Building a Zero-Copy DSP Offload Framework in Linux Using RPM...
Talkat Open Source Summit India 2026 · on Tuesday, 2026-06-16· at 14:00· Lotus Ballroom 2 (Level 3)·
summary
A practical upstream-friendly framework uses remoteproc, rpmsg and ALSA with DMA-backed shared memory and a zero-copy ping-pong buffer design to avoid redundant copies; an RPMsg DMA library handles pointer-based sync and notifications. An 8‑channel, 48 kHz demo runs real-time on C7X, lowering host CPU from ~37% to ~21% and latency from ~2 ms to ~0.33 ms versus ARM-only processing. Works without kernel patches and reuses existing DSP kernels; repository and examples are open source.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:50Problem statement
- 6:49Proposed solution
- 13:09Software architecture
- 16:01Demo: audio file processing
- 18:10Performance ARM vs DSP
- 20:18Demo: live audio streaming
- 23:33Upstream-friendly design
- 25:22Status and future plans
- 28:12Open source resources
- 29:14Q&A
related talks
51mUsing GStreamer for Seamless Off-loading Audio Processing
27mBuilding a Zephyr-Native Audio Framework With Sound Open Firmware
39mSoftware Defined Hardware on Heterogeneous SoCs Using OpenAMP
32mHeterogeneous Linux and RTOS Software Architecture for Low-Price RISC-V...
19mRPMsg to Accelerate Transition Between Multi-SoC and Multi-processor SoC Solutions
22mEnhancing OpenAMP: Making Linux and Zephyr Work Better Together
37mThe Story of PulseAudio and Compress Offload
43mLessons Learnt from Reusing QDMA NIC to Base Band PMD
