Rockchip HW acceleration not working on Orange Pi 5 plus 16GB #95
Comments
|
It looks like you opened the issue in the wrong repo? This is |
|
Sorry, I'm a bit confused where to open this. There is a long thread about this problem in the frigate repo and @MarcA711 (seems like he is supporting the frigate fork for the rockchip) advised to open the issue in this repo. Should I submit it there as well? I may be wrong but looks like @MarcA711 is using your fork of ffmpeg for RKNN support in Frigate. |
|
It looks like your RKNN runtime library or driver is out of date, causing inference failure. |
|
https://github.com/Joshua-Riek/ubuntu-rockchip |
|
Ok thanks, will try, many thanks! |
|
Hey, Other users also reported issues with frigate version 13.2 and newer kernels. You could try to upgrade to v14 release candidate 2. |
|
@MarcA711 Hi Marc, thanks, I will try that too and let know the results. |
|
In the meantime I found this thread. According to @Link164735
If this is related to the model dimensions and NPU capacity, shouldn't the fix be on the model itself? |
Hi @MarcA711 , FYI - I tried the ghcr.io/blakeblackshear/frigate:0.14.0-rc2-rk on the same system I have (Linux orangepi5plus 6.1.43-rockchip-rk3588) and it happened to be unusable also. Seems like it is still under development. I will try to re-flash the host OS as @nyanmisaka suggested and will probably go back to your earlier image patched with the latest ffmpeg-rockchip from this repo. Will let you know the outcome. |
|
Hey, you have to adjust two options in your config for v14. I made these changes bold: You can also select one of the models here instead of yolonas_m: https://deploy-preview-11419--frigate-docs.netlify.app/configuration/object_detectors#choosing-a-model I will post this issue here later this day: blakeblackshear/frigate#11502 |
|
I posted a comment about this potential error here: blakeblackshear/frigate#11502 (comment) This issue is not related to ffmpeg, so it can be closed. Moreover, I believe that this issue is just an incompatibility between the older yolov8 model and newer kernels. Since yolov8 is not supported anymore, I will not dig deeper into this problem. Please upgrade to Frigate v14 instead. |
Thanks for the confirmation. |
|
@MarcA711 Hello Marc, I tested this on my old host setup, did not re-flash the OS yet. The issue is still the same as before - very slow speed of rknn, and same error messages: I also noticed that the models are proprietary, this is also not good. The performance did not change much: 
|
|
Hey, Yes, you are using rknn driver 0.9.3. You should update your os. yolonas is not proprietary. It is open source and available here. |
|
Thank you, noted. Will re-flash tomorrow and let know the results. Many thanks again. |
Hello,
I'm experimenting with the Orange PI5 Plus box and the frigate version with rknn but it doesnt seem to work on my setup at all. The inference seems to be very poor and also Capture FPS is close to zero. Hence all the detects are skipped:
Please kindly help me to understand what I'm missing.
Here is my setup:
OS/Kernel: Linux orangepi5plus 6.1.43-rockchip-rk3588 #1.0.8 SMP Sat Feb 3 00:04:38 CST 2024 aarch64 aarch64 aarch64 GNU/Linux
docker-compose.yml
config.yml
The frigate starts normally with no obvious issues but then I see the log full of messages like:
is there an image where this is fixed already? Or I need to update the host?
The text was updated successfully, but these errors were encountered: