Awesome
FFTX Project
This is the public repository for the FFTX API source, examples, and documentation.
License
This project is licensed under the BSD License (available here)
Copyright (c) 2018-2023, Carnegie Mellon University. All rights reserved.
Building FFTX
Prerequisites
There are several pre-requisites that must be installed prior to building and using FFTX. You will need:
- CMake, minimum version 3.14; however, a more-modern version is preferred as some deprecated features must be used with older versions.
- if on Linux/Unix, gcc and make.
- if on Windows, Visual Studio, and an ability to run bash shell scripts. You can use the Git Bash shell available with git, but other shells such as Cygwin or msys will also work.
- if on macOS, version 10.14 (Mojave) or later of macOS, with a compatible version of Xcode and Xcode Command Line Tools.
- python, version 3.6 or higher. On some systems, both python (usually version 2.7) and python3 exist. The scripts used to create the FFTX library source code check the version of python, and if it is version 2.X it will try to run python3 instead. A user, therefore, should not have to worry whether python or python3 comes first in the user's path.
Summary
A brief summary of the steps necessary to install FFTX follows; these steps outline the most straight forward case (building for CPU only), for more complex installs and to utilize GPUs please follow the more detailed instructions below.
1. Make a directory to hold the download code and cd to it (e.g., mkdir ~/work ; cd ~/work)
2. git clone https://github.com/spiral-software/fftx
3. cd fftx
4. export FFTX_HOME=`pwd`
5. unset SPIRAL_HOME
6. source ./get_spiral.sh
7. ./config-fftx-libs.sh
8. mkdir build
9. cd build/
10. cmake -DCMAKE_INSTALL_PREFIX=$FFTX_HOME -D_codegen=CPU ..
11. make install -j
To build and use FFTX, including how to utilize GPUs, please follow these steps:
- Install SPIRAL and associated packages.
- Clone the FFTX repository.
- Generate library source code.
- Compile library source code and examples.
1. Install SPIRAL and associated packages
If you already have SPIRAL installed (and have the SPIRAL_HOME environment variable set) FFTX will use that installation and you can skip to step 2, "Clone the FFTX repository".
If you want to manually install SPIRAL follow the steps below; alternatively, skip to step 2, "Clone the FFTX repository" and have FFTX pull down the necessary SPIRAL repositories and perform the build steps.
Clone spiral-software (available here) to a location on your computer. E.g.,:
cd ~/work
git clone https://www.github.com/spiral-software/spiral-software
This location is known as SPIRAL HOME, and you must set an environment variable
SPIRAL_HOME (here, ~/work/spiral-software) to point to this location later.
FFTX requires the SPIRAL packages fftx, simt, mpi, and jit.
You need to download these separately, as follows:
cd $SPIRAL_HOME/namespaces/packages
git clone https://www.github.com/spiral-software/spiral-package-fftx fftx
git clone https://www.github.com/spiral-software/spiral-package-simt simt
git clone https://www.github.com/spiral-software/spiral-package-mpi mpi
git clone https://www.github.com/spiral-software/spiral-package-jit jit
NOTES:
- The SPIRAL packages must be installed under directory $SPIRAL_HOME/namespaces/packages and must be placed in folders with the prefix spiral-package- removed.
- If you already have spiral-software installed, please refresh the installation to ensure you're up-to-date, especially for the SPIRAL packages.
- It is preferable to download the SPIRAL packages before performing the SPIRAL build steps.
Follow the build instructions for spiral-software (see the README here ).
2. Clone the FFTX repository.
Clone FFTX to a location on your computer. E.g.,
cd ~/work
git clone https://www.github.com/spiral-software/fftx
cd fftx
Set the environment variable FFTX_HOME to point to the directory where you want to install FFTX (which is not necessarily the same directory where you have cloned FFTX; you may want to have separate installation directories for different backends).
If you have not already installed SPIRAL and the SPIRAL packages required for FFTX, you can have them downloaded and built from the repositories by sourcing the get_spiral.sh shell script now. This script checks the definition of the SPIRAL_HOME environment variable, and if undefined it will get the SPIRAL code, build it and export a definition for SPIRAL_HOME. Make sure you source (vs. run) the script:
. ./get_spiral.sh
or
source get_spiral.sh
3. Generate library source code.
FFTX builds libraries of transforms for a set of different sizes. The library source code is generated from SPIRAL script specifications, and must be created before building FFTX itself. For the sizes that are pre-built the code will be added to libraries (the pre-defined fixed-size libraries). Alternatively, sizes not defined can have the code generated and compiled at run-time (RTC).
Before creating the library source code consider if you will be running on CPU only, or also utilizing a GPU. If you create all the source code (and related cmake scripts and library APIs) for GPU and then try building for CPU only you may encounter compiler errors or unexpected results. It is recommended that you have separate installs for CPU and GPU.
The shell script config-fftx-libs.sh is a utility script in the FFTX home directory that marshals resources for building the libraries and examples. There is a flag that enables or disables the building of examples (enabled by default). NOTE: All libraries should be created; creating the library is a pre-requisite to creating its API files (files which are required later when building the RTC code). Building the examples may be turned off if you only need to build the libraries for an external application). The script is self-documenting; typically it should not be edited (unless you need to disable building the examples).
