Transformations to FPGAs : https://doi.org/10.1109/FPL50879.2020.00065
- ConvFPGA - an OpenCL based library for FFT-based convolution on FPGAs
- FFTFPGA-eval - archives reports and measurements from FFTFPGA and ConvFPGA
- Marius Meyer for code reviews, testing and discussions.