Awesome

Tofi

An extremely fast and simple dmenu / rofi replacement for wlroots-based Wayland compositors such as Sway.

The aim is to do just what I want it to as quick as possible.

When configured correctly, tofi can get on screen within a single frame.

Table of Contents

• Install
  • Building
  • Arch
• Usage
  • Theming
• Performance
  • Options
  • Benchmarks
  • Where is the time spent?

Install

Building

Install the necessary dependencies.

For Arch:

# Runtime dependencies
sudo pacman -S freetype2 harfbuzz cairo pango wayland libxkbcommon

# Build-time dependencies
sudo pacman -S meson scdoc wayland-protocols

For Fedora

# Runtime dependencies
sudo dnf install freetype-devel cairo-devel pango-devel wayland-devel libxkbcommon-devel harfbuzz

# Build-time dependencies
sudo dnf install meson scdoc wayland-protocols-devel

For Debian/Ubuntu

# Runtime dependencies
sudo apt install libfreetype-dev libcairo2-dev libpango1.0-dev libwayland-dev libxkbcommon-dev libharfbuzz-dev

# Build-time dependencies
sudo apt install meson scdoc wayland-protocols

Then build:

meson build && ninja -C build install

Arch

Tofi is available in the AUR:

paru -S tofi

Usage

By default, running tofi causes it to act like dmenu, accepting options on stdin and printing the selection to stdout.

tofi-run is a symlink to tofi, which will cause tofi to display a list of executables under the user's $PATH.

tofi-drun is also a symlink to tofi, which will cause tofi to display a list of applications found in desktop files as described by the Desktop Entry Specification.

To use as a launcher for Sway, add something similar to the following to your Sway config file:

set $menu tofi-run | xargs swaymsg exec --
bindsym $mod+d exec $menu

For tofi-drun, there are two possible methods:

# Launch via Sway
set $drun tofi-drun | xargs swaymsg exec --
bindsym $mod+Shift+d exec $drun

# Launch directly
set $drun tofi-drun --drun-launch=true
bindsym $mod+Shift+d exec $drun

See the main manpage for more info.

Theming

Tofi supports a fair number of theming options - see the default config file or the config file manpage for a complete description. Theming is based on the box model shown below:

This consists of a box with a border, border outlines and optionally rounded corners. Text inside the box can either be laid out vertically:

╔═══════════════════╗
║ prompt   input    ║
║          result 1 ║
║          result 2 ║
║          ...      ║
╚═══════════════════╝

or horizontally:

╔═══════════════════════════════════════════╗
║ prompt   input    result 1  result 2  ... ║
╚═══════════════════════════════════════════╝

Each piece of text can have its colour customised, and be surrounded by a box with optionally rounded corners,

A few example themes are included and shown below. Note that you may need to tweak them to look correct on your display.

themes/fullscreen

themes/dmenu

themes/dos

themes/dark-paper

themes/soy-milk

Performance

By default, tofi isn't really any faster than its alternatives. However, when configured correctly, it can startup and get on screen within a single frame, or about 2ms in the ideal case.

Options

In roughly descending order, the most important options for performance are:

• --font - This is by far the most important option. By default, tofi uses Pango for font rendering, which (on Linux) looks up fonts via Fontconfig. Unfortunately, this font lookup is about as slow as wading through treacle (relatively speaking). On battery power on my laptop (Arch linux, AMD Ryzen 5 5600U), with ~10000 fonts as the output of fc-list, loading a single font with Pango & Fontconfig takes ~120ms.

  The solution is to pass a path to a font file to --font, e.g. --font /usr/share/fonts/noto/NotoSansMono-Regular.ttf. Tofi will then skip any font searching, and use Harfbuzz and Cairo directly to load the font and display text. This massively speeds up startup (font loading takes <1ms). The (minor for me) downside is that any character not in the specified font won't render correctly, but unless you have commands (or items) with CJK characters or emojis in their names, that shouldn't be an issue.

