This is a simple and easy extensible benchmark system to answer the question, which FFT library performs best under which conditions. Conditions are given by compute architecture, inplace or outplace as well as real or complex transforms, data precision, and so on. This project is still in development.
If you want to just browse our results, see the raw benchmark data or our online visualisation and comparison tool.
gearshifft was published at ISC2017. The respective preprint is available on the arxiv.
- cmake 3.7+
- C++14 capable compiler
- CUDA FFT library cuFFT 8.0+ or clFFT 2.12.0+ (OpenCL) or FFTW 3.3.4+
- Boost version 1.59+
- should be compiled with same compiler version or ...
- ... disable the C++11 ABI for GCC with the
-DGEARSHIFFT_CXX11_ABI=OFFcmake option
- half-code by Christian Rau for float16 support (currently used for cufft half precision FFTs)
Go to the gearshifft directory (created by git clone):
mkdir release && cd release
cmake ..
make -j 4CMake tries to find the libraries and enables the corresponding make targets.
After make is finished, you can run e.g. ./gearshifft/gearshifft_cufft.
If the FFT library paths cannot be found, CMAKE_PREFIX_PATH has to be used, e.g.:
export CMAKE_PREFIX_PATH=${HOME}/software/clFFT-cuda8.0-gcc5.4/:/opt/cuda:$CMAKE_PREFIX_PATHEnable float16 support for cuFFT:
cmake -DGEARSHIFFT_FLOAT16_SUPPORT=1 ..
make gearshifft_cufft # automatically downloads half librarygearshifft's dependencies can be build automatically by using the cmake's superbuild options.
Superbuild mode is currently OFF by default, so you have to enable it via cmake option GEARSHIFFT_USE_SUPERBUILD=ON.
An example for downloading and building Boost, clFFT and FFTW::
cmake -DGEARSHIFFT_USE_SUPERBUILD=ON \
-DGEARSHIFFT_SUPERBUILD_EXT_INBUILD=OFF \
-DGEARSHIFFT_SUPERBUILD_EXT_DIR=$HOME/sources/deps \
-DGEARSHIFFT_SUPERBUILD_EXT_DOWNLOAD_Boost=ON \
-DGEARSHIFFT_SUPERBUILD_EXT_DOWNLOAD_CLFFT=ON \
-DGEARSHIFFT_SUPERBUILD_EXT_DOWNLOAD_FFTW=ON \
-DGEARSHIFFT_SUPERBUILD_EXT_DOWNLOAD_ROCFFT=OFF \
-DGEARSHIFFT_USE_STATIC_LIBS=ON \
..
Notes:
GEARSHIFFT_SUPERBUILD_EXT_INBUILD=OFFinstalls the dependency builts into a separate directory ($GEARSHIFFT_SUPERBUILD_EXT_DIR), otherwise builts are located in the binary directory ($CMAKE_BINARY_DIR}).- If
GEARSHIFFT_USE_STATIC_LIBS=OFFthen you have to take care of environment setting such asLD_LIBRARY_PATH GEARSHIFFT_USE_STATIC_LIBS=ONprobably requires to build Boost manually- To avoid clashes Boost's unit test suite main method is not build (see Boost doc)
- Boost's utf main is added with
BOOST_TEST_DYN_LINKdefinition (automatically done in our cmake)
- Use cmake option
GEARSHIFFT_VERBOSE=ONto get more information during the cmake's build generation process
If the libraries have been build separately, just provide the correct path. An example:
BOOST_VER=1.67.0
FFTW_VER=3.3.8
BOOST_ROOT=${HOME}/sw/boost-${BOOST_VER}
FFTW_ROOT=${HOME}/sw/fftw-${FFTW_VER}/
export CMAKE_PREFIX_PATH==${BOOST_ROOT}/lib:${BOOST_ROOT}/include:${FFTW_ROOT}:${CMAKE_PREFIX_PATH}
if [ -d "CMakeFiles" ]; then
make clean
rm -rf CMakeFiles CMakeCache.txt cmake_install.cmake Makefile
fi
cmake ..
makeUse cmake superbuild or follwing compile script.
## compile_boost.sh
# boost version to download
BOOST_VERSION=1.67.0
# where to download
BOOST_SRC="${HOME}/Downloads/boost"
# boost install dir
BOOST_ROOT="${HOME}/sw/boost-${BOOST_VERSION}"
# if install dir is empty, then download && build
if [[ -z "$(ls -A ${BOOST_ROOT})" ]]; then
mkdir -p ${BOOST_SRC}
wget http://sourceforge.net/projects/boost/files/boost/${BOOST_VERSION}/boost_${BOOST_VERSION//\./_}.tar.bz2 -nc -O "${BOOST_SRC}/../boost.tar.bz2"
(cd ${BOOST_SRC}/../; tar jxf boost.tar.bz2 --strip-components=1 -C ${BOOST_SRC})
(cd ${BOOST_SRC}; ./bootstrap.sh --with-libraries=program_options,filesystem,system,test)
(cd ${BOOST_SRC}; ./b2 --prefix="$BOOST_ROOT" -d0 install --variant=release)
fiUse cmake superbuild or follwing compile script.
