SMAUG: Simulating Machine Learning Accelerators Using gem5-Aladdin

SMAUG is a deep learning framework that enables end-to-end simulation of DL models on custom SoCs with a variety of hardware accelerators. SMAUG is designed to enable DNN researchers to rapidly evaluate different accelerator and SoC designs and perform hardware-software co-design. Simulation is powered by the gem5-Aladdin SoC simulator, allowing users to easily write new hardware accelerators and integrate them into SMAUG for testing and exploration.

SMAUG provides stable Python and C++ APIs, allowing users to work at varying levels of abstraction. For example, researchers can:

• Focus on building and evaluating models by working with the high-level Python API.
• Focus on evaluating different SoC configurations through gem5.
• Evaluate new tiling strategies and configurations through th C++ tiling optimizers.
• Build new accelerators and integrate them into SMAUG.

If you are using SMAUG in research, we would appreciate a reference to:

Sam (Likun) Xi, Yuan Yao, Kshitij Bhardwaj, Paul Whatmough, Gu-Yeon Wei, and David Brooks. SMAUG: End-to-End Full-Stack Simulation Infrastructure for Deep Learning Workloads. ACM Transactions on Architecture and Code Optimization, 17, 4, Article 39 (November 2020). https://doi.org/10.1145/3424669.

API documentation and tutorials are available at https://harvard-acc.github.io/smaug_docs.

Installation

SMAUG requires use of a Docker image, available on Docker Hub here. Users who cannot use Docker can follow the commands in the Dockerfile to set up a local environment. However, due to our limited resources, we can only offer support to Docker users.

To install SMAUG, first install Docker, then pull the Docker image:

docker pull xyzsam/smaug:latest

Then, run the following command to create a Docker volume that hosts your workspace and start the container. The local volume (aka smaug-workspace below) will store all your source code, local changes, and build artifacts, so that you can start/stop the Docker container without losing any of your work.

docker run -it --rm --mount source=smaug-workspace,target=/workspace xyzsam/smaug:latest

The Docker container already contains all the source code repositories you need, but they are probably out of date. You will need to update them. Go into your /workspace directory and run the following commands:

cd gem5-aladdin && git pull origin master && git submodule update --init --recursive && cd ..
cd LLVM-Tracer && git pull origin master && cd ..
cd smaug && git pull origin master && git submodule update --init --recursive && cd ..

Building

We need to build gem5-Aladdin. The -j parameter controls how many CPUs are used. Increase this value to speed up the build, but keep in mind that you may run out of memory before you run out of CPUs. Running out of memory or disk space can cause mysterious build failures.

cd /workspace/gem5-aladdin
python2.7 `which scons` build/X86/gem5.opt PROTOCOL=MESI_Two_Level_aladdin -j2

And then SMAUG:

cd /workspace/smaug
make all -j8

You are now ready to work with SMAUG. Read on to learn how to run your first model.

Running your first model

Run the 4-layer Minerva model with our NVDLA-like backend codenamed SMV:

cd /workspace/smaug/experiments/sims/smv/tests/minerva
sh run.sh

Tutorials and APIs

Documentation for writing new models and building on top of SMAUG can be found at https://harvard-acc.github.io/smaug_docs.