Linux driver submission adds H.264 hardware encoding for Rockchip RK3576

The submission of a mainline driver for the RK3576 moves the industry closer to using low-power Rockchip silicon for edge-encoding and specialized distribution without proprietary vendor kernels. For the open-source hardware ecosystem, this bridge enables standardized media pipelines via V4L2 rather than fragmented vendor APIs. However, the P-frame stall highlights the ongoing difficulty of reverse-engineering hardware-specific memory management and internal timing. Watch to see if kernel maintainers identify an RK3576-specific 're-arm' quirk that resolves the inter-frame stalling bug, which would unlock high-efficiency mobile and edge encoding use cases.

The RK3576 is a mid-range SoC positioned between the legacy RK3399 and the high-end RK3588, featuring four Cortex-A72 and four Cortex-A53 cores. Per CNX Software in early 2024, the chip was designed to balance price and performance for AIoT and industrial applications, including a 6 TOPS NPU and a focus on video versatility. While Rockchip provides its own 'MPP' (Media Process Platform) for downstream kernels, the quest for a mainline-compatible driver reflects a shift toward long-term support for single-board computers like the Radxa ROCK 4D, which rely on standard Linux kernel features for security and compatibility.

Developing mainline drivers for Rockchip's encoder blocks has historically been a multi-year effort due to undocumented register behaviors. Per reports from the Collabora engineering blog in 2023, similar efforts for the Hantro and Verisilicon blocks on earlier Rockchip SoCs required extensive community collaboration to reach stable stateful encoding support. This latest submission for the VEPU510 follows that trajectory, aiming to avoid the 'MPP-service' model which many developers find difficult to integrate with standard GStreamer or FFmpeg pipelines. If finalized, this driver would allow the RK3576 to serve as a high-density, low-wattage transcoding node in decentralized streaming architectures.