Samuel Onoja
e22f0ebef9
Panel appearance now applies to the whole panel, with seven styles: Floating, Capsules, Notch, Edge to edge, Fused islands, Adaptive and Morph. - Share composition, geometry, and motion through the renderer-independent `ferese-panel` crate, including native materials, input regions, and Settings previews. - Add the Morph hub with media, calendar, notifications, controls, workspace previews, hover feedback, and attached mode. Keep recording and low-battery status visible and respect window clearance. - Cache workspace preview geometry, draw thumbnails directly, centralize Morph transitions, and keep rendering settings local to each shell instance. - Put Record Screen in the Morph header and show the display control only when an external monitor is connected. Update panel documentation, the desktop screenshot, and native check style names. Local validation included 660 compositor tests, 191 shell tests, isolated preview-cache checks, GPU Morph captures at 1×/1.25×/1.5×, formatting, and Clippy with warnings denied. The final header change passed all 23 Morph tests. Release workflow jobs are skipped for this PR.
3.2 KiB
Nested resize checks
Build and launch a preview with protocol and render logging:
cargo build --release --locked -p ferese -p ferese-shell
WAYLAND_DEBUG=1 FERESE_TRACE_PERFORMANCE=1 \
target/release/ferese --backend nested --grant-effects --grant-shell-control -- \
target/release/ferese-shell 2>/tmp/ferese-resize.log
Set a wallpaper, then drag each edge of the host window and repeatedly maximize and restore it. The wallpaper and bar should resize without long stalls. Check this at fractional scaling too.
Compare renderers
Repeat the check with ICED_BACKEND=wgpu and ICED_BACKEND=tiny-skia to compare
renderers. The shell prefers tiny-skia with a wgpu fallback for its small
surfaces, while Settings prefers wgpu with a tiny-skia fallback for larger
scrolling pages. An explicit ICED_BACKEND selects the requested backend, and
the shell's automatic preference is not passed to Settings when it launches it.
Follow the trace from host resize through layer configure, client acknowledgement,
new buffer commit and compositor rendering. A quick acknowledgement does not mean
the client has finished drawing. FERESE_TRACE_PERFORMANCE measures compositor
work, not shell rasterization.
Historical check
A KDE Wayland check at 175% scale used debug binaries and protocol tracing. Median configure acknowledgement was 1.35 ms with tiny-skia and 2.27 ms with wgpu. Wallpaper buffer intervals were 5919 ms with tiny-skia (only two buffers) and 9.85 ms median with wgpu. The captures had different durations, so these observations do not establish comparative throughput. GPU intervals also had stalls up to about 200 ms.
Repeat on the current build. A nested result does not establish direct-session resize or frame-pacing behavior.
Controlled slow client
Build the compositor and control tool, then run on a Wayland host:
cargo build --release --locked -p ferese -p feresectl
FERESE_TEST_RESIZE=1 FERESE_TEST_BINARY=target/release/ferese \
FERESE_TEST_CTL=target/release/feresectl \
python3 scripts/tests/test_resize_dependencies_isolated.py
The test creates a private nested desktop and two SHM clients. One acknowledges a resize immediately but withholds its new buffer until instructed. It checks that an independent first-column viewport target moves both windows together before the 300 ms deadline, while the slow client's raster and presented size stay held, then commits an actually resized raster. This is a state/coordinate check on the selected host renderer, not a smoothness or hardware benchmark.
For deterministic commit, supersession, multiple-client, timeout, unmap, workspace, swipe-cancel, stacked-column, prediction, hit-testing, snapshot and capture privacy coverage without a host desktop:
cargo test --locked -p ferese resize_dependencies_protocol_and_pixels -- --ignored
That test requires an offscreen EGL device, including a software device. Layout and dependency unit tests run without EGL in the ordinary workspace test suite. Viewport dependencies retain at most 32 relative target recipes; a longer burst falls back to a conservative linked wait until commit, deadline, or an absolute retarget. Existing resize snapshots and their handoff lifetime are reused.