analyze-kernel-bottleneck - Analyze CUDA kernel bottlenecks for optimization
Classifies a CUDA kernel as compute-, memory-, or latency-bound using performance, roofline, occupancy, tile-ratio, and SASS analysis.
Tags
Updated: 2026-10-01Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Measure baseline kernel performance
- Classify roofline bottlenecks
- Calculate GPU occupancy
- Compute tile load ratios
- Inspect SASS instruction mix
- Analyze instruction stall codes
- Check shared-memory occupancy cliffs
- Select optimization strategies
Inputs
- Compiled CUDA kernel
- CUDA kernel source
- Kernel build command
- CUDA benchmark harness
- Problem dimensions
- Target GPU architecture
- Expected peak utilization
- Prior profiling data
Outputs
- Baseline timing and throughput metrics
- Bottleneck classification
- Occupancy analysis table
- Compute-to-load ratio
- cp.async recommendation
- SASS instruction count table
- Stall code summary
- Optimization decision matrix
Requirements
- CUDA toolkit with nvcc
- cuobjdump utility
- CUDA-capable NVIDIA GPU
- CUDA event timing support
- Permission to compile and run kernels