What problem does it solve? SHMEM communication and fused compute-communication operators on Ascend NPU often underperform their HCCL baselines, and manual bottleneck analysis across communication bandwidth, synchronization, core utilization, memory efficiency, and pipeline depth is slow and error-prone. This Skill runs a structured five-round optimization loop that locates bottlenecks, delegates code changes, and verifies each step with correctness and performance re-measurement. ## Core Features & Use Cases - Bottleneck Diagnosis: Reads performance reports and design docs to classify bottlenecks across five categories (communication bandwidth, synchronization overhead, core utilization, memory efficiency, pipeline depth), including algorithm-topology mismatch checks. - Delegated Optimization Rounds: Outputs structured change proposals and delegates design changes to shmem-ops-design, code changes to shmem-ops-code-gen, builds to shmem-ops-compile-debug, and correctness re-tests to shmem-ops-correctness-eval. - Round-by-Round Verification: Each round re-runs performance evaluation via shmem-ops-performance-eval, records step-level metrics (kernel_bus_bandwidth_GBps, e2e/kernel latency, compute utilization), and makes keep/revert decisions with delta-percent comparisons against baseline. - Use Case: After a SHMEM allgather or matmul_allreduce operator passes correctness but reaches only 65% of the HCCL baseline bandwidth, run this Skill to execute five mechanism-optimization rounds (e.g., concurrent peer communication, ping-pong double buffering, barrier-to-signal replacement) and produce a complete performance_report.md. ## Quick Start Ask the agent to run the SHMEM performance optimization loop for the operator that finished Phase 6 performance collection but missed its target, and complete all five optimization rounds with a final performance report.
