Enabling HW Video Accel on RPi 4b + Coral TPU #3677
-
|
I have been struggling to get hw video decode running on my RPi4b (8GB) + Coral Frigate setup... I read all the threads I could find, but have not been able to get things to work... Would cherish any guidance... Background
mqtt:
host: 192.168.100.234
port: 1883
topic_prefix: frigate
client_id: frigate
user: USER
password: PW
stats_interval: 60
detectors:
coral:
type: edgetpu
device: usb
cameras:
garage_front:
ffmpeg:
hwaccel_args:
- -c:v
- h264_v4l2m2m
inputs: # subtype =01 Main, 0 sub
- path: rtsp://XXXXXXXX@192.168.100.119:554/cam/realmonitor?channel=1&subtype=0
roles:
- detect
- rtmp
rtmp:
enabled: False # <-- RTMP should be disabled if your stream is not H264
detect:
width: 640 # <---- update for your camera's resolution
height: 480 # <---- update for your camera's resolution
fps: 10
garage_back:
ffmpeg:
# hwaccel_args:
# - -c:v
# - h264_v4l2m2m
inputs: # subtype =01 Main, 0 sub
- path: rtsp://XXXXXXXX@192.168.100.145:554/cam/realmonitor?channel=1&subtype=0
roles:
- detect
- rtmp
rtmp:
enabled: False # <-- RTMP should be disabled if your stream is not H264
detect:
width: 640 # <---- update for your camera's resolution
height: 480 # <---- update for your camera's resolution
fps: 10
|
Beta Was this translation helpful? Give feedback.
Replies: 8 comments 17 replies
-
|
Have you not updated the amount of gpu memory in the boot config? The logs are calling out the issue:
|
Beta Was this translation helpful? Give feedback.
-
|
https://www.raspberrypi.com/documentation/computers/configuration.html |
Beta Was this translation helpful? Give feedback.
-
|
I have gpu_mem=512 set in the RPi config.txt |
Beta Was this translation helpful? Give feedback.
-
|
What is the memory usage? Seems like maybe too much is already being used and there's not enough to be allocated for the GPU |
Beta Was this translation helpful? Give feedback.
-
|
I am running HASS 8.4 on a 8GB RPi4. Sorry... I don't know an easy way to answer your question about mem usage... Is there a way to extract the mem usage from within HA? |
Beta Was this translation helpful? Give feedback.
-
|
Have continued to mess with the RPi4 config.txt. Using the gpu_mem_1024 command, I get this log file when I start Frigate.... [2022-08-18 19:22:56] frigate.app INFO : Starting Frigate (0.11.0-37325c7) |
Beta Was this translation helpful? Give feedback.
-
|
I think just realized the issue, your config is using subtype 0 which is the main stream of amcrest not the substream. The main streams are too high resolution for the pi to decode |
Beta Was this translation helpful? Give feedback.
-
|
Good idea! I used this article to configure the RTSP stream. It says substream=0 RTSP line in my frigate.yml...
[2022-08-18 19:50:41] frigate.app INFO : Starting Frigate (0.11.0-37325c7) I don't have a second RPi4b, but wondering if I need to get one to see if the current one has some strange issue... |
Beta Was this translation helpful? Give feedback.
-
|
My amcrest the subtype 0 is the main stream. Try it as just |
Beta Was this translation helpful? Give feedback.
-
|
I used this path...
Log... Same result... I again confirmed, if I comment out the hwaccel_args, the garage_front camera still works.... |
Beta Was this translation helpful? Give feedback.
-
|
Might not be your final solution, but another fix you'll want is to lower your detect FPS in Frigate. 5 is the 'default' and recommended upper limit. The detect resolution needs to match your camera, but the detect FPS is how often Frigate reviews objects. Realistically 1 or 2 FPS most people would be happy with the response speed of and beyond 5FPS all you're doing is pushing the hardware and getting no appreciable benefit. This does not cause your video stream to get choppy or anything, this is just object detection related. It could be that you're calling for reviews of the same stream faster than the pi can handle them by using 10FPS. The Coral can handle ~100FPS total, but if the Pi can't feed each stream fast enough then the Coral's capabilities don't really matter. As a testing fix to see if it's performance related, I'd drop the substream quality down even lower and drop the detect FPS to 1. If everything starts working (albeit not optimally obviously) then you can start turning them up until you find the limits. If it DOESN'T help then performance is less likely the cause. detect: |
Beta Was this translation helpful? Give feedback.
-
|
@Meowx3 - Good suggestion. I had tried 5 fps on the substream yesterday, and slowed it down to 1 fps this morning. Sadly, I still get the same log results as pasted above... |
Beta Was this translation helpful? Give feedback.
-
|
Whats the camera device? Are u sure you have a proper h264 stream (check with vlc or else and post a screenshot)? |
Beta Was this translation helpful? Give feedback.
-
|
Yes... Positive I have a proper h264 stream... If I enable the hwaccel_args (see below) in frigate.yml I get the failing logs above... If I comment out the hwaccel_args lines in frigate.yml, the cameras work fine, the RPi CPU is just overwhelmed. I am using two Amcrest cameras - on 8MP, one 5MP. |
Beta Was this translation helpful? Give feedback.
-
|
What resolution and codec specific details are? |
Beta Was this translation helpful? Give feedback.
-
|
Both cameras have their sub streams to encode H.264, VGA resolution and I have tried 10, 5, and 1 fps. |
Beta Was this translation helpful? Give feedback.
-
|
What OS are you running? DietPi? |
Beta Was this translation helpful? Give feedback.
-
|
Running HASSOS v8.5 and HA 2022.8.6. running 64-bit version |
Beta Was this translation helpful? Give feedback.
-
|
Possibly related #847 |
Beta Was this translation helpful? Give feedback.
-
|
Could be... Though nicw was using a 32-bit HassOS. |
Beta Was this translation helpful? Give feedback.
-
|
I upgraded to HA 2022.8.7 and my error messages changed... HW accelerated decode is still not working, but I do get different error messages (see below)... I have not changed my frigate.yml from what I posted above, just upgraded HA... [2022-08-25 20:15:16] frigate.app INFO : Output process started: 217 |
Beta Was this translation helpful? Give feedback.
-
|
Update: After trying lots of different options, and searching the web, I have not found a reliable working description/configuration of anyone using hardware video decode from a RPi4 while running HAOS, and Frigate. So... I decided to buy a refurbished $200 Intel Gen 7 SFF (small form factor) PC and install HAOS on that, and use that for my Home Assistant + Frigate platform. I followed the Frigate docs and now have video decode running on the Intel CPU. I really wanted to RPI4 to work, but I have moved on. |
Beta Was this translation helpful? Give feedback.
-
|
I'm trying to do something very similar on a 2GB Pi4 (64bit, docker, no HASS) and also can't get hardware accelerated decoding to work. Is the consensus that it's not possible? Perhaps this should be added to the documentation. |
Beta Was this translation helpful? Give feedback.
-
|
It is possible, if it's not working then you may be experiencing #3780 |
Beta Was this translation helpful? Give feedback.
-
|
Think you're right, thanks. Will wait and test a fix for that issue. |
Beta Was this translation helpful? Give feedback.
Update: After trying lots of different options, and searching the web, I have not found a reliable working description/configuration of anyone using hardware video decode from a RPi4 while running HAOS, and Frigate.
So... I decided to buy a refurbished $200 Intel Gen 7 SFF (small form factor) PC and install HAOS on that, and use that for my Home Assistant + Frigate platform. I followed the Frigate docs and now have video decode running on the Intel CPU. I really wanted to RPI4 to work, but I have moved on.