[HW Accel Support]: Beginners question - seem that hwaccel isnt't working on my system

[Helmut1965]

Describe the problem you are having

Dear All,

I set up frigate for the first time. I try to enable hwaccel_args: preset-rpi-64-h264 in the config.yml. Overal there isn't any change in the CPU consumption. In case I comment the hwaccel statement or not - always the same load
15% CPU consumption for 1024x576 (Stream rtsp://XXX.XXX.XXX.XXX:554/s2 of my camera)
78% CPU in case I change to 1920x1080 (for sure this is another stream Stream rtsp://XXX.XXX.XXX.XXX:554/s0 of my camera)

Any idea what is happen here or is it quite normal that there is no change if the hwaccel statement is in/out?

Version

0.14.1-f4f3cfa

Frigate config file

mqtt:
  host: xx.xx.xx.xx
  topic_prefix: frigate
  client_id: frigate
  stats_interval: 300

ffmpeg:
    hwaccel_args: preset-rpi-64-h264

detectors:
  coral:
    type: edgetpu
    device: usb

detect:
  fps: 5

cameras:
  garage:
    ffmpeg:
      inputs:
        - path: rtsp://XXX.XXX.XXX.XX:554/s2 #The Stream you want to monitor
          roles:
            - detect
            - record
    detect:
      enabled: True
      width: 1024
      height: 576

    motion:
      threshold: 21
      contour_area: 10
      improve_contrast: 'true'
    zones:
      Eingang:
        coordinates: 0.236,0,0.002,0.345,0.002,0.995,0.998,0.997,0.996,0.006
        loitering_time: 0
    review:
      alerts:
        required_zones: Eingang

    snapshots:
      enabled: true
      timestamp: true
      bounding_box: false
      retain:
        default: 90
        mode: active_objects

record:
  enabled: true
  retain:
    days: 3
    mode: motion
  events:
    retain:
      default: 30
      mode: motion
version: 0.14

docker-compose file or Docker CLI command

services:
  frigate:
    container_name: frigate
    privileged: true # this may not be necessary for all setups
    restart: unless-stopped
    image: ghcr.io/blakeblackshear/frigate:0.14.1-standard-arm64
    shm_size: "128mb"

    devices:
      - /dev/video11:/dev/video11 #Map the Pi hardware acceleration

    volumes:
      - /etc/localtime:/etc/localtime:ro
      - ./config.yml:/config/config.yml
      - ./storage:/media/frigate
      - /dev/bus/usb:/dev/bus/usb # passes the USB Coral, needs to be modified for other versions
      - type: tmpfs #Remove this if using a Pi with 2GB or Less RAM
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "xxx.xxx.xxx.xxx:8971:8971"
      - "xxx.xxx.xxx.xxx:554:554" # RTSP feeds
      - "xxx.xxx.xxx.xxx:8555:8555/tcp" # WebRTC over tcp
      - "xxx.xxx.xxx.xxx:8555:8555/udp" # WebRTC over udp
      - "xxx.xxx.xxx.xxx:1883:1883/tcp" # Mqtt
    environment:
      FRIGATE_RTSP_PASSWORD: "XXXXXXXX"

Relevant Frigate log output

2025-01-15 16:18:58.742461974  [INFO] Preparing Frigate...
2025-01-15 16:18:59.150288844  [INFO] Starting Frigate...
2025-01-15 16:19:09.476492653  [2025-01-15 16:19:09] frigate.app                    INFO    : Starting Frigate (0.14.1-f4f3cfa)
2025-01-15 16:19:09.478063023  [2025-01-15 16:19:09] frigate.util.config            INFO    : Checking if frigate config needs migration...
2025-01-15 16:19:09.551406079  [2025-01-15 16:19:09] frigate.util.config            INFO    : frigate config does not need migration...
2025-01-15 16:19:09.731372523  [2025-01-15 16:19:09] peewee_migrate.logs            INFO    : Starting migrations
2025-01-15 16:19:09.731822097  [2025-01-15 16:19:09] peewee_migrate.logs            INFO    : There is nothing to migrate
2025-01-15 16:19:09.756873190  [2025-01-15 16:19:09] frigate.app                    INFO    : Recording process started: 357
2025-01-15 16:19:09.780326245  [2025-01-15 16:19:09] frigate.app                    INFO    : Recording process started: 358
2025-01-15 16:19:09.796986153  [2025-01-15 16:19:09] frigate.app                    INFO    : go2rtc process pid: 98
2025-01-15 16:19:09.973170856  [2025-01-15 16:19:09] frigate.app                    INFO    : Output process started: 388
2025-01-15 16:19:10.017189856  [2025-01-15 16:19:09] detector.coral                 INFO    : Starting detection process: 386
2025-01-15 16:19:10.017206449  [2025-01-15 16:19:09] frigate.detectors.plugins.edgetpu_tfl INFO    : Attempting to load TPU as usb
2025-01-15 16:19:10.147714504  [2025-01-15 16:19:10] frigate.app                    INFO    : Camera processor started for garage: 406
2025-01-15 16:19:10.153010504  [2025-01-15 16:19:10] frigate.app                    INFO    : Capture process started for garage: 407
2025-01-15 16:19:14.170300335  [2025-01-15 16:19:14] frigate.detectors.plugins.edgetpu_tfl INFO    : TPU found

Relevant go2rtc log output

I'm not sure how I can get this. Is it needed here for the question?

