Core Principles · Celerity

The Celerity API and runtime system follow a small number of core principles, which help us to focus on the goals of the project and guide our design decisions.

Data flow over control flow - dependencies in computations using the Celerity API should be defined by their respective data flow rather than explicit control flow. Avoid creating two potentially redundant synchronization paths.
Everything is asynchronous - in the runtime implementation, all command generation and data transfers should happen asynchronously. This applies both across distributed memory nodes and between individual runtime threads on each node.
When in doubt, stay close to SYCL - when introducing new API functionality and concepts, try to stick as closely as possible to SYCL, to ease the migration path for new users. Of course, some functionality needs to be reconsidered for distributed memory.