Files
wl-tab/README.md
T
Milad Alizadeh 33d75222cf Make the previews live
Capture sessions stay open and tiles refresh, so the grid shows what the
windows are actually doing rather than a snapshot from when it opened.
--live all|current|none picks how much of that happens, --fps caps it.

Three things keep it cheap:

The protocol is damage-driven. After a session's first frame the
compositor only produces another once the content changes, so a request
left outstanding on an idle window costs nothing. Measured with one
animating window out of ten: 52,52,1,1,1,1,1,1,13,1 frames over 4s — the
static windows delivered their first frame and then nothing.

The overlay's own wl_surface.frame callbacks are the clock, so refreshes
stop when it isn't being presented and no timer or poll loop is needed.
Per-tile throttling on top of that measured 12.1/s at --fps 12.

Each window gets two buffers, since a capture must not write into one the
compositor is reading, and wl_buffer.release says which is free. That
release is the whole contract: with wl_shm the compositor copies the
pixels at commit and hands the buffer straight back, so the slot on
screen is usually free too. Waiting for it to stop being the displayed
slot instead — which is what this first did — deadlocks after two frames,
with both slots stuck busy (687 blocked attempts, 2 frames per tile).

Live mode doubles the shm handed to the compositor (110MB for ten windows
here, against 55MB with --live none); our own RSS is unaffected because
those pages are still never mapped. --verbose now reports frames, ticks,
releases, blocked attempts and pool size, which is what localised the
release bug.
2026-08-23 10:54:06 +01:00

5.8 KiB

wlgrid

A window switcher for wlroots compositors: a grid overlay of live window previews that looks like a rofi theme, and focuses the window you pick.

It replaces a wlthumbs | rofi pipeline. The difference is that no thumbnails exist: each window is captured straight into a wl_shm buffer that is handed to its own wl_subsurface, and wp_viewporter tells the compositor which rectangle to scale it into. There is no image encoding, no scaler, and no full-resolution bitmap in this process — which is also why it appears in about 60 ms and holds ~18 MB of RSS however many windows are open.

sway-tree       0.6ms     window list + con_ids over sway IPC
toplevels       0.2ms     ext-foreign-toplevel-list handles
constraints     1.5ms     every capture session's buffer size, in one roundtrip
capture        52.5ms     8 windows, all frames in flight at once
labels          0.0ms     shaped on a worker thread while the captures ran
mapped          4.9ms     layer surface + subsurfaces on screen

The capture phase is the compositor reading full-resolution window pixels out of the GPU. It is bandwidth-bound (~1.1 GB/s here) and unaffected by how large the thumbnails are — which also makes it a free window to do other work in. Loading a font and rasterising its first glyphs costs ~20ms, so labels are shaped on a worker thread started before the captures and joined after them, and cost nothing in wall clock.

Status

Working: a labelled grid of live previews with keyboard navigation. Type-to-filter is the one thing the rofi version had that this doesn't — see the roadmap.

Usage

wlgrid [--print] [--verbose] [--hide-labels] [--font FAMILY] [--font-size PX]
       [--live all|current|none] [--fps N] [--timeout SECS]
  • --print writes the selected sway con_id to stdout instead of focusing it
  • --verbose prints phase timings and how many windows were captured
  • --hide-labels draws an icon-only grid
  • --font FAMILY label font family (default Berkeley Mono)
  • --font-size PX label size in logical px
  • --live all|current|none which tiles keep updating (default all)
  • --fps N cap on updates per tile per second (default 12)
  • --timeout SECS exits after a deadline (an escape hatch: the overlay takes an exclusive keyboard grab)

Bind it in sway:

bindsym $mod+Tab exec wlgrid
key
→ ← / Tab Shift+Tab next / previous window
↑ ↓ move a row
Home End first / last
Enter focus the selection
Escape / q cancel, leaving focus alone

Navigation reads raw evdev keycodes, so it is layout-independent — but it also means virtual-keyboard clients such as wtype (which invent their own keymap) cannot drive it. That goes away with xkb support, which filtering needs anyway.

Live previews

Capture sessions stay open, so a tile can be refreshed. Three things keep that from being expensive:

  • It is damage-driven. After a session's first frame the compositor only produces another once the window content changes, so a request left outstanding on an idle window costs nothing. Measured over 4s with one animating window out of ten: 52,52,1,1,1,1,1,1,13,1 frames — the static windows delivered exactly their first frame and nothing more.
  • Frame callbacks are the clock. Re-captures are driven by the overlay's own wl_surface.frame callbacks, so they stop when it isn't being presented, and --fps throttles per tile on top of that (12 fps measured as 12.1).
  • Two buffers per window, alternating. A capture must not write into a buffer the compositor is reading, so each window gets two and wl_buffer.release decides which is free. Note that release is the entire contract: with wl_shm the compositor copies the pixels out at commit and hands the buffer straight back, so the slot on screen is usually free too — waiting for it to stop being displayed instead deadlocks after two frames.

The cost is memory and bandwidth: two full-resolution buffers per window (110 MB of shm for ten windows on this display, versus 55 MB with --live none) and a readback per refreshed frame. --live current refreshes only the selected tile, which is much cheaper and still reads as alive.

Look

Colours, font metrics and grid geometry come from the rofi theme this replaces (gruvbox dark, a yellow selection filling the element padding, ceil(sqrt(n)) columns capped at 4, 16:9 tiles, title · app centred underneath) and live in src/theme.rs. They will move to a config file so they can't drift from the .rasi.

The font is looked up by family name. Your own font directories are scanned first because they are small; the full system scan (~37ms) happens only if the family isn't found there, and an unknown family then falls back to whatever cosmic-text picks rather than failing. Long titles are ellipsised to the cell.

Requirements

A wlroots compositor advertising ext-image-copy-capture-v1, ext-image-capture-source-v1 (with the foreign-toplevel source manager), ext-foreign-toplevel-list-v1, wlr-layer-shell-unstable-v1 and wp_viewporter — sway 1.11+, and in principle Hyprland, labwc and jay, though only sway is tested. sway is also the source of truth for the window list and for focusing, over its IPC socket.

Known upstream issue: holding per-toplevel capture sessions open makes windows blurry on fractionally scaled outputs (sway#9113). Integer scales are unaffected. It matters more once previews are live.

Roadmap

  • type-to-filter with fzf-quality fuzzy matching (and the xkb keyboard input it needs, which would also let virtual-keyboard clients drive the overlay)
  • dmabuf capture, so the pixels never leave the GPU at all — and live previews stop costing a readback per frame

Building

cargo build --release
cargo test          # grid geometry, ellipsising, glyph output