LibreRT: Portable Heterogeneous Real-Time Programming for the Embedded Computing Continuum

The LibreRT project aims to address the lack of portable system software that can efficiently target the full spectrum of embedded devices, ranging from ultra-low-power micro-edge micro-controllers to deep- and meta-edge systems with embedded GPUs. It is based on three technological pillars: (1) Portable programming models and OS support for highly responsive real-time processing on micro-edge devices; (2) Portable programming for high-throughput realtime processing on heterogeneous deep- and meta-edge systems; (3) Real-time processing leveraging approximate computing and autotuning. LibreRT proposes a number of new technology advances: the programming model, based on the SYCL standard, has been extended with two novel extensions: SYrtos, which introduces priority- and deadline-based scheduling into SYCL with preemptive multitasking on OS integration; and SYprox, which provides approximate computing techniques integration into SYCL language semantics. The LibreRT software stack integrates programming models and compiler with complementary operating system technologies. Specifically, LibreRT proposes a new operating system kernel architecture called fluid kernel, designed for embedded system support, and a multi-core real-time scheduler for micro-edge devices and a high-resolution timing subsystem. Both features have been implemented in the Miosix OS. Overall, the LibreRT software stack has been validated on several benchmarks, including AI and image processing applications, and demonstrated in a real-world real-time scenario using a thermal camera prototype.