What problem does it solve? Developing SHMEM-based communication operators and fused compute-communication operators for Ascend NPU involves many error-prone stages: requirement intake, design documentation, test case generation, kernel coding, compilation debugging, correctness verification, PyTorch binding, and performance benchmarking. This Skill orchestrates that entire pipeline as a gated Phase 0-7 workflow so no stage is skipped or delivered incomplete. ## Core Features & Use Cases - Phase-Gated Orchestration: Drives nine specialized sub-skills (design, testcase-gen, code-gen, compile-debug, correctness-eval, code-review, torch-bind, performance-eval, performance-optim) through strict serial phases with hard gates between each stage. - Environment & Intake Contract: Probes CANN, Docker, SHMEM_REPO, and idle NPU devices before asking users only for missing business options (Torch, performance, auto-optimization). - Fused Compute-Communication Support: Enforces dedicated gates for fused operators (CATLASS BlockMmad + SHMEM CoC), baseline selection priority (aclnn fused > stitched > metric_only), and mandatory 5-round performance optimization loops. - Use Case: A user needs a custom AllToAllV SHMEM operator with PyTorch binding and HCCL baseline comparison. The Skill confirms the environment, generates design.md and case matrix, produces and compiles the kernel, runs full correctness and 8-PE Torch tests, collects S/L-scale bandwidth and latency tables, and iterates optimization rounds until delivery. ## Quick Start Ask the agent to use the shmem-ops-dev skill to develop a new SHMEM communication operator end to end, providing the operator name, target SoC, dtype, and whether Torch binding and performance benchmarking are required.
