Contents

• 1. Preparing An Application For Profiling
  • 1.1. Focused Profiling
  • 1.2. Marking Regions of CPU Activity
  • 1.3. Naming CPU and CUDA Resources
  • 1.4. Flush Profile Data
  • 1.5. Profiling CUDA Fortran Applications
• 2. ​Visual Profiler
  • 2.1. Getting Started
    • 2.1.1. Setting up Java Runtime Environment
    • 2.1.2. Installing JRE
    • 2.1.3. Modify Your Application For Profiling
    • 2.1.4. Creating a Session
    • 2.1.5. Analyzing Your Application
    • 2.1.6. Exploring the Timeline
    • 2.1.7. Looking at the Details
    • 2.1.8. Improve Loading of Large Profiles
  • 2.2. Sessions
    • 2.2.1. Executable Session
    • 2.2.2. Import Session
      • 2.2.2.1. Import Single-Process nvprof Session
      • 2.2.2.2. Import Multi-Process nvprof Session
      • 2.2.2.3. Import Command-Line Profiler Session
  • 2.3. Application Requirements
  • 2.4. Visual Profiler Views
    • 2.4.1. Timeline View
      • 2.4.1.1. Timeline Controls
      • 2.4.1.2. Navigating the Timeline
      • 2.4.1.3. Timeline Refreshing
      • 2.4.1.4. Dependency Analysis Controls
    • 2.4.2. Analysis View
      • 2.4.2.1. Guided Application Analysis
      • 2.4.2.2. Unguided Application Analysis
      • 2.4.2.3. PC Sampling View
      • 2.4.2.4. Memory Statistics
      • 2.4.2.5. NVLink view
    • 2.4.3. Source-Disassembly View
    • 2.4.4. GPU Details View
    • 2.4.5. CPU Details View
    • 2.4.6. OpenACC Details View
    • 2.4.7. OpenMP Details View
    • 2.4.8. Properties View
    • 2.4.9. Console View
    • 2.4.10. Settings View
    • 2.4.11. CPU Source View
  • 2.5. Customizing the Profiler
    • 2.5.1. Resizing a View
    • 2.5.2. Reordering a View
    • 2.5.3. Moving a View
    • 2.5.4. Undocking a View
    • 2.5.5. Opening and Closing a View
  • 2.6. Command Line Arguments
• 3. ​nvprof
  • 3.1. Command Line Options
    • 3.1.1. CUDA Profiling Options
    • 3.1.2. CPU Profiling Options
    • 3.1.3. Print Options
    • 3.1.4. IO Options
  • 3.2. Profiling Modes
    • 3.2.1. Summary Mode
    • 3.2.2. GPU-Trace and API-Trace Modes
    • 3.2.3. Event/metric Summary Mode
    • 3.2.4. Event/metric Trace Mode
  • 3.3. Profiling Controls
    • 3.3.1. Timeout
    • 3.3.2. Concurrent Kernels
    • 3.3.3. Profiling Scope
    • 3.3.4. Multiprocess Profiling
    • 3.3.5. System Profiling
    • 3.3.6. Unified Memory Profiling
    • 3.3.7. CPU Thread Tracing
  • 3.4. Output
    • 3.4.1. Adjust Units
    • 3.4.2. CSV
    • 3.4.3. Export/Import
    • 3.4.4. Demangling
    • 3.4.5. Redirecting Output
    • 3.4.6. Dependency Analysis
  • 3.5. CPU Sampling
    • 3.5.1. CPU Sampling Limitations
  • 3.6. OpenACC
    • 3.6.1. OpenACC Options
    • 3.6.2. OpenACC Summary Modes
  • 3.7. OpenMP
    • 3.7.1. OpenMP Options
• 4. Remote Profiling
  • 4.1. Remote Profiling With Visual Profiler
    • 4.1.1. One-hop remote profiling
  • 4.2. Remote Profiling With nvprof
    • 4.2.1. Collect Data On Remote System
    • 4.2.2. View And Analyze Data
• 5. NVIDIA Tools Extension
  • 5.1. NVTX API Overview
  • 5.2. NVTX API Events
    • 5.2.1. NVTX Markers
    • 5.2.2. NVTX Range Start/Stop
    • 5.2.3. NVTX Range Push/Pop
    • 5.2.4. Event Attributes Structure
    • 5.2.5. NVTX Synchronization Markers
  • 5.3. NVTX Domains
  • 5.4. NVTX Resource Naming
  • 5.5. NVTX String Registration
• 6. MPI Profiling
  • 6.1. Automatic MPI Annotation with NVTX
  • 6.2. Manual MPI Profiling
  • 6.3. Further Reading
• 7. MPS Profiling
  • 7.1. MPS profiling with Visual Profiler
  • 7.2. MPS profiling with nvprof
  • 7.3. Viewing nvprof MPS timeline in Visual Profiler
• 8. Dependency Analysis
  • 8.1. Background
  • 8.2. Metrics
  • 8.3. Support
  • 8.4. Limitations
• 9. Metrics Reference
  • 9.1. Metrics for Capability 5.x
  • 9.2. Metrics for Capability 6.x
  • 9.3. Metrics for Capability 7.x
• 10. Warp State
• 11. Migrating to Nsight Tools from Visual Profiler and nvprof
• 12. Profiler Known Issues
• 13. Changelog
• 14. Notices
  • 14.1. Notice
  • 14.2. OpenCL
  • 14.3. Trademarks