MICRO22 Sparseloop Artifact

This repository provides the evaluation setups for MICRO22 artifact evaluation for the paper Sparseloop: An Analytical Modeling Approach to Sparse Tensor Accelerators. We provide a docker environment and a jupyter notebook for the artifect evlauation. Please contact timeloop-accelergy@mit.edu if there are any questions.

System requirement

---

• Please install docker app

Perform artifact evaluation

Recursively clone repo

git clone --recurse-submodules git@github.com:Accelergy-Project/micro22-sparseloop-artifact.git
cd <cloned repo>
ls docker/

You should see the subdirectories in docker/ populated with actual source code instead of submodule pointers.

Step 0: Prepare your docker-compose.yaml

---

• Create a docker compose file by making a copy of the docker compose tempalte file cp docker-compose.yaml.template docker-compose.yaml
• Examine the instructions in docker-compose.yaml to setup the docker correctly, i.e., setup the correct UID and GID.

Step 1: Get the docker

---

We provide two options for obtaining the docker image. Please choose one of the methods listed below.

Option 1. Pull pre-built image from docker hub

• We provide a pre-built image on dockerhub

  docker-compose pull
  

Option 2. Build docker image from source

• Build the timeloop-accelergy infrastructure docker

  cd docker/accelergy-timeloop-infrastructure
  make
  

• Build the pytorch docker (which uses the timeloop-accelergy infrastructure as a basis)

  cd ../timeloop-accelergy-pytorch
  make
  cd ../../  # go back to root directory of the repo
  

To check if the image is obtained successfully, please do docker image ls and you should see mitdlh/timeloop-accelergy-pytorch with a tag name micro22-artifact listed.

Step 2: Start the docker

---

• Run docker-compose up. You should see the docker being setup.
• This docker uses Jupyter notebooks, and you will see an URL once the docker is up. Please copy and paste the 127.0.0.1 URL to a web browser of your choice to access the workspace.
• If you are experiencing any issues when bringing the page up, please try the tourble shooting notes in docker-compose.yaml.

Step 3: Run experiments in the docker

---

We provide a jupyter notebook for the experiments. Please navigate to workspace/2022.micro.artifact/notebook to run the experiments. Each cell in the notebook provides the background, instructions, and commands to run each evaluation with provided scripts.

For each experiment, we give a very conservative estiamtion of how long the sweeping will take. The input specifications and related scripts can be found in workspace/2022.micro.artifact/evaluation_setups. The easiest way to validate the outputs is to compare the generated figure/table to the figure/table in the paper, but we do provide a ref_outputs folder for each evaluation for more detailed comparison of results if necessary.