What problem does it solve? Building correct SHMEM (symmetric memory) operators requires rigorous testing across multiple PE counts, data scales, dtypes, and boundary conditions, but manually designing case matrices, golden references, and precision checkers is error-prone and often skipped until after implementation. This Skill converts the correctness contract in a design.md file into an executable test plan and validation scripts before any kernel code is written. ## Core Features & Use Cases - Case Matrix Generation: Produces a test matrix covering XS/S/M/L scale tiers, boundary conditions (chunk tails, PE indivisibility, non-power-of-two shapes, UB capacity limits), stress repeats, and performance cases, with a minimum of 20 cases. - Golden & Checker Scripts: Generates gen_data.py with fixed-seed deterministic inputs and high-precision golden outputs, plus check_result.py implementing dual-statistic judgment (precision_percent and eb bias) with per-OpTypes rtol/atol selection. - Multi-PE Run Scripts: Creates scripts/run.sh that launches one process per PE, sets up the SHMEM runtime environment with dynamic endpoint allocation, enforces a 120-second timeout, and passes explicit tolerance arguments. - Use Case: Given a design.md for a matmul_allreduce fused operator, produce the full test suite including rank-pattern golden construction, fp16 tolerance tables, and an 8-PE performance case pair (perf_s and perf_l) ready for Phase 6 evaluation. ## Quick Start Generate the complete test plan, case matrix, gen_data.py, check_result.py, and run.sh for my SHMEM operator based on the correctness contract in design.md.
