README.md

Build and install oneDNN:

wget https://github.com/oneapi-src/oneDNN/archive/refs/tags/v2.5.1.tar.gz    # get v2.5.1 source code
tar -xzvf v2.5.1.tar.gz && cd oneDNN-2.5.1/ # extract the source code
mkdir build && cd "$_"    # setup a build dir
cmake -DCMAKE_INSTALL_PREFIX=../install .. && make -j && make install    # build and install
export oneDNN_HOME=`pwd`/../install
export CPLUS_INCLUDE_PATH=${CPLUS_INCLUDE_PATH}:${oneDNN_HOME}/include
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${oneDNN_HOME}/lib64

build TVM

get tvm

git clone --recursive https://github.com/yangulei/tvm -b dev_byoc
cd tvm

build for tvm benchmark

mkdir build_release_native && cd "$_"
cp ../cmake/config.cmake .
sed -i '/USE_LLVM/ s/OFF/ON/g' config.cmake    # Enable LLVM
cmake .. -DCMAKE_BUILD_TYPE=Release && make -j

build for byoc benchmark

cd .. && mkdir build_release_gnu && cd "$_"
cp ../cmake/config.cmake .
sed -i '/USE_LLVM/ s/OFF/ON/g' config.cmake    # Enable LLVM
sed -i '/USE_DNNL_CODEGEN/ s/OFF/ON/g' config.cmake    # Enable oneDNN
sed -i '/USE_OPENMP/ s/none/gnu/g' config.cmake    # needed for oneDNN-BYOC, but slow for native TVM codegen
cmake .. -DCMAKE_BUILD_TYPE=Release -DEXTERN_LIBRARY_DNNL=${oneDNN_HOME}/lib64/libdnnl.so
make -j && cd ..

setup env

export TVM_HOME=`pwd`
export PYTHONPATH=${TVM_HOME}/python:${PYTHONPATH}

Setup python environment

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh && source ~/.bashrc
conda create -n tvm python=3.8
conda activate tvm
pip install numpy decorator attrs tornado psutil xgboost cloudpickle pytest
pip install mxnet==1.8.0.post0 gluoncv

benchmark

cd bench_oneDNN_BYOC/
bash bench.sh