## compile_fftw.sh
# fftw version to download
VERS=3.3.8
# we compile in separated directories, so binaries do not clash
VERS_single=${VERS}_single
VERS_double=${VERS}_double
# the install directory for fftw and fftwf
FFTW_ROOT=${HOME}/sw/fftw-${VERS}/
# "./compile_fftw.sh clean" removes directories
if [ "$1" == "clean" ]; then
echo "clean sources"
rm -rf fftw-${VERS_single} fftw-${VERS_double}
fi
# if directories do not exist, create them and unpack fftw_**.tar.gz
if [ ! -d "fftw-${VERS_single}" ] && [ ! -d "fftw-${VERS_double}" ]; then
wget http://www.fftw.org/fftw-${VERS}.tar.gz
mkdir fftw-${VERS_single}
echo "unpacking"
tar xfz fftw-${VERS}.tar.gz -C fftw-${VERS_single}
cp -r fftw-${VERS_single} fftw-${VERS_double}
fi
IFLAG_STD="--enable-static=yes --enable-shared=yes --with-gnu-ld --enable-silent-rules --with-pic"
IFLAGS="--prefix=${FFTW_ROOT} --enable-openmp --enable-sse2 -q $IFLAG_STD"
# float
cd fftw-${VERS_single}/fftw-${VERS}
./configure $IFLAGS "--enable-float"
make -j8
make install
# double
cd ../../fftw-${VERS_double}/fftw-${VERS}
./configure $IFLAGS
make -j8
make installSet CMAKE_INSTALL_PREFIX as you wish, otherwise defaults are used.
mkdir release && cd release
cmake -DCMAKE_INSTALL_PREFIX=${HOME}/gearshifft ..
make -j 4 installSuperbuild mode in cmake can be used to automatically download and compile the dependencies. It is also intended for creating packages together with static linking of the dependencies. It also supports multi-arch packaging of gearshifft targets, which have been compiled with different compiler.
cmake -DGEARSHIFFT_USE_SUPERBUILD=ON ..
make -j
make gearshifft-packageThe tests can be executed by make test after you have compiled the binaries.
cmake -DBUILD_TESTING=ON ..
make -j
cd gearshifft-build/ # because superbuild generates a build subfolder for gearshifft
make testSee help message (pass --help|-h) for the command line options.
-h [ --help ] Print help messages
-e [ --extent ] arg Specific extent (eg. 1024x1024) [>=1 nr. of
args possible]
-f [ --file ] arg File with extents (row-wise csv) [>=1 nr.
of args possible]
-o [ --output ] arg (=result.csv) output csv file, will be overwritten!
-t [ --add-tag ] arg Add custom tag to header of output file
-v [ --verbose ] Prints benchmark statistics
-V [ --version ] Prints gearshifft version
-d [ --device ] arg (=gpu) Compute device = (gpu|cpu|acc|<ID>). If
device is not supported by FFT lib, then it
is ignored and default is used.
-n [ --ndevices ] arg (=0) Number of devices (0=all), if supported by
FFT lib (e.g. clfft and fftw with n CPU
threads).
-l [ --list-devices ] List of available compute devices with IDs,
if supported.
-b [ --list-benchmarks ] Show registered benchmarks
-r [ --run-benchmarks ] arg Run specific benchmarks (wildcards
possible, e.g. ClFFT/float/*/Inplace_Real)
... further backend specific options ...
gearshifft comes with preset files which contain a broad range of FFT extents and are located in:
/gearshifft/share/gearshifft/*.conf
For FFTW benchmarks with, e.g., FFTW_MEASURE runtimes grow very quickly. You can delete or comment out the unwanted lines in the configuration file with #.
The extents configurations are loaded by the gearshifft command-line interface:
./gearshifft_fftw -f myextents.confRuns complete benchmark for clFFT (also applies for cuFFT, FFTW, ..)
./gearshifft_clfft
# equals
./gearshifft_clfft -f myextents.conf -o result.csv -d gpu -n 0List compute devices
./gearshifft_clfft -l
$ "ID",0:0,"ClFFT Informations","Device","Tesla K20Xm","Hardware","OpenCL 1.2 CUDA" <snip>
$ "ID",1:0,"ClFFT Informations","Device","Intel(R) Xeon(R) CPU E5-2450 0 @ 2.10GHz" <snip>
Select compute devices by id returned by --list-devices|-l
./gearshifft_clfft -d 1:0
512x512-point FFT, single and double precision, all transforms, print result to console.
