Buddy-MLIR is an MLIR-based compiler framework designed for a co-design ecosystem, bridging domain-specific languages (DSLs) to domain-specific architectures (DSAs). It targets engineers and researchers seeking to streamline the development and deployment of specialized hardware accelerators.

How It Works

The project leverages MLIR's infrastructure to build a unified compiler toolchain. It enables the translation of high-level DSLs into optimized code for specific hardware architectures (DSAs), facilitating deep hardware-software co-design. This approach aims to provide flexibility and performance by targeting diverse architectures.

Quick Start & Requirements

• Primary install/run: Requires cloning the repository, initializing Git submodules, and setting up a Python environment (conda recommended).
• Prerequisites: LLVM/MLIR dependencies, flatbuffers-compiler, libflatbuffers-dev, libnuma-dev, and Python 3. The build process involves cmake and ninja for LLVM/MLIR/CLANG and buddy-mlir itself.
• Build Targets: Supports host, RISCV, and optionally NVPTX for NVIDIA GPUs with CUDA runner enabled.
• Environment: PYTHONPATH must be configured to include LLVM/MLIR and buddy-mlir Python packages.
• Python Packages: Wheels for x86_64 and riscv64 can be built using scripts/release_wheel_manylinux.sh.
• Links: Examples documentation is available