Immich•2y ago

Trying Out GPU Transcoding

Hi I wanted to know if there is any way to try the GPU transcoding feature early. I know it's still in development but I have about 10,000 videos I want to import into immich and if I would have to CPU transcode them it would take way too long. I would really appreciate any help anyone can give me 🙂

77 Replies

bo0tzz•2y ago

You could try to use the dev image for your microservices container ghcr.io/immich-app/immich-server:pr-3171, but absolutely zero guarantee - it could horribly malfunction and garble all your videos or something like that, so at your own peril 😛

SamTheRubyOP•2y ago

so I just do a docker pull command? how do I set that up? will it automatically use my nvidia gpu?

bo0tzz•2y ago

I think so But it's probably better to wait until this is released with actual documentation and such ;)

SamTheRubyOP•2y ago

any eta?

bo0tzz•2y ago

nope

SamTheRubyOP•2y ago

ah ok I may have to try it since I'm mostly importing my videos not only using it as my storage, I can at least let you know how it goes.

brighteyed•2y ago

It is hard to try it on your Immich instance. As I remember you need (1) to modify docker-compose.yml (see PR). Since hardware acceleration is disabled by default, you (2) will need also to pull the web image to change this setting

SamTheRubyOP•2y ago

isin't it like 2 lines in the docker compose file

SamTheRubyOP•2y ago

brighteyed•2y ago

Not sure, but I believe also a couple of lines in the devices section

SamTheRubyOP•2y ago

of the docker compose file?

brighteyed•2y ago

Yep

SamTheRubyOP•2y ago

@brighteyed @bo0tzz In the docker compose file Im assuming I replace the immich-microservices line

 immich-microservices:
    container_name: immich_microservices
    image: immich-microservices:latest

 immich-microservices:
    container_name: immich_microservices
    image: immich-microservices:latest

with immich-server:pr-3171 or do I replace this part?

container_name: immich_server
    image: immich-server-dev:latest

container_name: immich_server
    image: immich-server-dev:latest

brighteyed•2y ago

Both immich-server and immich-microservices containers use immich-server image

SamTheRubyOP•2y ago

so I replace both of the images with immich-server:pr-3171? @brighteyed @bo0tzz

brighteyed•2y ago

I would suggest to setup a separate Immich instance to try this experimental feature

SamTheRubyOP•2y ago

SamTheRubyOP•2y ago

here are my docker-compose.yml hwaccel.yml and .env files

env

docker-compose.yml

hwaccel.yml

SamTheRubyOP•2y ago

here is my .env file as a txt file so it does not need to be downloaded

env.txt

SamTheRubyOP•2y ago

only thing I changed in the .env is the upload location

Kryptonian•2y ago

It also needs to be enabled in settings on Immich

SamTheRubyOP•2y ago

It is

SamTheRubyOP•2y ago

sogan•2y ago

thank you for testing this. the pr has been mostly done but there's so many hardware configurations it's hard to know if it will work for everyone can you share your hwaccel.yml as well as the microservices logs? oh, i see you posted it and it looks good. since you have nvenc enabled in settings, i'm guessing the hwaccel is failing and it's falling back to software do you have nvidia container runtime installed on this server?

SamTheRubyOP•2y ago

no I did not, let me see if that fixes the issue @sogan still no luck but I see this in the logs for the microservers container

[31m[Nest] 1  - [39m07/21/2023, 4:27:54 PM [31m  ERROR[39m [38;5;3m[MediaService] [39m[31mAn error occurred while configuring transcoding options: Error: ENOENT: no such file or directory, scandir '/dev/dri'[39m

[31m[Nest] 1  - [39m07/21/2023, 4:27:54 PM [31m  ERROR[39m [38;5;3m[MediaService] [39m[31mAn error occurred while configuring transcoding options: Error: ENOENT: no such file or directory, scandir '/dev/dri'[39m

and I see that none of the videos are actually being transcoded

sogan•2y ago

Ah, I think this is a bug. Can you add this to your hwaccel.yml and redeploy your microservices container?

devices:
  - /dev/dri:/dev/dri

devices:
  - /dev/dri:/dev/dri

SamTheRubyOP•2y ago

sure one sec THAT WORKED!!!

sogan•2y ago

Awesome!

