improve-cutile-kernel-perf

You can optimize cuTile kernel performance by profiling the baseline, analyzing memory and compute bottlenecks, applying one optimization per iteration, and logging incremental improvements in perf_results.md.

The best way to identify cuTile kernel bottlenecks is to profile the baseline and analyze memory, compute, and scheduling patterns to pinpoint specific performance limitations before applying tuning changes.

You track kernel optimization results by logging each iteration's performance metrics and numerical correctness verification in a reproducible perf_results.md file to maintain a continuous improvement history.

No, you should apply one optimization from the playbook per iteration to the cuTile kernel, verifying numerical correctness and checking for regressions before proceeding to the next change.

You need reference materials such as an optimization playbook, a performance knobs catalog, and IR/DUMP guides to direct cuTile kernel tuning changes and avoid performance regressions.