All Tools
M
Dev ToolsFreeOpen Source
MLIR
Multi-level IR and compiler infrastructure inside the LLVM project
Apache-2.0
ABOUT
ML and hardware compilers used to invent a new IR for every framework and chip, so optimizations could not be shared. MLIR gives a common multi-level IR with dialects and progressive lowering so tensor graphs, GPU kernels, and custom accelerators reuse passes instead of rewriting a compiler from scratch.
INTEGRATION GUIDE
1. Lower tensor and linalg dialects into GPU or CPU kernels for ML runtimes
2. Build a domain-specific compiler with custom MLIR dialects and passes
3. Target AI accelerators by adding a hardware dialect on top of existing IRs
4. Share optimization passes across PyTorch, JAX, and TensorFlow compiler stacks
TAGS
compilerllvmmlirirgpuml-compilerhardwareopen-source