fast-kernels

CUDA kernel experiments, with a benchmark-first layout from the start.

The default path here is hand-written CUDA C++. Python is used for suite configs, environment capture, reporting, and CLI tooling.

Layout

• csrc/: native code, headers, launchers, and kernel-family directories
• src/fast_kernels/: CLI, schemas, registry, environment capture, and reporting
• benchmarks/: suite definitions, shapes, baseline adapters, and runners
• results/: versioned benchmark outputs

Quickstart

uv sync
uv run fk env
uv run fk verify benchmarks/suites/template_gemm.toml
uv run fk bench benchmarks/suites/template_gemm.toml

That benchmark command writes a run under results/template_gemm/<run-id>/.

CUDA build

Editable installs default to FK_ENABLE_CUDA=ON and build for the active GPU architecture.

Default CUDA setup:

uv sync --extra benchmark
uv run fk env

CPU-only opt-out:

CMAKE_ARGS=-DFK_ENABLE_CUDA=OFF uv sync

For direct CMake builds:

cmake --preset dev
cmake --build --preset dev