ascendc-operator-performance-eval

Benchmark Ascend-C operator performance by applying torch_npu.profiler with a fixed schedule of warmup=5 and active=5 to quantify custom versus baseline implementations, generating a Markdown report and per-case JSONL results.

Dual-path comparison for Ascend kernel performance evaluation measures custom operator implementations against native baselines using profiler metrics to quantify performance differences and output deterministic Markdown reports.

Generate JSONL test cases from design.md by deriving supported dtypes and shapes, ensuring full coverage across execution modes to validate the operator performance test suite before running the profiler workflow.

Yes, the torch_npu profiler supports fixed warmup and active iterations for benchmarking; this workflow enforces a specific schedule with warmup=5 and active=5 to ensure deterministic performance measurement across profiler runs.

Ascend operator performance statistics use op_statistic.csv metrics derived from torch_npu.profiler output, providing quantifiable data to compare custom and baseline implementations within the final Markdown report.