./gearshifft_clfft -e 512x512 -v
1048576-point FFT, only real outplace transform in single precision.
./gearshifft_clfft -e 1048576 -r */float/*/Outplace_Real
1024-point and 128x256x64-point FFTs, on CPU with 1 thread.
- If no device parameter is given, then ClFFT/OpenCL will use GPU, if available
./gearshifft_clfft -e 1024 128,256,64 -d cpu -n 1
1024x1024-point FFT, double precision inplace transforms.
--list-benchmarks|-bgives a list of available extents read in--file|-f(default is ../config/extents.csv)
./gearshifft_clfft -r */double/1024x1024/Inplace_*
gearshifft_fftw offers some additional options, e.g., for the plan rigor or the time limit of plans, just check ./gearshifft_fftw -h.
FFTW itself allows you to pregenerate FFT plans by the so-called wisdom files. These can be generated by FFTW tools, e.g.:
./fftw-3.3.5_single/tools/fftwf-wisdom -v -c -n -T 24 -o wisdomf # single precision
./fftw-3.3.5_double/tools/fftw-wisdom -v -c -n -T 24 -o wisdom # double precisionIt is recommended to compile double-precision (fftw-...) and single-precision (fftwf-...) to separate directories to avoid linker issues.
The wisdoms settings must match the gearshifft_fftw configuration (number of cores, precision, extents).
Most of the times you do not benefit from a multi-core setting, because the FFT is already computed almost in no time.
FFTW spends a lot of time in planning, except you use FFTW_ESTIMATE or time limits, which are also only lower borders.
The FFT scenario is a roundtrip FFT, i.e. forward and backward transformation. The result is compared with the original input data and an error is shown, if there was a mismatch. If a benchmark cannot be completed due to an error, it proceeds with the next benchmark. The library dependent FFT steps are abstracted, where following steps are wrapped by timers.
- buffer allocation
- plan creation (forward/backward plan)
- memory transfers (up-/download)
- forward and backward transforms
- cleanup
- total time (allocation, planning, transfers, FFTs, cleanup)
- device initialization/teardown (only once per runtime)
Furthermore, the required buffer sizes to run the FFT are recorded.
The results of the benchmark runs are stored into a comma-separated values file (.csv), after the last run has been completed. To ease evaluation, the entries are sorted by columns
- FFT transform kind and precision
- dimkind (oddshape, powerof2, radix357)
- oddshape: at least one extent is not a combination of a power of 2,3,5,7
- powerof2: all extents are powers of 2
- radix357: extents are combination of powers of 2,3,5,7 and not all are powers of 2
- extents and runs
See CSV header for column titles and meta-information (memory, number of runs, error-bound, hostname, timestamp, ...).
- linux (CentOS, RHEL, ArchLinux, Ubuntu, Fedora)
- gcc 5.3.0, gcc 6.2.0, gcc 7.1.1, gcc 8.1.1, clang 3.8 & 4.0.1 (FFTW threads disabled)
- CUDA 8.0.x, CUDA 9.x
- clFFT 2.12.0, 2.12.1, 2.12.2
- FFTW 3.3.4, 3.3.5, 3.3.6pl1, 3.3.8
- OpenCL 1.2 (Nvidia, AMD, Intel)
- Nvidia Pascal V100, P100, Kepler K80, K20Xm, GTX1080, Broadwell, Haswell and Sandybridge Xeon CPUs
- clFFT does not support arbitrary transform sizes. The benchmark renders such tests as failed.
- clFFT on CPU cannot transform the 4096-FFT and 4096x4096-FFTs (see this issue)
- clFFT seems to have lower transform size limits on CPU than on GPU (a complex 16384x16384 segfaults on clfft CPU, while it works on GPU). gearshifft marks these cases as "Unsupported lengths" and skips them.
- rocFFT is currently only supported on the HCC platform (no support for rocFFT->cuFFT path)
- At the moment this is for single-GPUs, batches are not considered
- if
gearshifftis killed before, no output is created, which might be an issue on a job scheduler system like slurm (exceeding memory assignment, out-of-memory killings) - in case the Boost version (e.g. 1.62.0) you have is more recent than your
cmake(say 2.8.12.2), usecmake -DBoost_ADDITIONAL_VERSIONS=1.62.0 -DBOOST_ROOT=/path/to/boost/1.62.0 <more flags> - Windows or MacOS is not supported yet, feel free to add a pull-request
- cufft float16 transforms overflow at >=1048576 elements