SWStenDSL

A stencil domain specific language supporting Sunway TaihuLight.

Build Instructions

SWStenDSL depends on a build of llvm including MLIR. So you need to build and install llvm first. The repository is developed based on LLVM project commit e2dee9af8db.

Requirement

Dependency	Version
cmake	3.19.7
ninja	1.10.2
gcc	10.2.0
python	3.9.6

we also have tested on Ubuntu 20.04.1 LTS

Build LLVM project

We assume that you build LLVM and MLIR in build directory and install them to <INSTALL_DIR>.

git clone https://github.com/llvm/llvm-project.git
cd llvm-project
git checkout e2dee9af8db
mkdir build && cd build
cmake -G Ninja ../llvm -DLLVM_BUILD_EXAMPLES=OFF -DLLVM_TARGETS_TO_BUILD="host" -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> -DLLVM_ENABLE_PROJECTS=mlir -DLLVM_OPTIMIZED_TABLEGEN=ON -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_BINDINGS=OFF -DLLVM_INSTALL_UTILS=ON #-DLLVM_PARALLEL_LINK_JOBS=2
cmake --build . --target install

Build SWStenDSL project

git clone https://github.com/JackMoriarty/SWStenDSL.git
cd SWStenDSL
export LLVM_INSTALL=<path_to_llvm_install>
export LLVM_BUILD=<path_to_llvm_build_directory>
mkdir build && cd build
cmake -G Ninja .. -DMLIR_DIR=$LLVM_INSTALL/lib/cmake/mlir -DLLVM_EXTERNAL_LIT=$LLVM_BUILD/bin/llvm-lit
cmake --build . --target stenCC

The execuatable binary will be generated in path SWSten/build/bin, named stenCC.

Build SWStenDSL parallel communication library

cd SWStenDSL/utils
make

The library will generated in current directory SWsten/utils, named libswstenmpi.a

How To Use

simple run

After building all project, You can run the example program provided by SWStenDSL. There is a sample command.

cd SWStenDSL/build
./bin/stenCC ../test/Examples/StencilDialect/laplaceAddInputIteration/laplaceAddInputIteration.dsl --emit=sw > laplaceAddInputIteration.sw 2>&1
python3 ../tool/translate.py laplaceAddInputIteration.sw

After running those command, three C source file will be generated. We also provide driver program and build script in ${SWStenProject}/test/Examples/StencilDialect/laplaceAddInputIteration/.

You need create a directory on Sunway Taihulight supercomputer, then copy three source file, driver program, buildscript and the ${SWStenDSL}/utils into the directory you created.

cd <path_to_directory_you_created>
make
make run

Then, the executable program will be generated and be submited to Sunway TaihuLight supercomputer.

MPI large-scale computing support

If you need run large-scale stencil computing, you need modify the dsl file and build script.

In dsl file, you need add mpiTile and mpiHalo keyword. For more information please reference ${SWStenDSL}/doc/StencilDSL.md.

In build script, you need declare a marco SWStenMPI, the build script in ${SWStenDSL}/examples/2d9pt_box/makefileMPI is an example.

enable optimization option

Now, we provide two optimization options: --kernel-fusion, --enable-vector=<vector-width>.

kernel-fusion will merge the kernels if one kernel depends on another kernel result. This option will reduce the number of spe launching, and reduce the number of data transport bewteen main memory and LDM.

enable-vectorwill enable vectorizable computation on spe.

There is a sample command.

./bin/stenCC --kernel-fusion --enable-vector=4 ../test/Examples/StencilDialect/laplaceAddInputIteration/laplaceAddInputIteration.dsl --emit=sw > laplaceAddInputIteration.sw 2>&1