performance-summary

Generating a performance comparison table requires aggregating result.json files from verify_<op_name> directories. The summary process extracts latency_ms, throughput, memory_usage_mb, and speedup metrics to build the comparison table and detect anomalies.

Metrics extracted for kernel performance comparison include latency_ms, throughput, memory_usage_mb, and speedup. These metrics identify best and worst results and detect abnormal performance points across operators.

Yes, optimization recommendations are automatically generated from operator benchmark data. The performance summary outputs analysis notes and optimization suggestion templates based on the aggregated speedup and latency metrics.

The performance summary requires the verify_<op_name>/result.json file structure for benchmarking workflows. It reads these specific result files to extract performance metrics and generate the final comparison report.

The best way to summarize multiple operators' performance results is using an automated aggregation tool that reads result.json files. It extracts key metrics, generates a comparison table, and outputs ready-to-use analysis notes and optimization recommendations.