perf(cef): opt-in BeginFrame from the frame clock and pin the OSR rate (#429)

* perf(cef): drive BeginFrame from the frame clock and allow pinning the rate

Pins the bridge revision that carries the presenter knobs, and applies the two
Dumber-side frame-path changes:

- CEF produces frames on BeginFrame ticks driven by the GTK frame clock instead
  of its own timer, so frame production follows the widget's cadence;
- `DUMBER_CEF_WINDOWLESS_FRAME_RATE` pins the OSR frame rate and takes
  precedence over the configured value and over adaptive monitor polling, so a
  run can be held at one cadence without editing the config file.

Non-positive and unparsable frame rates fall back to the configured behaviour,
and `DUMBER_CEF_EXTERNAL_BEGIN_FRAME=0` restores CEF's own timer.

The rate is parsed with a 32-bit bound rather than converted from Atoi, which
the linter rejects as a potential overflow.

This is not an ownership change: the bridge still presents a borrowed DMA-BUF
that CEF releases to its pool when the callback returns.

* chore(deps): pin the render-path bridge revision

Updates purego-cef2gtk to the revision that gates GtkGraphicsOffload on
the runtime GTK version, reports the effective offload, import-priority
and retire-limit settings in Diagnostics, and delivers batched wheel
motion once per frame tick.

* fix(cef): keep the external BeginFrame opt-in

The env flag had been inverted into a default-on switch. Restore the
canonical opt-in semantics: only an explicit true value drives CEF
BeginFrame from the GTK frame clock, while unset, empty, invalid and
false values keep CEF's own cadence. The tests now exercise
externalBeginFrameEnabled itself instead of the generic boolean helper.

The pinned OSR rate override stays: a valid positive
DUMBER_CEF_WINDOWLESS_FRAME_RATE wins over the configured value and
disables adaptive monitor-refresh polling, with invalid, non-positive and
overflowing values falling back to the configured behaviour. That pin
selects how often CEF is asked to capture frames; it is independent of
the external BeginFrame request rate.

The render diagnostic snapshot now also reports the effective GDK
offload request, import priority and retire limit.

* chore(deps): repin the bridge to the reviewed render-path revision

Moves the pin to the bridge revision that reports the offload request and
its installation separately, keeps offload recovery to the constructor, and
delivers the accrued wheel share when a stall retires a burst. The stale
go.sum lines for superseded revisions are dropped.

* fix(cef): log the pinned OSR rate and cover the BeginFrame call site

A pinned OSR rate silently suspended adaptive polling without saying so in
the startup log, so the cef: configured engine line now reports
windowless_frame_rate_pinned next to the rate it resolves to.

The BeginFrame opt-in was only ever asserted through the boolean helper,
which is how an inverted default shipped; a call-site test now drives
configureNativePopupWindow over absent, empty, false, invalid and true
values and checks the WindowInfo field itself. The render diagnostic
snapshot also reports gdk_offload_installed next to the request.

* chore(deps): pin purego-cef2gtk v0.11.0

The released tag carries the same content as the pseudo-version this branch
pinned (the bridge pull request was squash-merged), so this swaps a
branch-scoped revision for the released one without changing behaviour.

* feat(cef): enable external BeginFrame by default

The embedder's frame clock is the cadence the compositor samples, and the
tick implementation already exists, so CEF now waits for a BeginFrame
request on each tick instead of running its own timer. The environment
variable stays as the opt-out: 0, false, no or off restores CEF's internal
cadence.

This is a deliberate default under observation rather than a measured win.
The tick loop requests a frame on every tick regardless of the configured
capture rate, and no frame-timing measurement shows an improvement over
CEF's own timer yet, so the default is being dogfooded and can be reverted
to opt-in without touching the implementation. The comment on the helper
and the call-site test say so, and both now assert the default rather than
an opt-in.

commit

81c09fd1c5fd5608be698f9b56692adb04875e16

author

bnema · 41d ago · 2026-09-11 10:24 UTC

committer

GitHub · 41d ago · 2026-09-11 10:24 UTC

parent

62dc92c

signature

GPG UnverifiedB5690EEEBB952194 The key is unknown