tune

What problem does it solve?

This guide provides a repeatable methodology to measure, analyze, and tune GPU kernel performance in TileOPs, ensuring reported metrics reflect GPU-only execution and that autotune trial runs do not contaminate profiler traces.

Core Features & Use Cases

• Authoritative Benchmarking: Run the benchmarks/ops bench_xxx scripts to obtain median GPU-only latencies for fair comparisons between kernel variants and configs.
• Clean Tracing Workflow: Disable autotune and fix the best config before profiling with nsys so aggregated kernel stats reflect steady-state behavior.
• Deep Metric Analysis: Use ncu (Nsight Compute) for per-metric inspection (memory throughput, L1/L2/HBM hit rates, occupancy, stall reasons) and apply practical TMPDIR and launch-skip/launch-count workarounds.
• Tuning SOP & PR Rules: Standardized workflow from benchmarking to profiling to PR submission, including required performance tables and autotune config disclosures.

Quick Start

Run the benchmark to collect median GPU-only latency, disable autotune to fix the config, and then run nsys or ncu to capture clean kernel traces for analysis.