Build PyTorch with MKL & OpenMP threading on macOS

build_pytorch_openmp_mkl_macOS.sh

# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# This install used the following development environment:
# - We have installed xcode command line tools
# - We have installed cmake via homebrew (my version is 3.20.2)
# - We have created a fresh conda environment "pytorch-dev" and installed 
#   python=3.8 and all of the packages from pytorch/README.md except cmake 
#   (including the torch.distributed deps)
# - We have copied the relevant OpenMP header files into a fresh directory at ~/openmp
# - We have cloned the pytorch github repo to ~/pytorch
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

# conda environment where we have installed dependencies
conda activate pytorch-dev

# root directory of our conda environment
CONDA_ENV=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}

# directory where we have placed the intel openmp header files
OMP_INCLUDE=$HOME/openmp/include

# these two exports are necessary
export CMAKE_PREFIX_PATH=$CONDA_ENV
export CMAKE_INCLUDE_PATH=$OMP_INCLUDE

# not 100% sure these two exports are necessary, but worked for me
export CMAKE_SHARED_LINKER_FLAGS="$CONDA_ENV/lib/libiomp5.dylib -Wl,-rpath,$CONDA_ENV/lib"
export CMAKE_EXE_LINKER_FLAGS="$CONDA_ENV/lib/libiomp5.dylib -Wl,-rpath,$CONDA_ENV/lib"

export MACOSX_DEPLOYMENT_TARGET=11.0
export CC=clang
export CXX=clang++
export MAX_JOBS=9

export BLAS=MKL
export USE_OPENMP=1
export MKL_THREADING=OMP
export ATEN_THREADING=OMP
export MKLDNN_CPU_RUNTIME=OMP

# directory where we have cloned pytorch
cd $HOME/pytorch

python setup.py install