Lloyd's algorithm implementation using DPC++

Implementation closely following https://github.com/soda-inria/sklearn-numba-dpex to compare performance of numba_dpex-based implementation and DPC++ implementation.

DPC++ implementation can be compiled using intel/llvm open-source DPC++ compiler targeting CUDA, or HIP backends.

This is a work-in-progress work.

Building

1. Create a Python 3.9 conda environment with ninja as a build system.

conda create -n env_name python=3.9 cmake ninja

2. Activate this environment.

conda activate env_name

3. Install packages from PyPi

pip install --no-cache-dir packaging setuptools distro scikit-build pytest numpy

4. Activate oneAPI compiler and TBB by exporting environment variables:

source /opt/intel/oneapi/compiler/latest/env/vars.sh
source /opt/intel/oneapi/tbb/latest/env/vars.sh

5. Install dpctl

   • Build it from source

     git clone https://github.com/IntelPython/dpctl
     pushd dpctl
     python scripts/build_locally.py
     popd

   • Alternatively you can also install dpctl from the dppy/label/dev conda channel, but you need to ensure that oneAPI compiler's version matches the version of the dpcpp-cpp-rt conda package installed as dependency of the dpctl conda package.

6. Install kmeans_dpcpp (this project) from source

git clone https://github.com/oleksandr-pavlyk/kmeans_dpcpp.git
cd kmeans_dpcpp
python setup.py develop -- -DCMAKE_CXX_COMPILER:PATH=$(which icpx) -DDPCTL_MODULE_PATH=$(python -m dpctl --cmakedir)

Notice that -DCMAKE_CXX_COMPILER setting is necessary for find_package(IntelDPCPP REQUIRED) to work.

One can alternatively use CXX environment variable to specify the compiler:

CXX=icpx python setup.py develop -- -DDPCTL_MODULE_PATH=$(python -m dpctl --cmakedir)

Running tests

python -m pytest -s tests/

Building fat binary for both SPIRV64 and NVPTX devices

Assuming that CodePlay's plugin enabling support of NVidia GPUs is installed, as well as CUDAToolkit and the drivers:

CXX=icpx python setup.py develop -- -DDPCTL_MODULE_PATH=$(python -m dpctl --cmakedir) -DCUDA=ON

dpctl compiled with CUDA support is also needed.