ncu-report-skill

To profile CUDA kernels, you use Nsight Compute to collect metrics and diagnose bottlenecks. This generates an evidence-backed optimization report with prioritized recommendations for fixing identified performance issues.

Generating a GPU kernel optimization plan for Blackwell B200 involves profiling with Nsight Compute to extract metrics. The profiling results map signals to actionable fixes across six analysis dimensions to guide your changes.

Yes, you need the ncu_report Python module to parse and compare Nsight Compute reports. It enables Python-based analysis to extract metrics from profiling runs and map those signals to concrete optimization fixes.

To start an end-to-end profiling workflow, create a new run directory under profile/<run_name>/ and build a standalone harness with -lineinfo. You then profile the kernel with Nsight Compute to gather reports for analysis.

Yes, you can compare Nsight Compute profiling runs using the ncu_report module to identify kernel regressions. This Python-based analysis extracts metrics from different runs to map performance signals to required fixes.