gpullama-ncu-analysis

# GPULlama3 Nsight Compute Analysis ## When To Use Use `ncu` only for a specific hot CUDA kernel identified by `nsys`, TornadoVM profiler output, or generated-kernel investigation. Do not run broad `ncu` collection over an entire decode loop without a kernel target. ## Workflow 1. Identify the target kernel name or regex. 2. Use a focused GPULlama3 command that launches the kernel repeatedly. 3. Skip warmup/JIT launches. 4. Start with SpeedOfLight. 5. Add targeted sections based on the first classification. Start with: ```bash ncu --section SpeedOfLight --csv --kernel-name regex:"<kernel>" --launch-skip <N> --launch-count <M> -- <command> ``` Escalate: - memory-bound: `MemoryWorkloadAnalysis` - compute-bound: `ComputeWorkloadAnalysis` - latency/occupancy issue: `LaunchStats`, `Occupancy` - scheduler/warp issue: `WarpStateStats`, `SchedulerStats` ## Classification Use measured SOL metrics: - high memory throughput, low compute throughput: memory-bound - high compute throughput, lower memory throughput: compute-bound - both low with short kernels or low occupancy: latency/occupancy-bound For the target hot kernel, pay special attention to: - DRAM throughput - L1/L2 behavior - achieved occupancy - register count - shared/local memory usage - global load width and instruction count ## Report Include: - target kernel regex - command - raw relevant `ncu` output or exported metrics path - SOL classification - occupancy/register/shared-memory observations - recommended kernel or TornadoVM codegen change Do not present invented or estimated metrics.