pto-kernels

A collection of high-performance custom kernels for Ascend NPUs, built on top of pto-isa — the Parallel Tile Operation virtual instruction set architecture designed by Ascend CANN.

PTO focuses on tile-level operations, enabling efficient, composable kernel development targeting Huawei's Ascend AI processors.

---

Prerequisites

• A configured torch-npu environment
• Ascend toolkit installed at /usr/local/Ascend/ascend-toolkit

Run the one-time setup before building:

make setup_once

Install repository using pip

The repository is "pip installable", i.e.,

export CMAKE_GENERATOR="Unix Makefiles" && pip install -v git+https://github.com/huawei-csl/pto-kernels.git

---

Build

source /usr/local/Ascend/ascend-toolkit/set_env.sh
pip3 install -r requirements.txt
make build_wheel

This produces an installable Python wheel:

pto_kernels-X.Y.Z-*.whl

---

Installation

pip install --force-reinstall pto_kernels-*.whl

---

Testing

make test

---

Repository Structure

pto-kernels/
├── csrc/                  # C++ kernel source files
├── python/pto_kernels/    # Python bindings and utilities
├── examples/jit_cpp/      # JIT compilation examples
├── tests/                 # Test suite
├── scripts/               # Helper scripts
├── doxygen/               # API documentation config
└── CMakeLists.txt         # CMake build configuration

---

Contributing

Contributions are welcome! Please read CONTRIBUTING.md before opening a pull request.

---

License

BSD-3-Clause-Clear — see LICENSE for details.