[iMX8MP] V4L2 Buffer Copy Time

[DIM]

Hello All,

I need to iterate over an image captured from the camera subsystem on the i.MX8M Plus. The image is greyscale 1920×1200 and copying the buffer currently takes about 15 ms. This suggests that the buffer memory is mapped as non-cacheable.

On previous ARM systems I have worked with, DMA buffers could be mapped as cache-coherent (for example using the dma-coherent device-tree property or similar mechanisms), which reduced a similar copy operation to roughly 2 ms.

Currently the V4L2 capture buffers appear to be uncached, so both copying the buffer and iterating over it directly (zero-copy processing) are quite slow.

Is there a mechanism on the i.MX8M Plus to enable cache-coherent mappings for these buffers (for example via a device-tree configuration), or another recommended approach to improve CPU access performance?

Thanks

[joanxie]

you can try the dmabuf, The dmabuf uses buffers of a hardware DMA in order to perform a zero-copy pipeline, as shown below:
$ gst-launch-1.0 v4l2src device=/dev/video0 num-buffers=300 io-mode=dmabuf ! \
'video/x-raw,format=(string)NV12,width=1920,height=1080,framerate=(fraction)30/1' ! \
queue ! v4l2h264enc output-io-mode=dmabuf-import ! avimux ! filesink location=test.avi

[DIM]

Hi joanxie, thank you for your reply.

I understand I can pass the data using zero-copy via DMABUF, however, it is very slow to access that data when I go to process it. I'd ultimately like minimal latency in processing time. I believe gstreamer doesn't exactly represent my use case because it is parallelizing the processing of the buffers, so the latency doesn't matter in that case.

Is there some way to allow processing of the buffer with minimal latency? I believe this would require cache-coherence, typically implemented on ARM processors via the Accelerator Coherency Port. Does the i.MX8MP have this feature?

Thanks,