Programming Model

The programming model for oneCCL describes how to:

• Set up distributed computations.

• Perform collective communication operations (for example, ALLREDUCE, BROADCAST, GATHER).

See also

See oneCCL specification that oneCCL is based on.

oneCCL supports a single rank/process per GPU device. The current implementation does not yet support a single process opening multiple devices.

Review the oneCCL generic workflow <https://uxlfoundation.github.io/oneAPI-spec/spec/elements/oneCCL/source/spec/generic_workflow.html> in the specification before getting started with the communication operations.

You can quickly get started with:

• Host Communication
• Device Communication