Skip to content

gpu.md

Latest commit

 

History

History
77 lines (60 loc) · 3 KB

gpu.md

File metadata and controls

77 lines (60 loc) · 3 KB
Raw

GPU acceleration

The box can use the GPU of the host to decode and encode video (ffmpeg with VA-API), and for OpenGL and Vulkan: Intel, AMD and the Raspberry Pi's, through /dev/dri. Everything about it lives in one command, devbox gpu.

devbox gpu            # the state, and the next step when something is missing
devbox gpu install    # the drivers of the GPU found, through devbox pkg
devbox gpu test       # 10 s of 1080p encoded on the GPU, then on the CPU
devbox gpu remove     # take the drivers out again

devbox status sums it up on one line and points to devbox gpu.

Passing the GPU to the box

The box sees no GPU until the host hands it over. The setup script asks when the machine has a /dev/dri; otherwise, on the host, add it to compose.override.yaml (the GPU block of compose.override.example.yaml) and restart the container:

services:
  dev-box:
    devices:
      - /dev/dri

With the podman block, /dev/dri goes in its devices list: a second devices key is refused. A host without /dev/dri (Docker Desktop on macOS) refuses to start the container with that line, which is why it stays out of compose.yaml.

The nodes of /dev/dri belong to the video and render groups of the host, whose GIDs mean nothing in the image. At every start the entrypoint adds the user to the group that owns each node, created as dri<gid> when the image has none with that GID: no group_add to write, the same override works on any host. docker logs dev-box says which groups were added.

Drivers

devbox gpu install picks the packages from the kernel driver of the GPU and installs them through devbox pkg, so they come back after a rebuild:

GPU Packages
Intel (i915, xe) intel-media-driver, vulkan-intel
AMD (amdgpu, radeon) vulkan-radeon
every GPU libva-utils (vainfo), vulkan-tools (vulkaninfo)

mesa, which holds the VA-API drivers of AMD and the others, comes as their dependency. DEV_ENVS=gpu in .env does the same at start (the setup script writes it when the GPU is asked for); devbox dev-env gpu is kept as a hidden alias for that.

Checking and using it

devbox gpu shows the device and its kernel driver, whether your user can open it, the VA-API driver with the codecs it decodes and encodes, and the Vulkan device. devbox gpu test needs ffmpeg (devbox dev-env media): on an Intel UHD 620 class iGPU, the GPU encodes the 10 s of 1080p in about 1.2 s against 5 s for the CPU.

vainfo --display drm --device /dev/dri/renderD128   # the VA-API profiles
ffmpeg -hwaccel vaapi -hwaccel_device /dev/dri/renderD128 \
  -hwaccel_output_format vaapi -i in.mp4 -c:v h264_vaapi out.mp4

What does not use it

  • NVIDIA: its GPU goes through the NVIDIA Container Toolkit, not /dev/dri; not covered.
  • Headless Chromium (devbox dev-env browser): it renders in software (SwiftShader) even with the GPU and the drivers there, whatever the flags. Screenshots and DOM dumps do not need more.