Profiling investigates where simulation time and hardware resources are spent. Select a workload and tools in the repository's config file, then run make profile:

# config: example GPU profile settings
PROFILE_TOOLS="nsys ncu"
PROFILE_BACKEND="cuda"
PROFILE_PERF_SIZE=1024
PROFILE_PERF_STEPS=200

These settings use Nsight Systems to trace API calls and GPU work, and Nsight Compute to inspect kernel metrics. Results are saved under results/profiles/.

For CPU work, select tools such as perf-stat, perf-record, callgrind, or cachegrind. The profile build includes symbols, frame pointers, and CUDA line information to relate observations to code.

Profiling instrumentation changes execution cost, particularly under Valgrind. Use benchmark timings for performance comparisons and profiles to explain them.

Source: config.example.