Skip to main content

​
Static Analysis

Real-time CUDA code analysis shows performance metrics directly above your kernels.

​
What It Does

CodeLens automatically analyzes your CUDA code and shows:
  • Execution time and GPU efficiency
  • Register and shared memory usage
  • Occupancy calculations
  • Performance bottlenecks

​
How to Use

1

Write Your CUDA Code

Open a .cu file and write your kernel.
2

Save the File

Press Ctrl+S. CodeLens updates automatically.
3

See Performance Metrics

Look above your kernel for two lines of performance data.
4

Fix Any Issues

If you see warnings, the CodeLens tells you what’s limiting performance.

​
What You See

CodeLens Metrics

​
CodeLens Display

Two lines appear above each kernel: Line 1 - Runtime Performance: 
✅ addVectors: 18.5ms • SM:95.3% • Occ:87.5% • Mem:245.8GB/s
Line 2 - Static Analysis: 
⚠️ Static: Registers: 64 • Shared: 8KB • Max occupancy: 50% (reg-limited)

​
Performance Indicators

  • ✅ Green: Excellent performance
  • ⚠️ Yellow: Can be optimized
  • ❌ Red: Performance issue needs attention

​
Understanding the Metrics

​
Runtime Metrics

  • Time: How fast your kernel runs (ms)
  • SM: GPU utilization percentage
  • Occ: Thread occupancy percentage
  • Mem: Memory bandwidth (GB/s)

​
Static Metrics

  • Registers: Per-thread register count
  • Shared: Shared memory per block
  • Limiting factor: What’s preventing better occupancy

​
Common Issues and Fixes

​
”reg-limited”

Too many registers per thread:
  • Simplify calculations
  • Use shared memory for arrays
  • Reduce local variables

​
”smem-limited”

Too much shared memory:
  • Reduce shared array sizes
  • Use dynamic shared memory
  • Process data in smaller tiles

​
Register Spilling

CodeLens shows spill warnings:
  • Reduce register pressure
  • Split complex kernels
  • Use compiler flags to limit registers

​
Advanced Build Settings

Click “Advanced Build Settings” in editor toolbar to:
  • ✅ Generate .ptx files
  • ✅ Generate .cubin files
  • ✅ Run after build
These options help with deeper optimization analysis.

​
Tips

Best Practices
  • Keep occupancy above 50%
  • Fix red warnings first
  • Save file to update metrics
  • Use build settings for PTX analysis
Remember CodeLens shows estimates. Always benchmark to verify actual performance.

​
Screenshots Needed

  1. CodeLens two-line display - Both performance lines above kernel
  2. Static analysis warnings - Yellow/red indicators with messages
  3. Advanced Build Settings - Checkbox panel with PTX/CUBIN options

​
Next Steps