Home

CUTracer is an NVBit-based CUDA binary instrumentation tool. It cleanly separates lightweight data collection (instrumentation) from host-side processing (analysis). Typical workflows include per-warp instruction histograms between clock-cycle markers and kernel hang detection.

🎯 Start here:

• Quickstart
• Concepts
• Analyses
• Post-processing: IPC Merge

📚 Configuration and reference:

• Configuration
• Instrumentation Modes
• Outputs and File Formats
• API and Data Structures

🔧 Architecture and development:

• Architecture
• Developer Guide

🧪 Examples and help:

• Examples
• Triton_Proton Integration
• Troubleshooting
• FAQ
• Build, Test, and CI
• Code Formatting

If you only have 5 minutes, read Quickstart and Post-processing: IPC Merge.

Welcome to the CUTracer wiki!