Voyager DNN Accelerator Generator

This repository contains the Voyager DNN accelerator generator.

It is written in SystemC/C++ and can be used for simulation (which allows for fast verification) and for generating Verilog (RTL code). The RTL code is then used for RTL-level (cycle-accurate) simulations and further synthesis of real HW.

CI

We are using the built-in GitLab CI. It triggers whenever someone pushes to a branch.

Runners

• We are using a single GitLab runner on r8cad-tsmc40r.
• They can be configured using /etc/gitlab-runner/config.toml.
• The working directory of the runners is /sim/gitlab-runner/.

Setup

0. Preliminary Step

Install conda with miniforge.

1. Clone the repository

git clone [email protected]:tsmc40r/brainpower/accelerator.git

2. Environment Requirements

Install direnv, and then create your own .envrc file in the project root directory that sets up the conda environment and sources the env.sh file. A sample configuration might look like this:

layout anaconda ./.conda-env
export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
source env.sh

After creating the .envrc file, ensure the conda environment is activated.

• Additional Requirements:
  • You need to have git lfs installed.
  • You need g++ with at least C++17 support.
  • You need a python3 environment with required packages installed (try running the tests and you will be notified of the missing packages).

3. Initialize the submodules

git submodule update --init --recursive

• After initialization, install required packages:
  • In the submodule quantized_training, install with pip install -r requirements.txt
  • Further, run pip install -e .
  • From the root directory, execute pip install quantized_training

Running Examples

Once the setup is complete, you can test the environment with the following command to compile and run ResNet18 through the compiler:

make test/compiler/networks/resnet18/CFLOAT/params.pb

This will generate the folder test/compiler/networks/resnet18 with files such as a parameter list, a log file, and a tensor_files breakdown folder, among others.

Note: You may ignore command-line output messages like:

make: lsb_release: Command not found

To run a regression test for ResNet18, you can execute:

DATATYPE=CFLOAT IC_DIMENSION=16 OC_DIMENSION=16 python run_regression.py --models resnet18 --sims gold_model --num_processes 32

Upon completion, check the regression_results folder for the latest subfolder, where a list of log files for each submodule will be generated. All layers should ideally pass the gold model check. Feel free to adjust parameters and run different simulations as needed.

Test

1. The easiest way to run tests is to check the CI files, .gitlab/ci/sysc.yml and .gitlab/ci/rtl.yml, for SystemC and RTL simulation, respectively. The command to run regression can be found in the script field. The result will be written to the regression_results directory.
2. You can also manually build and run individual configuration and layer. The instruction will be added later.

Structure

• /data: Test data for simulation.
• /lib: External libraries.
• /models: DNN models used for verification.
• /scripts: Various .tcl scripts for RTL generation.
• /src: SystemC accelerator implementation.
• /test: Testing infrastructure, primarily SystemC/C++ files that invoke the accelerator.
• Makefile: Builds the source code.
• run_regression.py: Main script to invoke regression tests.

Adding New Models

To try a new model, you’ll need to modify this script:

• run_compiler.py: Reference existing models to understand the required compilation steps.

Add the network in the following file:

• codegen.mk: Specify the model and set up any required quantization.