NOTE: It is required that you build all libraries (even if you add few or no fixed size entries to the library). This is because the API header files are created automatically at build time. At run time the API can check if a specific transform of the requested size exists in the library and use it if it exists. If it doesn't exist then the desired transform can be generated and compiled on-the-fly (RTC).
Run the script as follows:
./config-fftx-libs.sh <platform>
## where <platform> is one of { CPU [default] | CUDA | HIP }
If no argument is provided, then the platform defaults to CPU. This script runs the build-lib-code.sh script in the src/library directory and will marshal the resources and options needed for the set of libraries. This step can take quite some time depending on the number of transforms and set of sizes to create. The code is targeted to run on a CPU or a GPU (either CUDA or HIP) depending on the platform specified with the script. By default, only a small number of fixed sizes will be created for the fixed-size libraries.
The text file build-lib-code-failures.txt will contain a list of any library
transforms that failed to generate in this step.
Running config-fftx-libs.sh also creates a file called options.cmake. The options defined in this file are used by CMake to determine what is actually compiled at build time.
4. Compile library source code and examples.
From your FFTX home directory, set up a build folder (which can be given any name, and you may want to have separate ones for different backends). When you configure using CMake you must specify the install prefix that CMake should use (the system default location for CMake may be a directory for which you do not have write privileges). Do that by setting the environment variable FFTX_HOME and specifying either the directory path or the environment variable on the CMake command line.
NOTES:
- You will need FFTX_HOME set in order to use or reference FFTX artifacts externally.<br>
- FFTX_HOME may be, but does not have to be, the same as your FFTX home directory.
- Some tips for building on a supercomputing platform at NERSC or OLCF are available here.
export FFTX_HOME=~/work/fftx
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=~/work/fftx -D_codegen=CPU .. # build for CPU, *or*
cmake -DCMAKE_INSTALL_PREFIX=$FFTX_HOME -D_codegen=CUDA .. # build for CUDA, *or*
cmake -DCMAKE_INSTALL_PREFIX=~/work/fftx -DCMAKE_CXX_COMPILER=hipcc -D_codegen=HIP .. # build for HIP
make install
When FFTX is built, the final step (of make install) creates a tree structure (at the location specified by CMAKE_INSTALL_PREFIX). The following directories will be created/populated:
| Directory Name | Description |
|---|---|
| ./CMakeIncludes | CMake include files and functions to ease integration with FFTX |
| ./bin | Executables for example programs built as part of the FFTX distribution |
| ./lib | FFTX libraries, that can be called by external applications |
| ./include | Include files for using FFTX libraries |
| ./cache_jit_files | Folder containing the RTC code generated for any transform not <br>found in a fixed-size library |
Building on Windows
FFTX can be built on Windows, however, you need to be able to run a [bash] shell script as mentioned above to build the library source code. To build FFTX, open a shell and do the following:
cd fftx
./config-fftx-libs.sh <platform> where <platform> is one of { CPU [default] | CUDA | HIP }
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=~/work/fftx -D_codegen=CUDA ..
cmake --build . --target install --config Release
This shows an example building for CUDA on Windows, you can also build for CPU or AMD HIP as shown above (under Building for Linux).
Running FFTX Example Programs
After building FFTX, to run the programs that are in the examples subtree, simply do:
cd $FFTX_HOME/bin
./testcompare_device
./testverify
etc. Since we set RPATH to point to where the libraries are installed, you likely will not need to adjust the library path variable, typically LD_LIBRARY_PATH.
NOTE: On Windows RPATH does not work correctly. If you encounter "missing library" errors when trying to run add the directory containing the installed libraries to your PATH.
The README in the examples folder here contains a list of examples, how to run them, and how to add a new example. More details about individual examples may be found in the individual example folder's README file.
Libraries
The libraries built and copied to the $FFTX_HOME/lib folder can be used by external applications to leverage FFTX transforms. To access the necessary include files and libraries, an external application's CMakeLists.txt should include CMakeInclude/FFTXCmakeFunctions.cmake. A full example of an external application linking with FFTX libraries is available in the fftx-demo-extern-app.
