This demo program encodes frames from the DVP camera into H.264 NAL units, and is tested with mainline Linux (5.19.3). This demo is best used as a package within the v3s3 buildroot tree, since it relies on a few patches to the kernel to work.
Allwinner S3 support is unverified, due to the lack of test hardware; however, it uses the same die as V3s, so I wouldn't expect any differences.
- Uses the latest mainline software (Linux v5.19) with minimum modifications. No dependency on Allwinner's old Linux 3.4 BSP kernel.
- Uses cedar and libcedarc to interface with the CedarVE hardware video codec block. Unfortunately, some binary blobs are still used (in libcedarc).
- 1920x1080 @ 30FPS video encoding, tested on the LicheePi Zero development board with an OV5640 image sensor
- Zero-copy: frames captured by DVP camera is stored in a buffer shared with the video encoding engine (achieved by a hack detailed below)
- Efficient memory usage; CPU consumption during video encoding is nearly 0%.
The proper way to achieve zero-copy would be to utilize V4L2_MEMORY_DMABUF or V4L2_MEMORY_USERPTR to share buffers. However, the former is unsupported by libcedarc, while the latter is unsupported by sun6i-video. This leaves us with V4L2_MEMORY_MMAP, but there is no way to obtain the physical addresses of the allocated buffers, which the video engine needs if we want to share our pre-allocated buffers with it.
This issue was resolved by a dirty hack: I added a custom ioctl() to the existing sun6i-video driver and borrowed some logic from deep within the videobuf2 and dma-buf frameworks, which allowed me to find the DMA address (i.e. dma_addr_t) associated with the buffer allocated by V4L2_MMAP and convert it to a physical address (phys_addr_t) via virt_to_phys(). I'm 99% sure this is NOT the proper way to do this, but I've tested it many times and it seems to work... If anyone knows a better way to achieve zero-copy video encoding on this platform, I'd love to hear about it.