Fun facts

Things you only learn by digging through git log, git blame, and the corners of the source tree.

The first commit is from 2018

The repo started on 2018-08-26 with a single line in Initial commit from msiglreith. The first non-trivial commit, Setup crate structure and wrap first interface, lands two days later. The repo has since accumulated 10,636 commits from 663 unique authors (or pseudonyms — humans, bots, and at least one collective handle named "Inner Daemons" all count).

Naga is older than its presence here

naga was a separate repository for years before being merged into wgpu on 2023-10-25 in PR #4296, "Integration of Naga into Repo." The git history rewrote its filesystem path under naga/, but the commits stretch back to early 2020. That's why the all-time top committer list mixes the original gfx-rs/wgpu folks (Dzmitry Malyshau, Connor Fitzgerald) with naga regulars (Jim Blandy, Andy Leiserson, teoxoy).

The largest source file in the repo translates Metal shaders

naga/src/back/msl/writer.rs is 349 KB of Rust. The HLSL writer is half its size at 214 KB. Combined, the naga/src/back/ tree alone is over 800 KB of Rust — most of it pattern-matching on Naga IR variants and emitting target-language text. A close runner-up is naga/src/proc/constant_evaluator.rs at 189 KB, which AGENTS.md warns has tricky borrow-checker patterns.

Almost 300 TODOs and FIXMEs are still on the books

grep -rE "TODO|FIXME|HACK" --include="*.rs" finds 298 occurrences across 142 files. That's roughly one tagged comment per three source files. The naga directories carry a disproportionate share, mostly under naga/src/back/ and naga/src/front/spv/.

Two xtask crates

The repo has two cargo xtask entry points: the workspace-level one in xtask/ (with subcommands cts, test, run-wasm, changelog, miri, vendor-web-sys, install-warp, install-agility-sdk) and a naga-only one in naga/xtask/ (with one subcommand: validate <backend>, which feeds checked-in snapshot output to external compilers like SPIRV-Tools, glslang, DXC, and Xcode CLI). Run them by cding into the right directory first.

Mozilla left fingerprints in the tag list

Between December 2019 and February 2020 the repo was tagged with names like mozilla-646613d3969290c44c08b7a43ee00a6dc5cff4ff — three of them, all referencing the exact wgpu-core revision Firefox vendored at the time. They're still in git tag -l. The tags are unsigned but they record the versions that shipped in early Firefox WebGPU previews.

bors[bot] merged 11 % of all commits

The historical merge bot, bors[bot], has 1,172 commits to its name — the second-most prolific "author" in the repo, and only because everything that went through bors was attributed to it. The project later moved to GitHub's merge queue, which doesn't leave a bot author behind. Combined with dependabot[bot] (280) and renovate[bot] (145), bots account for roughly 15 % of the recorded commits.

wgpu-hal exists because gfx-hal got too big

wgpu-hal/README.md is unusually candid: "wgpu_hal is a spiritual successor to gfx-hal, but with reduced scope, and oriented towards WebGPU implementation goals." The crate consolidates per-API backends that used to ship as separate gfx-hal-vulkan, gfx-hal-metal, etc. crates. The current in-tree HAL has no validation and no state tracking — those moved up to wgpu-core.

The Metal backend doesn't use the metal crate

It uses objc2 and objc2-metal (and objc2-quartz-core, objc2-foundation, etc.) directly. The reasoning: the metal crate adds a layer of Rust-friendly types over the bindings, but wgpu-hal wants the bindings directly so it can mirror Apple's headers exactly. Browse wgpu-hal/src/metal/ and you'll see raw MTLTexture, MTLBuffer, MTLRenderCommandEncoder references through objc2.

The repo has its own opinions about LLM-generated code

CONTRIBUTING.md says "Using LLMs and AIs to generate code that is part of a contribution is allowed. However, the author submitting the PR must fully adhere to Change Ownership rules." The undue-burden clause continues: maintainers may close PRs that cause "LLM slop" — code that looks plausible but cannot be justified by its author.

AGENTS.md is the more specific document for AI assistants; it asks for terse comments, no emoji, and explicit verification against the WebGPU/WGSL specifications rather than CTS expectations.

The CTS revision is pinned to a single hash

cts_runner/revision.txt is a single line — a SHA from the gpuweb/cts repository. The renovate[bot] runs a "update cts digest" job (visible in recent commits: chore(deps): update cts digest to 8640b42 (#9466)). When the digest moves, the test lists in cts_runner/test.lst and cts_runner/fail.lst may need updating to track new or renamed tests.

You can run the examples in your browser

cargo xtask run-wasm builds the example crate to wasm and serves it locally. The pre-built versions live at https://wgpu.rs/examples/. The same Rust code that drives the native examples — including the ray tracing examples — runs against navigator.gpu in a browser.

"Inner Daemons" is a real contributor

In the recent commit log, especially in wgpu-hal/ and naga/, you'll see Inner Daemons as the author. It's the handle of an active human contributor, not (despite the name) an AI agent or a bot. They're now top-five in commits over the past 90 days for wgpu-hal/.