Getting Started with oneCCL

Getting Started with oneCCL#

Note

This document describes the new C API that closely follows the NVIDIA Collective Communications Library (NCCL)* API standard. Documentation for the legacy C++ API can be found here.

Introduction#

Welcome to the getting started guide for oneCCL, the Intel oneAPI Collective Communications Library. This document guides you through the installation process and provides instructions for building and running the example code using our latest NCCL like C API.

Installation#

Linux build#

oneCCL provides support only for Linux and leverages the cmake build system. Different build options are available to suit CPU only and CPU+GPU targets. To build oneCCL with GPU support, specify -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx.

git clone https://github.com/uxlfoundation/oneCCL
cd oneCCL

mkdir build
cd build

cmake .. -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DCMAKE_BUILD_TYPE=debug -DCMAKE_INSTALL_PREFIX=$(pwd)/_install
make -j install

Usage#

Launching Example Application#

To experiment with oneCCL, use examples in the examples/sycl directory. The examples are built by default if the compiler supports SYCL (e.g. the icpx compiler). Before running the app, initialize the shell environment by sourcing the vars.sh script, installed in <install_dir>/env/vars.sh:
source <install_dir>/env/vars.sh

Warning

By default the vars.sh script will initialize the users environment using the Intel® MPI Library distribution bundled with oneCCL. To avoid overriding the configuration of custom systems (for example, systems using custom MPI impementation) use source vars.sh --ccl-bundled-mpi=no.

$ source <install_dir>/env/vars.sh
$ CCL_LOG_LEVEL=info mpirun -n 2 <install_dir>/examples/sycl/sycl
Contents