SamTheRubyOP•2y ago

TYSM!

sogan•2y ago

Glad I could help! Also if you have an Intel CPU, would you be able to try Quick Sync as well? You don't need to change the hwaccel.yml

SamTheRubyOP•2y ago

I dont have a linux setup but I have a windows pc would it work with that? If not I may be able to install linux on it on sunday

sogan•2y ago

Ah, I think quick sync doesn't support WSL2 unfortunately so it would need to be on linux lmk if you do end up trying it since it's the only backend I can't test atm

SamTheRubyOP•2y ago

ah ok Ill gladly try testing when I have a chance I also saw that VP9 is not supported on NVEC, doesn't nvidia support VP9 encode and decode? nope nvm sorry just saw that only decoding is supported not encdoing

sogan•2y ago

yeah I thought it did but unfortunately not. they basically skipped over vp9 since the latest ones support av1.

SamTheRubyOP•2y ago

yeah, shame but I guess expected I can also pretty much confirm that most amd gpus and APUS do not support VP9 encode @sogan do you know what command I should use to check if Intel quicksync is being used?

sogan•2y ago

you can check the microservices logs. if the -vcodec part says h264_qsv or hevc_qsv and there's no errors, it's using qsv

martabal•2y ago

@sogan tried it and it works well, nice work ! One concern, all theses hardware acceleration are available for arm devices too ?

Kryptonian•2y ago

No arm support for hwaccel

sogan•2y ago

not for arm itself, but if the device has a gpu then it should work e.g. jetson

Kryptonian•2y ago

Eg compute module 4 with an external GPU.

SamTheRubyOP•2y ago

@sogan do you also want me to test?

sogan•2y ago

if you can it’d be great. i made a few changes so qsv should work

SamTheRubyOP•2y ago

@sogan I get this error:

message.txt

sogan•2y ago

do you have this uncommented in the hwaccel.yml?

devices:
  - /dev/dri:/dev/dri

devices:
  - /dev/dri:/dev/dri

SamTheRubyOP•2y ago

yes I do

version: "3.8"

# Hardware acceleration for transcoding - Optional
# This is only needed if you want to use hardware acceleration for transcoding.
# Depending on your hardware, you should uncomment the relevant lines below.

services:
  hwaccel:
    devices:
      - /dev/dri:/dev/dri  # If using Intel QuickSync or VAAPI
    # volumes:
    #   - /usr/lib/wsl:/usr/lib/wsl # If using VAAPI in WSL2
    # environment:
    #   - NVIDIA_DRIVER_CAPABILITIES=all # If using NVIDIA GPU
    #   - LD_LIBRARY_PATH=/usr/lib/wsl/lib # If using VAAPI in WSL2
    #   - LIBVA_DRIVER_NAME=d3d12 # If using VAAPI in WSL2
    # deploy: # Uncomment this section if using NVIDIA GPU
    #   resources:
    #     reservations:
    #       devices:
    #         - driver: nvidia
    #           count: 1
    #           capabilities: [gpu]

version: "3.8"

# Hardware acceleration for transcoding - Optional
# This is only needed if you want to use hardware acceleration for transcoding.
# Depending on your hardware, you should uncomment the relevant lines below.

services:
  hwaccel:
    devices:
      - /dev/dri:/dev/dri  # If using Intel QuickSync or VAAPI
    # volumes:
    #   - /usr/lib/wsl:/usr/lib/wsl # If using VAAPI in WSL2
    # environment:
    #   - NVIDIA_DRIVER_CAPABILITIES=all # If using NVIDIA GPU
    #   - LD_LIBRARY_PATH=/usr/lib/wsl/lib # If using VAAPI in WSL2
    #   - LIBVA_DRIVER_NAME=d3d12 # If using VAAPI in WSL2
    # deploy: # Uncomment this section if using NVIDIA GPU
    #   resources:
    #     reservations:
    #       devices:
    #         - driver: nvidia
    #           count: 1
    #           capabilities: [gpu]

sogan•2y ago

hmm, this is tricky. there are two possibilities i'm thinking here: 1. the host system is missing a dependency for something, possibly libmfx or mesa? these are installed in the container but maybe there's something that needs to be installed in the host as well? 2. the qsv device is actually not renderD128 for this system but maybe something else? could you try the same settings with vaapi? it will internally use quick sync if it's available. also share: 1. the cpu model and whether this server has a gpu. 2. the output of ls /dev/dri on this server btw i managed to get the same error on another machine i tested. the pr has been updated to fix this issue so it should work now.

abbhishek_chakra•2y ago

I was thinking of trying out hw transcoding. But my system has an i3 4010u. As per specs it supports intel quick sync but as per documentation min 7th gen intel is needed. So just wanted to confirm if there is any way I can enable it?

