Setup

The contents of the repository can be directly deployed into an automatically configured Docker container through Visual Studio Code. To achieve this, the following requirements should be met:

• Installed (and running) Docker engine.

• Visual Studio Code (VS Code).

• Microsoft's VS Code Remote Development extension, available through the built-in extension marketplace.

With these requirements met, the following instructions can be followed to launch the container and attach VS Code to it:

1. Open VS Code and navigate to the Remote Explorer panel in the toolbar on the left.

2. In the Devvolumes panel, click on Clone Repository in Container Volume. If this option is not present, check if the Docker daemon is running and re-open VS Code before resuming from this step.

3. Paste https://github.com/phschaad/sc22_artifact.git into the input field and hit Enter.

4. Wait for the container to fully start.

5. Once an additional SDFG Optimization icon appears in the left toolbar, the container has been fully started and initialized.

A more in-depth step-by-step guide on how to launch Development Containers in VS Code can be found here.

Pre-Built VM

A pre-built VM Image is provided at https://doi.org/10.5281/zenodo.6424094 in cases where the approach described above can not be followed. The user is named sc22, and the user password is supercomputing. For this approach, import the VM image, and once the machine is booted, simply open VS Code.

NOTE: It is strongly discouraged to use the pre-built VM due to decreased performance and stability. We instead recommend following the steps outlined above.

Running

Analysis via Local View

The local view can be started with the shell script localview.sh through the integrated terminal in VS Code, connected to the container.

This starts a webserver on localhost:8080, accessible outside the container, on which the local view resides. A dropdown in the top navigation bar allows switching between the different applications mentioned or depicted in the paper using the local view.

HDIFF experiments

Horizontal diffusion can be run using the shell script run_hdiff.sh through the integrated terminal in VS Code, connected to the container.

This runs and times the following experiments using the dataset used in the original NPBench publication (paper dataset from NPBench, K=160, I=J=256):

1. Baseline: Unoptimized NumPy version provided by NPBench.

2. NPBench Best CPU Auto-Opt: Best automatically optimized version for CPUs provided by NPBench.

3. Reshaped / Reordered: A hand-tuned version, where data layouts and loop orders were optimized.

4. Aligned: A further optimized version, where post-padding was introduced to improve alignment.

Each experiment run is executed 100 times, and the median time is reported.

BERT experiments

The BERT encoder can be run using the shell script run_bert.sh through the integrated terminal in VS Code, connected to the container.

This runs and times the following experiments:

1. Baseline: Unoptimized, naive NumPy implementation.

2. Fused: Optimized version with some loop fusion.

3. Fully Fused ('Final'): Fastest obtained version with maximal degree of loop fusion.

Each experiment run is executed 10 times, and the median time is reported.

File Structure

• README.md: This file.

• .devcontainer: VS Code development container settings, including Dockerfile.

• .vscode: VS Code settings and configurations.

• BERT: Contains three versions of the BERT encoder.

  • 01_original.sdfg: Unoptimized, naive NumPy implementation.

  • 02_fused.sdfg: Optimized version with some loop fusion.

  • 03_final.sdfg: Fastest obtained version with maximal degree of loop fusion.

• HDIFF: Contains four versions of the horizontal diffusion (hdiff) program.

  • 01_original.sdfg: Unoptimized, naive NumPy implementation from NPBench.

  • 02_npbench_best_cpu.sdfg: Best auto-optimized CPU version obtained from NPBench.

  • 03_reshape_reorder.sdfg: Optimized version with optimized data layout and loop order.

  • 04_aligned.sdfg: Further optimized version with post-padding to improve alignment.

• local_view: Source code for the local analysis view.

• npbench: The NPBench version used for the experiments, as a git submodule.

• _*.py: Helper scripts to run the experiments, don't execute directly.

• localview.sh: Shell script to start the local analysis view in a webserver on port 8080.

• run_bert.sh: Shell script to run all BERT encoder experiments.

• run_hdiff.sh: Shell script to run all HDIFF experiments.