tiny-torch-backend

A minimal, custom MLIR backend pipeline that lowers PyTorch models directly to native OpenBLAS execution.

This project demonstrates how to write a custom MLIR pass to intercept linalg.matmul operations and swap them with optimized C-calls to OpenBLAS, completely bypassing LLVM's default nested loops.

Prerequisites

• LLVM & MLIR
• Torch-MLIR
• OpenBLAS (sudo apt-get install libopenblas-dev)

Project Structure

• custom-opt/: Contains the C++ source for the custom MLIR pass (ConvertMatmulToBlas.cpp).
• compile.sh: The progressive lowering pipeline that translates .mlir to LLVM IR, compiles to an object file, and links the native libraries.
• main.cpp: The C++ runner that initializes the MemRefDescriptor, calls the compiled MLIR function, and handles I/O.

Commands

export PATH=/path/to/torch-mlir/build/bin:$PATH export PYTHONPATH=/path/to/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir:$PYTHONPATH