sogan•2y ago

You can enable it and see if it works. Let me know how it goes if you do

SamTheRubyOP•2y ago

@sogan Im sorry for the delay, Im now on the latest version and when I try to transcode this happens:

message.txt

SamTheRubyOP•2y ago

this is my hwaccel.yml file:

version: "3.8"

# Hardware acceleration for transcoding - Optional
# This is only needed if you want to use hardware acceleration for transcoding.
# Depending on your hardware, you should uncomment the relevant lines below.

services:
  hwaccel:
    # devices:
    #   - /dev/dri:/dev/dri  # If using Intel QuickSync or VAAPI
    # volumes:
    #   - /usr/lib/wsl:/usr/lib/wsl # If using VAAPI in WSL2
    environment:
      - NVIDIA_DRIVER_CAPABILITIES=all # If using NVIDIA GPU
      # - LD_LIBRARY_PATH=/usr/lib/wsl/lib # If using VAAPI in WSL2
      # - LIBVA_DRIVER_NAME=d3d12 # If using VAAPI in WSL2
    deploy: # Uncomment this section if using NVIDIA GPU
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

version: "3.8"

# Hardware acceleration for transcoding - Optional
# This is only needed if you want to use hardware acceleration for transcoding.
# Depending on your hardware, you should uncomment the relevant lines below.

services:
  hwaccel:
    # devices:
    #   - /dev/dri:/dev/dri  # If using Intel QuickSync or VAAPI
    # volumes:
    #   - /usr/lib/wsl:/usr/lib/wsl # If using VAAPI in WSL2
    environment:
      - NVIDIA_DRIVER_CAPABILITIES=all # If using NVIDIA GPU
      # - LD_LIBRARY_PATH=/usr/lib/wsl/lib # If using VAAPI in WSL2
      # - LIBVA_DRIVER_NAME=d3d12 # If using VAAPI in WSL2
    deploy: # Uncomment this section if using NVIDIA GPU
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

SamTheRubyOP•2y ago

this is my docker compose file:

message.txt

SamTheRubyOP•2y ago

this is my .env file

SamTheRubyOP•2y ago

message.txt

sogan•2y ago

can you share the host os and version (e.g. Ubuntu 23.04)?

SamTheRubyOP•2y ago

yes it's ubuntu 23.04 server

sogan•2y ago

lol good guess try changing this line capabilities: [gpu] to capabilities: [gpu,video]

SamTheRubyOP•2y ago

hmm didn't seem to fix it @sogan

SamTheRubyOP•2y ago

message.txt

SamTheRubyOP•2y ago

^ from the immich_microservices container ^

sogan•2y ago

that did seem to fix the issue you had before, but this error is really weird. it normally means the video is corrupt. does software transcoding work?

SamTheRubyOP•2y ago

let me try

SamTheRubyOP•2y ago

@sogan

message.txt

SamTheRubyOP•2y ago

this is without the gpu

sogan•2y ago

yeah something's up with that video. if you can download 78b41b5a-aeef-4026-b17d-97901e586f36.mp4 to your pc, try playing it. if it plays, this is a bug with ffmpeg/ffprobe. if it doesn't, the video is corrupt.

SamTheRubyOP•2y ago

ok but isin't it happening with all the videos with the gpu

sogan•2y ago

i don't know how many of the videos it affects in this library but i think the logs you shared are for one video but there are several that fail in the last logs you showed so any of those would also work for testing overall though, this doesn't seem to be a hardware transcoding problem at this point. the error is about it not being able to decode the video, but decoding is always software-based (only encoding is hardware accelerated).

SamTheRubyOP•2y ago

oh ok thanks, Ill try to renable the gpu, leave it overnight and see what happens

SamTheRubyOP•2y ago

@sogan still getting issues with gpu enabled it seems like every video is failing to transcode:

message.txt

SamTheRubyOP•2y ago

here are my settings

SamTheRubyOP•2y ago

this is the bottom of the error in the immich_microservices container:

message.txt

sogan•2y ago

i think this is actually a bug with fluent-ffmpeg in node.js

sogan•2y ago

https://github.com/fluent-ffmpeg/node-fluent-ffmpeg/issues/932

GitHub

mp4 buffer as input fails to convert a video · Issue #932 · fluent-...

Version information fluent-ffmpeg version: 2.1.2 ffmpeg version: 4.1.3 OS: macos Code to reproduce A buffer comes from browser to node.js read by FileReader. // videoArrayBuffer is a buffer coming ...

sogan•2y ago

what a weird bug

Gaming

Programming

Trying Out GPU Transcoding

Did you find this page helpful?