Simple in-browser benchmark of different FFT implementations.
See the published page for more details on the benchmark and implementations. This README covers only development-related information.
This project is static HTML / CSS / JS, there's no build step or dependencies, just serve it and visit index.html.
npm install is only needed to get proper autocompletion / typechecking in the IDE.
The actual benchmark runs in a dedicated worker, see worker.mjs. The main thread (index.mjs) receives results and presents them in a basic UI.
The benchmarked implementations live in the impls folder. For each of them there's typically the compiled .wasm and the source code that generates it (except for JS-based ones which only have an .mjs). LLVM implementations are optimized with wasm-opt (some people have told me rustc should do that automatically if it's in PATH, but that's not the case for me, maybe because I use NixOS) and then its assembly is manually reviewed and improved a bit (mainly removing some unnecessary globals).
Note that since all current implementations are breadth order, the code at impls only implements a "depth phase" of the FFT (the hot part) while the rest of the FFT logic, like creating the twiddle factors and calling the depth phase function for all depths, is shared for all of them and can be found in worker.mjs.
Before the benchmark starts, all implementations are verified in all sizes against a reference 64-bit implementation in impls/reference.mjs. Those that differ by more than 0.1% on some of the tests get excluded from the benchmark and this is communicated in the UI. This is a very basic check, mostly meant to catch development mistakes.