skills.rest
yo-steven
yo-stevenCOMMUNITY
Explore all skills by this author →
GitHub

improve-cutile-kernel-perf

Profile and tune TileGym cuTile kernels with iterative latency optimization.

Updated May 23, 2026
Search Tags:#profiling#kernel optimization#autotuning#cutile#latency tuning#TMA#IR debugging
One-click install
npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill improve-cutile-kernel-perf-yo-steven
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: improve-cutile-kernel-perf
Source: https://github.com/yo-steven/skills-exploration-20260522/tree/main/skills/TileGym/improve-cutile-kernel-perf
Command: npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill improve-cutile-kernel-perf-yo-steven

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

It helps you systematically improve the performance of a cuTile kernel by profiling, diagnosing bottlenecks, and applying one optimization at a time without breaking correctness.

Core Features & Use Cases

  • Iterative profiling → bottleneck diagnosis → targeted tuning to reduce kernel latency (ms) while preserving numerical correctness.
  • Optimization loop with strict guardrails, including a correctness test each iteration and a latency regression cap of 2% vs baseline.
  • IR-level debugging support using TileIR/bytecode/SASS dumping and analysis to explain performance changes after parameter tweaks.
  • Use cases: when asked to optimize or tune a cuTile kernel (e.g., “optimize cutile kernel”, “make kernel faster”) in the TileGym project, especially for tile sizes, occupancy/num_ctas, TMA usage, persistent scheduling, and latency/flush/approxibility knobs.

Quick Start

Start an iterative optimization run to reduce latency for a specified TileGym cuTile kernel by creating a branch, collecting a baseline, then applying one optimization per iteration with correctness and performance re-checks logged in perf_results.md.

Frequently Asked Questions about improve-cutile-kernel-perf

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I optimize a cuTile kernel to reduce runtime latency safely?▼

To optimize a cuTile kernel safely, run an iterative workflow that profiles performance, diagnoses bottlenecks, and applies one tuning change per iteration while validating numerical correctness to prevent regressions.

What is the best way to tune TMA versus gather/scatter memory access in a cuTile kernel?▼

Tuning TMA versus gather/scatter requires profiling kernel latency, diagnosing memory bottlenecks, and iteratively applying targeted tuning changes while preserving numerical correctness to identify the optimal memory access pattern.

How do I debug cuTile performance changes using TileIR or SASS dumps?▼

Debugging cuTile performance changes with TileIR or SASS dumps involves comparing IR-level bytecode outputs after parameter tweaks to explain why specific optimization adjustments altered kernel latency.

Can I autotune cuTile tile sizes and occupancy without breaking numerical correctness?▼

Yes, you can autotune cuTile tile sizes and occupancy without breaking correctness by running a guarded optimization loop that applies one change per iteration and enforces strict correctness validation checks.

What latency regression limit should I set when tuning cuTile kernels?▼

When tuning cuTile kernels, you should set a latency regression cap of 2% versus the baseline benchmark to ensure each iterative optimization change yields measurable performance improvements without degradation.

Related Skills

Explore More Software Engineering Skills→
yo-stevenyo-steven

adding-cutile-kernel

Adds a cuTile GPU kernel operator to TileGm with full integration.

Community
Advanced
NVIDIANVIDIA
796

improve-cutile-kernel-perf

Profile cuTile kernels and apply one optimization per iteration.

Official
Advanced
yo-stevenyo-steven

perf-optimization

Coordinates CUDA Graph, Triton, TileIR, and CuTe DSL performance optimization workflows.

Community
Advanced
tensormuxtensormux
54

Choose Tile Size and Work Partitioning

Determine tile sizes and grid layouts for CUDA or Triton kernels.

Official
Advanced