Open Source Summit India 2026

Modern embedded audio devices increasingly run mainline Linux, but achieving deterministic low-latency multi-channel audio processing remains a challenge. While heterogeneous SoCs include DSP accelerators capable of handling signal processing workloads efficiently, integrating these accelerators into a standard Linux audio stack without proprietary middleware is non-trivial.
This talk presents a practical, upstream-friendly approach to building a zero-copy DSP offload framework using mainline Linux components such as remoteproc, rpmsg, and ALSA. Instead of relying on custom kernel patches or vendor-specific frameworks, the solution leverages DMA-backed shared memory and rpmsg-based signaling to enable efficient inter-processor communication between ARM application cores and a DSP.
A key focus of this session is eliminating redundant memory copies across kernel and user space boundaries. By designing a ping-pong buffer architecture with shared memory mapping and pointer-based synchronization, we achieved deterministic real-time streaming with significantly reduced CPU utilization and improved latency characteristics.