MWE using cmake to compile MEX and/or binary with or without CUDA support and additional libraries

This work is build upon the following two helpful links: CUDA-MEX-CMAKE and MEX-CMAKE

This is a CMake code template to demonstrate

• how to compile a Matlab MEX file with or without CUDA capability AND
• how to compile a binary with or without CUDA capability at the same time.

Furthermore, libraries can be included that

• are used by the binary only (e.g. OpenCV)
• are used by CUDA only (e.g. cuBLAS & cuFFT)
• are used by the binary and the mex file (e.g. Eigen).

Requirements

• CMake (mandatory)
• CUDA (optional)
• MATLAB (optional)
• OpenCV (optional dummy)
• Eigen3 (optional dummy)

Tested

This code has been tested under

• Ubuntu 16.04; CMake 3.5.1; CUDA 8.0; MATLAB R2016b
• TODO: Test using MacOS (without CUDA support) and Windows (with and without CUDA support)

Getting started

• Set MATLAB_ROOT environment variable in to your installed matlab path, such as
  export MATLAB_ROOT='/usr/local/MATLAB/R2016b' in ~/.bashrc (Linux)
  export MATLAB_ROOT='/Applications/MATLAB_R2016b.app' in ~/.bash_profile (MacOS)
  Add system variable MATLAB_ROOT=D:\\Program Files\\MATLAB\\R2016b (Windows) (not yet tested)
• (Optional) Copy the Eigen directory from Eigen into the third_party directory.
• In Terminal do
  cd path/to/mweCmakeMexCppCuda/build
cmake ..
make
• in build/lib is the mex file and in build/bin is the binary
• cd ../example
• execute the matlab script and/or the bash script

Information

• The cmake file is supposed to be generic enough to easily integrate this to existing code and/or to easily built up on it to generate a versatile cpp/mex/cuda project.

Possibilities

• Decide whether to generate a mex file: SET(MEXF {TRUE,FALSE})
• Decide whether to generate a binary file: SET(EXEF {TRUE,FALSE})
• Decide whether to have CUDA support: SET(CUDA {TRUE,FALSE})
• Decide whether you would like to use OpenCV: SET(OPENCV_INSTALLED {TRUE,FALSE})
• Decide whether you would like to use Eigen: SET(USE_EIGEN {TRUE,FALSE})

File linking

There are multiple integrated ways to link source code e.g.,

• CPP_MAIN_FILES are the main files for the binary
• MEX_MAIN_FILES are the main files for the mex support
• COMMON_SRC_FILES are files that are needed for both, cuda and non-cuda support (supposed to be some middlesman files to easily split cuda from non-cuda support
• CU_SRC_FILES are all the cuda files that replace the corresponding cpp files
• CPP_SRC_FILES are all the cpp files that replace the cuda files

Library setup

Additionally, this framework uses OpenCV and Eigen as some dummy libraries to show the versatility of this framework.

• OpenCV is used only for the binary build
• Eigen is used in the case of building MEX and binary files with or without CUDA support
• CUBLAS and CUFFT libraries are linked and can be easily extended

Tips & Troubleshooting

• It can happen that the building process during make fails. This can be due to the fact that the gpuadd.o is build twice (haven't found a way to fix this yet). Run make a couple of times to overcome this issue or first build with SET(MEXF TRUE) and afterwards with SET(EXEF TRUE). For me this error happend every ~10th building process, so I didnt' see the urge of fixing it.

• If Matlab throws an error like

Invalid MEX-file '/mweCmakeMexCppCuda/build/lib/mweAddMEX.mexa64':
/MATLAB/R2016b/bin/glnxa64/../../sys/os/glnxa64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found (required by /mweCmakeMexCppCuda/build/lib/mweAddMEX.mexa64)

then start Matlab from terminal and preload the the /usr/lib/x86_64-linux-gnu/libstdc++.so.6 library:

LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libstdc++.so.6" matlab

• If you get some linker error of memcpy within string.h, uncomment -D_FORCE_INLINES in CMakeLists.txt
  This can be due to using an older CUDA version.

• If during the compile process with CUDA support there is a reference to unknown error and you use cuda<8.0, uncomment the line calling cublasAxpyEx in src/lib/cuda/gpuadd.cu