---

Running models using IREE

Dependencies

• IREE's ONNX importer (The option [onnx] might not work when running pip on ZSH)
• IREE compiler
• IREE runtime

Running the models

The script iree.sh is responsible to convert a model from ONNX to torch MLIR, and compile and run this converted model using the IREE compiler. The script assumes the IREE compiler is installed inside a Python environment located in a folder called venv in this repository. If this is not the case, you must:

• If IREE is installed inside a Python environment, set the variable PYTHON_VENV_PATH
• If IREE is built from source out of the $PATH environment variable, set the variables IREE_IMPORT_ONNX, IREE_COMPILE and IREE_RUN_MODULE

The script runs the first model by default, but you can change it by setting variable MODEL_IDX to some value between the available models before running the script.

---

Running models using MLIR

Dependencies

• torch-mlir
• MLIR

Running the models

The running pipeline consists of 3 scripts:

• create_linalg.sh: converts the ONNX model to torch MLIR, then lower it to linalg dialect.
• lower.sh: run the lowering pipeline for nasbench models, which consists in adding a main function and run passes to lower from linalg to llvm dialect.
• run.sh: execute the lowered model.

The run script assumes the MLIR runner binary is in the $PATH environment variable. If this is not the case, define the variable MLIR_RUNNER to its respective path. Also, the script assumes that some MLIR libraries are within the /usr/lib/llvm/lib directory. Set the variables MLIR_UTILS and MLIR_C_UTILS to the path to the files libmlir_runner_utils.so and libmlir_c_runner_utils.so, respectively, if they are in a different path.

---

Generating memory allocation instances and outputs

Dependencies

• torch-mlir
• MLIR

Generating instances and outputs

The generation of instances and outputs for memory allocation is managed with the following scripts:

• memory_allocation_instantiatior.sh: generate instances for the memory allocation problem based on the nasbench models.
• memory_allocation_instrumentation.sh: generate the output for the memory allocation problem by instrumenting malloc and free calls, in order to output the address and size of allocated buffers. There are also instrumented versions for the lower and run scripts.

---

Generating models from Nasbench

Dependencies (Python libraries)

• tensorflow
• tqdm
• tf2onnx
• onnx
• gurobipy

Generating the models

The Nasbench has over 423k benchmarks but, due to space constraints, only 100 were included in this repository. The script NASBenchConvert.py can generate the remaining models, although it will take some time to run (ONNX files are already included on the .gitignore). However, it's necessary to extract the file nasbench_full.json.tar.gz first.

Development

Running clang-format on commit

To have clang-format run on each commit, copy the bash script to the correct folder:

mkdir -p .git/hooks
cp .githooks/pre-commit .git/hooks/pre-commit

Building

• Python Virtual Environment

python3.11 -m venv venv

source venv/bin/activate

pip install tensorflow tqdm tf2onnx onnx gurobipy

pip install --pre torch-mlir torchvision \
  --extra-index-url https://download.pytorch.org/whl/nightly/cpu -f \
  https://github.com/llvm/torch-mlir-release/releases/expanded_assets/dev-wheels

• C++, LLVM, MLIR
  It's possible compiler your own MLIR, but the packages on apt repository will also work (on Ubuntu for example)

# Download and run the LLVM installation script (clang-22 not in Ubuntu repos)
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 22

# Install MLIR development packages for version 22
sudo apt install libmlir-22-dev mlir-22-tools

# Point CMake to Clang 22
cmake -B build -DCMAKE_C_COMPILER=/usr/lib/llvm-22/bin/clang \
               -DCMAKE_CXX_COMPILER=/usr/lib/llvm-22/bin/clang++ \
               -DMLIR_DIR=/usr/lib/llvm-22/lib/cmake/mlir