Related projects and thanks

This document lists open-source projects and ecosystems closely related to AscendNPU IR and thanks the LLVM/MLIR and other communities.

MLIR

MLIR originates from the LLVM community and provides reusable, extensible compiler infrastructure. AscendNPU IR is built on MLIR. We thank all developers and contributors in the LLVM/MLIR community. AscendNPU IR benefits from MLIR in these ways:

• Modular design: Define IR at different abstraction levels for progressive lowering.

• Reuse of infrastructure: Parsing, transformation, optimization, and code generation from MLIR.

• Ecosystem interoperability: Extend MLIR dialects to interact and convert with other dialects (e.g. TensorFlow, PyTorch IR) and integrate with upper-level frameworks.

Triton-Ascend

Triton-Ascend brings Triton programming to Ascend, so Triton code runs efficiently on Ascend hardware. AscendNPU IR serves as the compilation backend for Triton, enabling developers to write high-performance kernels for Ascend NPU with familiar Triton syntax and programming model and lowering the barrier for Python developers.

TileLang-Ascend

TileLang is a domain-specific language for tensor computation; TileLang-Ascend is its Ascend-oriented version. By using AscendNPU IR as the backend, TileLang-Ascend leverages AscendNPU IR’s Ascend-aware optimizations to generate high-performance Ascend operators.