2. Environment Setup

This chapter describes the development environment configuration. The code is compiled and run in docker.

2.1. Code Download

You can clone the source code of tpu-mlir from github, it needs to be compiled in docker. For specific steps, please refer to the following.

2.2. Docker Configuration

You can get the required image file tpuc_dev_v3.4.tar.gz from the SDK package:

$ docker load -i tpuc_dev_v3.4.tar.gz

If you are using docker for the first time, you can execute the following commands to install and configure it (only for the first time):

$ sudo apt install docker.io
$ sudo systemctl start docker
$ sudo systemctl enable docker
$ sudo groupadd docker
$ sudo usermod -aG docker $USER
$ newgrp docker

Make sure the installation package is in the current directory, and then create a container in the current directory as follows:

$ docker run --privileged --name myname -v $PWD:/workspace -it tpuc_dev:v3.4
# "myname" is just an example, you can use any name you want
# use --privileged to get root permission, if you don't need root permission, please remove this parameter

Note that the path of the TPU-MLIR project in docker should be /workspace/tpu-mlir

2.3. ModelZoo (Optional)

TPU-MLIR comes with the yolov5s model. If you want to run other models, you need to download them from ModelZoo provided in the SDK package.

After downloading, put it in the same directory as tpu-mlir. The path in docker should be /workspace/model-zoo

2.4. Compilation

In the docker container, the code is compiled as follows:

$ cd tpu-mlir
$ source ./envsetup.sh
$ ./build.sh

Regression validation:

# This project contains the yolov5s.onnx model, which can be used directly for validation
$ pushd regression
$ python run_model.py yolov5s
$ popd

You can validate more networks with model-zoo, but the whole regression takes a long time:

# The running time is very long, so it is not necessary
$ pushd regression
$ ./run_all.sh
$ popd

2.5. Code Development

To facilitate code readability and development, it is recommended to use VSCode as the editor. In VSCode, install the following extensions:

C/C++ Intellisense : Provides intelligent suggestions, code navigation, and formatting for C++ code.
GitLens : Assists with Git version control and code review.
Python : Provides intelligent suggestions and code navigation for Python.
yapf : Formats Python code.
shell-format : Formats shell scripts.
Remote-SSH : Enables remote connections to code on a server (essential when code is not local).

After writing your code, right-click and select “Format Document” to ensure a consistent code style.

Since TPU-MLIR uses llvm-project and relies heavily on its headers and libraries, it is recommended to install llvm-project for improved code navigation. Follow these steps:

At the same level as the TPU-MLIR repository, create a third-party directory and clone llvm-project into it:
```
$ mkdir third-party
$ cd third-party
$ git clone git@github.com:llvm/llvm-project.git
```

Inside the TPU-MLIR Docker environment, build llvm-project (you may be prompted to install missing components during the build—follow the prompts to install them):

$ cd llvm-project
$ mkdir build && cd build
# If prompted for missing components (e.g., nanobind), install them:
#    pip3 install nanobind
$ cmake -G Ninja ../llvm \
    -DLLVM_ENABLE_PROJECTS="mlir" \
    -DLLVM_INSTALL_UTILS=ON \
    -DLLVM_TARGETS_TO_BUILD="" \
    -DLLVM_ENABLE_ASSERTIONS=ON \
    -DMLIR_INCLUDE_TESTS=OFF \
    -DLLVM_INSTALL_GTEST=ON \
    -DMLIR_ENABLE_BINDINGS_PYTHON=ON \
    -DCMAKE_BUILD_TYPE=DEBUG \
    -DCMAKE_INSTALL_PREFIX=../install \
    -DCMAKE_C_COMPILER=clang \
    -DCMAKE_CXX_COMPILER=clang++ \
    -DLLVM_ENABLE_LLD=ON
$ cmake --build . --target install

After installation, you can link code navigation to the llvm-project sources.