OAO-Translator

A source-to-source translator that translates OpenMP CPU programs to OpenMP Offloading programs with data transmission optimization.
You can compile OAO and run the tests by following these steps.

Installation of OAO

Compiler Preparation

1. Install GCC-8.3.0 and CUDA-10.1. (Higher version of CUDA is not supported by LLVM-9.0.0.)

2. Download source code of LLVM-9.0.0 release. (Other versions may not be compatible with our code, because of defferent APIs.)

3. Compile LLVM for the first time

   You can refer to the documentation of LLVM.
   (https://llvm.org/docs/GettingStarted.html#getting-the-source-code-and-building-llvm)

   Here are the key cmake parameters:
   -DCMAKE_BUILD_TYPE=Release \
   -DCMAKE_C_COMPILER=gcc \
   -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
   -DLLVM_ENABLE_PROJECTS="llvm;clang;clang-tools-extra;compiler-rt;lld;openmp" \
   -DLLVM_TOOL_CLANG_BUILD=ON \
   -DLLVM_TOOL_COMPILER_RT_BUILD=ON \
   -DLLVM_TOOL_LLD_BUILD=ON \
   -DLLVM_TOOL_OPENMP_BUILD=ON \

4. Compile LLVM for the second time to support OpenMP Offloading

   You can refer to the documentation of LLVM.
   (https://llvm.org/docs/GettingStarted.html#getting-the-source-code-and-building-llvm)

   Here are the key cmake parameters:
   -DCUDA_HOST_COMPILER=gcc \
   -DLIBOMPTARGET_NVPTX_CUDA_COMPILER=clang++ \ # the clang compiled in step 3
   -DCLANG_OPENMP_NVPTX_DEFAULT_ARCH=sm_xx \ # compute capability of NVIDIA GPU
   -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=xx \ # Compute Capability of NVIDIA GPU
   -DLIBOMPTARGET_NVPTX_ENABLE_BCLIB=TRUE \
   -DCUDA_BUILD_CUBIN=ON \

5. Patch some header files in LLVM installation path. (You can copy the patches to the corresponding path.)

   • LLVM_installation_path/include/clang/AST/RecursiveASTVisitor.h

     cd LLVM_installation_path/include/clang/AST/
     Make a copy of RecursiveASTVisitor.h. (cp RecursiveASTVisitor.h RecursiveASTVisitorProtected.h)
     Modify "private:" to "protected:" in RecursiveASTVisitorProtected.h. (Two places in total.)

   • LLVM_installation_path/lib/clang/9.0.0/include/__clang_cuda_math_forward_declares.h

     Comment line 49 to 52

   • LLVM_installation_path/lib/clang/9.0.0/include/__clang_cuda_cmath.h

     Comment line 55 to 58

6. Set the environment variables with PATH and LD_LIBRARY_PATH.

Compile and install OAO

1. Open the OAO-Translator/build directory.
2. Modify the CLANG_INSTALL and INSTALL_PATH in makefile.
3. make all && make install

How to run the tests

OAO parameter format

    OAO_installation_path/OAO.bin -fopenmp <compilation_parameters> source_file  

Before using OAO, you should modify your source files according to the following notes.
Notes:

1. All source files should be merged into one unique source file. (The unique source file can include necessary header files.)
2. In the unique source file, functions should be defined at the same time as they are declared, except main() function.
3. The main() function should be declared just after system or compiler header files and before user header files. The main() function should be defined at the end of the unique source file. The OAO treats the declaration of main() as a trigger to start the analysis. Code after the declaration of main() is translated.

Simple demo

    cd OAO-Translator/benchmarks/demo && make translate_demo  
    make demo_offloading

PolyBench and Rodinia

1. Modify pathes in OAO-Translator/benchmarks/path.mk
2. Use OAO-Translator/benchmarks/makefile to translate benchmarks
   cd OAO-Translator/benchmarks && make all # translate all benchmarks
   cd OAO-Translator/benchmarks && make gemm # translate GEMM benchmark
3. Use .sh files in OAO-Translator/benchmarks/shell_PolyBench/ and OAO-Translator/benchmarks/shell_Rodinia/ to run benchmarks and get performance data.

Citation

If you use GPOEO in your research, please cite us as follows:

Wang F, Zhang W, Guo H, et al. Automatic translation of data parallel programs for heterogeneous parallelism through OpenMP offloading[J]. The Journal of Supercomputing, 2021, 77: 4957-4987. https://github.com/ruixueqingyang/OAO-Translator

BibTex:

@article{wang2021automatic,
  title={Automatic translation of data parallel programs for heterogeneous parallelism through OpenMP offloading},
  author={Wang, Farui and Zhang, Weizhe and Guo, Haonan and Hao, Meng and Lu, Gangzhao and Wang, Zheng},
  journal={The Journal of Supercomputing},
  volume={77},
  pages={4957--4987},
  year={2021},
  publisher={Springer}
}