webgpu/

gnark WebGPU backend

Experimental browser acceleration of gnark's Groth16 and PLONK provers over BN254, BLS12-377 and BLS12-381. The prover is gnark's native Go prover compiled to wasm; its heavy operations (multi-scalar multiplications, FFTs, the PLONK quotient numerator) are offloaded to WGSL compute shaders driven by a TypeScript runtime.

Disclaimer

This package is experimental. The APIs may change and the backend is not audited.

How it works

The native provers (backend/groth16, backend/plonk) expose an Accelerator interface on their proving keys: a handful of coarse-grained hooks (MultiExp, ComputeH for Groth16; MultiExp, ToCanonical, ComputeNumerator for PLONK). When no accelerator is attached the prover runs gnark-crypto on the CPU, so the CPU path is unchanged. This package provides WebGPU implementations of those hooks:

1. Prepare(pk) uploads the proving key bases to the GPU once, through syscall/js, and attaches the accelerator to the key.
2. groth16.Prove / plonk.Prove then run unchanged; each hook call crosses the Go to JavaScript boundary with the raw little-endian Montgomery limbs of gnark-crypto elements (no re-encoding on the Go side), the TypeScript runtime dispatches the WGSL kernels, and results come back as affine points or coefficient vectors.

KZG openings go through gnark-crypto's kzg.OpenWithCommitter and kzg.BatchOpenSinglePointWithCommitter, so the opening and folding logic is not duplicated here.

Layout

• groth16/, plonk/: Prepare / Release facades and, per curve, one generated accelerator.go implementing the native prover's Accelerator interface. The internal/wasmruntime/{native,webgpu} commands are the wasm entrypoints loaded by the TypeScript package (native keeps the CPU prover for comparison).
• internal/bridge/: Go client of the JavaScript bridge objects (gnarkGroth16WebGPU, gnarkPlonkWebGPU) and the byte protocol.
• internal/wasmruntime/: JavaScript-facing runtime (read keys and constraint systems, prove, verify) shared by the four entrypoints.
• internal/generator/: bavard generator. Produces the per-curve Go accelerators (templates/go), the per-curve WGSL shaders (templates/wgsl, with all field constants computed from gnark-crypto) and the Go test-vector builders. Run go run . in that directory after editing a template.
• shaders/: WGSL kernels. shaders/curves/* are generated, shaders/common/* are hand-written and curve-agnostic (MSM bucket sort, Pippenger stages).
• web/: TypeScript runtime, browser test pages, the npm build and scripts/e2e.mjs, which runs the test pages headlessly with Playwright.

The Go packages build only for GOOS=js GOARCH=wasm.

Build

cd backend/accelerated/webgpu/web
npm ci
npm run build:all        # lint, shaders bundle, TypeScript, Go wasm assets

Narrower targets: npm run build, build:shaders, build:wasm, build:wasm:groth16, build:wasm:plonk, lint.

Test

Test fixtures for the browser pages are produced with npm run build:test-fixtures:<suite> (api, groth16, plonk). The pages can be served from web/ and opened at tests/index.html, or run headlessly:

npm run test:e2e     # API suites + one verified proof per system and curve
npm run bench:e2e    # prover benchmark matrix (2^15 and 2^18 fixtures)

Set PW_CHANNEL=chrome to use the system Chrome instead of Playwright's Chromium. .github/workflows/webgpu.yml runs test:e2e on pull requests.