KernelGeneration

KernelGeneration is a repository created to evaluate GPU kernel metrics, including:

• correctness
• performance
• portability

Project Structure (Before Cloning)

kernelGen/
├── README.md
├── install.py
└── operators/
    ├── bf16_layernorm/
    └── bf16_matmul/

• install.py — script to install dependencies, copy operators, and clone TritonBench and KernelLLM
• operators/ — contains GPU kernel operators
• TODO/WIP: run_benchmarks.py — script to run all operators and collect results into a JSON file for analysis.

Evaluated Kernels

• MatMul
• LayerNorm - WIP add KernelLLM
• WIP: GELU

Backend Tested:

• TorchInductor
• Triton
• KernelLLM
• WIP: Mako (based on kernelLLM)

Tested on: TritonBench

Usage

Run the setup and operator installation script:

python install.py