FFTX Libraries Built
FFTX builds libraries for 1D and 3D FFTs for a single device. FFTs are built for a set of specific sizes, thus not all possible sizes will be found in the libraries. There are default files specifying the sizes to build for each of: 1D FFTs and 3D FFTs (default file names below). You can customize the set of sizes to build by either editing these files or provide your own files -- just override the default file name in the config-fftx-libs.sh script. If you provide your own file(s) just follow the format shown in the default file(s). The default files specifying the sizes to build for each group of transforms are (all files are in the src/library folder):
| Type | File Name | Description |
|---|---|---|
| 1D FFT | dftbatch-sizes.txt | Batch of 1D FFTs |
| 3D FFT | cube-sizes-cpu.txt | 3D FFTs for CPU |
| 3D FFT | cube-sizes-gpu.txt | 3D FFTs for GPU |
The following is a list of the libraries built:
| Type | Name | Description |
|---|---|---|
| 3D FFT | fftx_mddft | Forward 3D FFT complex to complex |
| 3D FFT | fftx_imddft | Inverse 3D FFT complex to complex |
| 3D FFT | fftx_mdprdft | Forward 3D FFT real to complex |
| 3D FFT | fftx_imdprdft | Inverse 3D FFT complex to real |
| 3D Convolution | fftx_rconv | 3D real convolution |
| 1D FFT | fftx_dftbat | Forward batch of 1D FFT complex to complex |
| 1D FFT | fftx_idftbat | Inverse batch of 1D FFT complex to complex |
| 1D FFT | fftx_prdftbat | Forward batch of 1D FFT real to complex (in development) |
| 1D FFT | fftx_iprdftbat | Inverse batch of 1D FFT complex to real (in development) |
Library API
The following example shows usage of the 3D FFT complex-to-complex transform (others are similar; just use the appropriate names and header file(s) from the table above). The user (calling application) is responsible for setting up memory buffers and allocations as required (i.e., host memory for serial code, device memory for GPU code). The following abstract assumes device memory buffers are already allocated/initialized as needed; error checking is omitted for clarity and brevity:
#include "fftx3.hpp"
#include "transformlib.hpp"
std::vector<int> sizes{ mm, nn, kk }; // 'cube' dimensions
std::vector<void *> args{ dY, dX, dsym }; // pointers to the Output, Input, and symbol arrays
MDDFTProblem mdp ( args, sizes, "mddft" ); // Define transform
mdp.transform(); // Run the transform
If the size specified with the transform definition is found in a library then that code is executed; however, if it is not found in a library then RTC is invoked to generate and compile the necessary code (this is also cached for future use).
Linking Against FFTX Libraries
FFTX provides a cmake include file, FFTXCmakeFunctions.cmake, that provides functions to facilitate compiling and linking external applications with the FFTX libraries. An external application should include this file ($FFTX_HOME/CMakeIncludes/FFTXCmakeFunctions.cmake) in order to access the following helper functions to compile/link with the FFTX libraries. Two functions are available:
- FFTX_add_includes_libs_to_target ( target ) -- This function adds the include file paths, the linker library path, and the library names to the specified target.
- FFTX_find_libraries() -- This function finds the FFTX libraries, linker library path, and include file paths and exposes the following variables:
| CMake Variable Name | Description |
|---|---|
| FFTX_LIB_INCLUDE_PATHS | Include paths for FFTX include & library headers |
| FFTX_LIB_NAMES | List of FFTX libraries |
| FFTX_LIB_LIBRARY_PATH | Path to libraries (for linker) |
An application will typically need only call FFTX_add_includes_libs_to_target(), and let FFTX handle the assignment of paths, etc. to the target. Only if an application specifically needs to access the named variables above is it necessary to call FFTX_find_libraries().
External Application Linking With FFTX
A complete example of an external application that builds test programs utilizing the
FFTX libraries is available at
fftx-demo-extern-app.
If you're interested in how to link an external application with FFTX please download
this example and review the CMakeLists.txt therein for specific details. As an
example, in brief, the required steps to add the Poisson test program (using cmake)
are:
set ( POISSON1_TEST poissonTest )
## FFTX_HOME must be defined in the environment or on the command line
if ( DEFINED ENV{FFTX_HOME} )
message ( STATUS "FFTX_HOME = $ENV{FFTX_HOME}" )
set ( FFTX_SOURCE_DIR $ENV{FFTX_HOME} )
else ()
if ( "x${FFTX_HOME}" STREQUAL "x" )
message ( FATAL_ERROR "FFTX_HOME environment variable undefined and not specified on command line" )
endif ()
set ( FFTX_SOURCE_DIR ${FFTX_HOME} )
endif ()
## Include FFTX CMake functions
include ( "${FFTX_SOURCE_DIR}/CMakeIncludes/FFTXCmakeFunctions.cmake" )
add_executable ( ${POISSON1_TEST} ${POISSON1_TEST}.cpp )
target_link_libraries ( ${POISSON1_TEST} PRIVATE dl )
FFTX_add_includes_libs_to_target ( ${POISSON1_TEST} )
The Poisson test program is a CPU only sample program using the FFTX library and RTC interfaces. When the program is run it accepts a single argument, nx, that specifies the dimension of a cube to test. If no argument is provided, a default value of 128 is used. The program may be run for different size cubes as follows:
poissonTest ## defaults to 128^8; runs codegen
poissonTest -nx 80 ## Size 80^3, is present in the FFTX libraries
poissonTest -nx 64 ## Size 64^3, is present in the FFTX libraries
poissonTest -nx 224 ## Size 224^3; runs codegen
For sizes in the FFTX libraries the program calls into the library and runs immediately. For sizes not in the libraries, Spiral is run to generate the required source code, which is then compiled into a temporary library and executed. The source code is cached (meaning that if the specific size is run again, Spiral is not required as the source code is reused).