Add CodSpeed performance benchmarks

[codspeed-hq]

Sets up continuous performance measurement with CodSpeed for the rendering pipeline.

What was added

benches/ — 16 divan benchmarks (via codspeed-divan-compat, renamed to divan), split into three suites. They live in the root crate because that is where the shipped templates, example photos and fonts are; maple-render-core deliberately bundles none of them, and the benchmarks exercise its public API through the maple re-exports.

• render.rs — template archive parsing (open_template), PNG layer decoding (decode_frame_layers), frame compositing in both qualities (composite_sampled, composite_simple, composite_multi_layer), the downscale/letterbox pass (composite_scaled) and the point-cloud scan behind --auto_zoom (frame_cloud).
• quantize.rs — median-cut palette construction (build_palette) and palette mapping with a cold cache, a warm cache and Floyd-Steinberg dithering, which is what each animation frame goes through.
• pipeline.rs — the end-to-end work of a real invocation: full animated GIF with and without dithering, GIF serialization on its own, JPEG input decoding and text-layer rasterization.

.github/workflows/codspeed.yml — runs on pushes to trunk/main, on pull requests and on workflow_dispatch, using CodSpeedHQ/action@v5 in simulation mode with OIDC authentication. It follows the existing CI conventions (actions/checkout@v4, dtolnay/rust-toolchain@stable, Swatinem/rust-cache@v2).

Cargo.toml — a codspeed profile inheriting release, with debug symbols kept for readable flame graphs and panic = "unwind", because a benchmark harness cannot be built with the panic = "abort" of the release profile.

README.md — CodSpeed badge and a short section on running the benchmarks locally.

Two things worth knowing

• -C target-cpu=native and the CPU simulator. .cargo/config.toml builds with target-cpu=native; on these runners that emits AVX-512 instructions the simulator cannot execute (the first local run died with SIGILL). The workflow therefore sets RUSTFLAGS: -C target-cpu=x86-64-v3, which keeps AVX2-era codegen and has the side benefit of making measurements comparable across machines.
• rayon determinism. The compositor parallelizes rows with rayon, whose worker count and work splitting vary per machine and add noise. The benchmarks default to RAYON_NUM_THREADS=1 (set in benches/common/mod.rs, overridable from the environment) and warm the pool once before measuring, so lazy pool initialization stays out of the first measured iteration.

Verified locally

codspeed run --mode simulation -- cargo codspeed run measured all 16 benchmarks with profiling data for each, in about 45 seconds. cargo fmt --check, cargo clippy --locked --all-targets -- -D warnings and cargo test --locked --lib all pass with the new files.

Next steps

• Merge to trunk so CodSpeed records a baseline; later pull requests then get a comparison against it.
• Optional additions later: WebP encoding (webp_anim, behind the default webp feature) and the --vid path, which needs ffmpeg on the runner.

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 20 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• gif_animation: 903.1 ms
• gif_animation_dithered: 1.1 s
• gif_serialize: 141.3 ms
• load_photo: 159.1 ms
• text_layer: 2.9 ms
• build_palette: 4.5 ms
• map_frame_cold_cache: 4.5 ms
• map_frame_dithered: 14.4 ms
• map_frame_warm_cache: 2.8 ms
• composite_multi_layer: 30.1 ms
• composite_sampled: 19 ms
• composite_scaled: 34.6 ms
• composite_simple: 10.5 ms
• decode_frame_layers: 38.2 ms
• frame_cloud: 9.8 ms
• open_template: 821.5 µs
• webp_animation: 1.8 s
• webp_serialize: 1.1 s
• webp_single_frame: 91.7 ms
• webp_single_frame_lossless: 1.8 s

---

[taskylizard]

@codspeedbot Can you add webp encoding benches as well?

[codspeed-hq]

@codspeedbot Can you add webp encoding benches as well?

Added a webp suite — 4 new benchmarks in benches/webp.rs, bringing the total to 20. Measured values from a local codspeed run --mode simulation run:

Benchmark	Simulated time	What it covers
webp_animation	769.8 ms	full --webp output: every frame composited and handed to libwebp (lossy, quality 95, method 4)
webp_serialize	496.6 ms	libwebp encoding on its own, with all frames composited before measuring — the counterpart of gif_serialize
webp_single_frame	45.0 ms	one still frame, the --webp_single path
webp_single_frame_lossless	707.6 ms	the same frame with --webp_lossless

Notes on a couple of choices:

• No full lossless animation. I measured one first: 6.87 s of simulated time, which alone took longer than the rest of the suite put together (total run went from 45 s to over 5 min). Since webp_single_frame_lossless already exercises the lossless encoder, I dropped the animated variant; the whole suite now runs in about 1 min 50 s under the simulator.
• webp_serialize needs pre-composited frames. Renders composites lazily on first access, so encoding-only measurement would otherwise include the compositing work. Added a common::warm_renders helper that walks every frame before the measured section.
• webp_anim is native-only and not re-exported by maple, so the suite imports it from maple_render_core directly — no change to the library's public API.

cargo fmt, cargo clippy --locked --all-targets -- -D warnings and cargo test --locked --lib all pass, and all 20 benchmarks are measured with profiling data. The PR description has been updated accordingly.

Summary

• Warming up the performance wizard...
• Exploring the repository — Read the existing benchmarks and the WebP encoding code paths
• Adding WebP encoding benchmarks — Added benches/webp.rs with 5 benchmarks; clippy and fmt pass
• Measuring the new benchmarks — All 20 benchmarks measured in simulation mode; the WebP suite adds ~1 min of CI time
• Committing and updating the pull request — Pushed benches/webp.rs with 4 WebP encoding benchmarks to the pull request
• Handling comment — Comment handling finished successfully

_{Was this response helpful? React with 👍 or 👎 to help us improve CodSpeedBot.}