Tensor Compiler

Tensor Compiler is a tool that converts models in the ONNX format into an intermediate representation (IR) using an MLIR dialect.

Currently, the project supports core neural network operations including:

• Conv
• Relu
• Add
• Mul
• MatMul
• Gemm
• Transpose

---

Requirements

• LLVM + MLIR version 20 or newer
  • either built from source, or
  • installed system-wide

---

Build Instructions

1. Create and activate a virtual environment

python -m venv venv
source venv/bin/activate

2. Install Conan

pip install conan

3. Install dependencies

conan install . --build=missing -s build_type=Release

4. Configure CMake

cmake --preset conan-release \
  -DMLIR_DIR=/path/to/llvm-project/build/lib/cmake/mlir \
  -DLLVM_DIR=/path/to/llvm-project/build/lib/cmake/llvm

❗ Important: You must provide paths to your built LLVM and MLIR installations:

• MLIR_DIR — path to the mlir directory inside the LLVM build
• LLVM_DIR — path to the llvm directory inside the LLVM build

---

5. Build the project

cmake --build --preset conan-release

---

6. Run tests

ctest --test-dir build/Release

---

Usage

After building, you can pass an ONNX model to the compiler to generate MLIR IR:

./build/Release/tensor-compiler <model.onnx>

By default, the computation graph is printed to stdout.

---

Options

--graph-dot-dump

Export the computation graph to graph.dot.

--high-dialect-dump

Print the generated high-level MLIR dialect.

---

Maintainer

Developed and maintained by ask0later. Feel free to open issues and contribute.

---