Efficient Direct Convolution Using Long SIMD Instructions: Reproducibility Artifact

This document describes the procedures to interact with the reproducibility artifact for the paper entitled "Efficient Direct Convolution Using Long SIMD Instructions," accepted at the Principles and Practice of Parallel Programming (PPoPP) conference in 2023.

This artifact contains a guide for readers to carry out the following activities:

1. Validate the convolution algorithms;

2. Collect experimental data from an SX-Aurora compute node;

3. Reproduce the main performance plots from the paper using the data collected in Step 2.

The artifact also contains Python notebooks to perform the convolution algorithms' performance analysis. The users can validate the main paper's claims, listed below, by subjecting their own experiment's data to these notebooks.

1. The novel algorithms obtain speed-ups on layers where the state-of-the-art direct convolution experience cache conflict misses;

2. The novel algorithms achieve similar performance to the state-of-the-art direct convolution on layers without incidence of cache conflict misses;

3. The novel algorithms provide speed-ups over the state-of-the-art direct convolution on ResNet convolution workloads;