Project Goal

Overview

CUDA Stencil Benchmark is a framework for systematically generating, validating, and optimizing CUDA kernels for 3D finite-difference stencil computations using LLM-guided code generation.

Primary Objectives

1. LLM-Guided Kernel Generation: Use language models to generate optimized CUDA kernels from task specifications and reference implementations.

2. Correctness-First Validation: Ensure numerical parity between generated CUDA kernels and CPU reference implementations before performance optimization.

3. Systematic Performance Analysis: Characterize kernel performance using roofline methodology to understand memory-bound vs compute-bound behavior.

4. Iterative Optimization Loop: Establish a feedback-driven workflow where kernel generation, correctness testing, and performance analysis inform subsequent optimization attempts.

Target Applications

• 3D finite-difference time-domain (FDTD) stencil computations
• High-performance scientific computing kernels
• GPU acceleration of memory-bound numerical methods
• Research into LLM-assisted code generation for HPC

Success Criteria

• Correctness: Generated kernels maintain numerical parity with CPU reference (tolerance: 1e-5)
• Performance: Achieve significant GPU speedup (target: 5-10×) over CPU baseline
• Reproducibility: Automated testing and benchmarking framework
• Documentation: Clear workflow and methodology documentation

Non-Goals

• General-purpose CUDA code generation (focused on stencil computations)
• PyTorch integration (pure C++/CUDA implementation)
• Production deployment (research/benchmarking framework)