kernel-trace-analysis

Analyze ATT traces to profile GPU kernels and identify stall hotspots.

260|104|Updated Nov 23, 2025

Search Tags:#profiling #gpu #kernel #mfma #hotspot-analysis #att-trace

Ready to run

npx skills add https://github.com/ROCm/FlyDSL --skill kernel-trace-analysis-rocm

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: kernel-trace-analysis
Source: https://github.com/ROCm/FlyDSL/tree/main/.claude/skills/kernel-trace-analysis
Command: npx skills add https://github.com/ROCm/FlyDSL --skill kernel-trace-analysis-rocm

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve?

Profile GPU kernels to identify stall hotspots from ATT traces and generate actionable optimizations.

Core Features & Use Cases

Hotspot analysis using rocprof ATT traces and hotspot_analyzer.py
Map stalls to source lines with code context for targeted optimizations
Generate an actionable optimization plan focusing on VMEM, LGKM, MFMA stalls

Quick Start

Run the kernel-trace-analysis tool on a target kernel or dispatch directory to produce a hotspot report and optimization plan.

Frequently Asked Questions about kernel-trace-analysis

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

FAQPage Schema

How do I profile GPU kernels to identify stall hotspots from ATT traces?▼

To profile GPU kernels for stall hotspots, parse rocprofv3 ATT trace output using a dedicated analyzer script that aggregates stalls by source location and generates an actionable optimization plan.

What types of stall hotspots can I map back to source lines during GPU profiling?▼

GPU profiling maps VMEM, LGKM, and MFMA stalls back to source lines. It aggregates stalls by source location and outputs top hotspots with contextual code details for targeted optimizations.

Do I need existing dispatch directories to analyze GPU kernel stalls?▼

You need either rocprofv3 ATT trace output or existing dispatch directories containing code.json and snapshots.json files to parse and aggregate stalls by source location for hotspot analysis.

What is the best way to optimize GPU performance when dealing with MFMA stalls?▼

The best way to optimize MFMA stalls is running hotspot analysis on ATT traces to map stalls to source lines, producing an actionable optimization plan focusing on specific stall categories.

How does hotspot analysis work when parsing GPU kernel trace files?▼

Hotspot analysis works by parsing code.json and snapshots.json files to aggregate stalls by source location, then outputting top hotspots with contextual source code details for optimization.