DCIR Artifact

This repository contains the artifact for "Bridging Control-Centric and Data-Centric Optimization", submitted to CGO'23.

Requirements

Running and building the Docker container requires an installation of docker and a running instance of the Docker daemon.

The requirements for a manual setup are listed in the Dockerfile.

Setup

There are three options to run the benchmarks (sudo is not always necessary):

Option 1: Pull the docker image

Pull the docker image with:

sudo docker pull berkeates/dcir-cgo23:latest

And run it:

sudo docker run -it --rm berkeates/dcir-cgo23

Option 2: Manually build the docker image

Check out the code

git clone [--recurse-submodules] --depth 1 --shallow-submodules https://github.com/Berke-Ates/dcir-artifact
cd dcir-artifact

Build the image

sudo docker build -t dcir-cgo23 .

Run the container

sudo docker run -it --rm dcir-cgo23

Option 3: Manual Setup

For a manual setup follow the instructions in the Dockerfile on your machine.

Running

To run all benchmarks execute the following script inside the docker container:

./scripts/run_all.sh <Output directory> <Number of repetitions>

The outputs will then be in CSV format in the output directory, organized by figure number or result name. The results are also plotted to PDF files.

If running in a container, you can mount a local folder to view the results outside it. To do so, use the -v flag when running the container, as follows:

sudo docker run -it --rm -v <local folder>:<container folder> berkeates/dcir-cgo23

For example:

$ mkdir results
$ sudo docker run -it --rm -v /path/to/local/results:/home/user/ae berkeates/dcir-cgo23
# ./scripts/run_all.sh ae 10

The raw results from the original paper can be found in the output subdirectory.

To run a single benchmark use the runners in the subdirectories:

• scripts/Polybench: For Polybench benchmarks
• scripts/pytorch: For PyTorch benchmarks
• scripts/snippets: For snippet benchmarks

An example command can be seen in this line:

./scripts/Polybench/dcir.sh ./benchmarks/Polybench/2mm/2mm.c <Output directory> <Number of repetitions>

The subdirectories also contain a run_all.sh, which executes all benchmarks of their group.

File Structure

• benchmarks: Folder containing all benchmarks files
  • Polybench: Folder containing all Polybench benchmarks. The subfolders contain the edited C benchmarks and the generated SDFG files for the DaCe benchmarks.
  • pytorch: Folder containing all PyTorch benchmarks.
  • snippets: Folder containing all snippet benchmarks. The subfolders contain the adjusted C benchmarks, the C benchmarks with timer attached (suffix: -chrono.c) and the generated SDFG files for the DaCe benchmarks.
• dace: Folder containing the DaCe project
• mlir-dace: Folder containing the MLIR-DaCe project
• mlir-hlo: Folder containing the MLIR-HLO project
• output: Raw outputs from the paper
• polybench-comparator: Folder containing a python script to compare Polybench outputs
• Polygeist: Folder containing the Polygeist project
• scripts: Folder containing all scripts to run the benchmarks. Every script prints its usage, if it's called without arguments
  • Polybench: Folder containing Polybench specific run scripts
  • pytorch: Folder containing PyTorch specific run scripts
  • snippets: Folder containing snippets specific run scripts
  • get_sdfg_times.py: Script retrieving the runtime of a specific SDFG run
  • mhlo2std.sh: Script to convert the MHLO dialect to builtin dialects
  • multi_plot.py: Script to plot multiple CSV files
  • opt_sdfg.py: Script to optimize a SDFG file
  • run_all.sh: Script to run all benchmarks and generate the plots
  • single_plot.py: Script to plot a single CSV file
• torch-mlir: Folder containing the Torch-MLIR project
• Dockerfile: Dockerfile to build the image
• README.md: This file