choreo-kernel-examples

Optimize CUDA kernels for machine learning by applying warp specialization and pipeline staging patterns to improve resource management and concurrency. This collection provides reference implementations demonstrating tiling strategies and best practices for modern GPU architectures.

Warp specialization in GPU kernel design assigns specific computation tasks to different warps to maximize resource utilization and concurrency. These reference patterns demonstrate how to structure warp-level work distribution for high-performance CUDA implementations.

Implement pipeline staging in CUDA kernels by structuring your kernel design to overlap computation and memory access across different stages. These example implementations in the Choreo DSL provide templates for adapting this pattern to custom GPU compute tasks.

Yes, you can adapt these GPU kernel patterns for different hardware architectures because the examples are optimized for compatibility with modern GPU architectures. They serve as templates to adjust for different problem sizes and resource management requirements.

Use the Choreo DSL to generate and analyze CUDA code by reviewing the example files in the skills directory and executing the provided build scripts. This workflow allows engineers to explore diverse kernel structures and replicate best practices for correctness.

Custom CUDA kernel implementations are needed when standard libraries lack the specific tiling strategies or warp specialization required for tailor-made machine learning compute tasks. These patterns assist GPU kernel engineers and performance analysts targeting specific architectural bottlenecks.