gpu-mode/reference-kernels

What it solves

This repository provides a set of reference implementations for GPU kernel competitions. It serves as the ground truth for evaluating the performance and correctness of custom GPU kernels written by participants in competitions hosted by KernelBot.

How it works

The project organizes kernels into problem sets (such as linear algebra or specific hardware competitions like NVIDIA Blackwell or AMD). Each problem includes a PyTorch reference implementation (reference.py), a problem specification (task.yml), and a schema for inputs and outputs (task.py). These files allow the system to generate test cases and verify that the user-submitted kernels produce the same results as the reference implementations.

Who it’s for

GPU programmers and researchers who participate in KernelBot competitions or students in parallel programming courses (e.g., at Stanford and Princeton).

Highlights

• Reference kernels for high-stakes competitions (including AMD and NVIDIA hardware challenges).
• Support for university parallel programming assignments.
• Standardized problem specifications for automated testing and evaluation.