capture: an in-process Wayland backend, and the measurements that justify it

The frame server could only capture by running a program once per frame
(grim). That works, but ~95% of a grim frame is waiting -- fork, connect,
handshake, round trip -- not encoding, and grim makes the compositor import
a *new* shared-memory buffer for every single frame.

This adds `--capture wayland[:OUTPUT]`: one long-lived zwlr-screencopy-v1
client per server, with one shm buffer reused for the whole run, capturing
inside the process. Measured here at 1280x720 (spike capture-speed, 200
frames; live re-check on Virtual-1, 60 frames):

    grim (subprocess)    117.3 ms/frame   8.1 fps   compositor 116.4 ms
    grim-rs (Rust crate) 124.6 ms/frame   7.5 fps   compositor 123.1 ms
    this backend          78.0 ms/frame  13.9 fps   compositor  67.5 ms

    live, PNG both sides: this 128.5 ms/frame against grim's 389.3 ms (3.0x)

Three findings, each of which shaped the code:

* The win is buffer reuse, not language. grim and grim-rs both hand the
  compositor a fresh buffer per frame, which is why their compositor CPU is
  nearly double ours. The per-frame allocate-copy-swizzle that grim-rs pays
  was measured in isolation at 2.4 ms of *client* time; the rest of its
  deficit lands inside the compositor.
* ext-image-copy-capture-v1 is no faster (78.0 vs 77.7 ms) and its
  create_frame path appears in a Hyprland segfault stack, so this keeps the
  deprecated protocol grim uses. The switch is one constant when that crash
  is fixed.
* PNG only, deliberately: PNG is the one encoding this package produces
  without a dependency, so --format jpeg is refused rather than publishing a
  Hello whose MIME type the frames do not have. A PNG frame is ~2.1 MB
  against ~120 KB for JPEG q80, which is why the command backend stays the
  default and why a libjpeg-turbo encoder (1.7 ms/frame, measured) is the
  obvious follow-up.

The timeout is the load-bearing part of the client: pywayland only reads
events from read()/dispatch(block=True), neither of which takes a timeout,
and the server captures in a worker thread where SIGALRM is unavailable. It
therefore polls the display fd itself, so a compositor that stops answering
produces a CaptureError instead of a capture thread that never returns.

The protocol bindings are vendored under src/hhstream/_protocol/ exactly the
way hhstream_pointer already vendors its two, with regenerate.py recording
where each XML came from. No new dependency: pywayland was already required
by the input driver.

Contract: the new server flag and the two new module options are documented
in docs/CONTRACT.md, and docs/ARCHITECTURE.md's prediction that a native
screencopy client would be "a swap rather than a rewrite" is now the record
of what the swap cost.

Unit suite: 977 passed, 2 skipped, 0 failed (nix build -L .#checks.unit).
The live numbers were taken on this VM against the real compositor.
This commit is contained in:
Hermenix committed 2026-10-11 20:34:33 +00:00
1 parent f3ac9ada27
commit a7d9357b43
37 files changed
+7556 -10

No files matched your search

+22 -2
View File
@@ -87,8 +87,28 @@ Capturing is a separate program (by default `grim`, a wlr-screencopy client) run
inside the compositor's session and pointed at stdout. That choice is what keeps
`hhstream` free of Wayland and GPU code — the awkward part lives in a tool that
already exists and is already tested — and it is why the capture command is a
module option: tests inject a fake, and a future backend (a native screencopy
client, a hardware encoder) is a command-line swap rather than a rewrite.
module option: tests inject a fake, and a different backend is a swap rather than
a rewrite.
That swap has now happened once. `captureBackend = "wayland"` captures in-process
(`src/hhstream/wayland_capture.py`): one long-lived `zwlr-screencopy-v1` client
with **one shared-memory buffer reused for every frame**, instead of a `grim`
process per frame. Measured at 1280×720 over 200 frames, that is ~1.5× the frame
rate (78 ms against 117 ms per frame) and — the number that matters on this
machine — roughly **half the compositor's CPU per frame** (67 ms against 116 ms),
because `grim` makes the compositor import a new buffer every time. It is
deliberately the *deprecated* protocol: `ext-image-copy-capture-v1` measured
identically and appears in a Hyprland segfault stack (`create_frame` →
`CScreenshareFrame::transform`). The price of the in-process route is that the
package now carries Wayland protocol bindings of its own (`src/hhstream/_protocol/`,
vendored the same way `hhstream_pointer` already vendors its two) and that it can
only publish **PNG**, since PNG is the one format it can encode without a
dependency. Both costs are why `command` is still the default.
Comments in the spike that produced these numbers live in the branch's
`capture-speed` worktree; the honest caveat that came with them is that this box
is saturated by *rendering* whatever is on screen, so the win is in how much room
the compositor keeps, not in headroom that did not exist.
## The hardware requirement, stated plainly