Skip to content

Repository files navigation

KernelIndex

Find the fastest verified GPU kernel for your exact workload.

KernelIndex is the public performance index for GPU software. It resolves an exact workload (operation, shape, dtype, layout, hardware, framework) to the fastest currently known compatible implementation, with source, license, environment, benchmark protocol, raw evidence, and an explicit trust level.

Results are ranked only inside comparable workloads and environments.

Development

Requires Node 24.19.0 (.node-version) and pnpm 11.21.0 (selected by Corepack from the root packageManager field).

pnpm install --frozen-lockfile
pnpm dev        # start apps/web on http://localhost:3000
pnpm check      # lint + typecheck + production build
pnpm test       # unit tests; database tests run when DATABASE_URL is set

Pages read the catalog through the read seam in apps/web/src/lib/catalog.ts. CATALOG_BACKEND=fixtures (the default, no configuration needed) serves a deterministic, visibly illustrative catalog; CATALOG_BACKEND=postgres serves real published records.

Database

Local PostgreSQL 18 comes from compose.yaml (docker compose up -d), or without the compose plugin:

docker run -d --name kernelindex-postgres -p 127.0.0.1:5432:5432 \
  -e POSTGRES_USER=kernelindex -e POSTGRES_PASSWORD=kernelindex \
  -e POSTGRES_DB=kernelindex -v kernelindex-postgres:/var/lib/postgresql \
  postgres:18.6

Copy apps/web/.env.example to apps/web/.env.local, then:

pnpm --filter @kernelindex/web db:migrate   # apply committed SQL migrations
pnpm --filter @kernelindex/web db:seed      # illustrative registry examples
pnpm --filter @kernelindex/web import:sol -- --dry-run   # review report only
pnpm --filter @kernelindex/web import:sol -- --publish   # real SOL records
pnpm --filter @kernelindex/web import:gpumode -- --group helion --dry-run
pnpm --filter @kernelindex/web import:gpumode -- --group helion --publish

The importer is idempotent: re-running with unchanged upstream data inserts nothing. Imported results are labeled Reported — never "verified" — and every record links its workload, protocol, environment, source snapshot, and raw source metrics.

docs/ENGINEERING_DESIGN.md is the source of truth for product semantics, data architecture, and sequencing.

License

KernelIndex-authored code is licensed under Apache-2.0. Imported source code, artifacts, benchmark data, and database content retain their upstream licenses.

About

An open performance index for GPU kernels. Search by operation, shape, dtype, and hardware to find the fastest known implementation.

Topics

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages