Static Analysis

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 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