← All resources

CUDA Tile IR

Auto-discovered

Library / Codebase Advanced

NVIDIA

Visit resource ↗

Description

CUDA Tile IR is an NVIDIA open-source GitHub repository for an MLIR-based intermediate representation and compiler stack aimed at tile-based CUDA kernel optimization on NVIDIA GPUs, especially tensor-core workloads. The release is aligned with CUDA Toolkit 13.1.

The README describes a CUDA Tile dialect with first-class tile operations and types, Python bindings for building and transforming IR, a bytecode format for serialization, and a conformance test suite. It covers tiling patterns, memory-hierarchy management, and GPU-specific transforms driven by the published CUDA Tile IR specification.

It is for advanced CUDA and compiler engineers who already know MLIR or LLVM-style IR work and want abstractions for tiled tensor-core kernels rather than hand-written low-level CUDA alone.

The project is free on GitHub with CMake/Ninja build instructions, optional Python bindings, lit-based tests, and links to NVIDIA CUDA Tile documentation for the formal dialect semantics.

It is useful when you need a shared IR and tooling layer for expressing tiled GPU computations instead of reinventing dialect and bytecode plumbing.