Performance

Area	What’s drawn	Per draw
200×50	glass, label, hint	4.8 µs
80×24	glass, label, hint	4.6 µs
40×2	label, hint	2.7 µs
12×1	label, cut	1.0 µs

A full draw takes under 0.03% of a 60 fps frame. The time doesn’t grow with the area: only the glass’s 66 cells and the label’s and hint’s characters are written.

How it’s measured

examples/bench.rs draws the widget into a ratatui Buffer a million times per size, 1 ms of elapsed time apart, so every frame and band position comes round, after a warm-up, and prints the mean time per draw as pito-compare data. From a clone:

cargo run --release --example bench > draw-time.toml

Measured once, on 2026-10-14, on an AMD Ryzen 9 9950X with Rust 1.97.1, a release build on one thread. The data behind the chart, with the machine line and note added by hand:

title = "HOURGLASS DRAW TIME"
unit = "NANOSECONDS PER DRAW"
better = "lower"
suffix = " NS"
subtitle = "RYZEN 9 9950X, RUST 1.97.1, RELEASE, ONE THREAD"
date = "2026-10-14"
notes = ["MEAN OF 1,000,000 DRAWS PER ROW, 1 MS APART, ZERO ALLOCATIONS."]
parties = [{ name = "PITO-HOURGLASS" }]
rows = [
  { label = "200×50: GLASS, LABEL, HINT", values = [4807.0] },
  { label = "80×24: GLASS, LABEL, HINT", values = [4621.5] },
  { label = "40×2: LABEL, HINT", values = [2689.8] },
  { label = "12×1: LABEL, CUT", values = [981.4] },
]

pito-compare’s pcmp draws the chart from it:

pcmp draw-time.toml --out .

Zero allocations

Drawing allocates nothing. tests/allocation.rs counts every allocation through a global allocator while 200 frames are drawn at six sizes, and holds the count at zero; it runs on Linux, Windows and macOS on every CI run. From a clone:

cargo test --test allocation