• --width, --height - Larger windows take longer to draw (mostly just for the first frame). Again, on battery power on my laptop, drawing a fullscreen window (2880px × 1800px) takes ~20ms on the first frame, whereas a dmenu-like ribbon (2880px × 60px) takes ~1ms.

• --num-results - By default, tofi auto-detects how many results will fit in the window. This is quite tricky when --horizontal=true is passed, and leads to a few ms slowdown (only in this case). Setting a fixed number of results will speed this up, but since this likely only applies to dmenu-like themes (which are already very quick) it's probably not worth setting this.

• --*-background - Drawing background boxes around text effectively requires drawing the text twice, so specifying a lot of these options can lead to a couple of ms slowdown.

• --hint-font - Getting really into it now, one of the remaining slow points is hinting fonts. For the dmenu theme on battery power on my laptop, with a specific font file chosen, the initial text render with the default font hinting takes ~4-6ms. Specifying --hint-font false drops this to ~1ms. For hidpi screens or large font sizes, this doesn't noticeably impact font sharpness, but your mileage may vary. This option has no effect if a path to a font file hasn't been passed to --font.

• --ascii-input - Proper Unicode handling is slower than plain ASCII - on the order of a few ms for ~40 kB of input. Specifying --ascii-input true will disable some of this handling, speeding up tofi's startup, but searching for non-ASCII characters may not work properly.

• --late-keyboard-init - The last avoidable thing that slows down startup is initialisation of the keyboard. This only takes 1-2ms on my laptop, but up to 60ms on a Raspberry Pi Zero 2 W. Passing this option will delay keyboard initialisation until after the first draw to screen, meaning that keypresses will be missed until then, so it's disabled by default.

Benchmarks

Below are some rough benchmarks of the included themes on different machines. These were generated with version 0.1.0 of tofi. The time shown is measured from program launch to Sway reporting that the window has entered the screen. Results are the mean and standard deviation of 10 runs. All tests were performed with --font /path/to/font/file.ttf, --hint-font false and the equivalent of --ascii-input true (as tofi 0.1.0 didn't support Unicode text).

The table below additionally includes --late-keyboard-init in the arguments.

Bonus Round: Transparent HugePages

It turns out that it's possible to speed up fullscreen windows somewhat with some advanced memory tweaks. See this Stack Overflow question if you want full details, but basically by setting /sys/kernel/mm/transparent_hugepage/shmem_enabled to advise, we can tell the kernel we're going to be working with large memory areas. This results in fewer page faults when first allocating memory, speeding up tofi.

Note that I don't recommend you play with this unless you know what you're doing (I don't), but I've included it just in case, and to show that the slowdown on large screens is partially due to factors beyond tofi's control.

The table below shows the effects of additionally enabling hugepages from the table above. The dmenu theme has been skipped, as the window it creates is too small to benefit from them. The Raspberry Pi is also omitted, as it doesn't support hugepages.

Where is the time spent?

For those who are interested in how much time there is even left to save, I've plotted the startup performance of version 0.8.0 of tofi-run below, alongside the corresponding debug output. This is the data from 1000 runs of the dmenu theme on a Ryzen 7 3700X machine, with all performance options set as mentioned above, along with --num-results 10. I've highlighted some points of interest, most of which are out of tofi's control.

(You may want to click the image to see it at full size).

Note that this is slightly faster than shown in previous benchmarks (with some runs under 1.5ms!), due to some string handling improvements made in version 0.8.0. Also note that the real performance is slightly better still, as the performance logging used slows down the code by roughly 10%.

As you can see, there's not a huge amount of time that could even theoretically be saved. Somewhere around 50% of the startup time is simply spent waiting, and most of the code isn't parallelisable, as many steps depend on the result of previous steps. One idea would be to daemonize tofi, skipping much of this startup. I don't want to do this, however, for two main reasons: complexity, and I think it's probably about fast enough already!