Bitorch Engine

Readme will be extended soon. This package contains layer to provide fast(er) layer implementations for BITorch.

Installation

Currently, the supported installation method is using pip:

• Without any special cuda requirements (to hide the build output remove -v):

pip install -e . -v

• With higher CUDA versions you may need to install a torch pre-release and/or add an extra index URL:

pip install --upgrade --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cu113

For example, for Cuda 11.6.124 torch==1.12.0.dev20220324+cu113 should work.

Cuda Device Selection

To select a certain CUDA device, set the environment variable BIE_DEVICE, e.g.:

export BIE_DEVICE=1  # use 2nd cuda device

Development

If building fails, adapt the options in cpp_extension.py/ cuda_extension.py.

While developing, a specific cpp/cuda extension can be (re-)build, by using the environment variable BIE_BUILD_ONLY, like so:

BIE_BUILD_ONLY="bitorch_engine/layers/qconv/binary/cpp" pip install -e . -v

It needs to a relative path to one extension directory.

To build for a different CUDA Arch, use the environment variable BIE_CUDA_ARCH (e.g. use 'sm_75', 'sm_80', 'sm_86'):

BIE_CUDA_ARCH="sm_86" pip install -e . -v

MacOS

You should install OpenMP (brew install libomp) with homebrew and make sure to add the corresponding environment variables:

export LIBRARY_PATH=$LIBRARY_PATH:"$(brew --prefix)/lib"
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"$(brew --prefix)/lib"
export CPATH=$CPATH:"$(brew --prefix)/include"
# during libomp installation it should something like this:
export LDFLAGS="-L$(brew --prefix)/opt/libomp/lib"
export CPPFLAGS="-I$(brew --prefix)/opt/libomp/include"