FFprobe output from your camera

Input #0, rtsp, from 'rtsp://10.0.110.10:554/s2':
  Metadata:
    title           : s2
  Duration: N/A, start: 0.000000, bitrate: N/A
    Stream #0:0: Audio: aac (LC), 11025 Hz, mono, fltp
    Stream #0:1: Video: h264 (Main), yuvj420p(pc, bt709, progressive), 1024x576 [SAR 189:190 DAR 168:95], 30 fps, 30 tbr, 90k tbn, 60 tbc

Install method

Docker Compose

Object Detector

Coral

Network connection

Wired

Camera make and model

UVC G3 Bullet Unifi

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

uname -r: 6.6.62+rpt-rpi-v8
Coral doesn't work with the device I found in the internet I have to do it in the volumes mapping, after this, Coral was found.
config setting in the firmware file: gpu_mem=128, arm_64bit=1, arm_boost=1
Rasperry PI4

my stats screen in UI shows nothing for the GPU graph

[NickM-27]

image: ghcr.io/blakeblackshear/frigate:0.14.1-standard-arm64

is the wrong image to use for RPi, you should just be using the default image

[NickM-27]

You should just use image: ghcr.io/blakeblackshear/frigate:0.14.1

you have pointed docker at the entirely wrong image for the RPi

[Helmut1965]

Hi Nick,
I delete all the existing images with docker rm $(docker ps -a -q), afterwards I changed the image line to:
image: ghcr.io/blakeblackshear/frigate:0.14.1 as ypu proposed an build it new with docker compose up -d

Overall, the behavior is still the same. There is no change with/without hwaccel, and still the same load

[ElectronicBattle]

I am not certain but I think you need to look in the documentation at go2rtc which allows the GPU to do things like encode and decode to different formats and resolutions. i struggled to understand it and I am still experimenting, but this might help a little bit:

#12376

otherwise read through https://docs.frigate.video/guides/configuring_go2rtc

At the moment, you have set

    hwaccel_args: preset-rpi-64-h264

but you aren't then doing anything with it, hence no GPU workload. I think someone who knows more than me will be able to add a lot more helpful advice.

(edit: prose struck through because I know even less than I thought!)

[NickM-27]

but you aren't then doing anything with it, hence no GPU workload.

that's not correct, hwaccel_args by default help with decoding the streams which always happens for detect cameras

[NickM-27]

in this case, the load is likely just not related to the GPU, unfortunately there is a lot of overhead on the args used for decoding on the Pi so the benefit for decoding is not always that impactful

[Helmut1965]

Dear Forum, at the end of the day, I was able to figure it out. I play a few hours with a native ffmpeg installation to find the problem

1. There is the need to map both video devices. One is encode the other one is decoding. You can check this with a journalctl | grep video an d you see the chanels
   devices:
   • /dev/video11:/dev/video10
   • /dev/video11:/dev/video11
2. I have to set the hwaccel_args like this
   hwaccel_args:
   • -c:v
   • h264_v4l2m2m
     I try auto and also preset-rpi-64-h264, both of them lead to software compression. This depends maybe on the type of the camera. So it makes sense to play with the values

Both activities brought the load down from 90% (high resolution) through 20%. Afterwards I changed the gpu_mem. Based on the information this is not needed anymore and also this is no longer possible in raspi-config (at least in my system). In case I leave this untouched and check it with vcgencmd get_mem gpu => it shows 76MB. I changed it manually in the firmware/config.txt to 512MB (highest possible value in my config). This gives me another 9% of load.
Now I have 11% ffmpeg, before it was 90.

As a next step I try to reduce the parameter in the detect clause through:
width: 1280
height: 720
Keep in mind, in case you do nothing here, the camera native resolution is needed.
Result -> CPU load went up from 11% through 50%, It looks through me, that ffmpeg starts a resizing with a software decoder and using the general hwaccel settings from the configuration. I found nothing in the documentation so far for hwaccel during the resampling for the detection stream, so I decide to leave the detection part on the native high resolution, for sure with fps: 5.

Maybe this helps somebody in the future to not waste time. By the way: System: PI4, Debian Bookworm with Google Coral for detection
Have a great day
Helmut