Add CUDA/NVIDIA support in image - #440
Conversation
|
|
biglittlebigben
left a comment
biglittlebigben
left a comment
There was a problem hiding this comment.
Thanks for the patch! A few comments. Please also update the README to indicate how to setup NVENC based encoding, particularly in a docker and kubernetes environment.
| # capability is required for NVENC and is NOT part of the runtime default | ||
| # (compute,utility), so set it explicitly. These are no-ops without the NVIDIA | ||
| # runtime, so the image still runs CPU-only elsewhere. | ||
| ENV NVIDIA_DRIVER_CAPABILITIES=compute,video,utility |
There was a problem hiding this comment.
I think these should be opt in, at least for now: setting the docker container to support nvenc should be gated behind a parameter passed in on container start.
| EncoderBackendAuto EncoderBackend = "auto" // use a hardware encoder if available, else software | ||
| EncoderBackendSoftware EncoderBackend = "software" // x264enc / vp8enc | ||
| EncoderBackendNVENC EncoderBackend = "nvenc" // NVIDIA NVENC (nvh264enc) | ||
| EncoderBackendVAAPI EncoderBackend = "vaapi" // VA-API (not yet implemented) |
There was a problem hiding this comment.
Please remove options that are not supported yet
| } | ||
|
|
||
| if c.VideoEncoderBackend == "" { | ||
| c.VideoEncoderBackend = EncoderBackendAuto |
There was a problem hiding this comment.
We should default to software, at least for now.
| if nvh264EncoderAvailable() { | ||
| resolved = config.EncoderBackendNVENC | ||
| } else { | ||
| log.Warnw("nvenc video encoder requested but nvh264enc is unavailable, falling back to software", nil) |
There was a problem hiding this comment.
If the encoder backend is set to NVENC, we should call nvh264EncoderAvailable on startup and fail to start the service if we fail to initialize NVENC.
3 participants
Summary
Adds an optional hardware H.264 video encoder (NVIDIA NVENC) to the transcode
pipeline. The encoder is selected per node via a new
video_encoder_backendconfig option, with automatic detection and a transparent fallback to the
existing software path.
Only the H.264 encode is moved to the GPU.
Frames stay in system memory (
videoconvert/videoscaleand decode remain onCPU); the NVENC element uploads internally. No GStreamer rebuild is required —
the
nvcodecplugin (nvh264enc) is already present in thelivekit/gstreamerbase image; only the NVIDIA container runtime is needed.How it works
video_encoder_backend: auto | software | nvenc | vaapi(defaultauto).autoprobes fornvh264enconce at startup. If available, H.264 layers useNVENC; otherwise they use
x264encexactly as before.backend"). If
nvencis requested but unavailable, it logs a warning andfalls back to software.
H.264 baseline byte-stream simulcast — the room/SFU and clients see no
difference.
NVENC is configured for real-time WebRTC:
rc-mode=cbr,preset=low-latency-hq,tune=ultra-low-latency,bframes=0,zerolatency=true(verified againstgst-inspect-1.0 nvh264enc, GStreamer 1.26).Configuration
node-level; GPU is a property of the node, not the ingress
video_encoder_backend: auto # auto | software | nvenc | vaapi