CUDA Tile Open Sourced

4 months ago

CUDA Tile IR is an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization.
It focuses on tile-based computation patterns and optimizations for NVIDIA tensor core units.
The project provides a comprehensive ecosystem for expressing and optimizing tiled computations for NVIDIA GPUs.
Components include CUDA Tile Dialect, Python Bindings, Bytecode, and a Conformance Test Suite.
Requirements include CMake 3.20.0+, C++17 compiler, Python 3.6+, MLIR/LLVM sources, and Ninja build system.
Build options include automatic download from GitHub, use of local LLVM sources, or pre-built LLVM libraries.
Python bindings are available for programmatic IR manipulation.
Testing is enabled by default using LLVM's lit testing infrastructure.
Integration into projects can be done via pre-built libraries or building from source.
An example demonstrates compiling and running a simple Tile IR kernel.
External contributions are not currently accepted, but feedback via GitHub Issues is encouraged.
Licensed under Apache License v2.0 with LLVM Exceptions.

Hasty Briefsbeta