Provide CUDA kernel design and performance optimization best practices.
npx skills add https://github.com/ar4mirez/samuel --skill cuda-guide
Please help me install this Agent Skill. Skill: cuda-guide Source: https://github.com/ar4mirez/samuel/tree/main/template/.claude/skills/cuda-guide Command: npx skills add https://github.com/ar4mirez/samuel --skill cuda-guide
This Skill provides essential guardrails, patterns, and best practices for developing high-performance CUDA and GPU-accelerated applications, ensuring efficient memory usage and maximum computational throughput.
Use the cuda-guide skill to review best practices for optimizing CUDA kernel performance.
High-intent search queries and answers about installing and using this skill.
To optimize CUDA kernel performance, you must address memory hierarchy, coalesced access, occupancy, and synchronization. This ensures efficient memory usage and maximum computational throughput for GPU-accelerated applications.
Implementing robust CUDA API error checking and kernel launch validation is the best way to manage memory and handle errors in .cu files. Following established patterns prevents common performance pitfalls and ensures correct execution.
Nsight Compute helps with GPU performance tuning by providing measurement tools to analyze kernel execution. Using these profiling tools alongside best practices for memory access allows developers to identify bottlenecks and maximize hardware occupancy.
You can use these parallel computing patterns for general GPU code, not only deep learning models. The guardrails address fundamental kernel design and synchronization principles applicable to any high-performance CUDA application requiring maximum throughput.
Low occupancy and poor memory throughput in CUDA applications usually stem from uncoalesced memory access and improper synchronization. Applying specific kernel design patterns and understanding the memory hierarchy resolves these performance bottlenecks.
pontschoNavigate CUDA symbols and diagnose issues in .cu/.cuh files via clangd.
m0atAnalyze CUDA kernel performance bottlenecks with Nsight Systems and Nsight Compute.
ovachieverApply PyTorch FSDP for sharded, mixed-precision distributed training of large models.
UtsavBalar1231Assist Linux kernel development for device drivers, modules, and subsystems.
