AMD Kernel Reflection Prompts
- Use after a kernel run/test to drive structured self-review and fixes.
- Load
references/prompt_for_reflection.pyfor the full reflection prompt and guidance.
How to use
- Summarize failures/perf gaps, then feed the reflection prompt to propose patches.
- Follow the checklist: correctness first, then performance and readability.
- Keep AMD-focused advice: wave64 occupancy, LDS/bank conflict avoidance, coalesced and vectorized memory access.
- Output schema should include proposed code changes plus rationale for downstream tools.
References
references/prompt_for_reflection.py: Reflection prompt definitions.