convDirect

The convDirect library provides different implementations of the direct convolution operation.

Compilation and installation

To compile and install all the convDirect implementations in the convDirect library, the next software should be available:

• the BLIS library https://github.com/flame/blis, and
• the Apache TVM compiler https://tvm.apache.org/.

As convDirect also provides a direct call to convGemm, after cloning this project, the corresponding git submodule should be initialized with:

git submodule update --init

In case you want to update the convGemm submodule to the most recent version, the next command should be executed:

git submodule update --remote --merge

Please note that CMake is used to build convDirect. In case you require a more recent version of CMake, you can download it from https://cmake.org/download/ To install CMake from a compressed CMake archive, just uncompress it and add the 'cmake-VERSION-ARCH/bin' directory to the system PATH.

To compile and install the convDirect library, please execute the next commands:

cd build
cmake -D CMAKE_PREFIX_PATH="BLIS_INSTALL_PREFIX;TVM_INSTALL_PREFIX" -D CMAKE_INSTALL_PREFIX=INSTALL_PREFIX ..
make                 # Alternatively:  cmake --build . --clean-first
make install         # Alternatively:  cmake --install .

where BLIS_INSTALL_PREFIX is the prefix PATH where BLIS is installed, TVM_INSTALL_PREFIX is the prefix PATH where TVM is installed, if it is installed on a different path than BLIS, and INSTALL_PREFIX is the prefix PATH where lib/libconvDirect.so and include/convdirect.h will be installed.

The -D CMAKE_PREFIX_PATH=BLIS_INSTALL_PREFIX option on the first cmake command only is necessary if:

1. BLIS is not installed in a system PATH,
2. the environment variable LD_LIBRARY_PATH is not defined or does not include the BLIS_INSTALL_PREFIX PATH, and
3. BLIS_INSTALL_PREFIX is different of INSTALL_PREFIX.

The same applies to TVM_INSTALL_PREFIX, but for Apache TVM.

As for the -D CMAKE_PREFIX_PATH=INSTALL_PREFIX option, it is only required if the convDirect library should be installed on a prefix PATH different of /usr/local.

For example, if BLIS is installed under ~/opt/hpca_pydtnn and the convDirect library should be installed also under that directory, the next commands are sufficient:

cd build
cmake -D CMAKE_INSTALL_PREFIX=~/opt/hpca_pydtnn ..
make                 # Alternatively:  cmake --build . --clean-first
make install         # Alternatively:  cmake --install . (this does not work with cmake older versions)

Performance evaluation of the different implementations

To evaluate all the provided implementations on convDirect, recompile convDirect adding the -D COMPILE_TESTS=ON option and execute the evaluation target. For example as in:

cd build
cmake -D CMAKE_INSTALL_PREFIX=~/opt/hpca_pydtnn -D COMPILE_TESTS=ON ..
make evaluation      # Alternatively:  cmake --build . --target=evaluation

A Python script is also provided to automatically process the evaluation results and generate the corresponding plots. Once the evaluation is completed, the instructions to run it locally or on another machine will be shown on screen. This script requires the next Python modules: numpy, pandas, tabulate, matplotlib.

Testing the different implementations

To test the different implementations against a reference convolution, recompile convDirect adding the -D COMPILE_TESTS=ON option and execute the all_close_test target. For example as in:

cd build
cmake -D CMAKE_INSTALL_PREFIX=~/opt/hpca_pydtnn -D COMPILE_TESTS=ON ..
make all_close_test  # Alternatively:  cmake --build . --target=all_close_test

Please note that the performance results shown while performing this test can not be accurate as only one iteration of each evaluated convolution is performed.