You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Host: zero-copy capture through VA-API
The compositor's DMA-BUF now goes straight into the encoder: it is mapped
as a VA surface (no copy) and the GPU's video processor scales, letterboxes
and converts it to NV12 (scale_vaapi + pad_vaapi) before VA-API encodes it.
No pixel crosses the CPU.
Measured on a 60 fps 1080p test pattern mirrored to a 1600x720 phone:
CPU per frame 20.9 ms (0.3.0) -> 8.1 ms (SIMD scaler) -> 1.1 ms
capture thread 16 ms -> 4.8 ms -> 0.04 ms
frame rate 31-33 -> 50 -> 58 fps; phone latency 36 -> 22 ms
The whole daemon streams at 6.6% of a core. Colours match the CPU path to
within 1-3/255 (the VPP needs matrix, primaries and transfer all set to
BT.709; with the matrix alone it converted with BT.601).
- lamco-pipewire: DMA-BUF passthrough is configurable from the builder, and
passthrough frames carry the negotiated fourcc (BGRx was labelled BGRA).
- capture: Frame/RawFrame carry an optional GpuFrame (fd, stride, modifier,
fourcc). Captures ask for passthrough only while a VA-API encoder has
zero-copy ready; any import failure turns it off and frames are read back
to the CPU (GpuFrame::read_linear), so a missing filter (FFmpeg 4.4 in the
AppImage has no pad_vaapi) degrades to the old path instead of failing.
- encoder: encoder/zerocopy.rs (DRM frames -> derived VA-API frames ->
filter graph); rebuilt when the source changes or a bitrate change reopens
the encoder. DISPLAYSWARM_ZEROCOPY=0 turns it off.
- CI: libavfilter-dev for the Debian/Ubuntu builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0136ntADhC7Bm2oH17uq1UTA