Convolutional

Reference implementations of 2-dimensional tensor convolution

Requirements

• CMake 3.5 or newer
• On Windows: Visual C++ 2017
• On Linux: gcc 5.4 or newer

Setting paths to BLAS

This code does not automatically detect the path to your computer's BLAS implementation. This allows you to explicitly choose which BLAS implementation to use. Before building, edit the file BlasConfig.cmake and set the paths manually. For example, for Intel MKL BLAS on Windows 10, the BLAS configuration looks like this:

if(${BLAS_VERSION} MATCHES "winmkl") 
    set(INTEL_ROOT "/Program Files (x86)/IntelSWTools/compilers_and_libraries/windows")
    set(BLAS_COPY_DLLS 
        ${INTEL_ROOT}/redist/intel64/mkl/mkl_rt.dll
        ${INTEL_ROOT}/redist/intel64/mkl/mkl_intel_thread.dll
        ${INTEL_ROOT}/redist/intel64/mkl/mkl_core.dll
        ${INTEL_ROOT}/redist/intel64/mkl/mkl_avx2.dll
        ${INTEL_ROOT}/redist/intel64/compiler/libiomp5md.dll
    )
    set(BLAS_HEADER_FILE "mkl_cblas.h")
    set(BLAS_INCLUDE_DIRS ${INTEL_ROOT}/mkl/include)
    set(BLAS_LIBRARIES ${INTEL_ROOT}/mkl/lib/intel64/mkl_rt.lib)
    set(USE_BLAS TRUE)
endif()

The relevant CMake variables that need to be set are:

• BLAS_COPY_DLLS - a list of DLLs that are copied into the executable directory.
• BLAS_HEADER_FILE - the name of the BLAS header file to include in the code.
• BLAS_LIBRARIES - a list of libraries to provide to the linker.
• USE_BLAS - must be set to true, otherwise BLAS is not used and the results are very slow and meaningless.

Build and execute on Windows

After cloning the repository, cd into the main repository directory, create a new directory named build and cd into that directory. Next, type the command

cmake -DBLAS_VERSION=winmkl -G "Visual Studio 15 2017 Win64" ..

where the BLAS_VERSION parameter (set to winmkl in the example above) matchs the configuration that you defined in BlasConfig.cmake (see instructions above). Finally, to build the executable, type cmake --build . --config Release. The new executable will appear as \build\bin\convolutional.exe. These instructions are summarized in build.cmd.

To run the test, cd back to the main project directory and type build\bin\convolutional.exe benchmarks.csv. Edit benchmarks.csv to control the filter and output shapes used in the test.

Build and execute on Linux

After cloning the repository, cd into the main repository directory, create a new directory named build and cd into that directory. Next, type the command

cmake -DBLAS_VERSION=ubuntuopenblas ..

where the BLAS_VERSION parameter (set to ubuntuopenblas in the example above) matchs the configuration that you defined in BlasConfig.cmake (see instructions above). Finally, to build the executable, type cmake --build . --config Release. The new executable will appear as build/convolutional. These instructions are summarized in build.sh.

To run the test, cd back to the main project directory and type build/convolutional.exe benchmarks.csv. Edit benchmarks.csv to control the filter and output shapes used in the test.